University of Southern California Dissertations and Theses

Neuroimaging markers of risk & resilience to brain aging and dementia

(USC Thesis Other)

Neuroimaging markers of risk & resilience to brain aging and dementia

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content NEUROIMAGING MARKERS OF
RISK & RESILIENCE TO
BRAIN AGING AND DEMENTIA
by
Elizabeth Haddad
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(NEUROSCIENCE)
December 2024
Copyright 2024 Elizabeth Haddad

Dedication
To my dear nephew, Bassam, for always being my inspiration. I love you, buddy.
ii

Acknowledgements
First and foremost, I would like to express my deepest gratitude to my advisor, Neda Jahanshad. Your
guidance, insight, and encouragement over the years has profoundly shaped both my academic and personal
growth. I would not be where I am today without your mentorship and the countless opportunities you
provided me throughout my graduate career. Thank you for your belief in my potential. I’m grateful to
continue our collaboration and I look forward to all our future research endeavors.
I’m also sincerely thankful to my committee chair, Paul Thompson, and my committee members,
Hosung Kim, Andrei Irimia, and Mark Shiroishi, for your invaluable feedback and guidance over the years.
Your expertise and valuable feedback have greatly contributed to the development of this research and my
overall scholarly experience.
Thank you to all my labmates — Sophia, Ravi, Shayan, Shruti, Alyssa, Iyad, Talia — for your friendship
and support. This journey would not have been nearly as fun without you all.
Thank you to my parents, brothers, and entire family, for always being there for me. I am forever
grateful for your unwavering support and encouragement.
Lastly, I would like to thank all of my co-authors and collaborators for your contributions. Chapters 2-5
are based on articles for which your contributions were integral. Additional information about funding and
support for each article is provided in the respective acknowledgments section of each chapter.
iii

Table of Contents
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
Chapter 1: Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Aging and dementia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Hallmarks of brain aging: interactions with dementia risk . . . . . . . . . . . . . . 1
1.1.2 Hallmarks of aging: an emphasis on vascular aging . . . . . . . . . . . . . . . . . . 3
1.1.3 Hallmarks of Alzheimer’s disease: a brief primer . . . . . . . . . . . . . . . . . . . 7
1.1.4 Cerebrovascular dysfunction in aging and ADRDs . . . . . . . . . . . . . . . . . . 8
1.2 Assessing risk and resilience to neurodegeneration: the role of risk factors . . . . . . . . . 13
1.2.1 Vascular and metabolic risk factors . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.2.1.1 Hypertension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.2.1.2 Type 2 Diabetes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.2.1.3 Hyperlipidemia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.2.2 Lifestyle factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.2.2.1 Physical activity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.2.2.2 Smoking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.2.2.3 Diet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.2.2.4 Obesity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.2.2.5 Sleep . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.2.2.6 Alcohol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.2.3 Other risk factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.2.3.1 Infections and systemic inflammation . . . . . . . . . . . . . . . . . . . . 23
1.2.4 Cardiovascular diseases: shared risk and pathophysiology . . . . . . . . . . . . . . 25
1.2.4.1 Atrial Fibrillation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.2.4.2 Heart Failure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.2.4.3 Coronary Artery Disease . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.3 Modifiers of brain aging and dementia risk profiles . . . . . . . . . . . . . . . . . . . . . . 28
1.3.1 Genetics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.3.2 Sex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
iv

1.3.3 Ethnicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.4 Magnetic resonance imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.4.1 MRI inference of pathology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.4.2 "Brain age" . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
1.4.2.1 Biological clocks & measuring the brain’s age . . . . . . . . . . . . . . . 36
1.4.2.2 Limitations and proposed solutions . . . . . . . . . . . . . . . . . . . . . 40
1.4.2.3 Techniques used to measure brain age . . . . . . . . . . . . . . . . . . . . 43
Chapter 2: Multisite test–retest reliability and compatibility of brain metrics derived from
FreeSurfer versions 7.1, 6.0, and 5.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.3.1 Data sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.3.2 FreeSurfer regions and metrics of interest . . . . . . . . . . . . . . . . . . . . . . . 53
2.3.3 Statistics and quality control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.3.4 Replication analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.3.5 ComBat analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.4.1 Between version compatibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.4.2 Within-Version Reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.4.3 Combat Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.4.4 Quality Control and Population-Level Analysis . . . . . . . . . . . . . . . . . . . . 65
2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.6 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
2.7 Appendix A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Chapter 3: P-wave duration is associated with aging patterns in structural brain networks . . . . . 94
3.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.2 Clinical Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.3 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.4 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.4.1 Study population . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.4.2 MRI acquisition and preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3.4.3 Regional brain age index (BAI) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.4.4 Cognitive outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
3.4.5 ECG measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
3.4.6 Statistical analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.5.1 Demographic characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.5.2 PWD and regional brain aging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
3.5.3 Regional brain aging and cognition . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.5.4 PWD and cognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.7 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
3.8 Appendix B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
v

Chapter 4: Modifiable lifestyle factors and their association with sex-specific risk and resilience to
brain aging and neurodegeneration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.1 Lifestyle factors that promote brain structural resilience in
individuals with genetic risk factors for dementia . . . . . . . . . . . . . . . . . . . . . . . 121
4.1.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.1.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.1.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.1.3.1 "Brain age" . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.1.3.2 Subject selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.1.3.3 Lifestyle factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
4.1.3.4 Association rule learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.1.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4.1.4.1 Brain age . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4.1.4.2 Resiliency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4.1.4.3 Association rule learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4.1.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.1.6 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
4.2 Causal sensitivity analysis for hidden confounding: modeling the sex-specific role of diet
on the aging brain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.2.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.2.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.2.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.2.3.1 Partial identification of dose responses . . . . . . . . . . . . . . . . . . . 137
4.2.3.2 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
4.2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
4.2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
4.2.6 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
Chapter 5: Future works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
5.1 A virtually uncharacterized population in brain imaging studies: individuals from the
Middle East and North Africa (MENA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
5.2 Leveraging large biobanks with immigrant populations . . . . . . . . . . . . . . . . . . . . 149
5.2.1 Preliminary analysis: Cerebral microhemorrhage associations in UK immigrants
from the Middle East and North Africa . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.2.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.2.1.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5.2.1.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5.2.1.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5.2.2 Preliminary analysis: Characterizing cerebral microhemorrhage associations in
dementia subtypes in the UK Biobank . . . . . . . . . . . . . . . . . . . . . . . . . 156
5.2.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
5.2.2.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
5.2.2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
5.2.2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
vi

List of Tables
2.1 Cohort demographics and scan parameters for test-retest data sets analyzed. . . . . . . . . 53
2.2 Runtimes for test-retest datasets across all three versions. . . . . . . . . . . . . . . . . . . . 93
2.3 Mean Euler numbers extracted from the nofix surface and the final surface. . . . . . . . . . 93
3.1 Performance metrics for regional BAI networks. . . . . . . . . . . . . . . . . . . . . . . . . 103
3.2 Demographic characteristics stratified by percentiles of P-wave duration with associated
interval limits listed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.3 Variables and respective data fields used for cardiovascular risk related covariates, cognition
variables, and scanner/bias (MRI) used for principal component analysis. . . . . . . . . . . 117
4.1 Lifestyle factors and respective UK Biobank data field IDs used for associations with brain
structural resiliency. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
4.2 Performance metrics of brain age CNN model. . . . . . . . . . . . . . . . . . . . . . . . . . 128
4.3 Top antecedent set with resiliency as a consequent based on lift from the combined, female,
and male models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.4 Lifestyle factors and respective UK Biobank data field IDs used to calculate conditional
average causal derivatives. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
5.1 Demographic and disease prevalences, dementia subtypes, and MARS classification results
of those with at least one definite or total CMB, and across the whole dementia population. 158
vii

List of Figures
1.1 Brain aging interactions with factors that promote risk and resilience to dementia. . . . . . 3
1.2 Hallmarks of aging and vascular outcomes. . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Blood brain barrier interactions with neurodegeneration. . . . . . . . . . . . . . . . . . . . 11
1.4 MRI inference of pathology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.1 Radar plots demonstrating regional inter-version agreement between FreeSurfer v5.3, v6.0,
and v7.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.2 Example brain segmentation and surface output from FreeSurfer v5.3, v6.0, and v7.1 . . . . 61
2.3 Radar plots demonstrating regional intra-version agreement between FreeSurfer v5.3, v6.0,
and v7.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.4 Cortical quality control results between FreeSurfer v5.3, v6.0, and v7.1 . . . . . . . . . . . . 66
2.5 Subcortical quality control results between FreeSurfer v5.3, v6.0, and v7.1 . . . . . . . . . . 67
2.6 Regional age associations in all subjects across FreeSurfer v5.3, v6.0, and v7.1 . . . . . . . . 68
2.7 Regional age associations in subjects with no segmentation quality issues across FreeSurfer
v5.3, v6.0, and v7.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.8 Regional inter-version agreement (compatibility) for the HCP dataset. . . . . . . . . . . . . 80
2.9 Regional intra-version agreement (reliability) for the HCP dataset. . . . . . . . . . . . . . . 81
2.10 Regional inter-version agreement (compatibility) for the KKI dataset. . . . . . . . . . . . . 82
2.11 Regional intra-version agreement (reliability) for the KKI dataset. . . . . . . . . . . . . . . 83
2.12 Regional inter-version agreement (compatibility) for the OASIS dataset. . . . . . . . . . . . 84
2.13 Regional intra-version agreement (reliability) for the OASIS dataset. . . . . . . . . . . . . . 85
viii

2.14 Regional inter-version agreement (compatibility) for the HNU dataset. . . . . . . . . . . . 86
2.15 Regional reliability ICC3 measures for the HNU dataset using a 10 test-retest design. . . . 87
2.16 Difference in z-statistics () comparing v7.1 to the two previous versions. . . . . . . . . . . . 88
2.17 Difference in z-statistics () comparing v6.0 to v5.3. . . . . . . . . . . . . . . . . . . . . . . . 88
2.18 Harmonized v7.1 vs. harmonized v7.1 with v5.3 cohort mixtures. . . . . . . . . . . . . . . . 89
2.19 Harmonized v7.1 vs. harmonized v7.1 with v6.0 cohort mixtures. . . . . . . . . . . . . . . . 90
2.20 Harmonized v7.1 vs. harmonized v7.1 with v5.3 cohort mixtures. . . . . . . . . . . . . . . . 91
2.21 Harmonized v7.1 vs. harmonized v7.1 with v5.3 cohort mixtures. . . . . . . . . . . . . . . . 92
3.1 Inclusion flowchart. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3.2 Methodological overview of BAI extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
3.3 P-wave duration plotted against age and stratified by sex . . . . . . . . . . . . . . . . . . . 108
3.4 Regional BAI association results with PWD as the variable of interest. . . . . . . . . . . . . 109
3.5 Regional BAI associations with composite cognitive measures. . . . . . . . . . . . . . . . . 111
3.6 Regional BAI association results with PWD as the variable of interest in those without
major cardiac conditions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
3.7 Regional BAI association results with PWD as the variable of interest in those with and
without hypertension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
3.8 Regional BAI association results with PWD as the variable of interest in those with and
without diabetes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
3.9 PWD associations results with commonly used structural MRI measures. . . . . . . . . . . 120
4.1 Scatter plot of predicted age plotted against chronological age with example subject MRI
scans. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4.2 Association rule learning results: antecedent frequency (lifestyle) with resiliency as a
consequent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.3 Plot depiction of partially identified causal derivatives. . . . . . . . . . . . . . . . . . . . . 139
4.4 Linear model, effect robustness, and casual effect maps of diet on cortical thickness in
males and females. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
ix

4.5 Linear model, effect robustness, and casual effect maps of diet on subcortical volumes in
males and females. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
5.1 Microbleed Anatomical Rating Scale (MARS) criteria. . . . . . . . . . . . . . . . . . . . . . 154
5.2 Disease prevalence and microbleed findings. . . . . . . . . . . . . . . . . . . . . . . . . . . 155
5.3 Dementia subtypes for cases with definite and total CMBs as well as their respective counts
for each CMB location . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
5.4 CMB associations with other imaging metrics. . . . . . . . . . . . . . . . . . . . . . . . . . 160
x

Abstract
Whether or not an individual develops late onset dementia is highly variable, as it depends on a wide range
of factors, many of which are modifiable. Quantifying the rate at which the brain ages and the factors
affecting it has the potential to inform treatments and recommendations, which may shape individual
outcomes. In this dissertation, I explore neuroimaging markers of risk and resilience to brain aging. Chapter
1 provides background on the factors which influence brain aging and dementia risk. I discuss ways in
which researchers can leverage in-vivo MRI to quantify markers of brain aging. Chapter 2 focuses on the
importance of the reliability of such metrics, as derived biomarkers must be robust in order to be replicable
and translatable to the clinic. Chapter 3 focuses on assessing risk for brain aging by exploring the impact
that subclinical atrial abnormalities may have on the structure of brain aging networks. Chapter 4.1 focuses
on brain aging resilience, where we investigate which lifestyle factors contribute most to brain structural
resiliency. Finally, I focus on the generalizability of imaging associations in those underrepresented in
research studies. Chapter 4.2 explores the causal role that diet has on sex-specific brain structure and chapter
5 focuses on disease prevalence and microhemorrhage associations in individuals from the Middle East
and North Africa, an uncharacterized population in brain imaging studies. Collectively, I hope to highlight
the utility that MRI biomarkers have on characterizing aging and disease mechanisms and informing
recommendations for those at risk for neurodegeneration.
xi

Chapter 1
Introduction
1.1 Aging and dementia
1.1.1 Hallmarks of brain aging: interactions with dementia risk
During aging, a progressive decline in cognitive function occurs, one which has a precipitous deterioration
nearing the age of 65, and results in impaired learning and memory, attention, decision-making speed,
sensory perception, and motor coordination. Over the past several decades, researchers have outlined
cellular processes which subserves this decline in multi-domain cognitive performance. To date, there
have been 10 established hallmarks of brain aging which include (1) mitochondrial dysfunction, where agerelated alterations include mitochondrial enlargement or fragmentation, oxidative damage to mitochondrial
DNA, impaired function of the electron transport chain, an increase in mitochondrial depolarization,
impaired Ca2+ signaling, and decreased apoptotic efficiency; (2) oxidative stress, where an imbalance occurs
between increased reactive oxygen species (ROS) and a reduction in antioxidant activity, which causes the
accumulation of protein aggregates, dysfunctional mitochondria, and vascular dysfunction; (3) impaired
molecular waste disposal, where autophagic and proteasomal degradation becomes dysfunctional leading
to the accumulation of autophagosomes containing undegraded materials, dysfunctional mitochondria,
and polyubiquitinated proteins leading to altered protein function; (4) dysregulated neuronal calcium
1

homeostasis, where aging results in a sustained elevation of intracellular Ca2+ leading to the damage and
death of neurons; (5) compromised adaptive cellular stress response, which is characterized by a decrease
in the expression and downstream signaling of beneficial neurotrophic factors (causing impaired neuronal
and mitochondrial function, Ca2+ handling, and antioxidant defenses) and an impairment in neuronal
plasticity (due to chronic uncontrolled stress and hyperactivation of the hypothalamic-pituitary adrenal
axis), ultimately leading to neuronal cell vulnerability to degeneration; (6) aberrant neuronal network
activity, where synaptic excitability imbalances and demyleination arising from aforementioned hallmarks
of brain aging and inflammation results in the degeneration and dysfunction of several neuronal cell types
responsible for learning, memory, decision making, and mood regulation; (7) impaired DNA repair, where
reductions in the expression and/or enzymatic activity of DNA-repair proteins are observed, causing an
increase in oxidative damage and compromised neuronal and mitochondrial function; (8) inflammation,
which when left unchecked, can leave the brain in an aberrant proinflammatory state, contributing to
synaptic degeneration and function impairment; (9) stem cell exhaustion, where age-related reductions in
neurogenesis occur; and (10) dysregulated energy metabolism, which not only occurs in the peripheral tissue,
but also can occur centrally, where neural cells may exhibit insulin resistance, glucose transport impairment,
and altered lipid metabolism. While none of these hallmarks occur in isolation, dysregulated energy
metabolism is known to have a particularly prominent interaction with all 9 other hallmarks, emphasizing
the importance of maintaining metabolic health throughout the lifespan (Mattson & Arumugam, 2018).
Importantly, there exists a great deal of variability in the rates at which brain aging occurs. Crucial
modifiers of brain aging trajectories include genetics, sex, and ethnicity. These modifiers also interact
with environmental and lifestyle factors which can slow or accelerate rates of brain aging. The complex
interplay between brain aging and its genetic and environmental influences is critical to understand if
we wish to curb rates of Alzheimer’s disease and related dementias (ADRDs), which are projected to rise
nearly 3 fold by the year 2050 (E. Nichols et al., 2022). The status of one’s brain health is the bedrock
2

for assessing and determining risk of or resilience to late onset dementias. In fact, the latest dementia
prevention, intervention, and care report by the Lancet Commission estimates that nearly half of dementias
(45%) are potentially modifiable if one were to eliminate risk factors and adopt healthy lifestyles (Livingston
et al., 2024). This demonstrates the importance of normal brain aging in the context of dementia risk
and prevention (Figure 1.1). By understanding the relationship between brain aging and its decelerators,
accelerators, and modifiers, we can better inform treatment and preventative measures that aid in mitigation
of brain functional decline and risk of developing neurodegenerative diseases.
Figure 1.1: Brain aging interactions with factors that promote risk and resilience to dementia. Figure
adapted from (Mattson & Arumugam, 2018) and (Livingston et al., 2024).
1.1.2 Hallmarks of aging: an emphasis on vascular aging
Aging is characterized by the time-dependent accumulation of cellular damage governed by several mechanisms that have traditionally evolved to cope with stressors (López-Otín et al., 2013; Walker et al., 2022).
The mechanisms contributing to aging were initially outline in 2013 to include stem cell exhaustion, cellular
senescence, mitochondrial dysfunction, dysregulated nutrient sensing, loss of proteostasis, epigenetic
3

alterations, telomere attrition, genomic instability, and altered intercellular communication (Ferrucci &
Fabbri, 2018; Li et al., 2020; López-Otín et al., 2013). As the field of aging is continuously evolving, these
hallmarks were recently updated to include chronic inflammation, dysbiosis, and disabled macroautophagy
as its own category (López-Otín et al., 2023). The downstream and additive effects of aging have been
linked with genetic, environmental, and lifestyle risk factors in an effort to understand disease etiology.
One particular downstream effect of age, a phenomenon known as inflammaging, has emerged as a major
contributor to the manifestation of the aging process as well as several age-related phenomena including
brain aging and neurodegeneration (Jian et al., 2020; Li et al., 2020). Inflammaging is broadly defined as the
chronic low-grade inflammation that occurs with age and in the absence of overt infection (Alberro et al.,
2021).
One of the hallmarks of aging includes the dysregulation and dysfunction of the immune system, also
known as immunosenescence. Immunosenescence is marked by dysfunction in lymphocytes originating
from the thymus. In the T cells, this includes a reduction in naive CD4+
and CD8+
counts, an imbalance
in T cell subsets, and decreased expression coupled with an overall degradation in function of T cell
receptors. B lymphocytes also experience diminished ability to respond to new antigens accompanied with
a reduction in antibody production and function (Li et al., 2020). There also exists a more general type
of senescence, known as cellular senescence. Cellular senescence is characterized by growth replication
arrest, resistance to apoptosis, chromatin remodeling, metabolic reprogramming, and morphological
changes to different cell types. This results in what’s known as senescence-associated secretory phenotype
(SASP), or the chronic release of proteases, growth factors, cytokines, chemokines, and extracellular
matrix (ECM) modifying components. SASP factors overall promote the upregulation of a variety of
different immune cells including natural killer cells, macrophages, monocytes, neutrophils, and T cells.
Overall these drive chronic inflammation either directly, or through damage associated molecular patterns
(DAMPS). Mitochondrial dysfunction is another hallmark of aging and is considered to play a central
4

role in inflammation itself as well as with its interactions with other hallmarks of aging. As it results in
excessive ROS and decreased adenosine triphosphate (ATP) production, mitochondrial dysfunction can
affect several other disease related processes. For example, DNA damage causes mutations resulting in
genomic instability. As DNA damage accumulates, it is sensed by poly(ADP-ribose) polymerase 1 (PARP1)
which signals DNA repair mechanisms. However, the prolonged activation of PARP1 causes NAD+ depletion
which ultimately leads to dysregulation of mitochondrial SIRT pathways and increased inflammation via
NLRP3 inflammasomes. Moreover, defective autophagy mechanisms, another hallmark of aging, are unable
to clear defective mitochondria and mitochondrial DAMPS (mtDAMPS) are activated further perpetuating
increased inflammatory pathways (Walker et al., 2022). Last but not least, a bidirectional relationship
exists between gastrointestinal mucosa dysfunction and systemic inflammation. Normally, immune cells
infiltrate tissues responding to insulin increases, such as fat and muscle. In the case of overnutrition
however, chronic inflammation ensues along with insulin insensitivity leading to “metainflammation” or
metabolic inflammation. This is mediated by microbial byproducts and in the case of gut microbial dysbiosis,
inflammatory pathways can be further exacerbated contributing to inflammaging (Franceschi et al., 2018;
Li et al., 2020). Ultimately, these mechanisms strongly converge to promote chronic systemic inflammation,
or, inflammaging.
More recently, the shared molecular mechanisms and resulting pathologies between vascular aging and
aging in general have come to light. For example, aging-induced ROS, DNA damage, and/or mitochondrial
dysfunction can promote vascular senescence through age-associated arterial secretory phenotype (AAASP)
(Ungvari et al., 2020; Wang et al., 2014). In fact, Hendrickx et al., (2021) even posit that inflammaging
underlies arterial stiffness – a hallmark of vascular aging (Hendrickx et al., 2021). This is thought to
be mediated by inflammaging-induced vascular inflammation. AAASPs, much like SASPs in fibroblasts,
secrete proinflammatory molecules such as IL-1, IL-6, IL-7, MCP-1, and TNF-α. This proinflammatory
state causes a “phenotypic switch” of vascular smooth muscle cells (VSMC) where they exhibit senescence,
5

enhanced proliferation, and exaggerated migration/invasion capacity. This eventually results in arterial wall
remodeling characterized by intima-media thickening and stiffening of the vessels. Moreover, AAASP may
affect adjacent cells through paracrine senescence further contributing to arterial stiffness in a feed-forward
loop (Gardner et al., 2015; Ungvari et al., 2020; Wang et al., 2014). Adjacent macrophages, VSMC, and
endothelial cells are primed towards an inflammatory phenotype and this oxidative environment leads to the
activation of extracellular matrix (ECM)-degrading enzymes of the matrix metalloproteinases (MMP) family.
Ultimately, this leads to adverse remodeling of the vascular ECM characterized by elastin fragmentation,
collagen deposition, and vascular calcification (Jiang et al., 2012; Ribeiro-Silva et al., 2021). Arterial stiffening
from AAASP-induced oxidative environment and breakdown of the ECM can further lead to endothelial
dysfunction, or impaired endothelium dependent nitric oxide (NO)-mediated vasodilation – a mechanism
that regulates vascular resistance and tissue perfusion (Ungvari et al., 2018). Together, the mechanisms of
vascular aging, including arterial stiffness and endothelial dysfunction, perpetuate a vicious cycle driven by
inflammaging that lead to increased vascular resistance and decreased tissue perfusion (Figure 1.2).
Figure 1.2: Hallmarks of aging and vascular outcomes.
6

1.1.3 Hallmarks of Alzheimer’s disease: a brief primer
Alzheimer’s disease (AD) is the most common form of dementia, and as such, much of ADRD research has
focused on understanding the mechanisms and associations that are specific to the AD subtype. Clinical
trials, for example, have largely focused on targeting one of the earliest hallmarks of AD, amyloid beta
(Aβ). The time course of AD and its biological correlates have been relatively established, though the
causal roles of pathology’s influence on clinical symptoms is less understood and the underlying cause
of AD remains up for debate. It is known, however, that the disease process of AD begins 20-30 years
in advance of cognitive deficits. The first pathological hallmark to appear in development of AD is the
accumulation of Aβ, caused by a subtle shift in the activity of transmembrane secretases which favor the
proteolytic cleavage of amyloid beta precursor (APP) protein, leading to the accumulation of amyloidogenic
peptides. Aβ deposits extracellularly in the form of plaques and can also accumulate in the blood vessels,
which is a phenomenon known as cerebral amyloid angiopathy (CAA). While the accumulation of Aβ is
necessary for the development of AD, it is not sufficient. The next pathological hallmark is the presence of
neurofibrillary tangles (NFTs) which consist of hyperphosphorylated tau proteins. Though still contested,
it is thought that Aβ pathophysiology may serve as a trigger or facilitator of downstream tau pathology
including tau misfolding, tau-mediated toxicity, accumulation into tangles, and tau spreading (Hampel et al.,
2021). Following Aβ deposition and tau-mediated neuronal injury and dysfunction, impairments in glucose
metabolism are observed and neurodegeneration ensues. It is at this time where clinical manifestation of
memory and cognitive impairments are thought to present. This prolonged disease progression highlights
the importance of understanding contributing factors that may accelerate or delay disease progression in
advance of clinical onset (Babapour Mofrad & van der Flier, 2019; G.-F. Chen et al., 2017; Jack et al., 2013;
Knopman et al., 2021).
7

1.1.4 Cerebrovascular dysfunction in aging and ADRDs
Brain aging and neurodegeneration are no longer viewed as an isolated, single system phenomena. As
increasing evidence mounts regarding the crosstalk between the periphery and central nervous system,
aspects of inflammaging are increasingly linked to hallmarks of brain aging. A key mediator between
systemic inflammation and brain aging is the breakdown of the blood brain barrier (BBB) – a process
thought to underlie multiple pathways leading to neurodegeneration. The BBB is a continuous endothelial
membrane that forms a tightly regulated barrier between peripherally circulating blood and brain tissue.
Under normal conditions, the BBB prevents blood-derived molecules from entering into the brain unless
specialized facilitation exists within the endothelium (i.e. receptors/carriers). Under pathological conditions
such as neurodegeneration and as a result of normal aging, the BBB can become disrupted resulting in the
brain becoming vulnerable to neurotoxic products (Farrall & Wardlaw, 2009; Sweeney et al., 2018). This is
the result of the breakdown of the neurovascular unit (NVU) — a collective network of cells of vascular and
neural origin each with a distinct role in regulating cerebral blood flow. The cerebral microvasculature is a
highly regulated site of gas, nutrient, waste, protein, hormone, and drug exchange between the blood and
brain tissue (Horton & Barrett, 2021). More broadly termed the “neurovascular complex” to include pial and
penetrating arteries and veins that feed the capillary network, this vast network includes endothelial cells,
a basement membrane, vascular smooth muscle cells in the case of the larger feeding vessels, and pericytes
in the case of the capillary bed. Under normal conditions, the endothelial cells are held together by tight
junctions, and together with the basement membrane and pericytes, form the BBB. At the capillary-tissue
level, astrocytic endfeet participate in forming a near continuous cover of the basement membrane and
this barrier functions to exclude large macromolecules in the interstitial fluid from entering the brain.
Neurons communicate with interneurons and astrocytes to signal vasoconstriction or vasodilation based on
oxygen necessity (Schaeffer & Iadecola, 2021). Astrocytes communicate signals regarding vessel tone with
endothelial cells, pericytes, or smooth muscle cells either directly through contact with their astrocytic
8

endfeet, or functionally through calcium signaling (Muoio et al., 2014). Coined “autoregulation”, this
myogenic response maintains constant cerebral blood flow despite changes in arterial pressure. The NVU is
also responsible for functional hyperemia or neurovascular coupling. Also independent of blood pressure,
neurovascular coupling alters blood flow to meet the energetic demands of local neuronal activity. An aged
NVU, however, is characterized by astrocytic endfeet detachment, pericyte injury and detachment, basement
membrane thickening, loss of tight junctions, and endothelial activation and dysregulation (Chagnot et al.,
2021; Faraco & Iadecola, 2013). The mechanisms thought to underlie BBB disruption are largely related to
vascular aging.
In fact, damage to the vasculature causing BBB dysfunction and cerebral hypoperfusion is “hit #1”
in a competing theory underlying Alzheimer’s disease (AD) – the two hit vascular hypothesis of AD.
This initial insult leads to neuronal injury which is further exacerbated by the accumulation of misfolded
amyloid-β (Aβ) proteins – “hit #2” (Zlokovic, 2011). The accumulation of Aβ has recently been linked
to a dysfunctional “glymphatic” (glial-lymphatic) system – a system robustly activated during sleep. The
glymphatic system functions as a result of two processes: 1.) convective influx of the cerebrospinal fluid
(CSF) from the subarachnoid space to the perivascular space to distribute throughout the brain parenchyma
and 2.) CSF then mixes with interstitial fluid (ISF) resulting in perivenous efflux of this fluid into the
meningeal and cervical lymphatic vessels – ultimately draining toxic metabolic substrates accumulated
during a wake state (Rasmussen et al., 2018). Although initially discovered in rodent models, magnetic
resonance imaging using intrathecal administration of contrast has outlined a closely resembling system in
humans (Ringstad et al., 2017). Animal models have shown that this process is at least in part, driven by the
polarized expression of aquaporin 4 (AQP4) water channels at the junction between astrocytic endfeet and
perivascular space (Rasmussen et al., 2018). Interestingly, postmortem data from both animal models of AD
and patients with AD show altered expression and mislocalization of AQP4 channels (J. Yang et al., 2011;
Zeppenfeld et al., 2017). AQP4 knockout mice as well as aged mice show reduced clearance of Aβ where
9

the latter has been attributed to altered AQP4 expression and reduced arterial wall pulsatility (Kress et al.,
2014).
Disruptions in the BBB also allow the infiltration of systemic inflammatory mediators and immune
cells which initiates the activation of resident glia triggering a proinflammatory state (Finger et al., 2022;
Walker et al., 2022; T. Yang et al., 2017). Here, astrocytes become reactive and microglia are activated
resulting in the release of cytokines and chemokines now produced by the brain. Prolonged exposure
to inflammatory signals can lead to abnormal priming of the microglia and further exaggerate cytokine
expression and eventually result in senescence and further neuroinflammation (Walker et al., 2022). A
dysfunctional BBB can also lead to increased transendothelial bulk flow via transcytosis of other blood
derivatives into the brain parenchyma. For example, extravasation of red blood cells, or microbleeds, lead
to the accumulation of toxic iron-releasing products that need to be broken down – subjecting neurons to
oxidative stress. Other influxing blood derived proteins including plasmin, thrombin, and fibrin can promote
neuronal injury/inflammation through degradation of matrix proteins, neurotoxicity and BBB disruption,
and axonal retraction, respectively (McLarnon, 2021; Sweeney et al., 2018). Neuroinflammation, red blood
cell extravasation, and infiltration of blood derived proteins all generate ROS, further promoting a vulnerable
environment and a vicious cycle of proinflammation. Moreover, influx of albumin can lead to edema and
result in impaired blood flow and hypoxia. Overall, there are several downstream consequences of BBB
breakdown that lead to neuronal injury, synaptic dysfunction, and loss of neurons and their connections.
Together these processes are thought to form the basis of neurodegeneration (Sweeney et al., 2018) (Figure
1.3).
10

Figure 1.3: Blood brain barrier interactions with neurodegeneration.
Cerebral small vessel disease (SVD) refers to a range of vascular disorders involving degenerative
alterations to the morphology of cerebral microvessels, including the small arteries, and arterioles. Similar
to the pathogenesis of atherosclerosis in the periphery, arteriosclerosis for example, is characterized by the
accumulation of small plaques that result in endothelial proliferation and the splitting of the lamina elastica
interna. These plaques consist of plasma proteins, lymphocytes, and macrophages and can be found in the
intracerebral and leptomeningeal arteries that range in 200-800µm in diameter. Arteriosclerotic plaques are
subject to thrombosis and hemorrhage development. Another subform of SVD, lipohyalinosis, occurs in
smaller arteries of 40-300µm in diameter often as a result of BBB breakdown and increased arterial pressure.
Originating from plasma protein leakage and fibrinoid necrosis, the affected arteries exhibit asymmetric
areas of fibrosis/hyalinosis. Histopathological studies have found macrophagic foam cells and astrocytes in
lipohyalinotic lesions. At the smallest scale (40-150µm in diameter), concentric hyaline thickening of the
arteries, termed arteriolosclerosis, is observed. Lipohyalinosis and arteriolosclerosis both occur in the small
arteries of the white matter (Grinberg & Thal, 2010).
11

Another association with vascular aging, AD, and neuroinflammatory diseases is neurovascular calcification, although currently it remains inadequately characterized. Arterial calcification can occur in the
intima – where calcium phosphate deposits are found in the plaque like formations described previously
in arteriosclerosis. Medial calcifications, on the other hand, are due to phenotypic changes of the VSMCs
which cause mineralization of the ECM. Triggers of this phenotypic change include cellular senescence
and inflammation, which are hallmarks of inflammaging, and medial calcification has been associated with
aging, diabetes, chronic kidney diseases, and hereditary calcification diseases. At the smallest level, capillary
calcifications have also been observed, and interestingly these protrude into the brain parenchyma from the
capillary wall and are characterized by “mineralized beads”. Capillary calcification is known to generate
strong astrocytic and microglial activation. Microglia encircling vascular calcification possess a distinct
molecular signature and are coined “calcification-associated microglia” (Zarb et al., 2021). Currently insights
into human pathophysiology of neurovascular calcification are largely gleaned from hereditary calcification
disorders, such as primary familial brain calcification (PFBC). Given neurovascular calcification is associated
with aging and a diverse array of brain pathologies, research invested in its quantification and spatial
mapping may yield important insights into disease pathogenesis and therapeutic targets (Maheshwari et al.,
2022).
12

1.2 Assessing risk and resilience to neurodegeneration: the role of risk
factors
Pharmacological treatments for AD are still in their infancy and do not address pathology stemming from
other disease processes and forms of dementia. While AD is the most common form of dementia, the
majority of AD cases have mixed pathology. For example, Kapasi et al., (2017) autopsied 447 people who
were believed to have AD at death and found only 3% to have pure AD, and that the majority had AD
pathology plus one other type of dementia (Kapasi et al., 2017). The second most common form of dementia,
which often coexists among other forms, is vascular cognitive impairment and dementia (Razek & Elsebaie,
2021). In fact, Kapasi et al. found that 87% of probable AD cases at autopsy were found to have a vascular
component, highlighting the need to address the widespreadness of vascular pathology in all forms of
dementia. Fortunately, there are ways to address and mitigate risk of developing vascular pathology, absent
the need for pharmacology, and that’s through adoption of healthy lifestyle behaviors. In fact, an estimated
45% of dementias are attributable to modifiable risk factors, that is, if all of these factors are eliminated,
nearly half of all dementias could be prevented or delayed. A recent update to the Lancet Commission’s
Dementia Prevention, Intervention, and Care Report (2024) has identified 14 risk factors that make up this
estimate. These factors have been outlined throughout the lifespan, where the earliest in life risk factor
is less education, midlife risk factors include hearing loss, high LDL cholesterol, traumatic brain injury,
depression, physical inactivity, diabetes, hypertension, smoking, obesity, and excessive alcohol intake,
and late life factors include social isolation, air pollution, and vision loss (Livingston et al., 2024). Other
factors that prove harder to quantify and control for properly in research studies include unhealthy diet
and insufficient sleep duration/quality. Given the lack of well-established treatments for ADRDs, lifestyle
interventions that can be implemented immediately and that target pathology that exists in most forms of
dementia, are currently our most effective way to lessen the risk of developing ADRDs. By understanding
13

which harmful or protective factors offer the greatest advantage when either minimized or maximized, it
may be possible to intervene to reduce the current forecasts and alleviate the burden of disease.
1.2.1 Vascular and metabolic risk factors
1.2.1.1 Hypertension
Uncontrolled hypertension, starting in midlife, is one of the main risk factors of cardiovascular and
cerebrovascular diseases as well as dementias. Chronic hypertension induces the deposition of extracellular
matrix proteins in the vascular wall resulting in remodeling and stiffening in order to cope with the
mechanical stress of elevated blood pressure. This remodeling is known as arteriolosclerosis and is ultimately
maladaptive, resulting in chronic changes to the regulation of cerebral circulation causing reduced cerebral
blood flow, neurovascular coupling dysfunction, and BBB permeability disruptions. These changes alter
cerebral autoregulation such that higher pressure is needed to maintain the same amount of perfusion. The
small arteries and arterioles that feed the deep white matter and basal ganglia are most susceptible to the
chronic effects of hypertension. This may be due to the short linear path of blood flow from the larger
vessels at the base of the brain to the small arteries and arterioles, making these vessels more vulnerable
to mechanical stress (Faraco & Iadecola, 2013; Iadecola & Gottesman, 2019). Overall, stenosis induced by
hypertension causes chronic ischemia in the periventricular and deep white matter of the brain, causing
leukoaraiosis and lacunar infarcts – both pathologies highly characteristic of small vessel disease which
may eventually result in vascular cognitive impairment and dementia. Clinically, executive dysfunction is
impaired as a result of subcortical lacunar infarcts and white matter lesions (Chang Wong & Chang Chui,
2022).
14

1.2.1.2 Type 2 Diabetes
While both macrovascular (i.e., cardiovascular and cerebrovascular disease) and microvascular (i.e., retinopathy, nephropathy, and neuropathy) diseases are well known outcomes of diabetes, reversible microvascular
dysfunction is thought to precede these clinical manifestations. Still, the exact initiating events of diabetes’
effects on the brain remain unknown, but preclinical work suggests that hyperglycemia and hyperinsulinemia induce oxidative stress and inflammation that have adverse consequences on several components of the
NVU. If sustained and uncontrolled, these changes result in permanent functional and structural remodeling
of the microvasculature comprising the NVU, including basement membrane thickening, pericyte loss,
endothelial cell apoptosis, and capillary rarefaction. Ultimately this results in functional changes including
impaired neurovascular coupling and autoregulation and increased BBB permeability in which chronic
inflammation is perpetuated due to the infiltration of inflammatory factors from the periphery. Overall,
diabetes has several consequences on the microvascular that create a milieu conducive to developing
cerebrovascular disease and vascular cognitive impairment and dementia (Edgerton-Fulton & Ergul, 2022;
Horton & Barrett, 2021; C. Yan et al., 2020).
1.2.1.3 Hyperlipidemia
Hyperlipidemia is the leading cause of atherosclerosis and lipid-lowering treatments remain the cornerstone
in managing cardiovascular disease risk (Libby et al., 2019). Maintaining optimal cholesterol levels is
often included in the cluster of vascular risk factor recommendations that mitigate risk of developing both
cardiovascular disease and dementia (Gorelick et al., 2017). Included as a risk factor in the latest Lancet
Commission assessment of modifiable risk factors, midlife low-density lipoprotein (LDL) has been linked to
all cause dementia. This is thought to be a result of excess brain cholesterol and its tendency to increase
stroke risk as well as risk for protein aggregation (Livingston et al., 2024). However, the specifics of the
causal relationship between cholesterol levels and specific cerebrovascular outcomes is not well understood.
15

Though efforts to establish this relationship have been made over the past two decades, outcomes relating
to various dementia subtypes and to MRI markers of cerebral small vessel disease have been mixed, where
positive, negative, and sometimes no associations have been found (Anstey et al., 2017; Inoue et al., 2023).
These inconsistent findings are likely due to differences in study design, age at assessment (mid vs late life),
treatment age, and length of followup assessment. Nonetheless, recommendations as they currently stand
which promote cardiovascular health and aim to prevent primary and secondary outcomes of cardiovascular
and cerebrovascular disease are appropriate (Appleton et al., 2017).
1.2.2 Lifestyle factors
1.2.2.1 Physical activity
One way to control and/or mitigate the vascular risk factors discussed thus far is to engage in physical
activity or in its structured form, exercise. Maintaining adequate cerebrovascular regulation and blood
flow is essential for the upkeep of brain metabolism and function. While physical activity plays a major
role in the prevention of cardiovascular disease and its risk factors, which impair circulation, its benefits
on vascular health may go beyond mere prevention of disease to enhance mechanisms beneficial to brain
health. One of the ways in which physical activity is thought to be neuroprotective is through its ability to
enhance cerebral blood flow by increasing the bioavailability of nitric oxide, which improves flow-mediated
vasodilation (Dao et al., 2024). Another way is through ‘exerkines’, or signaling molecules released in
response to exercise, that act through endocrine, paracrine, and/or autocrine pathways to influence many
organs in the body including the brain (Chow et al., 2022). For example, in response to exercise, proteins
such as brain derived neurotrophic factor (BDNF) and cathepsin B are not only produced by the skeletal
muscle where they can cross the BBB, but are also expressed directly in the brain. Although the extent to
which plasma levels of these proteins influence the brain is not clearly understood, direct expression of
these exerkines in the brain, particularly in the hippocampus, have been shown to influence important
16

processes including neurogenesis and synaptic plasticity. Hormones are another class of exerkines that are
important regulators of exercise’s effects on the brain. Irisin and insulin-like growth factor 1 (IGF-1) are
two examples of hormones secreted by muscle in response to exercise that can cross the BBB to influence
BDNF expression in the brain. IGF-1 is also known to upregulate the expression of vascular endothelial
growth factor (VEGF), a protein that promotes angiogenesis, or the growth of new vasculature, and has
been shown to be necessary for exercise-induced neurogenesis (Huuha et al., 2022). Metabolic byproducts,
including lactate, accumulate in muscle during anaerobic exercise of high intensities or duration, and
are also released into the bloodstream after crossing a particular threshold (Ghosh, 2004). Under normal
conditions, astrocytes convert glucose to lactate to provide energy for neurons. During intense physical
activity, however, systemic lactate can cross the BBB via monocarboxylate transporters (MCTs) and it can be
used as an additional fuel source for the brain. In mouse models, increases in lactate in response to exercise
have been shown to increase VEGF and BDNF expression and also enhance hippocampal mitochondrial
biogenesis and function (S. Lee et al., 2023). Ketone bodies are another metabolite that can cross the BBB
via MCTs and can be used to fuel the brain, particularly when glycogen stores in the body become depleted.
β-hydroxybutyrate, for example, is a ketone body that has been shown to act upon BDNF promoters. While
it is known that glucose uptake is impaired in ADRDs, Croteau et al. (2018) have shown that ketogenic
metabolism is relatively unimpaired in MCI and AD patients, which highlights the importance of other
energy substrates, including lactate and ketone bodies, to be used as fuel sources for the brain (Croteau
et al., 2018).
1.2.2.2 Smoking
Smoking has long been known to induce oxidative damage to multiple organ systems in the body by
elevating levels of free radical species (reactive oxygen species and reactive nitrogen species, i.e. ROS &
RNS). While cardiovascular disease, chronic obstructive pulmonary disease, and various forms of cancer
17

are among the leading causes of death from smoking, oxidative damage to essential cellular components of
neurons, glia, and vascular brain tissue has also been demonstrated in vitro, in-vivo, and through postmortem
human studies. Oxidative stress in the brain from smoking is further amplified by the upregulation of the
immune system and the release of inflammatory cytokines. This upregulates the activity of enzyme and
non-enzyme based antioxidants, including glutathione (the primary antioxidant in the brain), eventually
causing their depletion and inducing a chronic state of oxidative stress. While oxidative stress is implicated
in the amyloidogenic pathway to increase AD risk, it is also independently associated with cardiovascular
and cerebrovascular disease, which are both risk factors for developing vascular cognitive impairment and
dementia (Durazzo et al., 2014).
1.2.2.3 Diet
Studies of nutrition continuously demonstrate the relationship between diet and incidence of dementia
and cognitive function. Diets rich in fish, nuts, seeds, fruits, vegetables, and complex carbohydrates, such
as in the Mediterranean diet (MeDi), contain polyunsaturated fatty acids, polyphenols, antioxidants, and
fiber which help to improve gut health, insulin sensitivity, inflammation, and promote the production of
neurotrophic factors such as blood derived neurotrophic factor (BDNF). It is believed that these effects
aid in improving cognitive function and lowering dementia risk. In contrast, diets high in saturated fats
and simple sugars, including the Western diet (WD), increase gut permeability, promote inflammation not
only in the periphery but also in the brain, increase insulin insensitivity, and have been associated with
a reduction in neurotrophin content. Poor diet quality such as in the WD has been linked to cognitive
impairment and elevated risk for developing dementia (Baranowski et al., 2020).
18

1.2.2.4 Obesity
The effects of obesity on brain health are inextricably linked to outcomes of poor diet discussed in the
previous section. In addition to the metabolic and cognitive effects that a poor diet has on the brain and
body, excessive subcutaneous and visceral fat demonstrated in those with obesity due to excessive caloric
intake are closely related and linked to overlapping detrimental processes. Poor diet and overconsumption
of calories can result in gut microbial changes, adipose tissue expansion and inflammation, metabolic
dysfunction, and insulin resistance (Leigh & Morris, 2020). In healthy individuals, gut microbiota exist in a
symbiotic relationship with their host as they aid in digestion and degradation of indigestible products,
become part of the intestinal barrier which provides a mucosal layer of protection, produce bacteriocins
which are antimicrobial peptides used to inhibit or kill other harmful bacteria, generate metabolites
including short-chain fatty acids and bile acids which prevent the colonization of pathogens in the intestinal
epithelium, and also help to foster a healthy immune system. In cases of gut dysbiosis, however, often
resulting from poor diet and concomitant obesity, the healthy balance between the gut microbiota and
the host is disrupted. This produces an inflammatory milieu characterized by a compromised intestinal
barrier (i.e., “leaky gut”) due to pathogen invasion which allows for gut microbiota, toxins, and proinflammatory factors to enter the circulation, leading to system-wide consequences (Zhao et al., 2023).
Though causal mechanisms are still under investigation, Rosendo-Silva (2023) and colleagues highlight
associations between high fat diet induced gut dysbiosis and dysfunction of adipose tissue. Much of the
evidence investigates lipopolysaccharide (LPS), which is a bacterial metabolite derived from the WD and
is responsible for metabolic endotoxemia – of which one of the main outcomes is an increase in adipose
tissue. LPS is one of many bacteria that leak out from sites of compromised intestinal barriers and triggers
the production of several pro-inflammatory cytokines responsible for chronic low-grade inflammation. In
addition to inflammation, LPS can also modulate adipose tissue matrix remodeling/plasticity and metabolic
activity – possibly a cause of adipose tissue dysfunction seen in obesity (Rosendo-Silva et al., 2023). Adipose
19

tissue, metabolic dysfunction, insulin resistance, and gut microbiome changes all contribute to increased
systemic inflammation – which can lead to a compromised BBB and increased central inflammation – two
preceding events of dementia, highlighting the role that obesity and poor diet/overeating have on brain
function (Leigh & Morris, 2020).
1.2.2.5 Sleep
Sleep disturbances and disruptions in circadian rhythms are associated with increased levels of oxidative
stress, inflammation, metabolic disruption, and decreased metabolic clearance, all of which create an
environment conducive to developing dementia (Ahnaou & Drinkenburg, 2021; Mattis & Sehgal, 2016).
Whether a cause, consequence, or both a cause and consequence of dementia, sleep interactions with
ADRDs have increasingly gained attention (Carroll & Macauley, 2019). The emerging link lies, in part by,
the recently discovered “glymphatic” (glial-lymphatic) system in the brain – a system robustly activated
during sleep. The glymphatic system functions as a result of two processes: 1.) convective influx of the
cerebrospinal fluid (CSF) from the subarachnoid space to the perivascular space to distribute throughout
the brain parenchyma and 2.) CSF then mixes with interstitial fluid (ISF) resulting in perivenous efflux of
this fluid into the meningeal and cervical lymphatic vessels – ultimately draining toxic metabolic substrates
accumulated during a wake state (Rasmussen et al., 2018). Disruptions in sleep may contribute to decreases
in glympahtic clearance which may exacerbate existing AD pathology. This is supported by animal work
showing increased glymphatic clearance (doubling) of Aβ during sleep compared to being awake (Xie
et al., 2013). Using positron emission tomography (PET) scans, Shokri-Kojori et al. (2018) showed in 20
healthy human participants that a single night of sleep deprivation led to increased Aβ accumulation in
hippocampus, parahippocampus, and thalamus (Shokri-Kojori et al., 2018). Moreover, sleep disordered
breathing and other sleep disturbances have been linked to hypoxia and inflammation, which may alter
20

cerebral blood flow and contribute to neuronal damage and dysfunction, leading to vascular and/or AD
pathology (Ward & Pase, 2020).
1.2.2.6 Alcohol
Excessive alcohol intake is not only considered a risk factor for dementia, but some researchers believe that
chronic and heavy alcohol abuse can induce its own subcategory of dementia – alcohol related dementia
(ARD) (Cipriani et al., 2021). Though criteria for ARD has been outlined, it has not been fully adopted by
the community as its own clinical entity. Nonetheless, excessive alcohol intake remains a risk factor for
dementia, and the latest report by the Lancet Commission estimates that 1% of all dementias are attributable
to excessive alcohol intake in midlife (Livingston et al., 2024). Aside from being causally linked to over
200 health conditions including infectious diseases, cancer, mental/behavior disorders, and diseases of the
nervous, circulatory, and digestive systems, alcohol consumption has been linked to various neurotoxic
effects that may contribute to dementia pathology (Cipriani et al., 2021; World Health Organization, 2024).
A growing body of literature demonstrates that excessive alcohol intake induces loss of liver function, as
the liver is unable to cope with metabolism demands. This results in 1.) an increase in blood alcohol levels
and 2.) liver damage induced production of metabolic and inflammatory mediators. The increased levels
of alcohol, its metabolites, and other proinflammatory mediators are thus able to reach the brain from
circulation and have been linked to various neurotoxic mechanisms including oxidative stress, excitotoxicity,
mitochondrial damage, and apoptosis. In addition to the effects on neurons, alcohol has been shown to
induce glial inflammation through its ability to act as an agonist of Toll-like receptor 4 (TLR4), which further
leads to neuroinflammation and both gray and white matter degeneration. Moreover, chronic alcohol
abuse has been shown to compromise the BBB leading to an increase in inflammatory mediators and also
contribute to nutritional deficiencies, particularly vitamin B1, leading to regional atrophy in the thalamus,
midbrain, brainstem, and cerebellum (Cipriani et al., 2021).
21

1.2.3 Other risk factors
There are a number of other factors, many of which are less under an individual’s control, that are thought
to contribute to dementia risk throughout the lifespan. For example, education in early life and cognitive
stimulation maintained throughout the lifespan are thought to increase cognitive reserve – the ability to
withstand pathology before showing cognitive symptoms. The mechanisms involved in the increased risk
that hearing loss in midlife poses to dementia are thought to reflect a reduction in cognitive reserve as
a result of less environmental stimuli, a tax on cognitive resources due to overcompensation needed for
listening, resulting psychological factors such as depression, loneliness and isolation, and/or an interaction
of these mechanisms with existing brain pathology. The role of depression in dementia risk is thought to be
more bidirectional in nature and may increase risk at all adult ages. Depression at all adult ages may lead to
neglected self care, less social contact, and/or an overproduction of cortisol causing inflammation-mediated
hippocampal atrophy, while in late life, dementia itself may be a cause of depressive symptoms. Traumatic
brain injury (TBI) may play a more direct causal role in the development of dementia, as TBI results in direct
trauma to the brain, which may damage axons causing or exacerbating proteinopathies, microglial activation,
and ultimately neurodegeneration. Social contact, particularly in late life, is thought to build cognitive
reserve, promote healthy behaviors, and lower stress and inflammation. Along the same vein, social isolation
has been demonstrated to increase risk of dementia, potentially through mechanisms opposite to that of
social contact (Livingston et al., 2024). Air pollution is also linked to dementia risk, though research on how
this may occur is still ongoing. Possible mechanisms attributed to particulate matter include being able to
enter the circulation where it can cross the BBB and/or being able to directly enter the brain through the
olfactory system, both of which may promote oxidative-stress induced neurodegeneration (You et al., 2022).
The association between vision loss and dementia could be related to a number of mechanisms, including
the vision loss itself, coexisting factors such as diabetes, or shared pathological processes between the brain
and the retina (Livingston et al., 2024).
22

1.2.3.1 Infections and systemic inflammation
The following section is adapted from:
Nir TM, Haddad E, Thompson PM, Jahanshad N. Neuroimaging advances in diagnosis and differentiation of HIV, comorbidities, and aging in the cART era. Curr Top Behav Neurosci 2021;50:105–43.
https://doi.org/10.1007/7854_2021_221
Infections and systemic inflammation are also known to play a role in dementia susceptibilities. Peripheral immune cell infiltration into the brain can modulate microglial function centrally, creating an
inflammatory milieu that may exacerbate or initiate existing neuropathology. For example, HIV infection
may promote the appearance and progression of age-related neurodegenerative diseases, including AD.
Levels of phosphorylated- tau and beta-amyloid (Aβ) may be elevated in HIV+ individuals compared to
controls, particularly in older HIV+ adults with neurocognitive impairment (Anthony et al., 2006; Brew
et al., 2005; Clifford et al., 2009; Cohen et al., 2015; Green et al., 2005). Some studies suggest that the HIV
inflammatory cascade may lead to an overproduction of APP, as well as factors that degrade APP into
neurotoxic Aβ (Adle-Biassette et al., 1999; Forloni et al., 1992; Liao et al., 2004; Nebuloni et al., 2001; Stanley
et al., 1994).
HIV is not the only viral infection thought to play a role in neurodegeneration and risk for lateonset dementias. In fact, members of the Herpesviridae family have been linked at varying levels to the
pathogenesis of AD. Hypotheses first emerging in 1982 posited a role of HHV-1 (Human herpesvirus 1)
reactivation of latent infection in the development of AD (Ball, 1982; Gannicliffe et al., 1986). When infected
with HHV-1, the virus infects nerve endings and translocates to sensory or autonomic ganglia where it
establishes latency and the viral genome remains in an episomal state (Whitley et al., 2011). In the event
where the immune system is compromised, as it is in aging and HIV infection, HHV-1 is able to propagate
freely where it can infect other nerve cells, produce viral proteins, and activate inflammatory processes. This
cascade of events is thought to lead to the formation of Aβ plaques and the accumulation of tau, suggesting
23

a role for HHV-1 reactivation in the pathogenesis of AD (Sochocka et al., 2017). More recently, work
using human-induced pluripotent stem cell (hiPSC) technology reported a HHV-1 induced tissue model of
sporadic AD (Cairns et al., 2020). Whereas high levels of HHV-1 infection resulted in cell death, low-level
viral inoculations lead to the development of large, multicellular, dense Aβ+ fibrillar plaque-formations
(PLFs), upregulation of PSEN1 and PSEN2, reactive gliosis, and neuroinflammation. This suggests that
low-level HHV-1 infection leads to high levels over time, which is said to be evocative of the latent induction
of HHV-1 infections in patients and provides evidence for HHV-1 induced AD pathology. Moreover, AD
genetic risk scores have been shown to interact with HHV-1 antibodies in assessing AD risk, suggesting
a role for the host genetic background in HHV-1-associated AD (Lopatko Lindman et al., 2019). HHV-2,
another member of the alphaherpesvirinae family alongside HHV-1 and HHV-3, is another neurotropic virus
that establishes lifelong latent infections. Kristen et al., (2015) demonstrated that HHV-2 infection leads to
the prominent accumulation of hyperphosphorylated tau, Aβ40, and Aβ42 in human neuroblastoma cells
(Kristen et al., 2015). Bubak et al. (2020) also reported a possible effect on AD pathogenesis of HHV-3 – also
known as the varicella zoster virus, which causes chicken pox in children and shingles in adults. When
HHV-3 was used experimentally to infect quiescent primary human spinal cord astrocytes, the infection
also produced intracellular amyloid. This suggests that HHV-3 infection may increase toxic amyloid burden
and play a role in amyloid-associated disease progression (Bubak et al., 2020).
To complicate matters further, co-infection with HHV-5, also known as human cytomegalovirus (CMV),
is common in people living with HIV. Several mechanisms have been put forth on how HHV-5 might
promote HIV persistence. Christensen-Quick et al. (2017) describe how latent HIV-infected cells can
increase through HHV-5 associated inflammation, altered trafficking, inhibitory signaling, proliferation, or
inhibition of the apoptosis of HIV-infected cells (Christensen-Quick et al., 2017). HHV-5 can even directly
transactivate latent HIV by inducing ongoing HIV RNA expression. HHV-5 and other members of the
betaherpesvirinae family – including HHV-6 and HHV-7 (roseola viruses) – also have a postulated role in
24

AD. Barnes et al. (2015) found that HHV-5 seropositivity was associated with an increased risk of AD and a
faster rate of decline in global cognition in a diverse population adjusted for age, sex, education duration,
and race while also observing that HHV-5 seropositivity was higher in Black populations compared to
White, possibly partially accounting for racial differences seen in AD burden (Barnes et al., 2015). Bu et
al. (2015) reported a similar finding where HHV-5 seropositivity was associated with AD risk, even after
adjusting for other risk factors and comorbidities (Bu et al., 2015).
While long term effects of the COVID-19 infection on dementia risk have yet to be fully established,
we’ve also been witness to the interactive effects of such a pandemic on lifestyle habits. As such, indirect
effects of COVID-19 on dementia risk may also exist, as infection and pandemic circumstances have been
associated with less exercise and greater risk of obesity (Livingston et al., 2024). This highlights the need
for continued research on how infection may contribute to dementia susceptibility.
1.2.4 Cardiovascular diseases: shared risk and pathophysiology
ADRDs and cardiovascular diseases (CVDs) are both progressive and primarily age-related diseases that
take decades to manifest clinically, providing a window of opportunity in which prevention strategies may
be implemented. While our understanding of the causes of CVDs has advanced tremendously over the last
several decades resulting in effective therapeutics, heart disease still remains the leading cause of death
worldwide. Causes and treatments of ADRDs have been unfortunately much slower to progress, and the
rate of ADRDs is outpacing that of CVD, underscoring the need for establishing effective interventions. As
highlighted in the 2024 Heart Disease and Stroke Statistics report from the American Heart Association
(AHA), one of the main drivers of ADRDs is poor vascular health, emphasizing the overlapping nature
of modifiable risk factors between CVDs and ADRDs. The same lifestyle behaviors that lower CVD risk
also lower ADRD risk. In fact, studies assessing the AHA’s Life’s Simple 7 as an optimal cardiovascular
health score have found higher scores to be associated with a lower risk of cognitive decline and dementia.
25

Recently updated to include sleep, Life’s Simple 8 emphasizes optimal health behaviors in the domains of
physical activity, diet, nicotine exposure, and sleep, as well as optimal health factors including body mass
index (BMI), blood lipids, blood glucose, and blood pressure. As all of these factors are known to influence
CVD, primary prevention strategies targeting these vascular risk factors are critical in lowering current
and future forecasts of CVDs and ADRDs. Moreover, CVDs themselves may have independent influences
on ADRD risk, as highlighted below.
1.2.4.1 Atrial Fibrillation
One major cardiac condition strongly linked to brain health is atrial fibrillation (AF), the most common
type of cardiac arrhythmia (Bunch, 2020). In AF, the heart’s electrical conduction system is dysfunctional,
leading to fast and irregular heart rhythms (Lip et al., 2016). Not only is the presence of AF linked to
cognitive impairment and dementia (up to 40% increased risk) (Qiu & Fratiglioni, 2015), but AF and other
arrhythmias are risk factors for cardioembolic stroke, and also for subclinical brain injuries (Conen et al.,
2019; van der Velpen et al., 2017). Proposed mechanisms for these associations include hypoperfusion,
thromboembolism, and pathologies of the microvasculature including cerebral microbleeds and white
matter lesions (Madhavan et al., 2018).
1.2.4.2 Heart Failure
Heart failure (HF) is a condition where there is a diminished ability of the heart to pump blood to the rest of
the body and brain. HF is associated with dementia and also cognitive impairment where prevalences across
population studies range from 25-75%. Many studies even report an improvement in cognitive outcome
after therapies such as device implantation and pharmacological agents which results in an improved left
ventricular ejection fraction (Qiu & Fratiglioni, 2015; M. Yang et al., 2021). Mechanisms by which HF is
linked to brain health are thought to reflect hypoperfusion of the brain leading to hypoxia and increased
26

risk of emboli, and microvascular dysfunction including lesions of the white matter (de Bruijn & Ikram,
2014).
1.2.4.3 Coronary Artery Disease
Coronary artery disease (CAD), the most common type of heart disease, is characterized by atherosclerotic
occlusions or plaques of the coronary arteries (Malakar et al., 2019). The buildup of these plaques can lead
to a reduction in cerebral blood flow due to narrowing of the coronary artery and rupture of these plaques
can lead to myocardial infarction or stroke (de la Torre, 2012). A recent meta-analysis found that in people
with clinical coronary artery diseases, the risk for mild cognitive impairment and dementia was increased
by 50%. Some evidence suggests the severity of coronary artery calcium, a marker of subclinical CAD, is
positively associated with dementia risk as well (Xia et al., 2020). Similar to HF, CAD is thought to affect
brain health through diminished cardiac function, hypoperfusion, and risk of emboli (de Bruijn & Ikram,
2014).
27

1.3 Modifiers of brain aging and dementia risk profiles
1.3.1 Genetics
Late-onset Alzheimer’s disease (LOAD) and related dementias, which are considered to be those that occur
after the age of 65, are complex disorders, driven by a combination and wide range of genetic, environment,
and gene-environment interactions. People with a family history of dementia have twice the risk of
being affected themselves than individuals in the general population (Loy et al., 2014). One of the most
common genes known to be implicated in risk for LOAD is the apolipoprotein E (APOE) gene located on
chromosome 19. The APOE gene codes for the protein ApoE, which is the primary protein component
of lipoproteins critically involved in cholesterol and phospholipid redistribution used in the repair and
remodeling of membranes and synapses. ApoE is trafficked within and across cells via several established
pathways including in the secretory pathway where ApoE can bind to lipid droplets on the membrane
of the endoplasmic reticulum or on lipid droplets within the cytoplasm, in extracellular space where
they bind to heparan sulfate proteoglycans or to ATP-binding cassette (ABC) transporters for lipidation,
and through the endolysosomal system where the lipoproteins are delivered to the receiving cell. Under
homeostatic conditions, ApoE lipoproteins are trafficked from astrocytes to neurons for building membranes
and synapses and have also been reported to have a bidirectional relationship, where neurons secrete
peroxidated lipids to ApoE lipoproteins which are taken up by astrocytes for detoxification. However, when
astrocytes become reactive due to neurotoxic conditions, they produce highly saturated lipoprotein which,
when taken up by neurons, induce cell death. When neurons undergo stress, they can also produce ApoE,
which can escape the secretory pathway, and target mitochondria within the cell, causing mitochondrial
dysfunction (Mahley, 2016; Windham & Cohen, 2024).
Research on APOE and its mediated mechanisms is critical to understand in the context of AD and other
dementias. The APOE gene consists of 3 common isoforms, e3 being the most common and considered
28

the “wildtype” conferring no difference in AD risk, e2 which is the least common but is considered to be
protective against AD, and e4 which is the second most common and is associated with increased risk of
developing AD and other forms of dementia. One copy of e4 poses an approximate 3 − 5-fold increase
in lifetime risk, whereas two copies pose an 8 − 20-fold increase (Loy et al., 2014; Strittmatter, 2012),
depending on a person’s ancestry. Though the exact mechanisms by which e4 confers a greater risk of
developing AD remain unclear, its effects likely span multiple pathways – as evidence is accumulating
demonstrating its role in lipid dysregulation, inflammation, endocytic defects, mitochondrial dysfunction,
and Aβ accumulation.
1.3.2 Sex
In the United States, 6.7 million adults are living with AD. Almost 2
3
of them are women (Alzheimer’s
Association 2024, n.d.). One prevailing explanation of this disproportion is that differences in the incidence
of dementia is due to the fact that women live longer than men, however, research is emerging demonstrating
differences in lifespan only partially explains the variance in incidence. Dementia risk in women is also
influenced by a wide range of factors including biological factors such as genetic or hormonal differences,
geographical, psychosocial, and cultural factors such as access to education and employment, lifestyle
factors such as obesity and physical activity, and also inherent differences in the structure and physiology
of the brain itself. For example, at the same levels of disease stage, females are shown to have greater
neuropathological burden. While females have been shown to have subtle increases in Aβ burden, a more
prominent increase in tau burden occurs later in the disease trajectory compared to males. NFTs are more
tightly coupled with AD progression than Aβ is, which may be why females exhibit a more rapid decline
in cognitive function compared to males when presented with similar levels of Aβ. It’s worth noting,
however, that females show this greater tau burden at the same level of disease as men – suggesting that
they may have a higher cognitive reserve, or the ability to maintain cognitive function despite pathology
29

(Babapour Mofrad & van der Flier, 2019). In addition to tau tangles, Oveisgharan et al. (2018) also found
more severe arteriolosclerosis, or thickening/hardening of small vessel walls, in females compared to males
(Oveisgharan et al., 2018). Thus, it is critical that researchers continue to explore the mechanisms by
which sex confers differences in susceptibility and resilience to ADRDs in order to inform treatments and
recommendations that benefit both males and females.
Sex differences in AD also extend to genetic risk factors, particularly the strongest genetic risk factor
for ADRDs – ApoE4. Neu et al., (2017) found the heterogeneity of female heterozygotes compared to males
was localized to specific age ranges, particularly ages 65-75 years for AD odds and 55-70 years for mild
cognitive impairment (MCI) odds (Neu et al., 2017). Interestingly, the authors suggest a potential hormonal
basis for this finding as the age ranges roughly correspond with menopausal and perimenopausal ages
in females, respectively. It’s suggested that these effects are due to estrogen depletion during these life
stages, where in-vitro studies have shown anti-amyloidogenic effects of estrogen (Morinaga et al., 2007).
Epidemiological studies assessing hormone replacement therapies (HRT) and dementia risk are more mixed,
however (LeBlanc et al., 2001; Shumaker et al., 2003, 2004), where many attribute these discordant findings
(some beneficial, some harmful) to the timing of HRT administration. Many genetic and environmental
factors are known to impact brain structure, and sex-specific influences of these factors remain unclear.
This may be due to hidden confounders, or factors that are often difficult or impossible to measure and
influence both the treatment and outcome variables of interest. For example, menopause can impact both
lifestyle choices and brain structure. As most studies merely covary for sex, the influence of menopause on
both treatment and outcome variables is more often than not overlooked. Quantifying and accounting for
sex-specific confounders, through sex-stratified analysis for example, may help to explain variation that
otherwise leads to mixed results.
Yet another facet of AD complexity exists in sex-specific environmental interactions. As noted previously,
BDNF expression is downregulated as a result of AD and can be modified with physical activity. Interestingly,
30

studies have shown that estrogen is a regulator of BDNF (Sohrabji & Lewis, 2006), highlighting potential
mechanistic differences in response to lifestyle factors across the sexes. An additional genetic-sex interaction
has been observed in carriers of the BDNF Val66Met polymorphism (Bessi et al., 2020), where associations
between BDNF levels and cognitive function, AD risk, hippocampal blood flow, and dorsolateral prefrontal
cortex volumes were found – often only in females (Barha et al., 2019; Bessi et al., 2020; Fukumoto et al.,
2009; Weinstein et al., 2014). A few studies have examined sex-specific effects of physical activity on brain
structure but results are conflicting and inconclusive (Barha et al., 2020; Brown et al., 2022). Sex differences
further extend to another modifiable risk factor – obesity. It has been observed that indices of adiposity
and obesity are positively associated with measures of BBB disruption, greater inflammatory responses,
and increased odds of developing cardiometabolic disorders in females compared to males (Ahonen et al.,
2012; Gustafson et al., 2007; Khera et al., 2009; Thorand et al., 2007; Toro et al., 2019). Given that females
are disproportionately affected by ADRDs, more research investigating such lifestyle factors in women
is needed to quantify the extent to which they can mitigate AD pathology which may be exacerbated by
estrogen loss (Bagit et al., 2021).
Lastly, it should be considered that sex differences in dementia risk can also be influenced by sociodemographic, cultural, and behavioral contributions – further complicating the ability to disentangle specific
effects. For example, race/ethnicity has been found to be a significant independent predictor of the timing of
menopause, and this association persisted even after accounting for smoking status, age at menarche, parity,
and body mass index (Henderson et al., 2008). Future studies should consider cross-cultural differences
when measuring modifiable risk factors for ADRD.
1.3.3 Ethnicity
Genetic risk variation also exists across ancestry groups. As discussed previously, the most common isoform
of the widely studied ADRD gene, APOE, is APOE3 (e3). However, differences in its frequencies exist
31

across global regions, where it has 79% frequency in Europe, 69% in Africa, 85% in Asia, and 77% in South
America. e3 is thought to have been evolutionarily selected against the second most common isoform – the
ancestral APOE4 (e4) – whose frequencies are particularly enriched in indigenous populations of Central
Africa (40%), Oceania (37%), and Australia (26%) (Belloy et al., 2019; Huebbe & Rimbach, 2017). Evolutionary
theories have been put forth regarding the selective pressures for past and current distributions of the
APOE isoforms. For example, the e4 isoform may exhibit antagonistic pleiotropy effects. On the one hand,
coined by some as the “meat-adaptive” gene, the e4 isoform has enhanced function for host defense against
parasites and pathogens and increases the health status of newborns – favoring the short term and young
age. On the other hand, relative to the e3 isoform, e4 carriers have worse lipid profiles and have a higher
coronary artery disease risk most pronounced in those living a more modern sedentary lifestyle, thus,
having deleterious effects on the long term and old age (Finch & Sapolsky, 1999; Finch & Stanford, 2004).
This may be why the e3 isoform is more selected for in more “modern life styles” compared to higher e4
frequencies seen in indigenous populations – highlighting a potential reason for ethnic and environmental
differences seen in e4 frequency.
The advantages of the e4 isoform and its interactions with environmental factors can be most readily
seen in the Yorubans in Nigeria – where infectious disease may be more prevalent, whose diet has been
described as low fat and low calorie, and have overall lower blood cholesterol levels (Gureje et al., 2006;
Hendrie et al., 2014). When compared to an African American cohort in Indianapolis, Indiana, the e4
isoform had a weaker effect on incident AD and only existed for those carrying two copies as opposed
to just one – highlighting the dependence of e4 isoform detrimental effects on population and lifestyle
(Finch & Stanford, 2004). Other research has shown differential effects between ancestry and e4 carriage.
For example, even though African-Americans and Hispanics are at increased risk for AD compared to those
of European ancestry, associations between e4 and AD are less strong than in European and East Asian
populations (Farrer et al., 1997; Griswold et al., 2021; Tang et al., 1998). One of the most striking ethnic
32

differences exists when assessing AD associations with e4 homozygosity occurs among those of Japanese
ancestry. Whereas the odds of AD for caucasians are estimated around 12-15, African Americans around
5.7, and Hispanics around 2.2, Japanese ancestry conferred an odds ratio of 33.1 – which may indicate why
e4 allelic frequencies in this region are among the lowest (Farrer et al., 1997). Thus, understanding the
evolution of the APOE gene and how it differentially affects those of different ancestries and lifestyles can
aid in informing the management of ADRDs.
33

1.4 Magnetic resonance imaging
1.4.1 MRI inference of pathology
Neurodegeneration is often assessed using structural MRI sequences such as T1-weighted (T1w) imaging –
which makes use of longitudinal relaxation. The maximization of the contrast between gray and white
matter and signal to noise ratio (SNR) allows for the assessment of global or local atrophy patterns.
Characterizations include but are not limited to tissue volume, thickness, surface area, and gray/white
matter intensity ratios. These measures are thought to reflect cellular processes such as neuronal death
and synaptic pruning – hence their applicability to mapping neurodegeneration (M. E. MacDonald & Pike,
2021). T2-weighted (T2w) imaging is another type of MRI sequence that highlights the differences in
transverse or “spin spin” relaxation times among the tissue types. This is useful in localizing pathologies
that are associated with edema, as T2w imaging is sensitive to fluid extravasation. A widely used type
of T2w sequence is known as fluid attenuated inversion recovery (FLAIR) imaging, which suppresses the
free water CSF signal – effectively highlighting fluid extravasation and edema even more (Saranathan
et al., 2017). One of the more common types of lesions seen on FLAIR as “hyperintense” and in the white
matter, are aptly named, white matter hyperintensities (WMH). Also referred to as leukoaraiosis, WMH
are found in the periventricular and deep white matter regions of the brain and appear punctate and
patchy. They are presumed to be of vascular origin and are often observed in diseases of the small vessels,
although their characterizations are widely heterogeneous and more post-mortem pathological studies are
needed (Wardlaw et al., 2015). Nonetheless, WMH have been associated with several vascular and neural
pathological processes including arteriosclerosis, arteriolosclerosis, gliosis, demyelination, subcortical
infarcts, cerebral hemorrhages, breakdown of the ventricular lining, and a reduction in periventricular
vasculature – all of which appear to contribute to cognitive impairment (Dallaire-Théroux et al., 2017;
Hase et al., 2018). Lacunar infarcts, or ischemic lesions caused by vessel occlusion of the perforating or
34

lenticulostriate arteries, are also observed in SVD. In their acute phase, they are oval in shape and can be
readily seen in diffusion weighted imaging (DWI) sequences but in the late stages, these lacunes can be
visualized using FLAIR – where they exhibit a hypointense center and a hyperintense rim (Razek & Elsebaie,
2021). FLAIR, T2w, and T1w can also reveal enlarged perivascular spaces (PVS), another clinical feature of
SVD. Enlarged PVS has also been found to be a marker of age, other vascular risk factors, inflammation,
BBB leakage, and potentially impaired glymphatic clearance (Wardlaw et al., 2020).
Other imaging sequences widely used to assess SVD and related pathologies include T2* weighted
(T2*w) and susceptibility weighted imaging (SWI) imaging, which take advantage of the susceptibility
of paramagnetic and diamagnetic substances like iron and calcium. For example, both T2*w and SWI
can detect microbleeds/hemorrhages due to the paramagnetic properties of iron-containing products like
hemosiderin that is deposited in macrophages. Similarly T2* and SWI, among other sequences, can be used
to detect markers of cerebral amyloid angiopathy (CAA) – a cerebrovascular pathology affecting the cortex
characterized by amyloid burden in the walls of the small arteries (Sweeney et al., 2018). For example,
cortical superficial siderosis for which iron deposits in the superficial cortical layers can be observed by a
dark rim along the gyri in T2* and SWI sequences. Moreover, Quantitative Susceptibility Mapping (QSM) is
a more recently developed method derived from SWI sequencing. This quantitative method produces a map
that is able to distinguish between hemorrhage and calcification as these substances produce paramagnetic
and diamagnetic susceptibilities, respectively (Harada et al., 2022). Given the potential clinical relevance of
neurovascular calcification, QSM mapping offers a valuable tool to detect and characterize spatial patterns
of calcification. One of the more direct ways of measuring BBB leakage makes use of a relatively newer
sequence known as dynamic contrast enhancement (DCE) MRI. DCE MRI uses a gadolinium contrast
agent and measures the signal intensity in the blood plasma and extracellular extravascular space in the
brain. This is collected over time allowing for the calculation of the blood-to-brain time constant, K-trans,
which is a measure of BBB permeability (Chagnot et al., 2021). Lastly, reductions in blood flow caused by
35

mechanisms of vascular aging can be measured using an MRI technique called arterial spin labeling (ASL).
ASL uses the water content in arterial blood as an endogenous tracer, essentially tagging protons or spins
in a labeling plane and relying on the time it takes for these labeled spins to make it to the imaging voxel
(Vemuri et al., 2022). This technique is particularly useful as a biomarker for cerebrovascular reactivity – a
mechanism that is known to be impaired in vascular pathologies such as SVD (Figure 1.4).
Figure 1.4: MRI inference of pathology.
1.4.2 "Brain age"
1.4.2.1 Biological clocks & measuring the brain’s age
Aging is a complex process marked by the time-dependent accumulation of cellular damage eventually
leading to functional decline and increased risk of death (López-Otín et al., 2013). Though great strides
have been made in characterizing the hallmarks of aging over the last few decades, research regarding the
interrelation between these hallmarks, as well as the interactions with environmental factors is still needed
(and ongoing). As aging is the biggest risk factor for the leading causes of death (i.e. cancer, cardiovascular,
36

and neurodegenerative diseases), quantifying the rate at which it occurs has appealed to researchers in
multiple disciplines. Consequently, several "biological clocks" have been developed to quantify different
aspects of the various hallmarks of aging. For example, epigenetic alterations are considered to be one
of the hallmarks of aging. These alterations can be measured from blood and characterize predictable
changes in DNA methylation patterns that are associated with aging physiology. By quantifying the extent
to which a person’s DNA methylation patterns differ from the trajectory of their biological age, researchers
have developed a so-called "epigenetic age". This is quite appealing as reducing complex genetic changes
to a single number holds promising clinical implications. External validation of this index has shown
associations with markers of overall health such as cognitive ability, lung function, and walking speed as
well as diseases such as obesity, HIV, Down’s syndrome, Alzheimer’s Disease, and more generally mortality
(Cole et al., 2019; Higgins-Chen et al., 2021).
Molecular and cellular changes that occur as a result of aging and disease accumulate to eventually cause
macroscopic structural and functional changes in many organ systems. Aging of the brain, for example,
involves complex processes characterized by region-specific and nonlinear patterns of morphological
changes (Cole & Franke, 2017). These include cortical thinning/atrophy, ventricular enlargement, and
white matter pathology to name a few. Structural brain changes often have behavioral consequences,
most notably, cognitive impairment and decline. Many of these insights were derived from the use of
neuroimaging, most commonly - MRI. For example, in addition to whole brain measures, hundreds of
regional morphological metrics can be obtained with the use of a single T1w MRI image. These metrics have
been widely associated with aging, disease, genetics, environment, physical and lifestyle factors, and have
provided invaluable insight into the etiology of brain health and disease (Mufford et al., 2017; Thompson
et al., 2020). Yet, inferences regarding atrophy rates require longitudinal data in these association studies
and clinical translation using these measures has proven slower to progress. This could be, in part, due to
37

the high dimensionality of brain features. The development of "brain age" as a marker for brain health has
offered a way to infer rates of atrophy with cross-sectional data and ease clinical translation.
"Brain age" can be estimated using any desirable amount of features, and more recently with the use of
deep learning, it can even be estimated from the MRI image itself. This highlights a strength of using "brain
age" and that is dimensionality reduction. By deriving brain age from healthy subjects using a "training"
set, predictions of brain age based on brain features can be made in an independent or "test" sample. Since
its development, deviations between brain age and chronological age, also known as "brain age gap" or
BAG, have been used in a wide variety of contexts. Its clinical associations include but are not limited to:
Alzheimer’s disease, schizophrenia, psychosis, bipolar disorder, major depressive disorder (MDD), alcohol
dependence, traumatic brain injury, HIV infection, Down’s syndrome, epilepsy, diabetes, obesity, smoking,
and mortality (Baecker, Garcia-Dias, et al., 2021; Cole et al., 2018; Cole & Franke, 2017; Cole et al., 2019;
Franke & Gaser, 2019). Even in the absence of an overt clinical diagnosis, brain age has also been used in the
context of normal aging, whereby one can infer accelerated or slowed aging of the brain. By using brain age
as a marker of general health, clinicians can gain insight into whether a patient is undergoing normal brain
aging, or, if they deviate from the normal population. Furthermore, a single brain age number in relation
to a chronological age is more intuitive for a patient to understand as opposed to describing changes in
regional brain traits (Cole & Franke, 2017). A clinician could additionally make recommendations on how
one may "slow" the rate of brain aging - a topic of interest in studies studying brain resiliency or younger
than normal appearing brains (Dunås et al., 2021; Steffener et al., 2016).
What remains clear is that brain aging does not occur uniformly. Many different factors can affect the
rate of brain aging and assessing individual variability is becoming evermore important as personalized
medicine has started to emerge. This highlights another strength: individual risk assessment. One of the
most promising features of brain age is the ability to predict the progression of disease states, most notably
the conversion of dementia stages. For example, brain age was found to be a significant predictor of the
38

conversion between MCI and dementia within 3 years of an initial MRI scan (Cole et al., 2019; Franke &
Gaser, 2012; Gaser et al., 2013; Löwe et al., 2016). As Cole et al. (2019) point out, this is important because
it demonstrates the ability of brain age to detect subtle changes that occur prior to disease presentation -
contributing information otherwise unknown (Cole et al., 2019). Not only can prognostic insights and early
detection be achieved, but Baecker et al. (2021) also point out another feature - differential distinctions
between brain-based disorders (Baecker, Garcia-Dias, et al., 2021). The utility of this is most apparent in
attempting to differentiate psychiatric disorders, where oftentimes, symptoms and comorbidities overlap and
lead to misdiagnosis. This can occur between schizophrenia and bipolar disorder, for example. Interestingly,
brain age gap has been consistently associated with schizophrenia (up to 11.7 years) (Baecker, Garcia-Dias,
et al., 2021; Constantinides et al., 2022; W. H. Lee et al., 2021), whereas the associations with bipolar disorder
show less consistency (Baecker, Garcia-Dias, et al., 2021; Van Gestel et al., 2019).
Lastly, often seen as a limitation is the lack of specificity of brain age - given it is associated with a wide
range of clinical outcomes. But in fact, Cole et al. (2019) see this as an opportunity to map within-disease
variability (Cole et al., 2019). MDD is used as an example where, often, there exists a high level of variability
within the patient group. If there is a subgroup with a higher BAG, treatment may be prioritized for these
patients or alternatively, varying degrees of BAG could be used to help stratify clinical trial enrollment.
Brain age may also offer a way to evaluate treatment efficacy. Although brain aging may be seen as a
global phenomenon, Cole et al. (2019) point out brain aging may be measuring the same downstream
consequences of differing initial disease processes (Cole et al., 2019). This brings up an important direction
for future research - and that’s to combine measures of brain age with other biomarkers that reflect distinct
pathological outcomes. This may be most beneficial during the prodromal stages of a disease.
39

1.4.2.2 Limitations and proposed solutions
Despite the potential for the use of brain age clinically and in disease and basic science research, it’s
important to highlight its limitations. For demonstration, let’s consider the following equations (adapted
from (Butler et al., 2021)). For simplicity, we’ll assume that chronological age could be modeled for the
i-th subject from one brain imaging feature with the following linear regression equation calculated on a
training set:
Chronological Age i ∼ b0 + b1 Brain Feature i + ei
(1)
where b0 represents the intercept, and b1 the slope of the imaging feature and ei
is the prediction error,
also known as the residual. Thus, the predicted age is applied as a function of equation 1 to subjects in the
testing set:
Predicted Age i = ChronologicalAge \ i = fb( Brain Feature i) (2)
The brain age gap is then calculated as a linear transformation of the residual estimated from equation 1 :
BAG = ebi = Predicted Age i − Chronological Age i
(3)
It’s been noted, however, that BAG is dependent on age - complicating its interpretation. This is due to
a phenomenon known as regression towards the mean, which states that extreme samples will likely be
closer to the mean if sampled again (Stigler, 1997). As shown in Butler et al. (2021), even if brain features
are independent of age, measuring BAG between two groups, for example, would be equivalent to testing
the mean difference in age between the two groups. A more realistic outcome of this feature results in older
subjects having a younger predicted age (negative BAG) and younger subjects having an older predicted
40

age (positive BAG). Thus, researchers have begun to correct for BAG’s dependence on age (Beheshti et al.,
2019; Chung et al., 2018; Liang et al., 2019; Smith et al., 2019). In short, variations of this correction aim to
regress out age from the brain age gap. This may be demonstrated by the following equation:
BAGi ∼ β0 + β1 Chronological Age i + εi
(4)
where β0 represents the intercept, and β1 the slope of chronological age and εi
is the prediction error.
Therefore, the "modified" BAG is an estimation of the residual from equation 4:
Modified BAG i = εbi = BAGi −

βb0 + βb1 Chronological Age i

(5)
And because the modified BAG is interpreted as the age-corrected residual, the modified predicted age is
added (or equivalent in correlation - subtracted) to/from chronological age:
Modified Predicted Age i = Chronological Age i ± Modified BAG i
(6)
However, as Butler et al. (2021) point out, modified BAG, if treated as a "corrected deviation from age",
artificially inflates the variance explained (R2 = corr (Chronological Age i
, Modified Predicted Age i)
2
) by
the model. Butler et al. (2021) report on this inflation from several studies that calculated modified brain
age and present the correlations (and MAE for one study) between chronological age and predicted age
before and after regressing out age. A series of simulations were performed where the correlation between
chronological age and a brain imaging feature varied between 0 and 1. Modified predicted age was then
calculated and the simulation showed that the inflated correlations were a function of true correlations, and
that the most drastic inflations occurred in models that had the lowest correlation between chronological
age and predicted age. Butler et al. (2021) then perform a similar experiment on age and regional brain
volumes from the Philadelphia Neurodevelopmental Cohort (PNC). The results confirmed the outcome of
41

the theoretical simulation, where inflation of the modified predicted age occurred. Moreover, they showed
that complex cognitive tasks were highly associated with age, less associated with BAG, and not associated
with modified BAG - indicating that association between BAG and cognition is driven by the association
between age and cognition.
To address this issue, alternatives have been proposed. For example, studies that use age as a covariate
when testing group differences in BAG, as in Le et al., 2018, can circumvent the issue of inflated correlation
(Le et al., 2018). When Butler et al. (2021) added age as a covariate when modeling the relationship between
BAG and cognition, similar nonsignificant results were observed compared to modeling the modified BAG
with cognition. By modeling age as a covariate with BAG, misinterpretation of the modified BAG having
higher predictive accuracy can be avoided. Another way in which inflation can be avoided is by scaling
the predicted age by the slope and intercept from the regression between predicted age and age (equation
4) (Cole et al., 2018). This is often coined, "revised predicted age". This linear transformation may be
demonstrated using the variables derived from equation 4 :
Revised Predicted Age i =
Predicted Age i − β0
β1
(7)
Given that correlations are invariant to linear transformations, the revised predicted age has the same
correlation with age as the predicted age does. However, the resulting revised BAG (calculated below) is
uncorrelated with age:
Revised BAGi = Revised Predicted Age i − Chronological Age i
(8)
When testing the revised correction using the same analysis from the PNC cohort, the results are as expected
- the correlation between age and both the revised predicted age and predicted age were the same and the
revised BAG’s correlation with age was nearly zero.
42

Another pitfall often discussed with the use of brain age is the validity of using the residual, which is an
estimate of random error, as a marker for brain health given the lack of ground truth needed to assess true
validity (Bocancea et al., 2021). However, just as with the use of any other type of statistical output, the
ability to provide useful predictions lies in external validation. This includes ensuring generalizability by
preventing overfitting through techniques like cross-validation and testing in independent samples. Brain
age’s widespread associations with features of aging such as cognitive performance, fitness, and ultimately
mortality provides evidence that, although an error metric, brain age can still provide useful biological and
clinical insights (Cole & Franke, 2017; Cole et al., 2019).
1.4.2.3 Techniques used to measure brain age
Brain age could be derived from machine learning models that have varying degrees of complexity. Many
of these models use classical or "shallow" machine learning methods that rely on regression techniques
(ex: linear, ridge, lasso, elastic net, Gaussian process) that model brain features (ex: whole brain/regional
parcellation metrics, image data in similarity matrix form) as independent variables and chronological age
as the dependent variable. Several studies have compared model performance across these regressions
methods (Clausen et al., 2022; Couvy-Duchesne et al., 2020; W. H. Lee et al., 2021). In one study, the ENIGMA
PTSD and brain age working groups compared the performance between two region of interest based
methods using FreeSurfer segmentation: ENIGMA-Photon (ridge regression) and BARACUS (linear support
vector regression) and a voxelwise method - brainageR (Cole et al., 2018; L. K. M. Han et al., 2021; Liem et al.,
2017). brainageR takes in a raw T1w input, segments grey and white matter, is normalized and nonlinearly
registered to a common space, and the images are concatenated and converted into a similarity matrix that
is used as a predictor for age in a Gaussian process regression. They evaluated four commonly used metrics
of model performance which compare chronological age and predicted age. These include Pearsons’s r, also
known as the correlation coefficient (higher values suggest a better fit), R2 which quantifies the proportion
43

of dependent variable variance explained by the independent variables (higher values suggest a better fit),
root mean squared error (RMSE) which is the square root of the average of squared errors (measures the
prediction error across the sample, lower values suggest a better fit), and the mean absolute error (MAE)
which is the average of the absolute value of each residual (similar to RMSE where lower values suggest a
better fit) (de Lange et al., 2022). In line with other recent work (Couvy-Duchesne et al., 2020), Clausen et
al. (2022) found that the model with the best performance was the voxel-wise approach, which had the
highest Pearsons’s r & R2
, and the lowest RMSE & MAE (Clausen et al., 2022).
While brainageR does use a raw T1w image as input, it’s worth noting that processed features are
ultimately what is used to predict brain age. Preprocessing and processing of MRI images are timeconsuming and require technical pipelines that often come with a great deal of variability (Haddad et al.,
2023). More recent advances using deep learning approaches have offered a way to circumvent processing
time and biases stemming from preprocessing methods. Deep learning models can implement complex
network architectures that can learn global and local features of the brain in a data-driven fashion (Cole &
Franke, 2017). One of the most popular deep learning techniques, convolution neural networks (CNNs), are
often applied directly to the raw structural MRI data. This allows for brain age prediction models that are
less susceptible to biases arising from preprocessing decisions and, in theory, could be readily extracted
in the clinic. In fact, a study that compared a CNN-based approach to the technique used in brainageR
demonstrated similar performance metrics and heritability estimates across both methods (Cole et al., 2017).
A number of deep learning methods estimating brain age have been developed since then, and comparison
studies against shallow machine learning methods have largely shown better performance metrics in the
deep learning models (Amoroso et al., 2019; Jonsson et al., 2019).
Yet, deep learning methods are not without limitations. Many criticize deep learning approaches due to
their ’black box’ nature and disregard for anatomical and neuroscientific information (Cole & Franke, 2017).
This is why methods using "saliency maps" and other feature importance methods have been developed to
44

identify regions of the brain that are contributing most to changes in brain age predictions (Hong et al., 2021;
Kolbeinsson et al., 2020; Levakov et al., 2020; Popescu et al., 2021). Still, deep learning methods may not be
ideal in all cases. Methods like brainageR and ENIGMA-Photon that preprocess data may be less variant to
scanner signal and contrast to noise ratios than methods that use the raw images. For example, although
CNN-derived brain age performed similarly to brainageR in Cole et al. (2017), the authors additionally show
that brainageR had higher between-scanner reliability than the CNN method (Cole et al., 2017). It’s also
important to consider data accessibility in neuroimaging research, where often only extracted measures
like those used in ENIGMA-Photon are available for analysis. Additionally, it’s important to consider
the feasibility of deep learning methods as they require a sufficient sample size and are computationally
expensive. Baecker et al. (2021a) have recognized this and have created a decision tree for model selection
that considers computation resources and sample size (Baecker, Dafflon, et al., 2021). Lastly, it should be
noted that model performance metrics vary with respect to cohort and study-specific data characteristics
and should be considered when evaluating model selection - whether shallow or deep machine learning
methods (de Lange et al., 2022). Recent work also suggests that moderately fitting brain age models may
obtain greater differentiations between disease groups (Bashyam et al., 2020).
To conclude, even with all the aforementioned considerations, brain age prediction holds great promise
for both neuroscience research and clinical application. On that note, I recognize that brain age discussed
here is largely in the context of aging and disease. Brain age can and has been used across the lifespan
(Franke & Gaser, 2019), and its use during development and midlife should not be overlooked. Additionally,
I’ve only discussed brain age derived from T1w structural scans. Studies using other modalities such as
diffusion and functional MRI as well as multimodal applications have also been used and deserve recognition
(Beck et al., 2022; Cole, 2020; Liem et al., 2017).
45

Chapter 2
Multisite test–retest reliability and compatibility of brain metrics derived
from FreeSurfer versions 7.1, 6.0, and 5.3
The following section is adapted from:
Haddad E, Pizzagalli F, Zhu AH, Bhatt RR, Islam T, Ba Gari I, Dixon D, Thomopoulos SI, Thompson
PM, Jahanshad N. (2023). Multisite test–retest reliability and compatibility of brain metrics derived from
FreeSurfer versions 7.1, 6.0, and 5.3. Human Brain Mapping, 44(4), 1515–1532. https://doi.org/10.1002/hb
m.26147
2.1 Abstract
Automatic neuroimaging processing tools provide convenient and systematic methods for extracting
features from brain magnetic resonance imaging scans. One tool, FreeSurfer, provides an easy-to-use
pipeline to extract cortical and subcortical morphometric measures. There have been over 25 stable releases
of FreeSurfer, with different versions used across published works. The reliability and compatibility of
regional morphometric metrics derived from the most recent version releases have yet to be empirically
assessed. Here, we used test–retest data from three public data sets to determine within-version reliability
and between-version compatibility across 42 regional outputs from FreeSurfer versions 7.1, 6.0, and 5.3.
Cortical thickness from v7.1 was less compatible with that of older versions, particularly along the cingulate
46

gyrus, where the lowest version compatibility was observed (intraclass correlation coefficient 0.37–0.61).
Surface area of the temporal pole, frontal pole, and medial orbitofrontal cortex, also showed low to
moderate version compatibility. We confirm low compatibility between v6.0 and v5.3 of pallidum and
putamen volumes, while those from v7.1 were compatible with v6.0. Replication in an independent sample
showed largely similar results for measures of surface area and subcortical volumes, but had lower overall
regional thickness reliability and compatibility. Batch effect correction may adjust for some inter-version
effects when most sites are run with one version, but results vary when more sites are run with different
versions. Age associations in a quality controlled independent sample (N = 106) revealed version differences
in results of downstream statistical analysis. We provide a reference to highlight the regional metrics that
may yield recent version-related inconsistencies in published findings. An interactive viewer is provided at
http://data.brainescience.org/Freesurfer_Reliability.
2.2 Introduction
The reproducibility of research findings in the biological sciences has recently come to light as a major
problem, particularly for the neuroimaging-heavy fields of psychological and neurological-sciences (Boekel
et al., 2015; Bowring et al., 2019; Button et al., 2013; Hodge et al., 2020; Poldrack et al., 2020). Studies
on major depressive disorder (MDD), for example, have pointed out inconsistencies in results as well as
difficulties in drawing comparisons due to analytical and study design variability (Beijers et al., 2019; Dichter
et al., 2015; Fonseka et al., 2018; Kang & Cho, 2020; Müller et al., 2017; Stuhrmann et al., 2011). In one study,
using a more heterogeneous sample and rigorous statistical testing, Dinga et al. (2019) were unable to
replicate the statistical significance used to define MDD biotypes previously found in the literature (Dinga
et al., 2019). Inconsistent results investigating neuroimaging traits and diseases have also been found in
studies of insomnia (Spiegelhalder et al., 2015) and mild traumatic brain injury (mTBI). A meta-analysis of
14 reports of working memory in mTBI showed mixed findings of functional magnetic resonance imaging
47

(MRI) hyperactivity, hypoactivity, and some studies even report both hyper and hypo activity (Bryer et al.,
2013). Neuroimaging offers mechanistic insights into the variability that leads to risk for brain dysfunction,
yet these findings must be replicable in order to extend the use of MRI-derived biomarkers to a clinical
setting.
It is important to understand how and why these discrepancies occur, so that we can better understand
why certain findings are, or are not reproducible. For example, studies may be underpowered, or the variable
of interest might have different effects across populations. Experimental results can also be affected by
methodological factors such as the type of data collection (X. Han et al., 2006; Jovicich et al., 2009; S. Yan
et al., 2020), data processing and analysis (Bennett & Miller, 2013; Botvinik-Nezer et al., 2020; Carp, 2012;
Lindquist, 2020), tool version and selection (Bigler et al., 2020; Dickie et al., 2017; Gronenschild et al., 2012;
Meijerman et al., 2018; Perlaki et al., 2017; Tustison et al., 2014; Zavaliangos-Petropulu et al., 2022), and
even operating system environments (Glatard et al., 2015). The presence of pathological tissue has also
been reported to cause systematic errors in segmentation output (Dadar et al., 2021). If sample population
and methodology differ, it can be difficult to tease apart the main source of the discrepant findings.
Recent efforts in the neuroimaging community have heightened awareness and partially addressed
concerns surrounding reproducibility. Guides and tools for enhancing reproducibility have been published in
an effort to promote Open Science. Open science aims to provide transparency into research studies to better
understand the data collected, the code implemented and software used, the analysis performed, and the full
scope of results, including null findings (Gorgolewski & Poldrack, 2016; Gorgolewski et al., 2015; Kennedy
et al., 2019; T. E. Nichols et al., 2017; Poldrack & Gorgolewski, 2017; Vicente-Saez & Martinez-Fuentes, 2018;
Zuo et al., 2014). These efforts often include detailed documentation and containerization of analytical
software to ensure consistency of software version, and even operating system to the extent possible should
the study be replicated. Other efforts such as the Consortium for Reliability and Reproducibility (CoRR)
emphasize reliability and reproducibility in neuroimaging. This is demonstrated by their open-source
48

test-retest data sets which help facilitate these reliability and reproducibility assessments in both structural
and functional MRI (Zuo et al., 2014). Compared to sample size, these metrics are often overlooked, but it is
important to note that reliability is a key determinant of statistical power (Zuo et al., 2019). Large consortia,
such as the Enhancing NeuroImaging Genetics through Meta-Analysis (ENIGMA) Consortium, have also
addressed issues of low power and varying data processing pipelines by conducting large-scale harmonized
meta- and mega-analyses across international data sets (Thompson et al., 2020). Analytical protocols are
proposed and approved by the community in advance; they are then distributed and made readily available.
These protocols also include data quality control (QC) guidelines to improve analytic consistency across
heterogeneous data sets and populations.
Large, publicly available and densely phenotyped data sets that use these protocols have recently
become a powerful resource that has advanced the field of neuroscience (Horien et al., 2021). Studies like
the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and the UK Biobank collect data from 1000 to
10,000 of individuals (Littlejohns et al., 2020; Weiner et al., 2015) with some collecting longitudinal data
that spans well over a decade (Weiner et al., 2017). Automatic segmentation tools are widely used on such
data sets and have allowed for tens to hundreds of thousands of scans to be conveniently processed, thus
enabling neuroimaging traits to be used in a wide range of clinical and epidemiological studies. However,
these tools do not come without challenges and limitations.
Data processed from updated versions of these softwares are continuously released (http://adni.lon
i.usc.edu/2021/) and this leaves researchers questioning which version is most reliable or whether data
and results from work that used prior versions are compatible with those of later releases. If the detected
effects depend on the software version used, then that variability could threaten the reproducibility of
published research and compromise clinical translation. However, these version updates are often needed
to keep up with the many advancements made in the neuroimaging field. For example, version updates
may include added options or tools to work with higher resolution images, or more computational efficient
49

image processing pipelines (e.g., the use of GPUs for processing). As newer software releases are made
available, we often lack information on whether new results will be consistent with prior findings, and
the overall impact of a software upgrade. To understand sources of study variability, it is important to
understand how version upgrades may impact outcome measures.
One such automatic feature extraction and quantification tool that is widely used in neuroimaging is
FreeSurfer (Fischl, 2012). FreeSurfer is a structural MRI processing suite that allows researchers to obtain
brain parcellations and metrics from just a single T1-weighted image. Running the software involves
just a one command, but the process itself is quite extensive-where the single image undergoes over 30
step-wise processing stages (https://surfer.nmr.mgh.harvard.edu/fswiki/recon-all). Notably, more than
60 research papers have been published detailing FreeSurfer’s algorithms and workflows (https://www.zo
tero.org/freesurfer/collections/F5C8FNX8). The overall processing steps include: image preprocessing,
brain extraction, gray and white matter segmentation, reconstruction of the white matter and pial surfaces,
labeling of cortical and subcortical regions, and a spherical nonlinear registration of the cortical surface
using a stereotaxic atlas, allowing for a more accurate alignment of gyral and sulcal landmarks. Users can
then extract features, such as cortical thickness (defined as the distance between the white matter and pial
surfaces), surface area (or the area of all the triangles on the mesh representing the white matter surface),
and cortical and subcortical volumes, measured in cubic millimeters (Fischl, 2012).
A PubMed search of "freesurfer," in the year 2020 alone, results in a total of 344 publications, indicating
its wide use as a neuroimaging resource (https://pubmed.ncbi.nlm.nih.gov/?term=%28freesurfer%29&filte
r=years.2020-2020). It has been a popular tool for over 20 years throughout which over 25 different stable
releases have been disseminated (https://surfer.nmr.mgh.harvard.edu/fswiki/PreviousReleaseNotes).
Version release updates have included, for example, improvements in accuracy of the cortical labels or
a change/ addition in a preprocessing step such as denoising or bias field correction (https://surfer.n
mr.mgh.harvard.edu/fswiki/ReleaseNotes). These version changes may affect certain extracted measures.
50

Gronenschild et al. (2012) (Gronenschild et al., 2012) compared volumes and cortical thickness measures
across FreeSurfer v4.3.1, v4.5.0, and v5.0.0 and found many measurements differed significantly. After the
release of the next version, v5.3, Dickie et al. (2017) (Dickie et al., 2017) performed correlation analysis
between cortical thickness measures output from FreeSurfer v5.1 and v5.3, and found high compatibility
between the two versions. Such work helped inform protocols for consortia such as ENIGMA, where groups
that had run FreeSurfer versions older than v5.0, were asked to rerun their processing pipeline, whereas
both v5.1 and v5.3 were used for analyses within certain working groups. A more recent study, Bigler et al.
(2020) (Bigler et al., 2020), compared FreeSurfer v5.3 and v6.0 across a select set of volumes, finding low
compatibility between versions for the volume of the globus pallidus.
The latest stable release, v7.1, has yet to be thoroughly assessed for intraversion reliability and between
version compatibility. Here, we assessed the reliability and compatibility of the last three stable FreeSurfer
version releases-v5.3 (2013), v6.0 (2017), and v7.1 (2020)-across three publicly available test-retest data
sets. We set out to determine the (1) between-version compatibility and (2) within-version reliability, for
cortical thickness, surface area, and subcortical volumes. We also perform a replication analysis using an
independent data set and test how batch correction using a mixture of versions affects age associations in
these test-retest data sets. To further test how these version differences may influence population-level
findings, we ran all three FreeSurfer versions on a subset of cross-sectional data from the UK Biobank, a
cohort of middle-aged to older adults. We visually quality controlled and computed Dice overlap scores
between each pair of versions for all regional outputs. Finally, we determined the linear effect of age for
each region and metric of interest, to understand the stability of this effect across software versions.
51

2.3 Methods
2.3.1 Data sets
Test-retest data sets from the Human Connectome Project (HCP) (Van Essen et al., 2013), Kennedy Krieger
Institute (KKI) (Landman et al., 2010), and Open Access Series of Imaging Studies (OASIS-1) (Marcus et al.,
2007) were used to assess reliability within and between FreeSurfer versions. We limited the analysis to
76 healthy individuals with T1-weighted brain MRI scans aged 19-61. KKI includes test-retest data from
21 healthy volunteers with no history of neurological conditions; a test-retest subset of 35 healthy young
adults was provided by HCP, and OASIS-1 includes 20 non-demented subjects imaged twice. The maximum
inter-scan interval of 11 months in the HCP data set is longer than OASIS and KKI, yet we do not suspect
considerable changes in brain structure between sessions given that HCP is comprised of generally healthy
young adults between the ages of 22 and 35 years (Van Essen et al., 2013). See Table 2.1 for more details.
A subset of 106 neurologically normal individuals was selected at random from the UK Biobank (Miller
et al., 2016a) to test age association outcome differences between versions. This included 56 females with a
mean age and standard deviation of 62.3 (7.2) years and 50 males with a mean age and standard deviation
of 61.2 (7.7) years. The age ranged from 46 to 78 years of age. In this case, being neurologically normal was
defined based on the following exclusion criteria: cancers of the nervous system, diseases of the nervous
system, aortic valve diseases, head injuries, and schizophrenia/bipolar disorders; as a large number of
individuals had either anxiety or depressive episodes, the entire mental disorders category was not excluded.
While the UK Biobank has over 40,000 individual scans, we selected a relatively small subset, with a sample
size more in line with most single-site current neuroimaging studies.
52

Table 2.1: Cohort demographics and scan parameters for test-retest data sets analyzed. HCP is a familybased data set including up to four individuals per family, so we limited our ICC investigations to one
randomly chosen individual per family. *indicates the maximum duration between any two consecutive
scans; the maximum duration between the baseline scan and the final retest is 40 days. Abbreviations: HCP,
Human Connectome Project; HNU, Hangzhou Normal University; ICC, intraclass correlation coefficient;
KKI, Kennedy Krieger Institute; OASIS, Open Access Series of Imaging Studies
Cohort Age range;
mean(SD)
No. of
subjects (%F)
Max interscan interval
in days (mean)
Manufacturer/
Field strength
Voxel
size [mm]3
HCP 22–35; 30.7(2.97) 35 (44%) 330 (144) Siemens 3T [0.7×0.7×0.7]
KKI 22–61; 31.8(9.47) 21 (48%) 14 Philips 3T [1×1×1.2]
OASIS 19–34; 23.4(4.03) 20 (60%) 90 (20.6) Siemens 1.5T [1.0×1.0×1.25]
HNU 20-30; 24.4(2.41) 30 (50%) 10 (3.7)* GE 3T [1.0×1.0×1.0]
2.3.2 FreeSurfer regions and metrics of interest
All scans were run through the same recon-all pipeline provided by FreeSurfer for stable v5.3, v6.0, and v7.1
releases on the USC Mark and Mary Stevens Neuroimaging and Informatics Institute’s high performance
computing cluster using a Linux-centos6 operating system, ensuring the same OS and environment. For
runtimes, please see supplementary Table 2.2. Cortical parcellations were computed based on the DesikanKilliany (DK) atlas (Desikan et al., 2006), where 34 distinct regions on each cortical hemisphere are labeled
according to the gyral patterns. For each cortical region, FreeSurfer outputs the average cortical thickness,
surface area, and volume. We focus our analyses on cortical thickness and surface area, as these are largely
independent measures (Winkler et al., 2010) and volume is a composite of the two. We also extract and
evaluate the FreeSurfer derived measures of total intracranial volume (ICV) and volumes of eight subcortical
regions: the nucleus accumbens, amygdala, caudate, hippocampus, lateral ventricle, pallidum, putamen,
and the thalamus. These metrics are all ones that have been repeatedly used throughout multinational
ENIGMA projects, and are therefore of particular interest to many collaborative investigators invested in
reproducible findings. For all of our intraclass correlation coefficient (ICC) analysis here, we report left and
right measures, as well as average cortical thickness, total surface area, and average subcortical volumes.
We also include hemisphere and whole brain cortical thickness and surface area. Euler values from the
53

test-retest data sets for the final surfaces, as well as the surfaces before topological defect correction, are
made available in the supplementary materials (Table 2.3).
2.3.3 Statistics and quality control
ICCs were calculated using the psych library in R (https://CRAN.Rproject.org/package=psych). The
following three compatibility comparisons were evaluated: v7.1 versus v6.0, v7.1 versus v5.3, and v6.0
versus v5.3. Only the first time points from the test-retest data were selected for these comparisons. ICC2
was used to compute between-version compatibility measures to account for any systematic errors using
the following formula:
ICC2 =
BMS − EMS
BMS + (k − 1)EMS + k(JMS − EMS)/n′
where BMS is the between-targets mean square, EMS is the residual mean square, k is the number of
judges, JMS is the between-judges mean square, and n
′
is the number of targets (in our context, the judges
would correspond to different software versions used to compute the measures).
Within-version reliability measures were performed on within-subject test-retest data for FreeSurfer
versions v7.1, v6.0, and v5.3. ICC3 was used to measure within-version reliability using the following
formula:
ICC3 =
BMS − EMS
BMS + (k − 1)EMS
where BMS is the between-targets mean square, EMS is the residual mean square, and k is the number
of judges. ICCs were computed for each site and a weighted average was also computed, where the reported
ICC2 and ICC3 measures represent a weighted average to account for the number of participants in each
data set. ICC interpretation was based on Koo and Li (2016)(Koo & Li, 2016): ICCs < 0.50 are considered
54

poor; between 0.50 and 0.75 are moderate, between 0.75 and 0.90 denote good agreement; and values greater
than 0.90 indicate excellent reliability.
To test if FreeSurfer version affects population level findings in studies of modest sample size, age
associations were performed in a cross-sectional subset of the UK Biobank using linear regressions. Sex was
used as a covariate; ICV was added as a covariate for subcortical volumes. In that same subset, detailed QC
was performed using the ENIGMA QC protocol (http://enigma.ini.usc.edu/protocols/imagingprotocols/)
to test differences in regional fail rates across the versions. Then, 54 subjects were assigned to rater #1 and
52 to rater #2. Each rater QC’ed the same subset across all three versions. Rater #3 then reviewed all QC
fails for consistency. All subcortical QC was performed by rater #3 where a fail constitutes any notable
overestimation or underestimation of volume for any structure. Age associations were also performed in
this QC’ed subset, where subjects were excluded if the QC of any ROI was inconsistent across versions.
If subjects had consistent regional fails, they were kept in the analysis, but those regions were excluded.
While many studies of such sample size may perform manual segmentation corrections, there is no way to
ensure consistent manual editing across the outputs of all software versions. We therefore opted to exclude
QC fails to ensure our reported differences were due to changes in software version.
For each set of regressions within a version, statistical significance was determined after controlling
the false discovery rate (FDR) at q < 0.05 across 234 measures, which included all bilateral, unilateral, and
full brain measures. FDR (Benjamini & Hochberg, 1995) corrected p-values and z-statistics were plotted on
brain surfaces for comparison. All values, including uncorrected p-values, are tabulated on our web-viewer.
Dice coefficients (Dice, 1945) were also calculated in the UK Biobank subset to assess the extent of spatial
overlap of ROls across versions, for all regions in the DK atlas.
55

2.3.4 Replication analysis
To ensure replicability of our results, we calculated reliability and compatibility measures on the Hangzhou
Normal University (HNU) cohort. This data set is a valuable resource to assess reliability with its 10
test-retest design (Zuo et al., 2014): 30 participants were scanned 10 times, all within 40 days of their
baseline scan with a mean of 3.7 days between two consecutive scans (see Table 1 for more details). Upon
visual inspection of the FreeSurfer outputs, we excluded three subjects (subject IDs: 25434, 25440, 25438)
due to an error in the brain extraction that cut off a superior segment of the brain in at least one of that
subject’s sessions. Out of 300 scans, we note that this was observed three times in v5.3, two times in v6.0,
and four times in v71. Radar plots of ICCs are available in the supplementary materials (Figures 2.14 and
2.15).
2.3.5 ComBat analysis
To test the effects of batch correction, we performed an additional set of age associations on all the testretest data sets-harmonizing for site using ComBat (Fortin et al., 2018). We limited our analysis to the first
timepoint with a max age of 35 years old as the KKI data set only had a few participants with ages beyond
this point. Here, we used all subjects from HNU given that errors in baseline scans were not observed. We
compare harmonized v7.1 results to a mixture of v7.1 and other versions, where we change the version
of one (75% v7.1) or two data sets (50% v7.1) to v5.3 or v6.0. Differences in bilateral significant z-statistics
before and after FDR correction are available in the supplementary materials (Figures 2.18, 2.19, 2.20, 2.21).
2.4 Results
The full set of our reliability, compatibility, and association results are available through an interactive
3D brain viewer here: http://data.brainescience.org/Freesurfer_Reliability/. Cohort specific ICCs and
associated statistics are also available in the supplementary material (Figures S1-S8).
56

2.4.1 Between version compatibility
Version compatibility results between FreeSurfer v5.3, v6.0. and v7.1 for all cortical and subcortical metrics
are shown in Figure 2.1. Overall, the version compatibility across all versions for average cortical thickness
was good to excellent (ICCv7.1:v6.0 = 0.81; ICCv7.1:v5.3 = 0.85; ICCv6.0:v5.3 = 0.91). Similarly, left and
right hemispheric thicknesses were good for v7.1 comparisons (left: ICCv7.1:v6.0 = 0.80, ICCv7.1:v5.3 =
0.86; right: ICCv7.1:v6.0 = 0.81, ICCv7.1:v5.3 = 0.83), and excellent when comparing v6.0 to v5.3 (left:
ICCv6.0:v5.3 = 0.91; right: ICCv6.0:v5.3 = 0.90). Furthermore, version compatibility was excellent for v7.1
vs. v6.0 in several bilateral regional parcellations including the paracentral, postcentral, superior frontal,
transverse temporal, and superior parietal cortices (ICCv7.1:v6.0>0.90). The postcentral (ICCv7.1:v5.3 =
0.91) and superior parietal (ICCv7.1:v5.3 = 0.91) gyri also showed excellent compatibility between v7.1 and
v5.3. Additionally, v6.0 was highly compatible with v5.3 in the superior frontal, superior temporal, parahippocampal, supramarginal, pars orbitalis, and the banks of the superior temporal sulcus (ICCv6.0:v5.3 ≥0.90).
Several bilateral regions showed poor compatibility between v7.1 and other versions, however. In particular, the lowest ICCs were found for the isthmus (ICCv7.1:v5.3 = 0.37; ICCv7.1:v6.0 = 0.58), posterior
(ICCv7.1:v5.3 = 0.41; ICCv7.1:v6.0 = 0.55), caudal anterior (ICCv7.1:v5.3 = 0.46; ICCv7.1:v6.0 = 0.45), and
rostral anterior (ICCv7.1:v5.3 = 0.61; ICCv7.1:v6.0 = 0.50) subregions of the cingulate gyrus. An example
subject with notable differences in cingulate segmentations is displayed in Figure 2.2A. Other regions that
showed moderate agreement with v7.1 and either v6.0 or v5.3 included the entorhinal (ICCv7.1:v5.3 = 0.64;
ICCv7.1:v6.0 = 0.67), middle temporal (ICCv7.1:v5.3 = 0.68), and insular (ICCv7.1:v6.0 = 0.67) cortices, as
well as the temporal (ICCv7.1:v5.3 = 0.69) and frontal poles (ICCv7.1:v5.3 = 0.70; Figure 2.1A).
Total surface area showed excellent compatibility across all three versions (ICCv7.1:v6.0 = 0.99;
ICCv7.1:v5.3 =0.96; ICCv6.0:v5.3 = 0.99). Left and right hemispheric surface area compatibility between
versions were also excellent across all comparisons (ICCs>0.96). Overall, the two most compatible versions
were v7.1 vs. v6.0, where, notably, 29/34 bilateral regions had ICCs>0.90. Several regions also showed
57

excellent compatibility (ICC>0.90) across all three version comparisons: these included the caudal middle frontal, the inferior parietal, postcentral, posterior cingulate, rostral middle frontal, superior parietal,
and the supramarginal gyri. However, we did find surface area compatibility discrepancies not only in
regions mostly distinct from cortical thickness, but also between the pairs of versions being compared as
well. The lowest bilateral regional surface area compatibility ICCs were observed in frontal and temporal
areas when comparing newer versions to v5.3, where v7.1 showed lower compatibility to v5.3 than to
v6.0. Frontal regions included the medial orbitofrontal cortex (ICCv7.1:v5.3 = 0.51; ICCv6.0:v5.3 = 0.76),
pars orbitalis (ICCv7.1:v5.3 = 0.54; ICCv6.0:v5.3 = 0.66) and the frontal poles which were not compatible
between either v7.1 (ICCv7.1:v5.3 = 0.19) or v6.0 (ICCv6.0:v5.3 = 0.32). However, compatibility between
v7.1 and v6.0 was moderate for the medial orbitofrontal cortex (ICCv7.1:v6.0 = 0.71), excellent for the pars
orbitalis (ICCv7.1:v6.0 = 0.94), and moderate for the frontal pole (ICCv7.1:v6.0 = 0.63). Temporal regions
that followed similar trends included the parahippocampal gyrus (ICCv7.1:v5.3 = 0.61; ICCv6.0:v5.3 = 0.70)
and the temporal poles (ICCv7.1:v5.3 = 0.43; ICCv6.0:v5.3 = 0.66), where in contrast v7.1 has excellent
compatibility with v6.0 for the parahippocampal gyrus (ICCv7.1:v6.0 = 0.90) and moderate compatibility
for the temporal pole (ICCv7.1:v6.0 = 0.73; Figure 2.1D).
ICV was highly compatible across all versions (ICCs>0.97). All bilateral subcortical volumes showed
good to excellent compatibility when comparing v7.1 to v6.0 (ICCs>0.87). Good to excellent compatibility
was also found comparing v5.3 to the newer versions in the lateral ventricle, hippocampus, thalamus,
caudate, and amygdala (ICCs>0.82). Compatibility issues arose when comparing v7.1 and v6.0 against v5.3.
Poor to moderate regional compatibility was found in the pallidum (ICCv7.1:v5.3 = 0.34; ICCv6.0:v5.3 =
0.36), putamen (ICCv7.1:v5.3 = 0.56; ICCv6.0:v5.3 = 0.52; Figure 2.2B), and to a lesser extent, the nucleus
accumbens (ICCv7.1:v5.3 = 0.78; ICCv6.0:v5.3 = 0.73; Figure 2.1G).
Replication analysis using the HNU dataset showed the compatibility of surface area and subcortical
volumes to be largely in line with our main analysis (Figure 2.14D, 2.14G). Whereas in our main analysis,
58

we show the most discrepancy between cortical thickness measures from v7.1 and the previous versions,
mostly in the cingulate regions, our replication analysis showed lower compatibility more broadly between
cortical thickness from v5.3 and those from newer versions (Figure 2.14A).
59

Figure 2.1: Regional inter-version agreement (compatibility; estimated by ICC2). Bilateral, left, and right
ICC2 values comparing cortical thickness (A-C), cortical surface area (D-F), and subcortical volumes (G-I)
between versions. Outer concentric circles represent lower ICC2 values, truncated at 0.50 , while the center
represents ICC2 = 1. Regions with the lowest compatibility differ for cortical thickness and surface area
These compatibility estimates shown are a sample-size weighted average of results in each of Human
Connectome Project (HCP), Kennedy Krieger Institute (KKI), and Open Access Series of Imaging Studies
(OASIS) data sets.
60

Figure 2.2: (A) Axial slices from the same UK Biobank participant across versions. Arrows indicate posterior
and isthmus cingulate differences in v7.1 versus v5.3 and v6.0. (B) Coronal slices from the same subject
across versions. Arrows demonstrate v5.3 volume differences in the putamen and pallidum versus v6.0 and
v7.1. (C) Medial surface representations of two UK Biobank participants across versions. Arrows highlight
differences in the medial wall pinning, particularly in the entorhinal cortex, in v7.1 compared to the two
prior releases.
61

2.4.2 Within-Version Reliability
The meta-analyzed scan-rescan reliability for all cortical and subcortical metrics within each of FreeSurfer
v5.3, v6.0. and v7.1 are shown in Figure 2.3. All versions showed high reliability for average bilateral,
left hemispheric, and right hemispheric cortical thickness (ICC>0.90). Regional bilateral metrics with the
lowest thickness ICCs – but still considered moderate to good – included the temporal pole (ICCv7.1 =0.71;
ICCv6.0 =0.83; ICCv5.3 =0.74), rostral anterior cingulate (ICCv7.1 =0.83; ICCv6.0 =0.79; ICCv5.3 =0.78),
and the medial orbitofrontal cortex (ICCv7.1 =0.85; ICCv6.0 =0.88; ICCv5.3 =0.80; Figure 2.3A). Total
bilateral, left hemispheric, and right hemispheric surface area reliability was also high (ICC=0.99) for
all three FreeSurfer versions. The regions with the lowest surface area ICC were all still highly reliable,
but included the frontal poles (ICCv7.1 =0.88; ICCv6.0 =0.87; ICCv5.3 =0.77), insula (ICCv7.1 =0.91;
ICCv6.0 =0.86; ICCv5.3 =0.89), and entorhinal cortex (ICCv7.1 =0.92; ICCv6.0 =0.95; ICCv5.3 =0.88)
(Figure 3D). Regional bilateral subcortical volumes were all reliable for each of the three versions (ICC >
0.86; Figure 2.3G). ICV reliability was also very high (ICC>0.97) for all versions.
Replication analysis using the HNU dataset showed the reliability of surface area and most subcortical
volumes to be largely in line with our main analysis – similar to compatibility (Figure 2.15D, 2.15G). The
reliability of the accumbens, amygdala, and hippocampus in v5.3 showed the most discrepancy from
the main analysis with lower ICCs in the replication dataset. Cortical thickness reliability showed more
widespread and lower reliability overall, across all versions. Regions that were classified as having moderate
to poor reliability in HNU included the temporal pole, insula, entorhinal, inferior temporal, rostral anterior
cingulate, medial orbitofrontal, and lateral orbitofrontal cortices (Figure 2.15A).
62

Figure 2.3: Regional intra-version agreement (reliability; estimated by ICC3). Bilateral, left, and right ICC3
values comparing cortical thickness (A, B, C), cortical surface area ( D, E, F ), and subcortical volumes
(G, H, I) between versions. Outer concentric circles represent smaller ICC3 values, truncated at 0.70 .
Regions with the lowest reliability differ for cortical thickness and surface area. These reliability estimates
shown are a sample-size weighted average of results in each of HCP, KKI, and OASIS datasets.
63

2.4.3 Combat Analysis
When comparing multi-site age association results of harmonized v7.1 measures to a combination of either
harmonized v7.1 and v6.0 or harmonized v7.1 and v5.3 measures, we find results to be similar, but not
identical. When we harmonize a single cohort’s v6.0 or v5.3 measures in combination with v7.1, we find
that all measures that were considered statistically significant after FDR correction in the harmonized
v7.1 analysis were also significant in the analyses for which we swapped out a single cohort’s measures
to be from v6.0 or v5.3 (Figures 2.18 and 2.19). The only exception was the thickness of the banks of the
superior temporal sulcus, which was not significant using HCP’s v6.0 values in the harmonization, but
was when using all harmonized v7.1 and any of the KKI, OASIS, or HNU v6.0 swaps in combination of
v7.1. We find some regions were not significant in the harmonized v7.1 alone analysis, but were significant
when using a single cohort’s v5.3 or v6.0 results. For example, the thickness of the precuneus was not
significantly associated with age when using only v7.1 harmonized measures, but was when swapping
HCP and OASIS measures for those of v5.3 and for swapping OASIS and HNU measures for those of v6.0.
When we harmonize the measures of half the cohorts with v7.1 and half the cohorts with v5.3, we no longer
find the same age associations as in v7.1 for the banks of the superior temporal sulcus thickness and the
posterior cingulate surface area. When we harmonize with half the cohorts run through v6.0, again the
banks of the superior temporal sulcus banks did not always show a significant association. Similar to the
single cohort swaps, we see regions that are significantly associated with age in the mixtures but not in the
v7.1 only analysis. In addition to the precuneus surface area in both v5.3 and v6.0, we also see associations
with the pars opercularis and transverse temporal surface areas in combinations that included either v5.3 or
v6.0, and postcentral surface area associations for combinations with v6.0. All results, including uncorrected
associations are available in the supplementary materials (Figures 2.18, 2.19, 2.20, 2.21).
64

2.4.4 Quality Control and Population-Level Analysis
Figure 2.4 highlights regional cortical quality issues noted in the subset of UK Biobank participant scans
across each of the evaluated FreeSurfer versions. The region that showed the greatest difference in failure
rate was the left superior temporal gyrus - where v7.1 performed the best (5.7% fails) followed by v6.0 (7.5%
fails), and v5.3 performed the worst (12.3% fails; Figure 2.4A). In one subject with poor image quality, a
general underestimation occurred throughout the brain in v5.3 but not in v6.0 and v7.1 (see Figure 4B).
Other regions that failed at a relatively similar rate across all three versions included the left banks of the
superior temporal sulcus (v7.1=17%; v6.0=17%; v5.3=18.9%), the left (v7.1=14.2%; v6.0=13.2%; v5.3=12.3%) and
right (v7.1=11.3%; v6.0=12.3%; v5.3=13.2%) pericalcarine, the left middle temporal (v7.1=13.2% ; v6.0=12.3%
; v5.3=13.2%), the left cuneus (v7.1=13.2%; v6.0 =12.3% ; v5.3=10.4%), and the right cuneus (v7.1=8.5%;
v6.0=10.4%; v5.3=10.4%).
The highest overlap was between v7.1 and v6.0 where most regions had a Dice coefficient of 0.90 or
greater. The lowest overlap occurred when comparing v5.3 to both v7.1 and v6.0, particularly in the frontal
pole (left: DCv7.1:v5.3 = 0.74, DCv6.0:v5.3 = 0.76; right: DCv7.1:v5.3 = 0.79, DCv6.0:v5.3 = 0.81 ), entorhinal (left: DCv7.1:v5.3 = 0.78, DCv6.0:v5.3 = 0.79; right: DCv7.1:v5.3 = 0.75, DCv6.0:v5.3 = 0.77 ),
and right cuneus (DCv7.1:v5.3 = 0.75, DCv6.0:v5.3 = 0.77). Other regions with lower Dice coefficients
were the cingulate regions, temporal pole, pericalcarine, and the banks of the superior temporal sulcus
(Figure 2.4C).
65

Figure 2.4: Cortical quality control results. Results based on 106 neurologically healthy UK Biobank
participants. A. Manual cortical quality control results (percentage fail) based on the ENIGMA QC protocol
across versions. Gray regions indicate no failures. Note more widespread failures particularly in the
temporal and frontal regions due to a single subject (representative failure case) in B. We also note generally
higher rates of failure in the left temporal lobes across all versions. C. Dice scores across left and right
hemisphere Desikan-Killiany atlas labels. We note the lowest overlap in the cuneus, entorhinal, pericalcarine,
cingulate cortices, and temporal and frontal poles, particularly when comparing v5.3 to the newer versions.
Figure 2.5 highlights the subcortical quality issues noted across each of the evaluated FreeSurfer versions.
The most regional failures were detected in v5.3. Failures occurred more often in the left hemisphere
(Figure 2.5A). The most notable differences in failure rates were for the left pallidum (v7.1=0.9%; v6.0=0.9%;
v5.3=18.9%), left amygdala (v7.1=7.5%; v6.0=11.3%; v5.3=17.9%), and left putamen ( v7.1 = 0.9%; v6.0 01.9%;
v5.3=14.2%). Example outputs may be viewed in Figure 2.5B. The regions with the lowest overlap were in the
left and right pallidum (left: DCv71:1:5.3 = 0.66, DCv6.0:v5.3 = 0.67; right: DCv71:1v.3 = 0.78, DCv. 0:v.3 =
0.78 ) as well as the left and right nucleus accumbens (left: DCv7.1:v5.3 = 0.72, DCvc.0:v5.3 = 0.71; right:
DCv7.1:v5.3 = 0.70, DCv6.0:v5.3 = 0.69 ) when comparing v5.3 to both newer versions. Notably, the
segmentation of the left putamen often appeared larger and the left pallidum smaller in v5.3 compared to
the newer versions (Figure 2.5B).
66

Figure 2.5: Subcortical quality control results. Results based on 106 neurologically healthy UK Biobank
participants. A. Manual subcortical quality control results (percentage fail) across versions. Gray regions
indicate no failures. Note generally higher fail rates in the left hemisphere and when comparing v5.3 to the
newer versions. B. Example subcortical outputs. Arrows indicate the left putamen (cyan) and pallidum
(light green) mis-segmentation in v5.3. C. Dice scores across left and right hemisphere subcortical regions.
Note the lowest overlap when comparing v5.3 to v6.0 and v7.1.
Age associations are shown in Figures 2.6 and 2.7. Maps of the z-statistic differences are also made available in the supplementary materials (Figures 2.16 and 2.17). In the full set (106 UK Biobank scans) age associations for cortical thickness (Figure 2.6), v7.1 had 29 regions that survived FDR correction, less than both v6.0
with 32 and v5.3 with 43; all these regions showed lower thickness with age other than the right rostral anterior cingulate, which showed a positive association with age across all versions. The strongest associations
were in the left supramarginal
zv7.1 = −5.55, qv7.1 = 3 × 10−5
; zv6.0 = −6.24, qv6.0 = 1×10−6
; zv5.3 =
−5.90, qv5.3 = 4×10−6
) and left superior temporal gyrus
zv7.1 = −4.88, qv7.1 = 1 × 10−4
; zv6.0 = −5.21,
qv6.0 = 4 × 10−5
; zv5.3 = −5.33, qv5.3 = 2 × 10−5
) for all three versions. All regions that were significant in v7.1 and v6.0 were also significant in v5.3, except for the left frontal pole in v6.0 (zv7.1 = −2.19,
qv7.1 = 9 × 10−2
; zv6.0 = −2.55, qv6.0 = 4 × 10−2
; zv5.3 = −1.83, qv5.3 = 1 × 10−1
). Generally, v5.3
had the largest absolute z-statistics compared to v7.1 and v6.0. For the surface area age associations, no
regions survived FDR correction in v7.1, whereas in v6.0, the left frontal pole survived correction, and

in v5.3, the right paracentral, left banks of the superior temporal sulcus, right entorhinal, right lateral
orbitofrontal, and right temporal pole were considered significantly associated with age after correction.
For subcortical volumes, all regions were significantly associated with age, except for the left and right
caudate and pallidum for all three versions and the right amygdala for v7.1 (qv7.1 = 0.08).
Figure 2.6: Regional age associations in all subjects. Results based on 106 neurologically healthy UK
Biobank participants. A. FreeSurfer v5.3, B. v6.0, and C. v7.1. Top row indicates the z-statistic and bottom
indicates − log10(q < 0.05) for left and right surface area, thickness, and subcortical volumes. We note
that v5.3 generally has the largest z-statistics, particularly for cortical thickness, and the largest number of
statistically significant regions.
A total of 69 subjects remained for regression analysis in the cortical QC’ed subset (37F, mean age:
61.1 ± 7.11). In this subset, cortical thickness was associated with age in 13 regions for v7.1, 16 for v6.0,
and 22 for v5.3 (Figure 2.7). As with the full set above, all regions that survived FDR correction in v7.1
also survived in v6.0 and v5.3 and all regions that survived in v6.0 were also significant in v5.3. Cortical
thickness regions that had a considerable proportion of fails and no longer reached the significance threshold
in the QC’ed subset included the left banks of the superior temporal sulcus, left middle temporal, right
precentral, and the right superior parietal gyrus. The left lingual, left cuneus, right pericalcarine, and the
right banks of the superior temporal sulcus were all regions that had considerable quality issues and for
which cortical thickness associations met FDR significance criteria for v5.3 in the full subset, yet these
thickness associations were no longer significant in the QC’ed subset. The only surviving region in the
68

QC’ed subset for surface area was the right entorhinal cortex in v5.3, although it is worth noting this region
was not heavily QC’ed. The external surface in this area was apparently different in v7.1 compared to the
previous versions (Figure 2.2C) and the rate at which this occurred would have resulted in the majority of
participants being considered as a "fail" in the older versions.
61 subjects (35F, mean age: 62.6 ± 6.8 years) were found to have no quality issues in the subcortical
segmentations across any versions. Age associations with these subjects indicated that only the thalamic
volumes were significantly associated with age in v5.3 (both right and left) and v6.0 (left only).
Figure 2.7: Regional age associations in subjects with no segmentation quality issues. Results based
on n = 69 (cortical) and 61 (subcortical) of the 106 neurologically healthy UK Biobank participants. A.
FreeSurfer v5.3, B. v6.0, and C. v7.1. Top row indicates the z-statistic and bottom indicates − log10(q < 0.05)
for left and right surface area, thickness, and subcortical volumes. Several regions found to be significant in
the full sample of n = 106 did not survive FDR correction here.
2.5 Discussion
Our work has four main findings that may help explain how a FreeSurfer version upgrade can impact
results:
1. The compatibility between v7.1 and the previous version, v6.0, was largely good to excellent for
measures of cortical surface area and subcortical volume, with the exception of the medial orbitofrontal
69

cortex and the frontal/temporal poles. Similar trends were observed in our replication analysis. Most compatibility issues arose in regional cortical thickness estimates, where moderate or even poor compatibility
was seen in the thickness estimates of the cingulate gyrus (rostral anterior, caudal anterior, posterior, and
isthmus), entorhinal, insula, and orbitofrontal regions (medial and lateral). In our replication analysis,
mostly moderate compatibility was found in these regions. The exceptions were the entorhinal cortical
thickness with poor compatibility and the thickness of the isthmus cingulate with good compatibility.
2. There were substantial compatibility issues between v7.1 and v5.3, in cortical regional thickness, area,
and subcortical volumes. Thickness measures with low compatibility between v7.1 and v5.3 were the same
as those between v7.1 and v6.0. However, the replication data set showed more similarities between the
comparisons evaluating v5.3 against the newer versions. Regions with cortical surface area and subcortical
volume compatibility issues between v7.1 and v5.3 were the same as the regions that were less compatible
between v5.3 and v6.0, suggesting these area and volume differences were introduced with v6.0, not v7.1,
which was in line with our replication analysis.
3. The test-retest reliability for all v7.1 metrics evaluated here was good to excellent in our main analysis,
except for the thickness of the temporal pole. Replication analysis showed similar trends for surface area
and subcortical volumes, but cortical thickness intraversion reliability was lower overall across all versions.
4. Age associations revealed generally smaller absolute z-statistics in v7.1 compared to earlier releases,
where v5.3 had the largest absolute z-statistics overall. Quality issues were more prevalent in v5.3, particularly in the left superior temporal gyrus, pallidum, and putamen. Age associations did not meet the
statistical significance threshold in many of the heavily quality controlled regions.
The regions in which v7.1 had the lowest compatibility with the previous versions were along the
caudal-rostral axis of cingulate cortex. The subdivisions of the cingulate cortex play distinct roles in
large-scale brain networks including the visceromotor, ventral salience, dorsal executive/salience, and
default mode networks (Touroutoglou & Dickerson, 2019). Alterations in the subregions of the cingulate
70

cortex have been demonstrated throughout the lifespan and in association with different neuropsychiatric
disorders. For example, compared to controls, developmental delays in adolescents with attention deficit
hyperactivity disorder are seen most prominently in the thickness of the prefrontal regions including the
cingulate cortices (Vogt, 2019). In posttraumatic stress disorder (PTSD) studies, the anterior midcingulate,
and in some cases the posterior cingulate, show, on average, lower thickness in individuals with PTSD
compared to healthy controls (Hinojosa et al., 2019). Subregions of the cingulate cortex have also been
associated with age related cognitive performance. In "SuperAgers," or adults over the age of 80 years,
whose episodic memory is resistant to age-related decline, a preservation of the anterior cingulate thickness
is observed (de Godoy et al., 2021; Gefen et al., 2015; Harrison et al., 2012, 2018; Sun et al., 2016). Many of
these studies were performed using versions of FreeSurfer that precede v7.1, so possible replication issues
in future studies may be partially explained by the version incompatibility described in this work. Although
we tested within a very narrow age range, and more extensive evaluation may be needed, we find that
batch correction methods may adjust for these effects in the case where the large majority of the cohorts
are run through the same version of FreeSurfer. By simulating a multi-cohort analysis, where all but one of
the cohorts have run v7.1 and one site has been run on a version that precedes v7.1, we find similar cortical
thickness and surface area cingulate associations after multiple comparisons correction. However, prior
to multiple comparisons correction, differences exist across the cohorts for both v6.0 and v5.3suggesting
that mixing versions could possibly result in false positives. Furthermore, as we also tested different age
associations after ComBat harmonization across iterations with two cohorts run with v7.1, and two cohorts
with other versions, we notice that as more cohorts are run with different FreeSurfer versions, ComBat
harmonization is less effective and some regional variability in results, possibly false positives, may be
introduced by using a mixture of versions. This analysis is limited in the number of combinations that were
tested here, and additional extensive evaluation may be needed on a wider age range with more cohorts
run across multiple versions.
71

Other regions with lower thickness compatibility with v7.1 included the medial and lateral orbitofrontal,
entorhinal, and insular cortices. Inferior frontal regions such as the medial and lateral orbitofrontal cortices
are often susceptible to signal loss and bias field inhomogeneities. v7.1 uses an updated bias field and
denoising method that could affect the gray/white matter contrast in these areas. Temporal regions, such
as the entorhinal and insular cortex, which were less compatible with v7.1, could be due to an algorithmic
update that pins the pial surface in the medial wall to the white matter surface. This prevents a premature
cutoff through the hippocampus and amygdala, which may affect surrounding regions in earlier versions.
Notably, visual inspection of the external surface of the entorhinal cortex revealed an improvement of
the entorhinal pinning to the medial wall in v7.1-as opposed to prior versions (Figure 2c). This issue was
extremely prevalent, and considering these subjects as "QC-fails" would have resulted in the majority of
subjects failing; therefore, subject scans affected by this cutoff in v5.3 and v6.0 remained included in our
"error-free" subset. Downstream effects of this may be demonstrated in our age associations within the full
n = 106 sample. Here, the left insular thickness showed significant age effects in v5.3 and v6.0, as well as
the thickness of the right entorhinal cortex in v5.3, but neither showed associations with age in v7.1. The
entorhinal cortex plays an important role in mediating information transfer between the hippocampus and
the rest of the brain (Coutureau & Di Scala, 2009; Garcia & Buffalo, 2020). Measurements of its thickness
are widely assessed in Alzheimer’s disease, as it is one of the first regions to be impacted by the disease
process (Braak & Braak, 1991)and researchers have found associations between its thickness and markers
of amyloid and tau (Thaker et al., 2017). Entorhinal thickness is often a feature of interest in models that
are designed to predict progressive cognitive decline due to its early vulnerability and role in the prodromal
stages of Alzheimer’s disease. Although v7.1 may have a more anatomically accurate segmentation, we
advise caution when comparing the performance of predictive models that use earlier releases of FreeSurfer
for deriving this metric.
72

Compatibility issues between v7.1 and older versions were less frequent with surface area and did not
occur in the same regions as cortical thickness. This could be due to the relative independence of these
measures: surface area is calculated as the area of all the triangles on the white matter surface, and the
large area covered by many regions makes them more robust to slight variation in vertex counts. On the
other hand, cortical thickness is measured as the distance between the vertices of the white matter and
pial triangulated surfaces, and is often between 2 and 4 mm thick, a span of only two to four voxels; slight
variability in partial voluming may have a more dramatic effect on cortical thickness. Yet, as the thickness is
averaged in the entire area, a slight variation in the number of vertices on the surface will have little effect
on the averaged cortical thickness estimates. The independence of these measures has also been established
in relation to their genetic associations (Grasby et al., 2020; Winkler et al., 2010) overall suggesting that
our results are not unexpected. Measures of v7.1 surface area that had poor compatibility with v5.3 (and
moderate with v6.0) included the frontal and temporal poles. The release of v7.1 included a remeshing of
the white matter surface to improve its triangle quality-potentially impacting the most rounded points of
the frontal and temporal lobes. We find that v7.1 had the lowest fail rate in the temporal pole compared to
v5.3 and v6.0 suggesting an improvement in the parcellation.
Subcortical volumes are also another set of metrics derived from FreeSurfer that are of major interest to
neuroimaging researchers (Ohi et al., 2020; Satizabal et al., 2019). Efforts to provide references of normative
subcortical volume changes that occur as a result of aging have been put forth (Bethlehem et al., 2022; Coupé
et al., 2017; Dima et al., 2022; Miletić et al., 2022; Narvacan et al., 2017; Potvin et al., 2016). For example,
Potvin et al., 2016 (Potvin et al., 2016) pooled data from 21 research groups ( n = 2790 ) and segmented
subcortical volumes using FreeSurfer v5.3 to provide norms of volumetric estimate changes during healthy
aging. Although this study, along with many others, provides a valuable resource to researchers, we advise
caution with the newer versions when referencing normative data derived from v5.3, particularly in the
lentiform nucleus. The lentiform nucleus (i.e., the putamen and globus pallidus combined) has often been
73

found to be difficult to segment due to the high white matter content in the pallidum-making it more
difficult to distinguish gray-white matter contrast (Bigler et al., 2020; Makowski et al., 2018; Ochs et al.,
2015; Visser et al., 2016). We find poor compatibility in the pallidum and moderate in the putamen when
comparing v7.1 and v5.3. Visual QC of these regions revealed a higher failure rate and lower Dice overlap in
v5.3 compared to v7.1, particularly in the left hemisphere. However, we find the compatibility between v7.1
and v6.0 to be excellent and the Dice overlap was greater than 90% in the lentiform nucleus. This suggests
that changes made in the release of v6.0 contributed to v5.3 discrepancies. For example, the putamen does
not extend so far laterally in the two newer versions-a known issue noted in the release notes of v6.0.
The main goal of our work was to evaluate FreeSurfer’s latest stable release, v7.1, yet it is also worth
noting how v6.0 differs from v5.3. While compatibility was generally good for cortical thickness, regional
surface area estimates were more moderately compatible, with the frontal pole even showing poor compatibility, similar to v7.1 compared to v5.3. Temporal lobe regions showing moderate compatibility in surface
area between v6.0 and v5.3 included the entorhinal, insula, parahippocampal, and temporal pole. Updates
that accompanied the release of v6.0 that may contribute to these compatibility discrepancies include
improved accuracy of the cortical labels and an updated template (fsaverage) that "fixes" the peri/entorhinal
labels. As previously mentioned, v6.0 compatibility with v5.3 was poorest in the pallidum and putamen.
Our results coincide with Bigler et al. (2020) (Bigler et al., 2020) where the lowest agreement was also found
in the pallidum and putamen when comparing v5.3 to v6.0.
Overall, we note consistencies across sites. For example, HCP, KKI, and OASIS all showed the lowest
compatibility in the cingulate regions when comparing v7.1 to the previous versions, and overall lower
compatibility in thickness measures compared to surface area. However, some site differences were observed.
For example, more widespread lower compatibility in cortical thickness in temporal and frontal regions was
seen for HCP compared to KKI and OASIS, which both have larger and anisotropic voxels. The opposite
occurred for compatibility of surface area between v5.3 and the later versions where OASIS and KKI
74

showed more widespread compatibility issues compared to HCP, particularly in the temporal regions.
HCP also showed the highest overall reliability for cortical thickness and surface area compared to KKI
and OASIS with more widespread and lower reliability, likely due to HCP’s more advanced acquisition
protocol. KKI had the lowest compatibility for the accumbens, particularly when comparing v5.3 to the
newer versions, where compatibility was poor/ moderate as opposed to moderate/good for OASIS and HCP,
both of which are from Siemens scanners, compared to Philips for KKI. Reliability across subcortical regions
and cohorts showed generally consistent good to excellent reliability, although v5.3 was most variable for
the hippocampus and accumbens.
To assess if our main analysis generalizes to other data sets, we performed a replication analysis for
reliability and compatibility using the HNU cohort-a data set composed of 30 participants with a 10 testretest design within 40 days of the initial baseline scan. While the results for reliability and compatibility
of surface area and subcortical volumes were largely in line with our main analysis, we observed distinct
trends in the replication data set for cortical thickness. For example, whereas our main analysis showed
generally good to excellent reliability of all measures for cortical thickness, reliability assessed in the HNU
data set showed lower, more widespread differences across the versions, where some regions even had
poor reliability. Between version analysis in the replication data set did not show the same distinctly lower
compatibility between v7.1 and the earlier versions for the cingulate regions. Instead, the replication analysis
showed more widespread discrepancy between v5.3 and the later versions. Differences in compatibility
could be attributed to a smaller sample size ( n = 27 ), as we only performed this analysis using the first
time point to replicate the methods in our main data sets. (Individual data set variation can be observed in
the supplementary materials, Figures S1-S8, Tables S3-S6.) However, despite the smaller sample size, the
study design lends itself to more stable within-version reliability measures, due to the increased number of
repeated measures. This suggests that there may be another source of variation that is not accounted for,
such as voxel thickness or scanner manufacturer, particularly impacting cortical thickness. KKI and OASIS
75

have anisotropic voxel sizes, with thickness being 1.2-1.25 mm, while HNU has isotropic 1 mm voxels. Also,
while HCP, and OASIS were scanned on Siemens and KKI on Philips, HNU used a GE scanner. Interscanner
variability of local thickness between Siemens and GE scanners, for example, was found to be on average
0.15 mm in Han et al. (2006) (X. Han et al., 2006), and differences in volumetric measures between all three
different platforms was observed in Jovicich et al. (2009) (Jovicich et al., 2009).
One limitation of our study was that there was no available higher-resolution or postmortem ground
truth data to know which FreeSurfer version most represents true anatomical structure. However, given
that many of these measures have been widely studied regarding their relationship with age, even in the
absence of postmortem or higher resolution data (Fischl, 2012; Frangou et al., 2022; Salat et al., 2004), we
instead assess age associations to gauge the downstream consequences of version differences. Versionrelated differences in FreeSurfer metrics between cases and controls have been assessed in Filip et al., 2022
(Filip et al., 2022). In their work, Filip and colleagues assess group differences between nine preselected
cortical and subcortical volumes of patients with type 1 diabetes and those of controls across the latest
FreeSurfer versions. They found the statistical significance between groups was dependent on version;
notably, analyses run using v7.1 metrics did not replicate the results of older versions. Our compatibility
findings highlight specifically the regions for which effects differ. Our work also highlights the dampened
effects that might be expected with v7.1, suggesting larger sample sizes might be needed to find similar
effects, than what might be expected from power calculations using v5.3 results.
We also performed QC of regional parcellations to rule out any spurious associations with gross missegmentations. One example worth noting is that v7.1 and v6.0 may be better able to handle images with
lower quality and/or motion as evidenced by one subject in our UK Biobank subset that failed in v5.3 but
not in the newer versions (Figure 4b). This could be due to the improved error handling of the Talairach
registration: if one registration fails, v7.1 and v6.0 would try an older atlas. Another example involved the
left middle temporal gyrus, which is often susceptible to underestimations due to the spillage/overestimation
76

of the banks of the superior temporal sulcus into that gyrus. This occurred at approximately the same rate
across versions. When associating the thickness of both the left banks of the superior temporal sulcus and
the middle temporal gyrus with age before QC, all versions reveal significant associations for both regions.
After removing subjects encountering this issue, although the direction of the effects stayed the same,
neither region was associated with age in any of the versions. While this may be due to a reduced sample
size and study power, it is also possible that findings in these regions may not represent true anatomical
structure, and may instead be due to common segmentation errors. It is also worth noting that our results
are solely based on the DK atlas (Desikan et al., 2006) and translation to other atlases may not apply. We
chose the DK atlas as it consists of a set of coarse regions defined by anatomical landmarks that can be
reasonably quality controlled. Most other atlases, while possibly more precise, define finer parcellations
based on cortical function, connectivity, topography, myelin, or a combination thereof (Glasser et al., 2016;
Schaefer et al., 2018). Visual QC by region may not be readily possible when cortical parcellations are
finer and there are over 100 regions in each hemisphere, so version performance of segmentation accuracy
may be more difficult to compare. Our data sets were exclusively from adults without major neurological
abnormalities, so our findings may not necessarily generalize to cohorts of young children, adolescents, or
individuals with significant brain abnormalities. Finally, we recognize that repeatability is an important
metric, and differences in repeatability may be explained, in part, by differences in operating systems used.
While Tustison et al. (2014) (Tustison et al., 2014) found good repeatability for FreeSurfer v5.3, users of
multiple workstations should exercise caution when pooling data run on various machines, as differences in
floating point precision may affect reproducibility of these measures (Glatard et al., 2015). Containerization
packages, such as Docker or Singularity (Matelsky et al., 2018), help mitigate differences in environment
along with differences in version, which are quantified in this manuscript.
Overall, we find generally high within-version reliability across most versions and data sets, and
many advantages to using FreeSurfer v7.1 over older versions for adult neuroimaging studies. However,
77

considerable differences are observed when analyzing between-version compatibility for regional cortical
thickness, surface area, and subcortical volumes. It is important to consider these compatibility differences
when pooling data or statistical inferences across software versions, and when comparing findings across
published works, especially for those regions with lower compatibility. Understanding these differences
may help researchers to make informed decisions on study design and provide insight into reproducibility
issues.
2.6 Acknowledgements
This research was funded by NIH grants R01MH117601, R01AG059874, R01AG058854, R01MH121246, and
P41EB015922. HCP data were provided [in part] by the Human Connectome Project, WU-Minn Consortium
(Principal Investigators: David van Essen and Kamil Ugurbil; U54MH091657) funded by the 16 NIH Institutes
and Centers that support the NIH Blueprint for Neuroscience Research and by the McDonnell Center for
Systems Neuroscience at Washington University. KKI was supported by NIH grants NCRR P41 RR015241
(Peter C.M. van Zijl), R01NS056307 (Jerry Prince), R21NS064534 (Bennett A. Landman/Jerry L. Prince),
R03EB01246 (Bennett A. Landman). OASIS: Cross-Sectional: Principal Investigators: D. Marcus, R. Buckner,
J. Csernansky, J. Morris; P50 AG05681, P01 AG03991, P01 AG026276, R01 AG021910, P20 MH071616, and
U24 RR021382.
78

79

2.7 Appendix A
Figure 2.8: Regional inter-version agreement (compatibility) for the HCP dataset. Bilateral, left, and right
ICC2 values comparing cortical thickness (A, B, C), cortical surface area (D, E, F), and subcortical volumes
(G, H, I) between versions. Outer concentric circles represent lower ICC2 values, truncated at 0.50, while
the center represents ICC2=1. Regions with the lowest compatibility differ for cortical thickness and surface
area.
80

Figure 2.9: Regional intra-version agreement (reliability) for the HCP dataset. Bilateral, left, and right ICC3
values comparing cortical thickness (A, B, C), cortical surface area (D, E, F), and subcortical volumes (G, H,
I) between versions. Outer concentric circles represent smaller ICC3 values, truncated at 0.70. Regions with
the lowest reliability differ for cortical thickness and surface area.
81

Figure 2.10: Regional inter-version agreement (compatibility) for the KKI dataset. Bilateral, left, and right
ICC2 values comparing cortical thickness (A, B, C), cortical surface area (D, E, F), and subcortical volumes
(G, H, I) between versions. Outer concentric circles represent lower ICC2 values, truncated at 0.50, while
the center represents ICC2=1. Regions with the lowest compatibility differ for cortical thickness and surface
area.
82

Figure 2.11: Regional intra-version agreement (reliability) for the KKI dataset. Bilateral, left, and right ICC3
values comparing cortical thickness (A, B, C), cortical surface area (D, E, F), and subcortical volumes (G, H,
I) between versions. Outer concentric circles represent smaller ICC3 values, truncated at 0.70. Regions with
the lowest reliability differ for cortical thickness and surface area.
83

Figure 2.12: Regional inter-version agreement (compatibility) for the OASIS dataset. Bilateral, left, and
right ICC2 values comparing cortical thickness (A, B, C), cortical surface area (D, E, F), and subcortical
volumes (G, H, I) between versions. Outer concentric circles represent lower ICC2 values, truncated at 0.50,
while the center represents ICC2=1. Regions with the lowest compatibility differ for cortical thickness and
surface area.
84

Figure 2.13: Bilateral, left, and right ICC3 values comparing cortical thickness (A, B, C), cortical surface area
(D, E, F), and subcortical volumes (G, H, I) between versions. Outer concentric circles represent smaller
ICC3 values, truncated at 0.70. Regions with the lowest reliability differ for cortical thickness and surface
area.
85

Figure 2.14: Regional inter-version agreement (compatibility) for the HNU dataset. Bilateral, left, and right
ICC2 values comparing cortical thickness (A, B, C), cortical surface area (D, E, F), and subcortical volumes
(G, H, I) between versions. Outer concentric circles represent lower ICC2 values, truncated at 0.50, while
the center represents ICC2=1. Regions with the lowest compatibility differ for cortical thickness and surface
area.
86

Figure 2.15: Regional reliability ICC3 measures for the HNU dataset using a 10 test-retest design. Bilateral,
left, and right ICC3 values comparing cortical thickness (A, B, C), cortical surface area (D, E, F), and
subcortical volumes (G, H, I) between versions. Outer concentric circles represent smaller ICC3 values,
truncated at 0.70. Regions with the lowest reliability differ for cortical thickness and surface area.
87

Figure 2.16: Difference in z-statistics () comparing v7.1 to the two previous versions.
Figure 2.17: Difference in z-statistics () comparing v6.0 to v5.3.
88

Figure 2.18: Harmonized v7.1 vs. harmonized v7.1 with v5.3 cohort mixtures. FDR-corrected significant zstatistics for A. thickness, B. surface area, and C. subcortical volume associations with age. Top box indicates
the reference association results (all cohorts run through v7.1 before harmonization), middle box indicates
associations after a single cohort was run through v5.3 (75% v7.1) before harmonization, and bottom box
indicates association results after harmonizing two cohorts run with v5.3, and the other two with v7.1 (50%
v7.1). The accumbens association with age was only significant in the v71(KKI&OASIS)+v53(HCP&HNU)
comparison (z = -2.87; q = 0.04); not pictured.
89

Figure 2.19: Harmonized v7.1 vs. harmonized v7.1 with v6.0 cohort mixtures. FDR-corrected significant
z-statistics for A. thickness, B. surface area, and C. subcortical volume associations with age. Top box
indicates the reference association results (all cohorts run through v7.1 before harmonization), middle box
indicates associations after a single cohort was run through v6.0 (75% v7.1) before harmonization, and
bottom box indicates association results after harmonizing two cohorts run with v6.0, and the other two
with v7.1 (50% v7.1).
90

Figure 2.20: Harmonized v7.1 vs. harmonized v7.1 with v5.3 cohort mixtures. Uncorrected significant
z-statistics for A. thickness, B. surface area, and C. subcortical volume associations with age. Top box
indicates the reference association results (all cohorts run through v7.1 before harmonization), middle box
indicates associations after a single cohort was run through v5.3 (75% v7.1) before harmonization, and
bottom box indicates association results after harmonizing two cohorts run with v5.3, and the other two
with v7.1 (50% v7.1). Significant associations between the accumbens volume and age were found in the
v71+v53(HCP) comparison (z = -2.71; p = 0.008) and the v71(KKI&OASIS)+v53(HCP&HNU) comparison (z
= -2.87; p = 0.005); not pictured.
91

Figure 2.21: Harmonized v7.1 vs. harmonized v7.1 with v6.0 cohort mixtures. Uncorrected significant
z-statistics for A. thickness, B. surface area, and C. subcortical volume associations with age. Top box
indicates the reference association results (all cohorts run through v7.1 before harmonization), middle box
indicates associations after a single cohort was run through v6.0 (75% v7.1) before harmonization, and
bottom box indicates association results after harmonizing two cohorts run with v6.0, and the other two
with v7.1 (50% v7.1). Significant associations between the accumbens volume and age were found for the
v71+v60(HCP) comparison (z = -2.10; p = 0.04) and the v71(KKI&OASIS)+v60(HCP&HNU) comparison (z =
-2.22; p = 0.03); not pictured.
92

Table 2.2: Runtimes for test-retest datasets across all three versions.
Cohorts
v5.3
runtime
(hours)
v6.0
runtime
(hours)
v7.1
runtime
(hours)
Avg 6.64 8.90 6.27
HCP Min 5.44 6.57 5.46
Max 14.20 11.56 7.59
Avg 7.80 9.80 6.67
KKI Min 6.51 7.77 5.75
Max 9.61 11.50 7.53
Avg 3.93 9.41 6.50
OASIS Min 3.19 7.49 5.65
Max 6.46 11.11 7.81
Table 2.3: Mean Euler numbers extracted from the nofix surface and the final surface. Parenthesis indicate
the min, and max values.
HCP KKI OASIS Measure
v5.3 v6.0 v7.1 v5.3 v6.0 v7.1 v5.3 v6.0 v7.1
lh.orig.nofix
holes
20
(7,46)
17
(4,38)
8
(2,18)
21
(7,51)
19
(4,46)
11
(3,33)
21
(5,39)
20
(7,47)
10
(1,23)
rh.orig.nofix
holes
22
(7,50)
17
(5,33)
9
(1,20)
23
(10,51)
21
(9,42)
10
(2,21)
18
(5,40)
16
(3,36)
8
(1,20)
total orig.nofix
holes
42
(19,83)
34
(9,60)
17
(5,33)
44
(21,102)
40
(13,87)
20
(5,54)
40
(17,76)
36
(10,83)
18
(6,38)
lh.orig.nofix
euler
-38
(-90,-12)
-31
(-74,-6)
-15
(-34,-2)
-40
(-100,-12)
-37
(-90,-6)
-19
(-64,-4)
-40
(-76,-8)
-38
(-92,-12)
-17
(-44,0)
rh.orig.nofix
euler
-41
(-98,-12)
-32
(-64,-8)
-15
(-38,0)
-43
(-100,-18)
-39
(-82,-16)
-17
(-40,-2)
-35
(-78,-8)
-29
(-70,-4)
-14
(-38,0)
total orig.nofix
euler
-79
(-162,-34)
-63
(-116,-14)
-30
(-62,-6)
-84
(-200,-38)
-76
(-170,-22)
-36
(-104,-6)
-75
(-148,-30)
-67
(-162,-16)
-32
(-72,-8)
lh.orig holes 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0)
rh.orig holes 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0)
total orig holes 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0) 0 (0,0)
lh.orig euler 2 (2,2) 2 (2,2) 2 (2,2) 2 (2,2) 2 (2,2) 2 (2,2) 2 (2,2) 2 (2,2) 2 (2,2)
rh.orig euler 2 (2,2) 2 (2,2) 2 (2,2) 2 (2,2) 2 (2,2) 2 (2,2) 2 (2,2) 2 (2,2) 2 (2,2)
93

Chapter 3
P-wave duration is associated with aging patterns in structural brain
networks
This section is adapted from:
Haddad E*, Matloff W, Park G, Liu M, Jahanshad N, Kim H. P-wave duration is associated with aging
patterns in structural brain networks. Journal of the American Heart Association (Article Under Review)
3.1 Abstract
Impaired cardiac function is associated with cognitive impairment and brain imaging features of aging.
Cardiac arrhythmias, including atrial fibrillation, are implicated in clinical and subclinical brain injuries.
Even in the absence of a clinical diagnosis, subclinical or prodromal substrates of arrhythmias, including an
abnormally long or short P-wave duration (PWD), a measure associated with atrial abnormalities, have been
associated with stroke and cognitive decline. However, the extent to which PWD has subclinical influences
on overall aging patterns of the brain is not clearly understood. Here, using neuroimaging and ECG data
from the UK Biobank, we use a novel regional "brain age" method to identify the brain aging networks
associated with abnormal PWD. We find associations between short PWD and accelerated brain aging in
the sensorimotor, frontoparietal, ventral attention, and dorsal attention networks, even in the absence of
overt cardiac diseases. These findings contribute to our understanding of the relationship between PWD
94

and structural brain aging. This work emphasizes the need for continued study designs which consider
brain based outcomes related to abnormally short PWD.
3.2 Clinical Perspective
What is new? This study contributes to our understanding of the relationship between P-wave duration
and outcomes of brain aging. We find the most pronounced associations between abnormally short P-wave
duration and brain aging in the structure of the sensorimotor, frontoparietal, ventral attention, and dorsal
attention networks, even in the absence of overt cardiac disease.
What are the clinical implications? Research on the clinical outcomes related to abnormally short
P-wave duration is needed to establish causal factors involved in accelerated brain aging and to better
inform clinical practice aimed at prediction and prevention of brain aging.
3.3 Introduction
The brain accounts for only about 2% of the body’s mass, yet it is estimated to receive 12% of cardiac output,
and use 20% of the body’s oxygen (Raichle & Gusnard, 2002; Williams & Leggett, 1989). Given that this
cardiac output provides the means by which cerebral blood flow is able to perfuse brain tissue, inevitably
the heart and brain are intricately linked. A growing body of literature suggests that cardiovascular diseases
(CVDs)-are associated with cognitive decline and brain imaging-derived features of aging (Friedman et al.,
2014; Tublin et al., 2019). For example, heart failure (HF) is associated with dementia and also cognitive
impairment with prevalences of co-occurrence across population studies ranging from 25-75%, and many
studies report an improvement in cognitive outcomes after CVD treatment (Qiu & Fratiglioni, 2015; M.
Yang et al., 2021). A recent meta-analysis found a 50% increased risk of mild cognitive impairment and
dementia in individuals with coronary artery disease (CAD). Some evidence suggests the severity of
95

coronary artery calcium, a marker of subclinical CAD, is positively associated with dementia risk (Xia
et al., 2020), highlighting the potential of subclinical abnormalities in biomarkers detected in advance of
major disease outcomes to be used as early indicators for interventions. Another cardiac condition strongly
tied to brain health is atrial fibrillation (AF), the most common cardiac arrhythmia (Bunch, 2020). Besides
increasing cognitive impairment and dementia risk by up to 40% (Chang Wong & Chang Chui, 2022; Qiu &
Fratiglioni, 2015), AF and other arrhythmias increase the risk of cardioembolic stroke and also subclinical
brain injuries(Conen et al., 2019; van der Velpen et al., 2017). While these associations emphasize the role
of cardiac health on brain health, in reality, the heart-brain axis is bidirectional. Communication between
the immune system and the brain also influences the pathophysiology of cardiovascular risk factors and
disease via an integrated network of autonomic nerves, hormones, and inflammatory factors(Hu et al., 2023;
Saeed et al., 2023). Thus, understanding the relationship between markers of cardiovascular health and
markers of neurological health may help us to gain further insight into the heart-brain axis.
P-wave indices on an electrocardiogram (ECG), specifically abnormalities in P-wave duration (PWD),
offer valuable information of the electrical activity within the heart’s atria. The PWD represents the current
moving from the sinoatrial node to the atrioventricular node and characterizes atrial depolarization (Lip
et al., 2016). Anomalies in PWD, whether shorter or longer than the norm, are believed to reflect structural
or physiological alterations in the left or right atria. These subtle variations thus provide markers of cardiac
pathology and can serve as early indicators of various cardiovascular diseases, including AF and other
arrhythmias, as well as cardiovascular risk factors and related deaths (He et al., 2017; Kosar et al., 2008;
Magnani et al., 2009, 2015; Nielsen et al., 2015; Uyarel et al., 2005). PWD abnormalities have also been
associated with brain outcomes. A recent study, supported by a large community-based cohort study
spanning 25 years, has shown a 1.6-fold heightened risk of dementia associated with prolonged PWD,
indicative of interatrial block (Power et al., 2022), independently of stroke and AF (Gutierrez et al., 2019).
A subsequent cross-sectional study examining the link between abnormal P-wave indices and brain MRI
96

outcomes showed elevated odds of cerebrovascular-related injuries, including cortical and lacunar infarcts,
lobar, and subcortical microhemorrhages (Reyes et al., 2023). Thus, exploring the interaction between
abnormal PWD and measures of brain aging, and in particular, within different functional networks, stands
to further our understanding of how cardiac irregularities may influence cognitive outcomes, especially in
the absence of or during the prodromal phase of AF.
Sensory, motor, and cognitive tasks are subserved by the synchronous activity of specific sets of brain
regions, known as functional networks. These include networks responsible for movement and sensation
such as the sensorimotor, visual, and auditory networks, and also higher-level cognitive functions such as
attention, executive control, and language (also known as association networks) (Seitzman et al., 2019). A
basic tenet of systems neuroscience is that structure is related to function. Pioneering work by Segall et
al., (2012) demonstrated correspondence between spatial gray matter density and functional components
derived from independent component analysis (Segall et al., 2012; Sui et al., 2014). Another seminal study
provided evidence for the "network degeneration hypothesis", which posits that neurodegenerative diseases
can be characterized by neuropathology within specific brain networks - resulting in functional impairment
caused by network specific degeneration (Drzezga, 2018). Work by Seeley et al. (2009) provided support for
this hypothesis as they discovered a direct link between gray matter structure and intrinsic connectivity
by assessing this relationship across 5 different dementia subtypes, including Alzheimer’s disease (AD).
The authors first identified syndrome specific atrophy patterns by comparing patients to healthy controls
and then used the regions which had the maximum extent of atrophy as seed regions for both structural
covariance and intrinsic functional connectivity analysis in healthy controls. Considerable overlap was
found between gray matter atrophy patterns and intrinsic connectivity in areas involved in cognitive
and behavioral outcomes affected by each specific dementia subtype. This work highlights the utility of
investigating neurodegenerative patterns in functional networks known to be associated with cognitive
97

functions which may be impaired as a result of accelerated brain aging (Alexander-Bloch et al., 2013; K. Liu
et al., 2017; Seeley et al., 2009).
Brain aging involves complex processes characterized by region-specific and nonlinear patterns of
morphological changes(Cole & Franke, 2017). Structural magnetic resonance imaging (MRI) can be leveraged
to extract hundreds of regional morphological metrics, many of which are associated with specific functional
and disease related processes. For example, cortical thickness declines with normal aging with the most
pronounced changes occurring in the frontal lobe (M. E. MacDonald & Pike, 2021). In AD, circumscribed
patterns of cortical thinning are able to predict diagnoses up to a decade in advance of diagnoses (Dickerson
et al., 2011; Satizabal et al., 2024). Other markers, such as gray/white matter intensity ratios, are emerging as
sensitive markers of disease severity, even beyond traditional measures of atrophy such as cortical thickness
(Putcha et al., 2023). However, inferences regarding rates of decline typically require longitudinal data,
which can be costly and time consuming to acquire. Fortunately, the development of "brain age" methods
as a marker for brain health has offered a way to infer rates of degeneration with cross-sectional data,
allowing researchers to leverage large datasets to make inferences regarding "accelerated" or "slowed"
brain aging. Brain age is derived from machine learning models, which often rely on regression techniques
that model brain features as independent variables and chronological age as the dependent variable in a
neurologically healthy "training" set. These models are then applied to an independent or "test" sample and
the difference between an individual’s predicted age and chronological age, also known as the "brain age
index" or residual, is used as an aging biomarker (Baecker, Garcia-Dias, et al., 2021; Bashyam et al., 2020;
Bocancea et al., 2021; Cole & Franke, 2017; Franke & Gaser, 2019; Higgins-Chen et al., 2021). Advances in
network architectures, such as graph convolutional networks, that take rich surface based measures (i.e.
cortical thickness) as inputs, coupled with the ability to partition functionally relevant brain networks,
allows one to investigate influential factors on brain aging networks.
98

Given the elevated risk abnormalities in PWD pose on dementia outcomes and brain imaging-derived
features of aging, a further investigation into its effects on brain aging networks may help us gain insight
into network degeneration and functional outcomes. Here, in a subset of UK Biobank participants who
have both ECG and brain MRI available (Miller et al., 2016b), we aim to identify if and how abnormal PWD
may relate to accelerated brain aging, even in the absence of overt cardiac diseases. In particular, we use an
advanced deep learning method capable of estimating "regional brain age" to characterize brain structural
vulnerability related to atrial remodeling/dysfunction as measured by abnormal PWD.
3.4 Methods
3.4.1 Study population
The UK Biobank is a large population-based study that has collected deep phenotypic and genetic data
from approximately 500,000 community dwelling adults in the United Kingdom (Bycroft et al., 2018). Its
dense phenotyping has resulted in extensive information on health status and lifestyle, in addition to the
collection of biological, physical, and cognitive assessments. Here, data from 40,678 participants with
imaging and ECG were initially considered for this study. 6,778 ECG diagnoses related to poor readings
were excluded, resulting in a total N=33,900. Subjects with neurological disorders, putative sex chromosome
aneuploidy, excess relatives, and missing covariates were also excluded (N=29,364). A total of 12,819 of
these subjects had our outcome variable, regional brain age, available. Lastly, to assess subclinical effects of
PWD on brain aging, subjects with ECG readings indicating an acute myocardial infarction or bifascicular
block were excluded. This resulted in a total of 12,762 subjects for analysis. To assess if abnormal PWD
affects regional brain structure even in the absence of overt cardiac-related diseases, we performed the same
analysis in a CVD-control subset for which we excluded subjects with AF, CAD, HF, and chronic kidney
99

disease resulting in a subset total of 11,771. Conditions were categorized as in Khurshid et al (Khurshid
et al., 2018). Details about inclusion/exclusion criteria are provided in Figure 3.1.
Figure 3.1: Inclusion flowchart. Self reported non-cancer illness codes (1082, 1083, 1240, 1244, 1245, 1246,
1247, 1256, 1258, 1259, 1261, 1262, 1263, 1264, 1266, 1289, 1397, 1425, 1433, 1659) and ICD10 codes for
dementia and Parkinson disease.
3.4.2 MRI acquisition and preprocessing
We used the CIVET pipeline (Ad-Dabbagh et al., 2006) on T1-weighted brain MRI to extract cortical
features including thickness and gray/white matter intensity. This pipeline includes the following serial
steps: non-uniform intensity correction (Sled et al., 1998) to correct for inhomogeneities produced by
100

scanner bias fields, brain extraction (Smith, 2002) to remove non-brain tissue (i.e. skull), registration to
a stereotaxic space (Collins et al., 1994) to ensure spatial correspondence between subjects, brain tissue
segmentation (Zijdenbos et al., 1998) to isolate tissue types, and reconstruction of inner and outer cortical
surfaces (Kim et al., 2005) to extract gray and white matter surfaces (resulting in 40,962 vertices in each
hemisphere). These cortical surface models underwent an iterative surface registration process to ensure
optimal correspondence at each vertex across individuals (Lyttelton et al., 2007; D. MacDonald et al., 2000).
The resulting surfaces represent the gray and white matter surfaces as triangular meshes in order to
extract commonly derived neuroimaging features. We selected cortical thickness given its well-documented
associations with aging and neurodegeneration (Dickerson et al., 2011; M. E. MacDonald & Pike, 2021;
Satizabal et al., 2024), as well as the gray/white matter intensity ratio (GW IR), which has been shown to
provide complimentary information regarding disease severity (Putcha et al., 2023). Cortical thickness
measurements were obtained by calculating the Euclidean distance between the vertices of the inner
cortical surface and their corresponding vertices on the outer cortical surface. In other words, cortical
thickness is defined as the distance between the inner (white matter) surface and the outer pial (gray matter)
surface of the brain. The GM/WM intensity ratio quantifies the signal contrast between the gray and white
matter by sampling 1 mm inside and 1 mm outside the inner white matter surface (Lewis et al., 2018). By
combining these two measures, our brain age model can leverage a comprehensive representation of brain
structure, potentially capturing subtle age-related changes that may not be apparent when either measure
is considered in isolation.
3.4.3 Regional brain age index (BAI)
To develop predictive models for regional brain age indices (BAI), we divided the cortical surface into ten
functional subregions based on Yeo et al (Yeo et al., 2011) and the AAL cortical parcellation atlas (TzourioMazoyer et al., 2002). These subregions include the sensorimotor, frontoparietal, dorsal attention, ventral
101

attention, default mode, salience, language, auditory, visual, and limbic networks. Two additional regions
have been defined - one based on the total cortical thickness and the other based on the cortical regions
associated with Alzheimer’s disease (AD signature region) (Dickerson et al., 2009). Regional brain ages
were extracted using Graph Convolutional Networks (GCNs) (Defferrard et al., 2016; Shuman et al., 2013),
exploiting the graph structure of the data. We chose a GCN over a convolutional neural network (CNN)
for brain age prediction because we used cortical thickness and gray/white matter intensity ratio derived
from surface models. The surface model of the brain is naturally represented as a graph, where vertices
(nodes) represent the intersection of three triangles on the tessellated 3D mesh, and the edges represent the
vectors of the aforementioned triangles. GCNs are explicitly designed to operate on graph-structured data,
making them the ideal choice for surface-based brain MRI data. In contrast, CNNs are primarily designed
for grid-like data (such as images) and may not fully capture the complex topological relationships inherent
in the 3D representation of brain structure. Therefore, GCNs provide a more natural and appropriate
framework for brain age prediction in the current study.
In our GCN model, cortical thickness (Thambisetty et al., 2010) and GW IR served as the signal at each
node. The GCNs used graph Fourier transforms, filtering, and pooling operations for feature aggregation.
The GCN architecture included a graph convolutional layer, a ReLU (Rectified Linear Unit) activation
function, a graph max pooling operation, and a fully connected layer for brain age prediction. The overall
flow of the GCNs model is shown in Figure 3.2. The training process involved mean square error as the loss
function, the Adam optimizer, 800 epochs, a learning rate of 10 e-6, L2 regularization to prevent overfitting,
and a batch size of 2. A 5-fold cross-validation was performed using 17,791 individuals of European ancestry
(52.7% female with a mean age and standard deviation of 63.15 ± 7.42 years) who were considered to be
neurologically healthy as defined by ICD and self report, resulting in an ensemble of five trained models.
BAls were calculated by subtracting the chronological age from the predicted brain age. A positive BAI
indicates accelerated aging, while a negative BAI suggests decelerated aging. Performance metrics are
102

shown in Table 3.1. To address regression dilution bias, we used the linear trend removal method proposed
by Smith et al (Smith et al., 2019). Linear trend removal involves regressing the trend of BAIs on age to
obtain corrected BAIs, eliminating age-related bias and ensuring relative brain health status independent of
age.
Table 3.1: Performance metrics for regional BAI networks. R-value (correlation coefficient); MAE: mean
absolute error.
Metric MAE
(years) R-value
Global 2.875 0.893
Sensorimotor network 2.977 0.881
Fronto-parietal network 2.903 0.887
Dorsal attention network 2.895 0.890
Ventral attention network 2.825 0.894
Default mode network 2.928 0.886
Salience network 3.000 0.883
Language network 2.893 0.892
Auditory network 2.934 0.888
Visual network 2.902 0.890
Limbic network 3.043 0.879
AD signature 3.086 0.886
103

Figure 3.2: Methodological overview of BAI extraction. T1w brain aging features including cortical thickness
and grey/white matter intensity ratio were inputs into a graph convolutional network resulting in a fully
connected layer and predicted "brain age" outputs. Chronological age was subtracted from predicted brain
age resulting in 12 regional brain age indices (BAI). Chronological age was regressed out from each BAI
and used for analysis. Linear modeling was performed with PWD as a predictor variable while correcting
for possible confounders by including them as covariates. PWD represents the current moving from the
sinoatrial node to the atrioventricular node and characterizes atrial depolarization.
3.4.4 Cognitive outcomes
Composite averages were calculated for four cognitive domains: memory, executive function, processing
speed, and reasoning. Memory was assessed using numeric memory and paired associate learning tasks.
104

Executive function was assessed using the trail making A and B tasks in addition to tower rearranging.
Processing speed used reaction times and symbol digit substitution and lastly, reasoning used fluid intelligence and matrix pattern completion tasks. Respective data fields for each of these domains may be found
in Table 3.3. All scores were standardized such that higher values indicate a better cognitive score (ex:
reaction time was multiplied by -1). Values from all four of these categories were additionally averaged to
test a fifth cognitive variable - total cognition.
3.4.5 ECG measures
Resting 12-lead ECGs and interval measurements were assessed using the Cardiosoft v6 program from GE
Healthcare. The PWD was measured as the duration from P-Onset to P-Offset (https://biobank.ndph.ox.ac
.uk/showcase/ukb/docs/CardiosoftFormatECG.pdf).
3.4.6 Statistical analysis
Principal component analysis (PCA) was run on mixed data, a combination of numerical and categorical
variables related to scanner biases affecting the T1w image (Alfaro-Almagro et al., 2021), using the PCAmixdata package (Chavent et al., 2014). Linear regressions for each regional BAI were run with the following
covariates: 4 MRI PCA components that cumulatively explained at least 80% of the variance, age, sex,
college attendance, the presence of hypertension, diabetes, hypercholesterolemia, sleep apnea, heel bone
mineral density T-score, smoking status, alcohol consumption, body mass index (BMI), ApoE4 carrier status,
systolic blood pressure, diastolic blood pressure, and PWD. We chose to model continuous PWD so as to not
bias our outcomes with previously defined "normal" ranges, as these largely vary across populations and by
age, sex, and ethnicity (Nielsen et al., 2015; Soliman et al., 2013). CAD, HF, AF, and chronic kidney disease
were also included in the models with all subjects. Respective datafield IDs may be found in Table 3.3.
Cognitive associations were performed between PWD and regional BAI. All models included age, sex, and
105

college education as a binary variable. Models assessing the relationship between cognition and regional
BAI included four MRI scanner PCA components. Benjamini & Hochberg correction for multiple testing
was performed across all 10 regional BAI measures, whole brain, and AD signature BAI for the following:
1) regional BAI associations within all covariates including PWD, 2) regional BAI associations across 5
cognitive measures (60 tests total), and 3) PWD associations with cognition. Exploratory analyses were
also performed to test the inclusion of a quadratic PWD term as well as PWD interactions with AF and sex.
Lastly, metabolic disease sensitivity analysis was performed where models were stratified by the presence
of hypertension and diabetes.
3.5 Results
3.5.1 Demographic characteristics
The 12,762 subjects selected for analysis were 52.6% female with a mean age and standard deviation of
63 ± 7.36 years and a range between 45.17 and 79.84 years. The subset without major cardiac conditions (
N = 11, 771 ) were 54.2% female with a mean age and standard deviation of 62.7 ± 7.33. Demographic
characteristics of all subjects stratified by percentiles of PWD may be found in Table 3.2. In this dataset,
the median percentile (40-60th) corresponded to a PWD of 96-100ms. In general, subjects with low PWD
were more likely to be female, whereas those with high PWD were more likely to be male (Wilcoxon test,
W = 16987621, p < 0.0001 ) (Figure 3.3). A higher incidence of AF, CAD, HF, and chronic kidney disease
was observed in the lowest (< 5%) and highest (> 95%) percentiles.
106

Table 3.2: Demographic characteristics stratified by percentiles of P-wave duration with associated interval limits listed. Continuous values represent
the mean and standard deviation in brackets. Discrete values represent the total N and percentage within group in parentheses.
Characteristic
<5th
(40-64ms)
N = 560
5-20th
(66-84ms)
N = 1892
20-40th
(86-94ms)
N = 2325
40-60th
(96-100ms)
N = 2256
60-80th
(102-108ms)
N = 2732
80-95th
(110-118ms)
N = 2112
>95th
(120-146ms)
N = 885
Overall
N = 12762
Age 63.95 [7.56] 62.49 [7.46] 62.13 [7.54] 62.59 [7.24] 63.15 [7.28] 63.64 [7.20] 64.42 [7.06] 62.97 [7.36]
Female 308 (55%) 1,143 (60%) 1,402 (60%) 1,232 (55%) 1,415 (52%) 910 (43%) 304 (34%) 6,714 (53%)
Male 252 (45%) 749 (40%) 923 (40%) 1,024 (45%) 1,317 (48%) 1,202 (57%) 581 (66%) 6,048 (47%)
College 232 (41%) 783 (41%) 983 (42%) 1,009 (45%) 1,220 (45%) 983 (47%) 422 (48%) 5,632 (44%)
Coronary Artery Disease 38 (6.8%) 108 (5.7%) 113 (4.9%) 94 (4.2%) 139 (5.1%) 130 (6.2%) 97 (11%) 719 (5.6%)
Heart Failure 4 (0.7%) 8 (0.4%) 14 (0.6%) 12 (0.5%) 10 (0.4%) 15 (0.7%) 11 (1.2%) 74 (0.6%)
Chronic Kidney Disease 7 (1.3%) 14 (0.7%) 11 (0.5%) 12 (0.5%) 19 (0.7%) 16 (0.8%) 14 (1.6%) 93 (0.7%)
Atrial Fibrillation 17 (3.0%) 45 (2.4%) 45 (1.9%) 30 (1.3%) 43 (1.6%) 51 (2.4%) 34 (3.8%) 265 (2.1%)
Hypertension 156 (28%) 441 (23%) 522 (22%) 506 (22%) 756 (28%) 622 (29%) 314 (35%) 3,317 (26%)
Hypercholesterolemia 84 (15%) 280 (15%) 337 (14%) 335 (15%) 428 (16%) 386 (18%) 208 (24%) 2,058 (16%)
Diabetes 33 (5.9%) 95 (5.0%) 108 (4.6%) 99 (4.4%) 156 (5.7%) 119 (5.6%) 64 (7.2%) 674 (5.3%)
Sleep Apnea 5 (0.9%) 19 (1.0%) 25 (1.1%) 13 (0.6%) 24 (0.9%) 24 (1.1%) 12 (1.4%) 122 (1.0%)
On Hypertension Medication 135 (24%) 379 (20%) 439 (19%) 406 (18%) 633 (23%) 549 (26%) 277 (31%) 2,818 (22%)
On Cholesterol Medication 131 (23%) 380 (20%) 444 (19%) 428 (19%) 575 (21%) 513 (24%) 269 (30%) 2,740 (21%)
Systolic Blood Pressure 136.12 [17.67] 134.82 [17.2] 135.62 [18.00] 136.17 [17.46] 137.59 [17.89] 138.76 [17.72] 139.11 [17.70] 136.80 [17.74]
Diastolic Blood Pressure 78.04 [9.45] 77.79 [9.69] 78.43 [9.95] 78.35 [9.78] 79.13 [10.22] 79.41 [9.57] 79.62 [10.27] 78.70 [9.90]
Bone Mineral Density T-score -0.23 [1.25] -0.22 [1.20] -0.33 [1.12] -0.27 [1.13] -0.24 [1.18] -0.19 [1.19] -0.04 [1.27] -0.24 [1.18]
Body Mass Index 26.89 [4.73] 26.54 [4.62] 26.26 [4.30] 26.18 [4.21] 26.59 [4.29] 26.90 [4.27] 27.50 [4.34] 26.58 [4.36]
Smoking Status
Never 347 (62%) 1,212 (64%) 1,495 (64%) 1,460 (65%) 1,692 (62%) 1,304 (62%) 529 (60%) 8,039 (63%)
Previous 192 (34%) 601 (32%) 711 (31%) 715 (32%) 916 (34%) 732 (35%) 314 (35%) 4,181 (33%)
Current 21 (3.8%) 79 (4.2%) 119 (5.1%) 81 (3.6%) 124 (4.5%) 76 (3.6%) 42 (4.7%) 542 (4.2%)
Alcohol Status
Infrequent 153 (27%) 546 (29%) 671 (29%) 613 (27%) 720 (26%) 530 (25%) 194 (22%) 3,427 (27%)
Occasional 321 (57%) 1,037 (55%) 1,283 (55%) 1,258 (56%) 1,512 (55%) 1,167 (55%) 518 (59%) 7,096 (56%)
Frequent 86 (15%) 309 (16%) 371 (16%) 385 (17%) 500 (18%) 415 (20%) 173 (20%) 2,239 (18%)
ApoE4 copies
0 418 (75%) 1,372 (73%) 1,673 (72%) 1,640 (73%) 1,975 (72%) 1,516 (72%) 652 (74%) 9,246 (72%)
1 130 (23%) 477 (25%) 601 (26%) 563 (25%) 690 (25%) 545 (26%) 220 (25%) 3,226 (25%)
2 12 (2.1%) 43 (2.3%) 51 (2.2%) 53 (2.3%) 67 (2.5%) 51 (2.4%) 13 (1.5%) 290 (2.3%)
107

Figure 3.3: P-wave duration plotted against age and stratified by sex. Colors denote P -wave duration
percentiles listed on the left hand side. Density plots across ages for each sex are depicted atop the scatters.
Density plot on the right displays P-wave duration for males (dashed line) and females (solid line). Note
higher densities of both high and low P-wave duration at older ages and across both sexes. Males tended to
have longer P-wave duration whereas females favored shorter.
3.5.2 PWD and regional brain aging
Regional BAI indices were negatively associated with PWD (i.e., shorter PWD associated with higher
BAI). Indices that survived multiple comparisons testing included the sensorimotor (full set: t = −3.06,
p = 0.002, q = 0.007; CVD-control subset: t = −2.64, p = 0.008, q = 0.025 ), fronto-parietal (full set:
t = −3.53, p = 4.1 × 10−04, q = 0.002; CVD-control subset: t = −3.07, p = 0.002, q = 0.010 ), dorsal
attention (full set: t = −4.11, p = 4.0 × 10−5
, q = 4.8 × 10−4
; CVD-control subset: t = −3.67, p = 2.4×
10−4
, q = 0.003 ), ventral attention (full set: t = −3.75, p = 1.8 × 10−4
, q = 0.001; CVD-control subset:
t = −3.03, p = 0.002, q = 0.010 ), and the language networks (full set: t = −2.92, p = 0.003, q = 0.008;
CVD-control subset not significant) (Figures 3.4 & 3.6).
108

Figure 3.4: Regional BAI association results with PWD as the variable of interest. A. T-statistics of regional
BAI (overlapping) associated with PWD. B. Standardized coefficients of PWD and all other regression
coefficients (with the exception of the four MRI PCA components explaining 88.3% of scanner related
variance) from all 12 models (whole brain BAI, AD signature region BAI, and 10 regional BAI). Shorter P-wave
duration was associated with a higher brain age. Regional BAI that survived multiple comparisons included
the fronto-parietal, dorsal attention, and ventral attention (middle panel), as well as the sensorimotor and
language networks (right panel). Covariates associated with a higher BAI include hypertension, diabetes,
smoking status, and frequent alcohol consumption. Covariates associated with a lower BAI include a higher
bone mineral density T-score, having a college education, and younger age. Filled points indicate effects
that survived multiple comparisons correction across all 12 BAls (within each covariate).
The factor that showed the highest BAI for all 12 networks in all subjects and the CVD-control subset
was diabetes. Hypertension and systolic blood pressure were also associated with a higher BAI for most
regions. Other factors associated with a higher BAI for most, if not all, regional BAI included currently or
previously smoking compared to never having smoked and frequent alcohol consumption compared to
infrequent consumption. Factors associated with a younger than actual brain age included a higher bone
mineral density T-score, having a college education, and lower chronological age. Biological sex showed
some regional variability where compared to females, males tended to have a higher BAI in the salience,
109

default mode, and dorsal attention networks. Conversely, males tended to have lower BAI in the ventral
attention and auditory networks compared to females.
Models including a quadratic PWD did not show any significant associations. Models including PWD
interactions with sex and atrial fibrillation similarly did not reveal any significant interactions.
3.5.3 Regional brain aging and cognition
The total number of subjects available for cognitive analysis from our initial set amounted to N = 4, 620
for average memory, N = 4, 447 for executive cognition, N = 4, 623 for processing speed, N = 4, 594 for
reasoning, and N = 4, 401 for total cognition. All regional BAI with the exception of the dorsal attention
network were associated with total cognition. All but the sensorimotor, dorsal attention, language, visual,
and AD signature regional BAI were associated with reasoning. All regional BAI were associated with
processing speed. All but the ventral attention, salience, and auditory networks were associated with
executive function. Lastly, only the default mode network was associated with average memory score
(Figure 3.5).
110

Figure 3.5: Regional BAI associations with composite cognitive measures - average memory, executive
function, processing speed, reasoning, and total cognition. Values displayed are standardized regression
coefficients. All models included age, sex, college attendance, and four MRI scanner PCA components
(explaining 88.3% of scanner related variance) as covariates. Higher cognitive scores were generally
associated with a lower brain age. Filled shape styles indicate varying levels of significance which survived
multiple comparisons correction across all 5 cognitive measures and all 12 BAl indices combined.
3.5.4 PWD and cognition
None of the composite cognitive measures were associated with PWD after correcting for age, sex, and
college education.
3.6 Discussion
Our study aimed to characterize the brain aging patterns and cognitive outcomes associated with PWD, an
ECG-surrogate of underlying atrial function. We found the following:
111

1.) A negative association between PWD and distinct patterns of regional brain aging exists. Several
brain networks, including the sensorimotor, frontoparietal, ventral attention, dorsal attention, and language
networks, were particularly vulnerable to short PWD. These networks were associated with global cognition
and also specific cognitive measures including executive function, processing speed, and reasoning.
2.) Our findings exist even when adjusting for cardiovascular risk factors, and when excluding major
cardiac conditions (with the exception of the language network), suggesting detrimental aging outcomes
may result from subclinically abnormal PWD.
3.) We did not find direct associations between PWD and cognition, but did find associations between
cognition and brain age measures. This suggests abnormal PWD may exert its effects on the brain in
advance of detectable cognitive decline, but longitudinal studies are needed to confirm this hypothesis.
Abnormal PWD is thought to reflect underlying atrial abnormalities, which can include atrial dilation,
atrial muscular hypertrophy, elevated atrial pressure, impaired ventricular distensibility, or delayed intra/inter atrial contraction (Hancock et al., 2009). Prolonged PWD, typically defined as greater than 120
ms , is often a result of inter-atrial block (Power et al., 2022). Although less is known about the correlates
of abnormally short PWD, in certain instances, short PWD may be considered an earlier and more subtle
feature than prolonged PWD. It has been hypothesized that shortened atrial repolarization and refractory
periods could indicate changes in atrial electrical properties preceding more overt abnormalities associated
with prolonged PWD, which can eventually result in arrhythmogenesis (Nielsen et al., 2015; Zhou et al.,
2023). Recent research has started to underscore that shorter PWD is also associated with adverse outcomes,
including AF, heart failure, stroke, and dementia (L. Y. Chen et al., 2022; Ostrowska et al., 2022a, 2022b;
Zhou et al., 2023).
Disruptions in atrial function could theoretically impact cerebrovascular health through hemodynamic
dysregulation, cerebral hypoperfusion, systemic inflammation, and hypercoagulation (van der Velpen
et al., 2017), but research related to the specific contribution of PWD is scant. It is clear, however, that
112

gradual disruptions of cerebrovascular hemodynamics can be a major contributor to the initiation of
vascular dementia pathology. Subclinical decreases in cardiac function and increases in arterial stiffness
can lead to cerebral hypoperfusion, blood brain barrier damage, and detrimental pulsatile blood flow. The
resulting neurovascular dysfunction and inflammation can impair glymphatic clearance and exacerbate
existing dementia-related pathology leading to synaptic dysfunction and eventually brain atrophy (Moore
& Jefferson, 2021). A recent study by Reyes et al found an association between prolonged PWD and various
cerebrovascular related injuries, but not with gross lobar volumes (Reyes et al., 2023). We, on the other
hand, found a significant association between short PWD and accelerated brain aging. Our enhanced
sensitivity may be explained by the fact that we used a specific component of volume (i.e. thickness) as
well as grey/white matter intensity ratio with a finer parcellation of brain networks as opposed to volume
measurement only in the whole brain and cortical lobes. While our results also suggest that longer PWD is
associated with lower brain age given the lack of significance in the models we tested with a quadratic
PWD term. This may be a result of fewer participants in the UK Biobank having what is often defined as
"prolonged PWD", which has been consistently associated with detrimental outcomes including increased
risk of atrial fibrillation, stroke, and dementia (Gutierrez et al., 2019; Hari et al., 2018; He et al., 2017; Nielsen
et al., 2015).
We did not a priori define abnormal PWD as dichotomous - prolonged or not - and instead investigated these associations in an unbiased manner using the continuous measure, enabling the analysis of
bidirectional PWD abnormalities (short and prolonged) without a threshold. Some studies prefer to use
percentile cutoffs to circumvent assumptions, although admittedly these can vary widely from population
to population and by other factors including age, sex, and ethnicity (Nielsen et al., 2015; Soliman et al.,
2013). For example, Nielsen et al (Nielsen et al., 2015) assessed a large primary care population from the
Copenhagen ECG Study and established that the optimal PWD with respect to the lowest risk of AF was
100 − 105 ms which fell between the 49th and 61th percentile in our study. Abnormally short PWD in
113

their study, or < 5 th percentile, was less than 90 ms compared to less than 65 ms in our study. Prolonged
PWD in their study, or > 95 th percentile, was greater than 131 ms compared to greater than 119 ms in our
study. Given the differences in percentile cutoffs across populations, we opted to use a continuous PWD.
Nonetheless, Nielsen et al. found an elevated risk of AF (and cardiovascular death and stroke although not
as robust) in those with abnormally short and long PWD. Using a "stopped" Cox model, they also showed
that the association between shorter PWD and AF risk was slightly stronger when evaluating short term
effects compared to long term. They hypothesized that short PWD, resulting in a more rapid conduction
time, may be a substrate for the early development of AF. While it is possible that the participants with
abnormally short PWD in our study will go on to develop AF, a longitudinal study design is needed to
evaluate this.
Although we detected associations between brain structure and PWD & brain structure and cognition,
we did not observe a direct association between PWD and cognition. This parallels previous longitudinal
work from the ARIC-NCS study which showed that abnormal PWD increased the risk of dementia, but
did not reduce cognitive scores. The authors suggest that atrial abnormalities may increase the risk of
sudden, but major, events such as stroke, as opposed to a more subtle and progressive decline (Gutierrez
et al., 2019). While this may be true, this study did not investigate the effects of abnormally short PWD, and
instead only investigated prolonged PWD associations with cognition. The brain areas that we found to be
associated with PWD include the frontoparietal network responsible for executive control, the sensorimotor
network in charge of processing bodily sensations and executing appropriate motor responses, and the
attentional networks (dorsal and ventral) responsible for top down and bottom up processing respectively
(Petersen & Posner, 2012). These networks are all in communication with one another to facilitate various
aspects of cognitive function tested in our study. The dorsal attention, frontoparietal, and the default mode
networks are reported to have reduced functional connectivity as a result of small vessel disease (Ter Telgte
et al., 2018) - a major risk factor for stroke and vascular dementia. Given that abnormalities in PWD are
114

associated with markers of small vessel disease (Reyes et al., 2023), it is plausible that abnormal PWD may
lead to cerebrovascular injury which in turn damages the structural integrity of the networks responsible
for attention and executive function, though this hypothesis needs formal testing. Overall, our results may
suggest that subtle brain alterations due to short PWD may impose accelerated aging effects to structural
networks that are responsible for important cognition tasks, but that these occur in advance of detectable
cognitive decline.
We note, however, that our study is cross-sectional in nature and a longitudinal study design is needed
to investigate whether short PWD predicts future cognitive deficits. While we attempted to account
for all possible confounders by including them as covariates in our linear models, we note that there
may be other contributing variables that we could not account for due to our retrospective study design.
Another limitation we acknowledge is that our brain age model only includes cortical features, even
though subcortical/deep gray matter and white matter features may be important predictors of brain aging,
particularly those vulnerable to vascular pathology (Wardlaw et al., 2013). In a preliminary investigation
into the relationship between PWD and hippocampal volume, we observe the same trends as our brain
age models - where shorter PWD is associated with less favorable brain outcomes (lower hippocampal
volumes, Figure 3.9). Future work will continue to investigate these associations. Lastly, it may also be the
case that the relationship between short PWD and cognitive decline is relatively weak in the preclinical
population we are studying, and that effects may be more pronounced in a clinical population of AD and
related dementia cases.
Nonetheless, we find, even in the absence of major cardiac conditions and controlling for cardiovascular
risk factors, that PWD still remains a significant predictor of many regional BAI. This suggests that early
detection and treatment of poor cardiac function, prior to manifestation of arrhythmias, may help curb
accelerated brain aging. Moreover, these associations may be helpful in understanding how the heart-brain
axis impacts brain health and cognitive aging trajectories.
115

3.7 Acknowledgements
This work was supported in part by National Institutes of Health (R01AG059874), National Institutes of
Health (P41EB015922), and Bright Focus Research Grant award (A2019052S).
UK Biobank Resource under Application Number: 11559
3.8 Appendix B
Figure 3.6: Regional BAI association results with PWD as the variable of interest in those without major
cardiac conditions. A. T-statistics of regional BAI (overlapping) associated with PWD in those without
major cardiac conditions. B. Standardized coefficients of PWD and all other regression coefficients (with
the exception of the four MRI PCA components explaining 88.3% of scanner related variance) from all
12 models (whole brain BAI, AD signature region BAI, and 10 regional BAI). Shorter P-wave duration
was associated with a higher brain age. Regional BAI that survived multiple comparisons included the
fronto-parietal, dorsal attention, and ventral attention (middle panel), as well as the sensorimotor network
(right panel). Covariates associated with a higher BAl include hypertension, diabetes, smoking status, and
frequent alcohol consumption. Covariates associated with a lower BAI include a higher bone mineral
density T-score, having a college education, and younger age. Filled points indicate effects that survived
multiple comparisons correction across all 12 BAls (within each covariate).
116

Table 3.3: Variables and respective data fields used for cardiovascular risk related covariates, cognition
variables, and scanner/bias (MRI) used for principal component analysis. Dx: diagnosis; Tx: treatment. Note: Table Position was initially explored but due to high correlation with Brain Z Position

r(12,762) = −0.92, p < 0.001
, it was not included in the MRI PCA.
Variable Datafield ID
Cardiovascular Risk Variables -
Hypertension Dx 41270, 20002, 6150
Hypercholesterolemia Dx 41270, 20002
Diabetes Dx/Tx 41270, 20002, 6153, 6177
Coronary Artery Disease Dx 41270, 41272, 20002, 20004, 6150
Heart Failure Dx 41270, 20002
Atrial Fibrillation Dx 41270, 41272, 20002, 20004, 12653
Chronic Kidney Disease Dx 41270, 41272, 20002, 20004
Sleep Apnea Dx 41270, 20002
Alcohol Consumption (infrequent, frequent, occasional) 1558, 41270
Smoking Status (never, previous, current) 20116
BMI 23104
Heel bone mineral density T-score 4106, 4125
APOE4 carrier -
Systolic Blood Pressure 4080, 93
Diastolic Blood Pressure 4079, 94
Cognitive Variables -
Memory 4282, 20197
Executive Function 6348, 6350, 21004
Processing Speed 20023, 23324
Reasoning 20016, 6373
Scanner/Bias (MRI) Variables -
Scan Site 54
Brain X Position 25756
Brain Y Position 25757
Brain Z Position 25758
Inverse SNR 25734
Inverse CNR 25735
Linear Registration Discrepancy 25731
11

Figure 3.7: Standardized coefficients of PWD and all other regression coefficients from all 12 models (whole
brain BAI, AD signature region BAI , and 10 regional BAI)stratified by those with and without hypertension.
All regional networks that survived in the main analysis survived in those without hypertension with the
exception of the language network. Filled points indicate associations that survived multiple comparisons
correction across all 12 BAls (within each covariate). 4 components from a PCA including all relevant MRI
scanner related variables (explaining 88.3% of the variance) were included in the model but omitted from
the figure for easier visualization of biological variables of interest.
118

Figure 3.8: Standardized coefficients of PWD and all other regression coefficients from all 12 models
(whole brain BAI, AD signature region BAI, and 10 regional BAI) stratified by those with and without
diabetes. All regional networks that survived in the main analysis survived in those without diabetes. Filled
points indicate associations that survived multiple comparisons correction across all 12 BAls (within each
covariate). 4 components from a PCA including all relevant MRI scanner related variables (explaining
88.3% of the variance) were included in the model but omitted from the figure for easier visualization of
biological variables of interest.
119

Figure 3.9: PWD associations results with commonly used structural MRI measures. Volumes of the
hippocampus and the AD signature ROI were positively associated with PWD (which also suggests that lower
volume is associated with lower PWD), in line with our main regional BAI results. Multiple comparisons
correction was performed across the 3 measures within each covariate. ICV and 4 components from a PCA
including all relevant MRI scanner related variables (explaining 88.3% of the variance) were included in
the model but omitted from the figure for easier visualization of biological variables of interest. Another set
of regressions was performed which included a quadratic PWD term but effects did not survive multiple
comparisons testing.
120

Chapter 4
Modifiable lifestyle factors and their association with sex-specific risk
and resilience to brain aging and neurodegeneration
4.1 Lifestyle factors that promote brain structural resilience in
individuals with genetic risk factors for dementia
This section is adapted from:
Haddad E, Javid S, Dhinagar N, Zhu AH, Lam P, Ba Gari I, Gupta A, Thompson PM, Nir TM, Jahanshad N
(2022). Lifestyle Factors That Promote Brain Structural Resilience in Individuals with Genetic Risk Factors
for Dementia. In: Abdulkadir, A., et al. Machine Learning in Clinical Neuroimaging. MLCN 2022. Lecture
Notes in Computer Science, vol 13596. Springer, Cham. https://doi.org/10.1007/978-3-031-17899-3_11
4.1.1 Abstract
Structural brain changes are commonly detectable on MRI before the progressive loss of cognitive function
that occurs in individuals with Alzheimer’s disease and related dementias (ADRD). Some proportion of
ADRD risk may be modifiable through lifestyle. Certain lifestyle factors may be associated with slower
brain atrophy rates, even for individuals at high genetic risk for dementia. Here, we evaluated 44,100
T1-weighted brain MRIs and detailed lifestyle reports from UK Biobank participants who had one or more
121

genetic risk factors for ADRD, including family history of dementia, or one or two ApoE4 risk alleles. In this
cross-sectional dataset, we use a machine-learning based metric of age predicted from cross-sectional brain
MRIs - or ’brain age’ - which when compared to the participant’s chronological age, may be considered
a proxy for abnormal brain aging and degree of atrophy. We used a 3D convolutional neural network
trained on T1w brain MRIs to identify the subset of genetically high-risk individuals with a substantially
lower brain age than chronological age, which we interpret as resilient to neurodegeneration. We used
association rule learning to identify sets of lifestyle factors that were frequently associated with brain-age
resiliency. Never or rarely adding salt to food was consistently associated with resiliency. Sex-stratified
analyses showed that anthropometry measures and alcohol consumption contribute differently to male vs
female resilience. These findings may shed light on distinctive risk profile modifications that can be made
to mitigate accelerated aging and risk for ADRD.
4.1.2 Introduction
Late-onset Alzheimer’s disease (LOAD) and related dementias, which are considered to be those that
occur after the age of 65 , are complex disorders, driven by a combination and wide range of genetic,
environment, and gene-environment interactions (Blennow et al., 2006). Recently, publicly available big
data initiatives and biobanks have taken a more holistic approach and collect a wide range of data to
help understand sources of risk factors for diseases, including ADRDs. Machine learning techniques
provide computationally efficient and effective means of data reduction and analyses that can promote
our understanding of neurodegenerative processes from the vast set of data features now available to
researchers; brain imaging and genetic data alone can contain trillions of univariate combinations of tests
(Medland et al., 2014). Here, we use machine learning approaches to discover lifestyle factors that may
contribute to brain resilience in individuals at heightened genetic risk for ADRDs.
122

People with a family history of dementia have twice the risk of being affected themselves than individuals
in the general population (Loy et al., 2014). Apolipoprotein (ApoE) E4 (e4) on chromosome 19 has consistently
been shown to be the commonly-occurring genetic factor most strongly associated with LOAD (Lambert
et al., 2013): one copy of e4 poses an approximate 3−5-fold increase in lifetime risk, whereas two copies pose
an 8 − 20-fold increase (Loy et al., 2014; Strittmatter, 2012), depending on a person’s ancestry. Modifiable
risk factors also contribute to risk for LOAD, and together are associated with approximately 40% of the
population attributable risk for dementia worldwide (Livingston et al., 2020; Nianogo et al., 2022). These
include less education, hearing loss, traumatic brain injury, hypertension, excessive alcohol consumption,
obesity, smoking, depression, social isolation, physical activity, air pollution, and diabetes. Other factors
that have been associated with Alzheimer’s pathology and cognitive decline include poor sleep and dietary
factors, respectively (Livingston et al., 2020). As LOAD is characterized by brain structural changes, such as
cortical and hippocampal atrophy (Jack, 2011), studying the impact of modifiable factors on these changes
may inform preventative, high yield recommendations for those at risk.
Typically, longitudinal data is required to infer rates of brain structural changes that occur as a result
of age or disease. Over the last decade, a surrogate cross-sectional marker for measuring "accelerated"
or "slowed" brain aging, more commonly referred to as "brain age", has been developed (Cole & Franke,
2017). Although not without limitations (Butler et al., 2021), brain age is derived from machine learning
models, which often rely on regression techniques that model brain features as independent variables and
chronological age as the dependent variable in a neurologically healthy "training" set. These models are
then applied to an independent or "test" sample and the difference between an individual’s predicted age
and chronological age, known as the "brain age gap" or residual, is used as an aging biomarker. Brain age
has been used to study associations of risk factors with accelerated aging, predict diagnostic outcomes, and
examine differences between clinical populations. Recently, with the use of deep learning techniques such
123

as convolutional neural networks (CNNs), researchers can now predict brain age with less overall error,
and more precise age estimates, directly from the full 3D MRI scan (Cole et al., 2017).
Most studies to date have used brain age in older adults to model accelerated aging, but investigating
associations with "slowed" brain aging is also important to better understand factors that can mitigate
accelerated brain aging. Continuous measures of brain age have shown protective associations with physical
exercise, education, and combined lifestyle scores (Bittner et al., 2021; Dunås et al., 2021; Smith et al., 2019;
Steffener et al., 2016) but no study to date has evaluated which combination of these and other modifiable
factors most consistently promotes brain resilience. The UK Biobank is a large, densely phenotyped epidemiological study that has collected health information from half a million UK participants, approximately
50,000 with neuroimaging (Miller et al., 2016a). This has allowed researchers to comprehensively examine
how genetic, sociodemographic, physical, and health outcomes may influence brain structure and function.
In such high dimensional datasets, data mining algorithms, including association rule learning (ARL) can
be used to discover associations between sets of features and a particular outcome (Agrawal et al., 1993).
By combining ARL with measures of brain aging in deeply phenotyped datasets, we can begin to make
inferences that incorporate the multifactorial nature of complex neurodegenerative processes and dementia
risk.
Using data from the UK Biobank, we identified high LOAD risk individuals - who are over age 65 and
either ApoE4 carriers or have a family history of dementia - who are resilient to brain aging, characterized
by a brain age younger than their chronological age; using ARL, we then examined which sets of lifestyle
factors were most often associated with this resiliency.
124

4.1.3 Methods
4.1.3.1 "Brain age"
44,100 T1-weighted (T1w) brain MRI scans from the UK Biobank (UKB) (Miller et al., 2016a) were preprocessed using a sequence of standard neuroimaging processing steps including N4 bias field correction for
non-parametric intensity normalization, skull stripping, and linear registration with 6 degrees of freedom
to a template. The preprocessed T1w scans from UKB were partitioned into non-overlapping training and
test splits using a 5-fold cross-validation approach. We used 20% of each training split as a validation set.
Our model architecture for brain age prediction is based on the 3D CNN originally proposed by (Peng et al.,
2021). The model consists of a sequence of five convolutional blocks with a [ConvBatchNorm-ActivationMaxPooling] structure as the backbone, followed by two fully connected layers. The architecture is relatively
lightweight for a 3D-CNN, with around two million parameters. We trained the CNN end-to-end for 30
epochs with a batch size of 5 using the Adam optimizer (Kingma & Ba, 2014) and learning rate of 0.00005.
We extracted the brain age measures for all 44,100 subjects with available T1w MRI at the time of writing.
We corrected for predicted age’s dependence on age to avoid artificially inflating performance metrics
(Butler et al., 2021). In brief, this involves scaling the predicted age by the slope and intercept from a
regression of predicted age on chronological age (Cole et al., 2018; de Lange & Cole, 2020).
4.1.3.2 Subject selection
UKB participants over age 65 with e3e4 or e4e4 genotype or a family history dementia (maternal, paternal,
or sibling) were included (4520M, 4493F). The mean absolute error (MAE) of our bias-corrected brain age
model was calculated and any subject having a brain age lower than their true age by a value larger than
the MAE was considered resilient.
125

4.1.3.3 Lifestyle factors
Lifestyle factors were derived and binarized from lifestyle and environment, anthropometry, and summary
diagnostic categories measures, documented at the time of imaging. Factors and their respective datafield
IDs are provided in Table 4.1. Body mass index classifications were categorized under CDC guidelines. Diet
quality was modeled using the same coding scheme as in (Said et al., 2018) and (Zhuang et al., 2021). Briefly,
an overall diet quality score was computed and binarized as adequate if the score was > 50 out of a total of
100. Individual diet components were also included and corresponded to adequate consumption of fruits,
vegetables, whole grains, fish, dairy, vegetable oil, refined grains, processed meat, unprocessed meat, and
sugary food/drink intake.
Table 4.1: Lifestyle factors and respective UK Biobank data field IDs. American Heart Association (AHA)
guidelines for weekly ideal ( ≥ 150 min/week moderate or ≥ 75 min/wk vigorous or 150 min/week
mixed), intermediate (1-149 min/week moderate or 1-74 min/week vigorous or 1-149 min/week mixed), and
poor (not performing any moderate or vigorous activity) physical activity. Supplementation was categorized
into any vitamins/minerals or fish oil intake. Salt added to food and variation in diet included the response
of "never or rarely". Smoking status included never having smoked, previously smoked, and currently
smokes. Alcohol frequency was categorized as infrequent ( 1 − 3 times a month, special occasions only, or
never), occasional ( 1 − 2 a week or 3-4 times a week), and frequent (self-report of daily/almost daily and
ICD conditions F10, G312, G621, I426, K292, K70, K860, T510). Social support/contact variables included
attending any type of leisure/social group events, having family/friend visits twice a week or more, and
being able to confide in someone almost daily.
Lifestyle Factor Features (Data Field ID)
Physical Activity/
Body Composition
AHA Physical Activity (884, 904, 894, 914); Waist to Hip Ratio (48,49);
Body Mass Index (BMI) (23104)
Sleep Sleep 7-9 Hours a Night (1160); Job Involves Night Shift Work (3426);
Daytime Dozing/Sleeping (1220)
Diet/
Supplements
Diet Quality Score and Components (based on Said et al., 2018;
Zuang et al., 2021) (1309, 1319, 1289, 1299, 1438, 1448, 1458, 1468,
1329, 1339, 1408, 1418, 1428, 2654, 1349, 3680, 1359,1369,1379, 1389,
3680, 6144); Fish Oil Supplementation (20084); Vitamin/Mineral
Supplementation (20084); Salt Added to Food (1478); Variation in Diet (1548)
Education College/University (6138)
Smoking Smoking Status (20116)
Alcohol Alcohol Intake Frequency (1558/ICD)
Social Contact/
Support
Attending Leisure/Social Group Events (6160); Frequency of Friends/Family
Visits (1031); Able to Confide in Someone (2110)
126

4.1.3.4 Association rule learning
ARL using the mlxtend library in Python (Raschka, 2018) was used to characterize sets of lifestyle factors
that co-occur with resilience across the older population, and after stratifying for sex.
Problem Formulation. Let I = {ii
, i2, . . . , im} be a set of binary attributes (items), here, being the
resilience classification, and the set of binary lifestyle factors described in Table 4.1. Let T = {t1, t2, . . . , tN }
be the set of subjects. Each subject is represented as a binary vector where tj [ik] = 1 if the subject j, tj ,
has that particular feature ik, and tj [ik] = 0 otherwise. An itemset, X, is defined by a collection of zero or
more items. In this context, an association rule is an implication of the form X ⇒ ik, where X is a set of
lifestyle factors in I and our consequent of interest, ik, is the item "resilience", with |X ∩ {ik}| = 0. ARL is
decomposed into two subproblems:
1. Frequent itemset generation, where the objective is to find itemsets that are above some minimum
support and
2. Rule generation, where the objective is to generate rules with high lift values (defined below) from
the frequent itemsets generated in step 1.
Frequent Itemset Generation. For the first step of the association rule generation, the Apriori principle
states that, if an itemset is frequent then all of its subset itemsets are also frequent (Tan et al., 2006). The
metric that can be used to generate frequent itemsets is support. Support is defined as the fraction of
subjects (i.e., transactions) in T that contain that itemset. That is
support(Z) = σ(Z)
N
(1)
where Z is an itemset, N is the total number of subjects, and σ(Z) is the support count of Z defined as
σ(Z) = |{ti
| Z ⊆ ti
, ti ∈ T}| (2)
127

where | · | represents the cardinality of the set. We generated all the frequent itemsets with a minimum
support threshold of 0.01 . The minimum support threshold was chosen based on the computational power
and memory available to us, we have chosen a very small value (1%) to incorporate as many itemsets as
possible for our association rule generation.
Rule Generation. For each frequent itemset fk generated, we would generate the association rules that
have "resilient" as their consequent. To measure the strength of a rule we used the lift value:
lift (X ⇒ ik) = support (X ∩ {ik})
support(X) × support ({ik})
(3)
A lift value greater than 1 implies that the degree of association between the antecedent and consequent is
higher than in a situation where the antecedent and consequent are independent. We generated all the
association rules with a minimum lift value of 1.2 .
4.1.4 Results
4.1.4.1 Brain age
The trained models were evaluated on the test set using the root mean squared error (RMSE), mean absolute
error (MAE), and Pearson’s r between the true age and predicted brain age as seen in Table 4.2. The 3D
CNN achieved an average MAE of 2.91(0.05) across the 5 splits. Following age bias correction, the MAE
was 3.21.
Table 4.2: Summary of the test performance of the 3D CNN on UKB for brain age prediction in the full
sample of N = 44, 100. RMSE: root mean squared error; MAE: mean absolute error.
Split 1 Split 2 Split 3 Split 4 Split 5
RMSE 3.747 3.676 3.712 3.604 3.584
MAE 2.972 2.905 2.958 2.864 2.84
Pearson’s r 0.883 0.878 0.888 0.885 0.887
128

4.1.4.2 Resiliency
1439/9013 (16%) of subjects over age 65 years with genetic risk factors for ADRD were considered to be
resilient, that is, having a brain age value > 3.21 years younger than their chronological age. 760/4520
(17%) of females and 679/4493 (15%) of males were resilient. Example subjects are featured in Figure 4.1.
Figure 4.1: A. Predicted (y-axis) age plotted against chronological (x-axis) age B. Brain structural T1-
weighted MRIs are shown for a resilient case (subject 1; actual age: 66; predicted age: 50) and a non-resilient
case (subject 2; actual age: 68; predicted age: 79). Note qualitative differences in ventricular size are visible,
even though subjects are close in chronological age.
4.1.4.3 Association rule learning
A total of 7,076 sets with a minimum support of 0.01 and lift value of at least 1.2 cooccurred with resiliency
in the combined set, 34,106 for females, and 5,883 for males. The top antecedent set for each model is shown
in Table 4.3. Frequencies of factors are visualized in Figure 4.2A for all three models. Corresponding lift
values are shown in Figure 4.2B.
The most frequent factor that appeared in antecedent sets for combined (69.7%) and female ( 61.6% ),
was "never or rarely adding salt to food", but also occurred with high frequency in males (49.5%). The most
129

frequent factor in males was having an adequate intake of whole grains ( 54.1% ) which also occurred with
high frequency in the combined set ( 48.3% ) and in the female set ( 36.3% ) although to a lesser extent.
Having an adequate diet quality score was in 39% of the combined sets, 45.4% of the female sets, and
28.4% of the male sets. Sleeping on average 7 − 9 h per night was in 60.4% of the female, 57.4% of the
combined, and 31.4% of the male sets. Participating in leisure or social activities appeared at a relatively
similar rate among combined (35.4%), female ( 38.8% ), and male (33.5%) models.
Factors that were in less than 1% of predictive sets in any of the combined, male and female models
included: obese or underweight BMI; frequent consumption of alcohol; night shift work; current smoking
status; having poor physical activity; and supplementation of vitamins or fish oil.
Table 4.3: Top antecedent set with resiliency as a consequent based on lift from the combined, female, and
male models.
Model Antecedents with resiliency
as a consequent
Antecedent
support
Consequent
support
Support Lift
Combined infrequent alcohol, never/rarely varies diet,
adequate fruit, never smoked 0.043 0.160 0.010 1.46
Female
never/rarely daytime dozes, never/rarely
varies diet, adequate fish, adequate dairy,
friend/family visits, infrequent alcohol
0.035 0.168 0.010 1.72
Male never/rarely varies diet, adequate dairy,
college, adequate processed meat, never smoke 0.042 0.151 0.010 1.59
130

Figure 4.2: A. Antecedent frequency across all three models: combined, females only, and males only. Left
side of the heatmap shows the absolute frequency of factors per model and the right side shows a ratio of
the absolute count over the total number of predictive sets. B. Corresponding lift values and means (dotted
line) for each respective model. Highest pair frequencies of lifestyle factors in antecedent sets in the C.
combined, D. female, and E. male models. Lower triangles are absolute frequencies truncated at 3500 and
upper triangles indicate ratio of absolute pair counts over the total number of predictive sets truncated at
0.35.
131

4.1.5 Discussion
As the prevalence of ADRDs continues to rise, understanding which lifestyle modifications can mitigate
brain aging risk is becoming increasingly important. Here, in those at elevated risk, we identified subjects
resilient to brain aging and examined the lifestyle factors most associated with this resiliency. The self
reporting of "never or rarely adding salt to food" was frequently associated with resilience across all of
our models tested. Dietary salt is associated with hypertension, a known vascular risk factor for dementia.
However, irrespective of hypertension, independent associations between a high salt diet and increased
dementia risk have been shown (Heye et al., 2016; Santisteban & Iadecola, 2018; Strazzullo et al., 2009).
We also found support for other modifiable risk factors such as diet quality, social contact, the absence of
smoking, physical activity, and sleep duration, as we detected associations with these factors and resilience.
As studies continue to show differential risk profiles across the sexes (Nianogo et al., 2022), we further
extend our work by building sex-specific models. Not only is the incidence of AD greater in females, but
carrying an e4 allele has a stronger effect in females compared to males (Moser & Pike, 2016; Podcasy
& Epperson, 2016; Rocca et al., 2014). Although e4 status and sex are non-modifiable, it is important to
investigate how they differentially interact with factors that are modifiable (Udeh-Momoh et al., 2021).
Interestingly, we found nearly 5 times the number of predictive sets in females compared to the combined
model and the males only model.
Our results are largely based on self-reported data from predetermined questionnaires, which may be
confounded by reporting and response biases. Moreover, ARL requires binary variables which reduces the
granularity of some of the features - a factor that may play an important role in predicting resiliency. We
also note limitations associated with the use of brain age gap, as a type of residual method, given the lack
of ground truth data needed to assess its validity (Bocancea et al., 2021). Our brain age method also lacks
spatial information regarding anatomical changes that occur as a result of brain aging. Future work will
address these limitations by modeling specific brain features as outcome variables in order to spatially map
132

which features most consistently contribute to brain aging resilience. Continuous lifestyle factors will be
modeled as predictors using methods that are robust to potential hidden confounders (Marmarelis et al.,
2023).
Nonetheless, with the ability to map the discrepancies between chronological age and what may be
more functionally important - biological age - we can begin to understand which modifiable factors are
more or less beneficial to brain health. More importantly, by building stratified models, we can shed light
on differential risk profiles that may inform respective populations on the most effective actions to take
to mitigate accelerated brain aging and dementia risk. Future work will continue to tease apart potential
differential effects by separately modeling e 4 carriage and familial history of dementia as they relate to
resilience, and modeling sex-specific brain age resilience.
4.1.6 Acknowledgements
This work was supported in part by: R01AG059874, U01AG068057, P41EB05922. This research has been
conducted using the UK Biobank Resource under Application Number ’ 11559 ’.
133

4.2 Causal sensitivity analysis for hidden confounding: modeling the
sex-specific role of diet on the aging brain
This section is adapted from:
Haddad E*, Marmarelis MG*, Nir TM, Galstyan A, Ver Steeg G, Jahanshad N. (2023). Causal Sensitivity
Analysis for Hidden Confounding: Modeling the Sex-Specific Role of Diet on the Aging Brain. In: Abdulkadir,
A., et al. Machine Learning in Clinical Neuroimaging. MLCN 2023. Lecture Notes in Computer Science, vol
14312. Springer, Cham. https://doi.org/10.1007/978-3-031-44858-4_9
*denotes equal contribution
4.2.1 Abstract
Modifiable lifestyle factors, including diet, can impact brain structure and influence dementia risk, but the
extent to which diet may impact brain health for an individual is not clear. Clinical trials allow for the
modification of a single variable at a time, but these may not generalize to populations due to uncaptured
confounding effects. Large scale epidemiological studies can be leveraged to robustly model associations
that can be specifically targeted in smaller clinical trials, while modeling confounds. Causal sensitivity
analysis can be used to infer causal relationships between diet and brain structure. Here, we use a novel
causal modeling approach that is robust to hidden confounding to partially identify sex-specific dose
responses of diet treatment on brain structure using data from 42,032 UK Biobank participants. We find
that the effects of diet on brain structure are more widespread and also robust to hidden confounds in
males compared to females. Specific dietary components, including a higher consumption of whole grains,
vegetables, dairy, and vegetable oils as well as a lower consumption of meat appears to be more beneficial
to brain structure (e.g., greater thickness) in males. Our results shed light on sex-specific influences of
134

hidden confounding that may be necessary to consider when tailoring effective and personalized treatment
approaches to combat accelerated brain aging.
4.2.2 Introduction
In individuals at risk for Alzheimer’s Disease and related diseases (ADRDs), changes in neuroimaging
derived biomarkers can occur in advance of noticeable cognitive decline. Identifying risk factors associated
with these changes can inform biological mechanisms, help to stratify clinical trials, and ultimately provide
personalized medical advice (Yassine et al., 2022). Modifiable risk factors throughout the lifespan that can
prevent or delay up to 40% of dementias worldwide have been identified and quantified in (Livingston
et al., 2020). To name a few, these include less education in early life; hearing loss, brain injury and
hypertension in midlife; and smoking, social isolation, and physical inactivity in late life. While diet and
dietary interventions have dedicated sections in this report, quantitatively assessing their contribution to
dementia risk through prodromal brain structural differences, proves challenging.
Inconsistencies in the neuroimaging correlates of nutrition may be due to heterogeneities in the way
dietary habits are collected or interventions assigned (e.g., supplementation of specific vitamins/oils versus
adherence to whole diets such as the Mediterranean diet), differences in study/analysis design, small samples,
largely cross-sectional associations, or even due to differences in the neuroimaging measures themselves.
While many studies include observed confounders as covariates (ex: age, sex, education, physical activity),
it is virtually impossible to account for all factors that affect both diet treatment and brain outcomes, as
complex lifestyle, environmental, and genetic interactions exist that may influence both. Also known
as hidden confounders, these factors may be a major reason why inconsistencies across findings exist
in nutritional research. Cross sectional studies generally reveal lower diet quality associated with lower
brain volumes but there are also studies that fail to replicate these findings (Drouka et al., 2022; Jensen
et al., 2021). Randomized control trials (RCTs) are still considered the gold standard, yet results from such
135

trials also remain inconclusive. Some RCTs show beneficial effects of a "better quality" diet on total brain
and hippocampal volumes, but not on cortical thickness, where associations have been found in some
cross-sectional studies (Townsend et al., 2022). Moreover, conducting RCTs is costly, thus limiting their
duration and perhaps the optimal time needed to detect noticeable changes in neuroimaging features due
to dietary interventions.
Fortunately, large scale epidemiological studies that include brain imaging, like the UK Biobank (Miller
et al., 2016a), can now be leveraged to build robust models and make inferences that can later be specifically
targeted and validated in such smaller RCTs. This can help target interventions to have more detectable
effects in a shorter time frame. For example, sex is rarely investigated as a variable of interest in nutrition
studies (Jensen et al., 2021), even though sex and sex hormones are known to confer differential effects
on glucose metabolism, insulin sensitivity, fat metabolism, adiposity, protein metabolism, and muscle
protein synthesis (Y. Chen et al., 2022). Large-scale studies allow for models to be sex-stratified, which can
inform how nutrition may differentially influence male and female brain structure, leading to more targeted
treatment interventions.
While observational studies are a powerful resource to study population level effects, correlation
is not sufficient to infer causality and can be subject to confounding factors (Calude & Longo, 2017).
Causal modeling approaches can account for confounders that are not known or modeled a priori. Causal
inference methods for observational studies may also attempt to predict counterfactuals, or alternative
outcomes as if they were properly randomized experiments (e.g., keeping all other variables equal, would
an outcome measure change if a patient had not been a smoker). Causal sensitivity analyses attempt to
answer the question of how a causal prediction might be biased if hidden confounders were affecting the
observational study. We recently developed a causal sensitivity analysis method for continuous-valued
exposures (Marmarelis et al., 2023) rather than the classical binary treatment vs control setting. The
use of continuous indicators reduces the degree of bias introduced by having to choose a threshold and
136

enables the estimation of incremental effects. Identifying marginal effects of incrementally increasing
exposures is perhaps more actionable than the broader proposition of completely including or eliminating
an exposure. It can also be easier to identify statistically because the predictive model would make smaller
extrapolations from the observed scores. Since the causal outcomes cannot be exactly identified in the
presence of hidden-confounding bias, these quantities are partially identified using ignorance bounds that
are computed using the sensitivity model.
Here, we seek to partially identify dose responses of several continuous dietary components on brain
cortical thickness and subcortical volumes using a novel methodology that is robust to hidden confounding.
We interrogate sex-specific models to decipher the heterogeneous dose responses of particular dietary
components on regional brain structure in a large, community dwelling, aging population.
4.2.3 Methods
4.2.3.1 Partial identification of dose responses
Causal inference with potential outcomes (Rubin, 1974) is classically formulated with a dataset of triples
(Y, T, X) that represent the observed outcome, assigned treatment, and covariates that include any observed
confounders. The first assumption in the potential-outcomes framework is the "stable unit treatment
value assumption" (SUTVA), which requires that the outcome for an individual does not depend on the
treatment assignment of others. We assume that our sample of observations is independently and identically
distributed, which subsumes the SUTVA. The second assumption is that of overlap, which simply means that
all treatment values have a nonzero probability of occurring for all individuals. Finally, the third assumption
is that of ignorability: {Yt ⊥ T} | X. This assumption requires that there are no hidden confounders, so that
the potential outcome Yt
is independent of the treatment assignment after conditioning on the covariates.
The potential outcome is conceptually different from the observed outcome. Every individual is assumed to
have a set of potential outcomes corresponding to the whole set of possible treatment assignments. The
137

treatment variable then controls which potential outcome is actually observed: Y = YT . In this paper, we
study bounded violations to the third assumption of potential outcomes by performing a causal sensitivity
analysis. Causal sensitivity analysis allows one to make reasonable structural assumptions about causal
effects across varying levels of hidden confounding.
Dose Responses. Treatment variables analyzed in our study were continuous diet scores. Formally, we
study the conditional average causal derivative (CACD) defined as ∂
∂tE [Yt
| X] at the observed treatment,
t = T. We present the CACD as the percent change from the observed outcome, with respect to the
observed diet-score treatment variable, which takes on continuous values from 1-10.
Approach to Partial Identification. By performing a causal sensitivity analysis using a suitable sensitivity
model, we may partially identify causal effects even under the influence of possible hidden confounders.
Our novel marginal sensitivity model for continuous treatments ( 8MSM ), like other sensitivity models,
makes reasonable structural assumptions about the nature of possible hidden confounders on the basis of
observed confounding. It differs from other sensitivity models in that it allows continuous-valued treatment
variables. The δ MSM exposes a single parameter, Γ ≥ 1, which controls the overall amount of assumed
hidden confounding. The lower and upper bounds on the estimated dose-response curves grow further
apart with Γ, since it makes identification less feasible. The fundamental structural assumption for hidden
confounders made by the δ MSM, stated in its most compact form, is

∂
∂τ log p (τ | yt
, x)
p(τ | x)

≤ log Γ
Here, we suppose that the logarithmic derivative of the ratio of two probability density functions is
bounded. We use the notation τ to indicate the observed treatment assignment, which may differ from the
treatment of the potential outcome that we are inquiring about. The numerator, p (τ | yt
, x), is termed the
complete propensity and is unobservable. The denominator, p(τ | x), is the observable analogue termed the
nominal propensity that measures association between treatment assignments and individuals as described
138

by covariates. The complete propensity diverges from the nominal propensity only in the presence of
hidden confounders
Figure 4.3: Partially identified causal derivatives. For each individual, we predict outcomes after lowering
a single diet score by a small amount, and also raising it by that amount. This enables approximation
of the causal derivative by finite differences. If the outcomes are partially identified (where Γ > 1 so
multiple outcomes and hence derivatives are admissible under the problem constraints) then we compute
the absolute value of the smallest possible derivative.
It relates via Bayes’ rule to a quantity that we call the counterfactual: p (yt
| τ, x). This conditional
density captures the counterfactual, or potential outcome of a treatment given that the individual was
assigned a different treatment (possibly based on confounding variables). The assignment only affects
the potential outcome when the ignorability assumption is violated. In such cases, knowledge of the
individual’s treatment assignment T = τ is informative of one or more hidden confounders, which by
139

definition affect the individual’s intrinsic potential outcomes. What is useful for actionable causal inference
is the distribution of a potential outcome conditioned just on covariates:
p (yt
| x) = Z
p (yt
| τ, x) p(τ | x)dτ
which informally integrates over all the possible individuals and therefore correctly averages out the
hidden confounding. Clearly, these counterfactuals cannot be inferred because the confounding influences
through the treatment assignment are unobservable. In (Marmarelis et al., 2023), we describe how to partially
identify that entire integral, therefore producing bounds for admissible potential-outcome distributions
and their expectations E [Yt
| X]. We learned two predictive models to achieve this: the observed outcome
p(y | τ, x) and the nominal propensity p(τ | x). Since the diet scores are bounded within 1-10, we rescaled
to the unit interval and used the Balanced Beta parametrization for the δMSM and the propensities, as
described in (Marmarelis et al., 2023). Here, we extend our previous work to evaluate diet effects on
sex-stratified models of brain structure.
Code Availability. The GitHub repository for our method may be found here: https://github.com/marma
relis/TreatmentCurves.jl.
4.2.3.2 Analysis
Data from 42,032 UK Biobank participants (mean age ± standard deviation: 64.57± 7.7) with structural
MRI were included. Females comprised 51.1% of this total and were on average slightly younger (64 ± 7.5)
compared to males ( 65.2 ± 7.8 ).
Lifestyle, dietary scores, and other covariates. Lifestyle factors (see Table 4.4) documented at the time
of imaging were included as covariates. Diet quality was calculated using the same coding scheme as in
(Said et al., 2018) and (Zhuang et al., 2021). Briefly, an overall diet quality score was computed as the sum of
individual diet components that corresponded to ideal (score of 10) and poor (score of 0 ) consumption of
140

various diet components. Ideal for fruit, vegetable, whole grains, fish, dairy, and vegetable oil generally
meant a higher consumption was better, whereas ideal for refined grains, processed meat, unprocessed
meat, and sugary food/drink intake meant lower consumption was better. Age at scan, ApoE4 (additive
coding), and Townsend deprivation index (TDI), a measure of socioeconomic status, (Phillimore et al., 1988)
were also included as covariates.
Table 4.4: Lifestyle factors and respective UK Biobank data field IDs. American Heart Association (AHA)
guidelines for weekly ideal (≥150 min/week moderate or ≥75 min/wk vigorous or 150 min/week mixed),
intermediate (1–149 min/week moderate or 1–74 min/week vigorous or 1–149 min/week mixed), and poor
(not performing any moderate or vigorous activity) physical activity. Supplementation was categorized into
any vitamins/minerals or fish oil intake. Salt added to food and variation in diet included the responses of
“never or rarely”, “sometimes”, “usually”, or “always/often”. Coffee, tea, and water intake were integer values
representing cups/glasses per day. Smoking status included never having smoked, previously smoked, and
currently smokes. Alcohol frequency was categorized as infrequent (1-3 times a month, special occasions
only, or never), occasional (1-2 a week or 3-4 times a week), and frequent (self-report of daily/almost daily
and ICD conditions F10, G312, G621, I426, K292, K70, K860, T510). Social support/contact variables included
attending any type of leisure/social group events, having family/friend visits twice a week or more, and
being able to confide in someone almost daily.
Lifestyle Factor Features (Data Field ID)
Physical Activity/
Body Composition
AHA physical activity (884, 904, 894, 914); waist to hip ratio (48,49);
body mass index (BMI) (23104); body fat percentage (23099)
Sleep sleep 7-9 hours/night (1160); job involves night shift work (3426);
daytime dozing/sleeping (1220)
Diet/
Supplements
diet quality scores (overall) and for the following components:
fruit (1309, 1319); vegetables (1289, 1299);
whole grains (1438, 1448, 1458, 1468); fish (1329, 1339);
dairy (1408, 1418); vegetable oil (1428, 2654, 1438);
refined grains (1438, 1448, 1458, 1468); processed meats (1349, 3680);
unprocessed meats (1369, 1379, 1389, 3680); sugary foods/drinks (6144).
fish oil supplementation (20084); vitamin/mineral supplementation (20084);
salt added to food (1478); variation in diet (1548); water intake (1528);
tea intake (1488); coffee intake (1498)
Education college/university (6138)
Smoking smoking status (20116)
Alcohol alcohol intake frequency (1558/ICD)
Social
Contact/Support
attending leisure/social group events (6160);
frequency of friends/family visits (1031);
able to confide in someone (2110)
MRI-derived features. Bilateral (left, right averaged) cortical thickness and subcortical volumes were
derived from FreeSurfer v7.1 (Fischl, 2012). Cortical parcellations from the Desikan-Killiany (DK) atlas
141

(Desikan et al., 2006) were used where 34 distinct regions on each cortical hemisphere are labeled according
to the gyral patterns. We also assessed 13 subcortical volumes.
Network architecture. To calculate CACD, we first implemented multilayer perceptrons with single-skip
connections with 40 inputs, which were the covariates and diet scores across 47 cortical and subcortical
outputs. We used Swish activations in the inner layers (Ramachandran et al., 2017) and trained with an
Adam optimizer (Kingma & Ba, 2014). We partitioned the entire dataset into 75/25 train/test splits four
times such that the test sets did not overlap and we could obtain out-of-sample predictions for all individuals.
The network had the following hyperparameters: layers:3; hidden units:32; learning rate: 5×10−3
; training
epochs: 104
; batch size: n/10; and ensemble size:16.
Presented outcomes. First, we present regional effect sizes for associations between 47 brain measures
and diet scores derived from linear regressions, covarying for the lifestyle factors listed in Table 4.4. This
will help serve as a comparison method given this is the most common statistical approach taken in the
literature. We also present regional brain effects as causal derivatives, specifically CACDs, expressed as
percentage changes from the individual’s observed outcome. Therefore, the CACD can be interpreted as,
"a one unit increase in diet score causes a X% increase in the outcome measure" for all measures except
ventricular volumes, which represents a decrease in outcome measure. In Figure 4.3, we show the CACD
values closest to zero that were admissible in the set of partially identified CACDs for each individual
and with each level of hidden confounding (Γ). False discovery rates (Benjamini & Hochberg, 1995) of all
regional one-sample Student t-test p-values across males and females were represented as q. We considered
a metric of robustness to hidden confounding as the largest Γ (out of a grid of tested values) for which the
diminished regional CACDs survived (i.e., were nonzero) with q < 0.05. CACDs with effect robustness
Γ ≥ 1.050 are reported as fully identified (where we assume no hidden confounding) percentage-change
estimates. In the rare cases where the fully identified, Γ = 1 estimates are insignificant with q ≥ 0.05 while
also being robust to some Γ > 1, we do not present those effects.
142

4.2.4 Results
Model Accuracy. All model errors were ≤ 1 (z-score normalized MSE).
Cortical thickness. The largest effect in the traditional linear model was in the transverse temporal
thickness in females for average diet score, however, this region did not withstand any amount of confounding in the causal sensitivity model. In males, the superior temporal and the postcentral gyrus had
relatively large effects for whole grain consumption that withstood high levels of confounding and had
significant CACD. A greater extent of significant causal effects robust to hidden confounding were observed
in males than females. The effects that were robust to the highest level of hidden confounding (Γ = 1.10)
were observed in males in the superior temporal (CACD = 0.042%; q = 3.6 × 10−80
, postcentral

CACD = 0.036%; q = 1.1 × 10−49
, and superior frontal thickness
CACD = 0.031%; q = 2.2 × 10−48 )
for whole grain intake, and the medial orbitofrontal
CACD = 0.015%; q = 8 × 10−7

and insular thickness (CACD = 0.009%; q = 5.7 × 10−5
) for vegetable intake. Causal derivatives that withstand non-zero
confounding were also observed across the brain in males for dairy, unprocessed meat (lower intake), and
whole grain intake (Figure 4.4).

Figure 4.4: A. Standardized β effects from a linear regression (LM) for thickness associations with diet
scores (thresholded at uncorrected p < 0.05 ); few effects survived FDR correction including: the medial
orbitofrontal for vegetables and whole grains, postcentral and superior temporal for whole grains in males,
and the cuneus for fish, and transverse temporal for average diet in females. B. Effect robustness, or the
extent of confounding that can be tolerated as causal derivatives are bounded away from zero. C. Causal
derivatives for regions with Γ ≥ 1.05 seen in B, i.e. percent difference between actual and predicted
thickness per unit increase in diet score.
Subcortical volumes. Several subcortical regions, particularly for the effects of whole grain consumption,
were linearly associated with diet scores, but did not have significant causal effects in females as opposed
to males. A greater extent of significant causal effects robust to hidden confounding were observed in
males compared to females. The effect of the accumbens volume
CACD = 0.05%; q = 1.4 × 10−9
) for
vegetable intake was robust to the highest level of confounding ( Γ = 1.075 ). Subcortical regions that had
significant causal derivatives given a confounding level of Γ = 1.05 are highlighted in Fig. 4.5C.
14

Figure 4.5: A. Standardized β effects from a linear regression (LM) for subcortical volume associations
with diet scores (thresholded at uncorrected p < 0.05 ); effects that survived FDR correction included:
the accumbens, cerebellum cortex & white matter, and ventral DC for fruits in females; the accumbens
for vegetables in males and females; the cerebellum white matter and thalamus for vegetables in males;
the accumbens, amygdala, cerebellum cortex & white matter, hippocampus, thalamus, and ventral DC for
whole grains in males and females; the caudate, lateral ventricle and putamen for whole grains in females;
and the accumbens, amygdala, cerebellum cortex, and hippocampus in males for vegetable oil. B. Effect
robustness, or the extent of confounding that can be tolerated as causal derivatives are bounded away from
zero. C. Causal derivatives for regions with Γ ≥ 1.05 seen in B, i.e., percent difference between actual and
predicted volume per unit increase in diet score. Nucleus accumbens not pictured.
4.2.5 Discussion
Here, we use a causal sensitivity model to study the causal effects of specific diet components on brain
structure from a large scale epidemiological study, as opposed to randomized control trials. Prior literature
on the relationship between cortical thickness and diet show contradicting findings, where some studies
fail to detect an effect while others find associations between higher thickness and adherence to a healthy
145

diet, particularly in the entorhinal and posterior cingulate cortices (Drouka et al., 2022). A healthy diet
has multiple components, and in this work, we evaluated the effect of the components of diet on male and
female brain structure.
We find that the causal effects of incremental changes of a better diet appeared more robust to confounding factors in males compared to females. Our model suggests that in males, a higher intake of whole
grains, vegetables, dairy, and vegetable oil, and a lower intake of unprocessed meat results in higher cortical
thickness and subcortical volumes. While some of the strongest effects in our causal model were also
detected in a standard linear model, other associations do not withstand a high degree of confounding,
particularly in females. For example, several subcortical regions, including hippocampal volume, which has
been shown to be positively associated with a better quality diet (Townsend et al., 2022), are associated with
a more favorable consumption of whole grains in the standard linear model, but in our casual sensitivity
model, do not withstand high levels of confounding. We caution against overinterpreting cross-sectional
results as these may be subject to hidden confounding, particularly in females.
Overall, our approach suggests causal effects of diet on brain structure can be identified despite some
degree of hidden confounding, allowing new computational approaches for modeling how lifestyle changes
may contribute to improved brain health and lower ADRD risk. Future work will continue to interrogate
causal sensitivity models with respect to other disease modifying effects across the lifespan.
4.2.6 Acknowledgements
Funding: R01AG059874, U01AG068057, P41EB05922. UK Biobank Resource under Application Number ’
11559 ’.
146

Chapter 5
Future works
5.1 A virtually uncharacterized population in brain imaging studies:
individuals from the Middle East and North Africa (MENA)
Understanding the effect of disease and aging pathophysiology in diverse ethnoracial backgrounds is critical
to provide efficacious and optimized healthcare to all populations. The Global Burden of Disease Study
(GBD) 2019 estimates that there will be a 166% increase in Alzheimer’s disease and related dementias
(ADRDs) cases over the next 25 years, with many high income countries of the Asia Pacific, Western Europe
and North America regions, having rates that are plateauing or declining. Nearly every single one of
the countries of the Middle East and North Africa (MENA) super region, however, will see over a 300%
growth in the incidence of ADRDs between 2019 and 2050. Overall, people from MENA countries will
see an increase that is more than double the global rate (367%), making it the region that will contribute
the most to total cases worldwide (E. Nichols et al., 2022). Yet, people from MENA countries are highly
underrepresented in health studies of aging and ADRDs. For example, individuals of MENA ancestry make
up nearly 10% of the global population, but have been left out of almost all genome-wide association studies
(GWAS) of human traits. These disparities contribute to the inability of largely European GWAS results to
replicate across ethnic groups and not only hinders translation to clinical practice, but could dangerously
147

influence public health policy (Sirugo et al., 2019). Given the complex and polygenic nature of late on-set
dementias, the current disparities in health studies of ADRD are paramount to address if we wish to change
the current forecasts of ADRDs.
Though a large percentage of the expected rise in ADRD rates among MENA countries is attributed to
a newly aging population and population growth, a larger proportion of increased rates are due to GBD
risk factors when compared to the global rate. These include metabolic and vascular risk related variables
such as high body mass index (BMI), high systolic blood pressure, high fasting plasma glucose, smoking,
low physical activity, low education, alcohol use, and air pollution exposure. A recent followup study from
the GBD 2019 showed that compared to the global rate, the MENA super region showed a 39% higher
potential preventable proportion of ADRDs, highlighting the importance of addressing modifiable factors
in this population (GBD 2019 North Africa and the Middle East Neurology Collaborators, 2024) (GBD
2019 North Africa and the Middle East Neurology Collaborators, 2024). Furthermore, ADRD risk is also
influenced by sex. In several MENA countries, there is also a sharp contrast in the prevalence of obesity
between men and women, with obesity being twice as prevalent in women than in men (NCD Risk Factor
Collaboration 2017). Obesity is strongly influenced by environmental factors such as diet and stress and is
associated with alterations in brain morphology and function and increased risk for ADRDs (Arnoldussen
et al., 2014). While obesity and other lifestyle factors are largely modifiable, they’re also complex, influenced
by both sociocultural-environmental and genetic factors. Thus, understanding differences in risk profiles
within ethnic populations is needed to provide a more precise assessment of the bio-social-environmental
determinants of risk for ADRDs in this uncategorized aging population. Research into these nuances can
lead to the curation of culture-specific interventions required to reduce the population-specific burdens of
preventable risk factors and ultimately to reduce the global burden of ADRDs.
Importantly, much of an individual’s risk based on these factors would not decline if the person migrated
to a lower-risk country. Over the last five decades, Western Europe and North America regions have seen
148

a growing influx of immigrants, refugees, and asylum seekers from MENA countries. Many of these
immigrants are motivated to do so due to political instability in their countries of origin and to obtain
economic opportunities. The MENA population experiences significant health disparities and barriers to
their unmet health needs, similar to the levels experienced by African and Hispanic immigrants. Yet there
is limited data on how environmental stressors and sociocultural factors (e.g., immigration circumstances,
language barriers, social isolation, xeno- and Islamophobia, discrimination, challenges acquiring secure
employment and housing) interact with biology to contribute to risk for negative health outcomes and
cognitive decline (Patel et al., 2022). A recent study comparing self-reported cognitive impairment rates,
suggested higher rates of cognitive impairment in Arab-Americans than in foreign born and native born
non-Hispanic Whites (Dallo et al., 2021). However, the mechanisms driving cognitive impariment and their
biological underpinnings have yet to be evaluated in this population. Large biobank datasets may be an
untapped resource to begin to do this.
5.2 Leveraging large biobanks with immigrant populations
The UK biobank (UKB) is a large, densely phenotyped study on aging which can provide insight into
health markers in immigrant groups in the UK (Bycroft et al., 2018). Its rich characterization of sociodemographic, disease, lifestyle, and genetic variables allows researchers to begin to disentangle ethnic and
culture specific tendencies which may be the cause of differing susceptibility to ADRDs. Furthermore,
the UKB also collects brain imaging data, which is an invaluable tool to noninvasively characterize brain
outcomes and their associations with ADRDs and their risk factors, as people from MENA countries remain
virtually uncharacterized in brain imaging studies of ADRD. In individuals at risk for ADRDs, changes in
neuroimaging derived biomarkers typically occur in advance of noticeable cognitive decline. Multimodal
magnetic resonance imaging (MRI), allows researchers to non-invasively map distinct features of ADRD
pathology. The most common MRI sequence used in ADRD research is T1 weighted (T1w) imaging, which
149

allows for the assessment of localized cortical atrophy patterns, hippocampal volume loss, and ventricular
enlargement – all characteristic features of AD that are used to track disease progression. However, personalized treatment regimens may wish to target aspects of pathology that are known to be more influenced by
modifiable risk factors such as lifestyle as opposed to non-modifiable factors such as deterministic genetic
effects. For example, vascular cognitive impairment and dementia has a more heterogeneous phenotype,
where oftentimes, hippocampal volumes may be relatively preserved. While heterogeneous, the vascular
contributions to dementia can be quantified using other MRI sequences sensitive to vascular pathologies.
T2w sequences such as fluid attenuated inversion recovery (FLAIR) imaging are useful for quantifying white
matter lesions (WML) and enlarged perivascular spaces – two consequences of hypertension-induced arteriolosclerosis. T2* GRE sequences such as susceptibility weighted imaging (SWI) can be used to detect iron
containing substances, such as cerebral microbleeds (CMB), for which extent and location are informative of
dementia subtype etiology. Moreover, CMBs are related to amyloid-related imaging abnormalities (ARIA),
particularly hemorrhagic ARIA (ARIA-H), reported in recent Alzheimer’s clinical trials of amyloid clearing
medications. Given that these trials are largely performed in those of European ancestry, identifying
differences in the extent of existing CMB load and susceptibilities may be an informative marker for clinical
trial stratification. Thus, analyzing multimodal MRI features may be the best way to help characterize
ethno-specific susceptibilities in-vivo which may inform appropriate treatment plans and management.
Below are two preliminary analysis which leverage the UK Biobank to explore 1.) differences in disease
prevalence and CMB associations in UK immigrants from MENA countries and 2.) further characterization of
CMB associations with demographic, disease, and other imaging markers in those who already have varying
subtypes of dementia. Future work will extend these analyses to compare CMB associations across other
ethnic groups in addition to extending our analyses to investigate more comprehensive biopsychosocial
and imaging associations across these ethnic groups.
150

5.2.1 Preliminary analysis: Cerebral microhemorrhage associations in UK immigrants
from the Middle East and North Africa
This section is adapted from:
Haddad E, Sheikh-Bahaei N, Javid S, Jahanshad N. (2023). Cerebral microhemorrhage associations in UK
immigrants from the Middle East and North Africa (MENA). Alzheimer’s & Dementia: The Journal of the
Alzheimer’s Association, 19(S17). https://doi.org/10.1002/alz.080629
5.2.1.1 Introduction
Understanding the effect of disease and aging pathophysiology in diverse ethnoracial backgrounds is
critical to provide efficacious and optimized healthcare to all populations. People from the Middle East and
North Africa (MENA) are highly underrepresented in health studies, yet, they are forecasted to contribute
the most to increasing projections of dementia over the next 30 years (E. Nichols et al., 2022). Cerebral
microhemorrhages, or microbleeds (CMBs) have been linked to small vessel disease, neurodegeneration,
and aging and may provide a proxy for intervention studies to identify those at higher risk (Charidimou &
Werring, 2011). For example, CMBs are related to amyloid-related imaging abnormalities (ARIA), particularly
hemorrhagic ARIA (ARIA-H), reported in recent Alzheimer’s clinical trials of amyloid clearing medications
(Sperling et al., 2011). Identifying those at elevated risk for ARIA-H can help refine inclusion criteria for
clinical trials and mitigate risk of complications from these medications. Fortunately, CMBs are readily
detected on T2*-GRE MRI sequences, such as susceptibility weighted imaging (SWI) – a modality used to
detect magnetic field distortions caused by products like iron (Haacke et al., 2004; S. Liu et al., 2017). The
UK Biobank is a large, densely phenotyped study on aging which can provide insight into health markers
in immigrant groups in the UK (Miller et al., 2016b). Here, we compared the prevalence of metabolic and
cardiovascular diseases in people born in MENA countries to other populations and investigated the effect
of these diseases and lifestyle on the presence of CMBs detected on SWI.
151

5.2.1.2 Methods
Analyses of non-imaging and imaging features were conducted. First, we compared the prevalence of
metabolic and cardiovascular disorders in a large sample of MENA immigrants (N=3557) to immigrants from
India (N=2959) and Germany (N=1103). In the imaged subset of MENA subjects who had SWI (N=185), we
assessed the effect of disease and lifestyle factors on CMBs. We used the MARS rating scale (Gregoire et al.,
2009) to identify CMBs on SWI and validated our findings using other sequences to exclude CMB ‘mimics’
(Figure 5.1). The number and position of CMBs were reported as well as their classification including
‘definite’, ’possible’ and ‘total’. The effects of disease and lifestyle were investigated using association
rule learning (ARL) (Agrawal et al., 1993), and Chi-squared/Fisher’s exact tests on the most prominent
associations.
5.2.1.3 Results
Compared to Germans, MENA and Indian groups showed higher rates of diabetes and hypercholesterolemia
(Figure 5.2). In our MENA imaging subset, 10 subjects had at least one definite CMB (70% strictly lobar) and
20 had at least one definite or possible CMB (85% strictly lobar). ARL showed combinations of male sex,
diabetes, obesity, and high waist-to-hip ratio have 100% likelihood for having a CMB in our population.
Post-hoc testing revealed significant associations between the presence of definite CMB and diabetes
(p=0.017) and high BMI (overweight) (p=0.017).
5.2.1.4 Discussion
A high prevalence of metabolic disorders (ie. diabetes) was found in MENA, which was associated with a
higher rate of CMBs. CMBs were associated with male sex, and abnormally high anthropometric measures
in the MENA group. The spatial distribution of the CMBs could reflect underlying pathologies, where
strictly lobar CMBs are associated with amyloid angiopathy, and deep/mixed CMBs are associated with
152

hypertensive arteriopathy and metabolic syndromes such as diabetes and hypercholesterolemia (Beaman
et al., 2022; Sperling et al., 2011), two conditions with a higher incidence in MENA. Nevertheless and
independent of their underlying pathology, both groups of CMBs can contribute to brain health and
development of dementia and can play a role as a potential risk for future treatment stratification. Further
work testing a larger population and teasing apart differential associations with CMB location and markers
of health is required, particularly in diverse ethnoracial background groups. Our findings start to shed light
on distinctive aging and disease risk profiles in underrepresented populations, including those from MENA
countries.
153

Figure 5.1: Microbleed Anatomical Rating Scale (MARS) criteria. A. MARS criteria used to classify the
presence of microbleeds in the UK Biobank. The consensus criteria had to be met in order to be marked as
a ‘definite’ microbleed. Possible microbleeds reflect signal that may be less hypointense and differentiation
from a mimic is less certain (van der Eerden et al., 2022). ‘Total’ microbleed counts reflect the combination of
definite and possible. B. Example ‘mimics’ of hemorrhagic and nonhemorrhagic nature. Other confirmatory
procedures include observing a darker signal on the second echo (TE2) compared to the first (TE1) and
excluding basal ganglia signal that mimic CMBs which are bilateral, as these are likely to be vasculature.
154

Figure 5.2: Disease prevalence and microbleed findings. A. Non-imaging population age histogram, sex
distribution by birth country group, and disease prevalences binned into 5-year age increments and stratified
by sex. Males from MENA countries and India showed distinctly higher rates of coronary artery disease and
heart failure with increasing age, compared to the German group. Females from MENA countries showed
the highest rates of dementia starting at age 65 and obesity was most prevalent in MENA populations –
particularly in females, which showed upwards of 30% prevalence at all age bins. B. Imaging population age
histogram colored by the presence of definite, total, and no CMBs, microbleed counts colored by certainty
for all subjects which were found to have potential CMBs, ARL frequencies and respective sets with 100%
confidence, and post-hoc results from the strongest ARL associations. 10 subjects were considered to have
at least one definite CMB (90% M; mean age (sd): 62.4 ± 10.19; 70% strictly lobar) and 20 were considered
to have at least one definite or possible CMB (75% M; mean age (sd): 65.8 ± 9.60; 85% strictly lobar). ARL
showed combinations of being male, diabetic, overweight/obese, and having a high waist-to-hip ratio to
have 100% likelihood for having a CMB (having all of these factors resulted in CMBs 100% of the time), and
post-hoc testing in these factors showed significant associations between definite microbleeds and having
diabetes (p=0.017) and being overweight (p=0.017).
155

5.2.2 Preliminary analysis: Characterizing cerebral microhemorrhage associations in
dementia subtypes in the UK Biobank
This section is adapted from:
Haddad E, Sheikh-Bahaei N, Jahanshad N (2023). Characterizing cerebral microhemorrhage associations in
dementia subtypes in the UK Biobank. Alzheimer’s & Dementia: The Journal of the Alzheimer’s Association,
19(S17). https://doi.org/10.1002/alz.080704
5.2.2.1 Introduction
Cerebral microbleeds (CMBs) are associated with neurodegenerative diseases (Charidimou & Werring,
2011) and have been identified as an adverse amyloid-related imaging abnormality event (ARIA-H) related
to amyloid clearing medications (Sperling et al., 2011). Identifying risk factors for CMBs can help refine
inclusion criteria for clinical trials and mitigate the risk of ARIA-H. CMBs are readily detected on T2*-GRE
MRI sequences, including susceptibility weighted imaging (SWI) – a modality used to detect magnetic field
distortions caused by paramagnetic products like iron (Haacke et al., 2004; S. Liu et al., 2017). We explore
CMB associations with demographic, disease, and imaging markers in a subset of participants with varying
dementia subtypes and complaints in the UK Biobank (Miller et al., 2016b).
5.2.2.2 Methods
A total of 75 participants who had clinically diagnosed or self-reported dementia and SWI imaging were
selected; 4 participants were excluded due to extreme motion. We used MARS criteria (Gregoire et al., 2009)
to identify CMBs on SWI and validated our findings using other sequences to exclude CMB ‘mimics’. The
number, classification, and position of CMBs were recorded. Logistic regressions modeling the outcome of
at least one ‘definite’ or ‘total’ (‘definite’ + ‘possible’) CMB were performed and effect sizes were calculated
156

across hippocampal volume, white matter hyperintensity load (WMH), and global WM diffusion MRI (dMRI)
microstructural measures.
5.2.2.3 Results
Participants were predominantly male (65%) and white (97%) with a mean age of 69.8 (± 7.1) years. The
most common ICD10 conditions were “unspecified dementia”, “Alzheimer’s disease”, and “Dementia in
Alzheimer’s disease”. 9 participants had at least one ‘definite’ CMB and 8 more had at least one ‘possible’
CMB for a ‘total’ of 17 individuals with CMBs. CMBs were mostly ‘strictly lobar’ for both definite (78%) and
total (71%) classifications. Individuals with CMB(s) were significantly older than those without. Definite
and total CMBs were found in mostly male populations as well (78% and 71% respectively). The prevalences
of hypertension, hypercholesterolemia, and stroke were notably higher in those with CMBs compared to
those without. Being an ApoE4 carrier was also more common in the definite (67% vs 48%) and total (59%
vs 48%) CMB groups. CMB mimics were found in 7% of the population (Table 5.1). The presence of CMBs
were found across a spectrum of dementia diagnoses (Figure 5.3). Microstructural dMRI measures including
global WM FA, MD, and ISOVF were significantly associated with presence of total CMBs (q<0.05) (Figure
5.4).
5.2.2.4 Discussion
Measures reflecting WM organization were sensitive to the pathophysiology of CMBs in those with dementia.
Future work should investigate whether WM microstructure can predict the development of CMBs in those
with dementia, particularly in individuals undergoing amyloid clearing treatments.
157

Table 5.1: Demographic and disease prevalences, dementia subtypes, and MARS classification results of
those with at least one definite or total CMB, and across the whole dementia population.
Definite Total All Dementia
MARS Classification No, N = 621 Yes, N = 91 No, N = 541 Yes, N = 171 N = 711
Demographics
Age 69.26 [7.25] 74.09 [4.20] 69.23 [7.55] 71.93 [5.09] 69.88 [7.10]
Sex: Female 23 (37%) 2 (22%) 20 (37%) 5 (29%) 25 (35%)
Sex: Male 39 (63%) 7 (78%) 34 (63%) 12 (71%) 46 (65%)
Race: White 60 (97%) 9 (100%) 52 (96%) 17 (100%) 69 (97%)
Race: Asian 1 (1.6%) 0 (0%) 1 (1.9%) 0 (0%) 1 (1.4%)
Race: Black 1 (1.6%) 0 (0%) 1 (1.9%) 0 (0%) 1 (1.4%)
Hypertension 35 (56%) 6 (67%) 30 (56%) 11 (65%) 41 (58%)
Diabetes 5 (8.1%) 1 (11%) 4 (7.4%) 2 (12%) 6 (8.5%)
Hypercholesterolemia 26 (42%) 6 (67%) 23 (43%) 9 (53%) 32 (45%)
Stroke 8 (13%) 2 (22%) 7 (13%) 3 (18%) 10 (14%)
Parkinson’s Disease 5 (8.1%) 0 (0%) 4 (7.4%) 1 (5.9%) 5 (7.0%)
Coronary Artery Disease 18 (29%) 2 (22%) 18 (33%) 2 (12%) 20 (28%)
Heart Failure 2 (3.2%) 0 (0%) 2 (3.7%) 0 (0%) 2 (2.8%)
Atrial Fibrillation 11 (18%) 3 (33%) 10 (19%) 4 (24%) 14 (20%)
Chronic Kidney Disease 5 (8.1%) 0 (0%) 5 (9.3%) 0 (0%) 5 (7.0%)
ApoE4: Carrier 29 (48%) 6 (67%) 25 (48%) 10 (59%) 35 (51%)
ApoE4: Not Available 2 0 2 0 2
Years of Education 16.2 [4.4] 15.2 [3.5] 16.3 [4.4] 15.3 [3.9] 16.1 [4.3]
Current Smoker 2 (3.2%) 1 (11%) 2 (3.7%) 1 (5.9%) 3 (4.2%)
Frequent Alcohol User 14 (23%) 0 (0%) 13 (24%) 1 (5.9%) 14 (20%)
Body Mass Index: Healthy 23 (39%) 3 (33%) 21 (40%) 5 (31%) 26 (38%)
Body Mass Index: Underweight 1 (1.7%) 0 (0%) 1 (1.9%) 0 (0%) 1 (1.5%)
Body Mass Index: Overweight 25 (42%) 4 (44%) 20 (38%) 9 (56%) 29 (43%)
Body Mass Index: Obese 10 (17%) 2 (22%) 10 (19%) 2 (13%) 12 (18%)
Body Mass Index: Not Available 3 0 2 1 3
Dementia Subtypes
Creutzfeldt-Jakob disease 3 (4.8%) 0 (0%) 3 (5.6%) 0 (0%) 3 (4.2%)
Dementia in Alzheimer’s disease 12 (19%) 2 (22%) 9 (17%) 5 (29%) 14 (20%)
Dementia in Alzheimer’s disease with early onset 3 (4.8%) 0 (0%) 3 (5.6%) 0 (0%) 3 (4.2%)
Dementia in Alzheimer’s disease with late onset 1 (1.6%) 0 (0%) 1 (1.9%) 0 (0%) 1 (1.4%)
Dementia in Alzheimer’s disease, atypical or mixed type 1 (1.6%) 1 (11%) 1 (1.9%) 1 (5.9%) 2 (2.8%)
Dementia in Alzheimer’s disease, unspecified 9 (15%) 1 (11%) 6 (11%) 4 (24%) 10 (14%)
Vascular dementia 2 (3.2%) 1 (11%) 2 (3.7%) 1 (5.9%) 3 (4.2%)
Other vascular dementia 1 (1.6%) 0 (0%) 1 (1.9%) 0 (0%) 1 (1.4%)
Vascular dementia, unspecified 2 (3.2%) 0 (0%) 2 (3.7%) 0 (0%) 2 (2.8%)
Dementia in other diseases classified elsewhere 4 (6.5%) 1 (11%) 4 (7.4%) 1 (5.9%) 5 (7.0%)
Dementia in Pick’s disease 2 (3.2%) 1 (11%) 2 (3.7%) 1 (5.9%) 3 (4.2%)
Dementia in other specified diseases classified elsewhere 2 (3.2%) 0 (0%) 2 (3.7%) 0 (0%) 2 (2.8%)
Unspecified dementia 22 (35%) 4 (44%) 19 (35%) 7 (41%) 26 (37%)
Delirium superimposed on dementia 4 (6.5%) 1 (11%) 4 (7.4%) 1 (5.9%) 5 (7.0%)
Amnesic syndrome 2 (3.2%) 0 (0%) 2 (3.7%) 0 (0%) 2 (2.8%)
Alzheimer’s disease 18 (29%) 4 (44%) 15 (28%) 7 (41%) 22 (31%)
Alzheimer’s disease with early onset 3 (4.8%) 0 (0%) 3 (5.6%) 0 (0%) 3 (4.2%)
Alzheimer’s disease with late onset 1 (1.6%) 0 (0%) 1 (1.9%) 0 (0%) 1 (1.4%)
Other Alzheimer’s disease 2 (3.2%) 1 (11%) 2 (3.7%) 1 (5.9%) 3 (4.2%)
Alzheimer’s disease, unspecified 15 (24%) 3 (33%) 12 (22%) 6 (35%) 18 (25%)
Circumscribed brain atrophy 3 (4.8%) 2 (22%) 3 (5.6%) 2 (12%) 5 (7.0%)
Other specified degenerative diseases of nervous system 6 (9.7%) 0 (0%) 5 (9.3%) 1 (5.9%) 6 (8.5%)
Self Report Dementia/Alzheimers/Cognitive Impairment 12 (19%) 3 (33%) 9 (17%) 6 (35%) 15 (21%)
MARS Location
Lobar 0 (0%) 7 (78%) 0 (0%) 12 (71%)
Deep 0 (0%) 1 (11%) 0 (0%) 3 (18%)
Infratentorial 0 (0%) 0 (0%) 0 (0%) 0 (0%)
Mixed - Lobar & Deep 0 (0%) 1 (11%) 0 (0%) 1 (5.9%)
Mixed - Lobar & Infratentorial 0 (0%) 0 (0%) 0 (0%) 1 (5.9%)
CMB Mimics 5 (7.0%)
1Mean [SD]; n (%)
158

Figure 5.3: Dementia subtypes for cases with definite and total CMBs as well as their respective counts
for each CMB location. The presence and location of CMBs were found across a spectrum of dementia
diagnoses.
159

Figure 5.4: A. Example imaging outputs for the respective measures. B. Data distributions across all measures
for both definite and total CMB presence. C. Effect sizes (Cohen’s d) between those with definite and total
CMBs compared to those without CMBs. Significance associations that survived multiple-testing correction
(q<0.05) are denoted by solid circles. Logistic regression covariates for hippocampal volume (mm3) and
white matter hyperintensities (mm3 on the log scale) included age, sex, and ICV, whereas diffusion metrics
only included age and sex.
160

Bibliography
Ad-Dabbagh, Y., Einarson, D., Lyttelton, O., Muehlboeck, J. S., Mok, I., Vincent, R. D. L., Lerch, J.,
Fombonne, E., & Evans, A. C. (2006). The CIVET Image-Processing environment: A fully
automated comprehensive pipeline for anatomical neuroimaging research.
Adle-Biassette, H., Chrétien, F., Wingertsmann, L., Héry, C., Ereau, T., Scaravilli, F., Tardieu, M., & Gray, F.
(1999). Neuronal apoptosis does not correlate with dementia in HIV infection but is related to
microglial activation and axonal damage. Neuropathol. Appl. Neurobiol., 25(2), 123–133.
Agrawal, R., Imieliński, T., & Swami, A. (1993). Mining association rules between sets of items in large
databases. Proceedings of the 1993 ACM SIGMOD international conference on Management of data -
SIGMOD ’93.
Ahnaou, A., & Drinkenburg, W. H. I. (2021). Sleep, neuronal hyperexcitability, inflammation and
neurodegeneration: Does early chronic short sleep trigger and is it the key to overcoming
alzheimer’s disease?
Ahonen, T., Vanhala, M., Kautiainen, H., Kumpusalo, E., & Saltevo, J. (2012). Sex differences in the
association of adiponectin and low-grade inflammation with changes in the body mass index from
youth to middle age. Gend. Med., 9(1), 1–8.
Alberro, A., Iribarren-Lopez, A., Sáenz-Cuesta, M., Matheu, A., Vergara, I., & Otaegui, D. (2021).
Inflammaging markers characteristic of advanced age show similar levels with frailty and
dependency. Sci. Rep., 11(1), 4358.
Alexander-Bloch, A., Giedd, J. N., & Bullmore, E. (2013). Imaging structural co-variance between human
brain regions. Nat. Rev. Neurosci., 14(5), 322–336.
Alfaro-Almagro, F., McCarthy, P., Afyouni, S., Andersson, J. L. R., Bastiani, M., Miller, K. L., Nichols, T. E., &
Smith, S. M. (2021). Confound modelling in UK biobank brain imaging. Neuroimage, 224, 117002.
Alzheimer’s Association 2024. (n.d.). 2024 alzheimer’s disease facts and figures.
Amoroso, N., La Rocca, M., Bellantuono, L., Diacono, D., Fanizzi, A., Lella, E., Lombardi, A., Maggipinto, T.,
Monaco, A., Tangaro, S., & Bellotti, R. (2019). Deep learning and multiplex networks for accurate
modeling of brain age. Front. Aging Neurosci., 11, 115.
161

Anstey, K. J., Ashby-Mitchell, K., & Peters, R. (2017). Updating the evidence on the association between
serum cholesterol and risk of late-life dementia: Review and meta-analysis. J. Alzheimers. Dis.,
56(1), 215–228.
Anthony, I. C., Ramage, S. N., Carnie, F. W., Simmonds, P., & Bell, J. E. (2006). Accelerated tau deposition in
the brains of individuals infected with human immunodeficiency virus-1 before and after the
advent of highly active anti-retroviral therapy. Acta Neuropathol., 111(6), 529–538.
Appleton, J. P., Scutt, P., Sprigg, N., & Bath, P. M. (2017). Hypercholesterolaemia and vascular dementia.
Clin. Sci., 131(14), 1561–1578.
Arnoldussen, I. A. C., Kiliaan, A. J., & Gustafson, D. R. (2014). Obesity and dementia: Adipokines interact
with the brain. Eur. Neuropsychopharmacol., 24, 1982–1999.
Babapour Mofrad, R., & van der Flier, W. M. (2019). Nature and implications of sex differences in AD
pathology. Nat. Rev. Neurol., 15(1), 6–8.
Baecker, L., Dafflon, J., da Costa, P. F., Garcia-Dias, R., Vieira, S., Scarpazza, C., Calhoun, V. D., Sato, J. R.,
Mechelli, A., & Pinaya, W. H. L. (2021). Brain age prediction: A comparison between machine
learning models using region- and voxel-based morphometric data. Hum. Brain Mapp., 42,
2332–2346.
Baecker, L., Garcia-Dias, R., Vieira, S., Scarpazza, C., & Mechelli, A. (2021). Machine learning for brain age
prediction: Introduction to methods and clinical applications. EBioMedicine, 72, 103600.
Bagit, A., Hayward, G. C., & MacPherson, R. E. K. (2021). Exercise and estrogen: Common pathways in
alzheimer’s disease pathology. Am. J. Physiol. Endocrinol. Metab., 321(1), E164–E168.
Ball, M. J. (1982). “limbic predilection in alzheimer dementia: Is reactivated herpesvirus involved?” Can. J.
Neurol. Sci., 9(3), 303–306.
Baranowski, B. J., Marko, D. M., Fenech, R. K., Yang, A. J. T., & MacPherson, R. E. K. (2020). Healthy brain,
healthy life: A review of diet and exercise interventions to promote brain health and reduce
alzheimer’s disease risk. Appl. Physiol. Nutr. Metab., 45(10), 1055–1065.
Barha, C. K., Best, J. R., Rosano, C., Yaffe, K., Catov, J. M., & Liu-Ambrose, T. (2020). Sex-Specific
relationship between Long-Term maintenance of physical activity and cognition in the health ABC
study: Potential role of hippocampal and dorsolateral prefrontal cortex volume. J. Gerontol. A Biol.
Sci. Med. Sci., 75(4), 764–770.
Barha, C. K., Liu-Ambrose, T., Best, J. R., Yaffe, K., Rosano, C., & Health, Aging and Body Composition
Study. (2019). Sex-dependent effect of the BDNF Val66Met polymorphism on executive functioning
and processing speed in older adults: Evidence from the health ABC study. Neurobiol. Aging, 74,
161–170.
Barnes, L. L., Capuano, A. W., Aiello, A. E., Turner, A. D., Yolken, R. H., Torrey, E. F., & Bennett, D. A.
(2015). Cytomegalovirus infection and risk of alzheimer disease in older black and white
individuals. J. Infect. Dis., 211(2), 230–237.
162

Bashyam, V. M., Erus, G., Doshi, J., Habes, M., Nasrallah, I., Truelove-Hill, M., Srinivasan, D., Mamourian, L.,
Pomponio, R., Fan, Y., Launer, L. J., Masters, C. L., Maruff, P., Zhuo, C., Völzke, H., Johnson, S. C.,
Fripp, J., Koutsouleris, N., Satterthwaite, T. D., . . . Davatzikos, C. (2020). MRI signatures of brain
age and disease over the lifespan based on a deep brain network and 14 468 individuals worldwide.
Brain, 143, 2312–2324.
Beaman, C., Kozii, K., Hilal, S., Liu, M., Spagnolo-Allende, A. J., Polanco-Serra, G., Chen, C., Cheng, C.-Y.,
Zambrano, D., Arikan, B., Del Brutto, V. J., Wright, C., Flowers, X. E., Leskinen, S. P., Rundek, T.,
Mitchell, A., Vonsattel, J. P., Cortes, E., Teich, A. F., . . . Alzheimer’s Disease Neuroimaging
Initiative. (2022). Cerebral microbleeds, cerebral amyloid angiopathy, and their relationships to
quantitative markers of neurodegeneration. Neurology, 98, e1605–e1616.
Beck, D., de Lange, A.-M. G., Pedersen, M. L., Alnaes, D., Maximov, I. I., Voldsbekk, I., Richard, G.,
Sanders, A.-M., Ulrichsen, K. M., Dørum, E. S., Kolskår, K. K., Høgestøl, E. A., Steen, N. E.,
Djurovic, S., Andreassen, O. A., Nordvik, J. E., Kaufmann, T., & Westlye, L. T. (2022).
Cardiometabolic risk factors associated with brain age and accelerate brain ageing. Hum. Brain
Mapp., 43, 700–720.
Beheshti, I., Nugent, S., Potvin, O., & Duchesne, S. (2019). Bias-adjustment in neuroimaging-based brain age
frameworks: A robust scheme. Neuroimage Clin, 24, 102063.
Beijers, L., Wardenaar, K. J., van Loo, H. M., & Schoevers, R. A. (2019). Data-driven biological subtypes of
depression: Systematic review of biological approaches to depression subtyping. Mol. Psychiatry,
24(6), 888–900.
Belloy, M. E., Napolioni, V., & Greicius, M. D. (2019). A quarter century of APOE and alzheimer’s disease:
Progress to date and the path forward.
Benjamini, Y., & Hochberg, Y. (1995). Controlling the false discovery rate: A practical and powerful
approach to multiple testing. J. R. Stat. Soc. Series B Stat. Methodol., 57(1), 289–300.
Bennett, C. M., & Miller, M. B. (2013). fMRI reliability: Influences of task and experimental design. Cogn.
Affect. Behav. Neurosci., 13(4), 690–702.
Bessi, V., Mazzeo, S., Bagnoli, S., Padiglioni, S., Carraro, M., Piaceri, I., Bracco, L., Sorbi, S., & Nacmias, B.
(2020). The implication of BDNF Val66Met polymorphism in progression from subjective cognitive
decline to mild cognitive impairment and alzheimer’s disease: A 9-year follow-up study.
Bethlehem, R. A. I., Seidlitz, J., White, S. R., Vogel, J. W., Anderson, K. M., Adamson, C., Adler, S.,
Alexopoulos, G. S., Anagnostou, E., Areces-Gonzalez, A., Astle, D. E., Auyeung, B., Ayub, M., Bae, J.,
Ball, G., Baron-Cohen, S., Beare, R., Bedford, S. A., Benegal, V., . . . Alexander-Bloch, A. F. (2022).
Brain charts for the human lifespan. Nature.
Bigler, E. D., Skiles, M., Wade, B. S. C., Abildskov, T. J., Tustison, N. J., Scheibel, R. S., Newsome, M. R.,
Mayer, A. R., Stone, J. R., Taylor, B. A., Tate, D. F., Walker, W. C., Levin, H. S., & Wilde, E. A. (2020).
FreeSurfer 5.3 versus 6.0: Are volumes comparable? a chronic effects of neurotrauma consortium
study. Brain Imaging Behav., 14(5), 1318–1327.
163

Bittner, N., Jockwitz, C., Franke, K., Gaser, C., Moebus, S., Bayen, U. J., Amunts, K., & Caspers, S. (2021).
When your brain looks older than expected: Combined lifestyle risk and BrainAGE. Brain Struct.
Funct., 226, 621–645.
Blennow, K., de Leon, M. J., & Zetterberg, H. (2006). Alzheimer’s disease. Lancet, 368, 387–403.
Bocancea, D. I., van Loenhoud, A. C., Groot, C., Barkhof, F., van der Flier, W. M., & Ossenkoppele, R. (2021).
Measuring resilience and resistance in aging and alzheimer disease using residual methods: A
systematic review and meta-analysis. Neurology, 97, 474–488.
Boekel, W., Wagenmakers, E.-J., Belay, L., Verhagen, J., Brown, S., & Forstmann, B. U. (2015). A purely
confirmatory replication study of structural brain-behavior correlations. Cortex, 66, 115–133.
Botvinik-Nezer, R., Holzmeister, F., Camerer, C. F., Dreber, A., Huber, J., Johannesson, M., Kirchler, M.,
Iwanir, R., Mumford, J. A., Adcock, R. A., Avesani, P., Baczkowski, B. M., Bajracharya, A., Bakst, L.,
Ball, S., Barilari, M., Bault, N., Beaton, D., Beitner, J., . . . Schonberg, T. (2020). Variability in the
analysis of a single neuroimaging dataset by many teams. Nature, 582(7810), 84–88.
Bowring, A., Maumet, C., & Nichols, T. E. (2019). Exploring the impact of analysis software on task fMRI
results. Hum. Brain Mapp., 40(11), 3362–3384.
Braak, H., & Braak, E. (1991). Neuropathological stageing of alzheimer-related changes. Acta Neuropathol.,
82(4), 239–259.
Brew, B. J., Pemberton, L., Blennow, K., Wallin, A., & Hagberg, L. (2005). CSF amyloid beta42 and tau levels
correlate with AIDS dementia complex. Neurology, 65(9), 1490–1492.
Brown, B. M., de Frutos Lucas, J., Porter, T., Frost, N., Vacher, M., Peiffer, J. J., & Laws, S. M. (2022).
Non-Modifiable factors as moderators of the relationship between physical activity and brain
volume: A Cross-Sectional UK biobank study. J. Alzheimers. Dis.
Bryer, E. J., Medaglia, J. D., Rostami, S., & Hillary, F. G. (2013). Neural recruitment after mild traumatic
brain injury is task dependent: A meta-analysis. J. Int. Neuropsychol. Soc., 19(7), 751–762.
Bu, X.-L., Yao, X.-Q., Jiao, S.-S., Zeng, F., Liu, Y.-H., Xiang, Y., Liang, C.-R., Wang, Q.-H., Wang, X.,
Cao, H.-Y., Yi, X., Deng, B., Liu, C.-H., Xu, J., Zhang, L.-L., Gao, C.-Y., Xu, Z.-Q., Zhang, M.,
Wang, L., . . . Wang, Y.-J. (2015). A study on the association between infectious burden and
alzheimer’s disease. Eur. J. Neurol., 22(12), 1519–1525.
Bubak, A. N., Como, C. N., Coughlan, C. M., Johnson, N. R., Hassell, J. E., Mescher, T., Niemeyer, C. S.,
Mahalingam, R., Cohrs, R. J., Boyd, T. D., Potter, H., Russ, H. A., & Nagel, M. A. (2020).
Varicella-Zoster virus infection of primary human spinal astrocytes produces intracellular amylin,
Amyloid-β, and an amyloidogenic extracellular environment. J. Infect. Dis., 221(7), 1088–1097.
Bunch, T. J. (2020). Atrial fibrillation and dementia. Circulation, 142(7), 618–620.
Butler, E. R., Chen, A., Ramadan, R., Le, T. T., Ruparel, K., Moore, T. M., Satterthwaite, T. D., Zhang, F.,
Shou, H., Gur, R. C., Nichols, T. E., & Shinohara, R. T. (2021). Pitfalls in brain age analyses. Hum.
Brain Mapp., 42, 4092–4101.
164

Button, K. S., Ioannidis, J. P. A., Mokrysz, C., Nosek, B. A., Flint, J., Robinson, E. S. J., & Munafò, M. R. (2013).
Power failure: Why small sample size undermines the reliability of neuroscience. Nat. Rev.
Neurosci., 14(5), 365–376.
Bycroft, C., Freeman, C., Petkova, D., Band, G., Elliott, L. T., Sharp, K., Motyer, A., Vukcevic, D.,
Delaneau, O., O’Connell, J., Cortes, A., Welsh, S., Young, A., Effingham, M., McVean, G., Leslie, S.,
Allen, N., Donnelly, P., & Marchini, J. (2018). The UK biobank resource with deep phenotyping and
genomic data. Nature, 562(7726), 203–209.
Cairns, D. M., Rouleau, N., Parker, R. N., Walsh, K. G., Gehrke, L., & Kaplan, D. L. (2020). A 3D human
brain-like tissue model of herpes-induced alzheimer’s disease. Sci Adv, 6(19), 8828.
Calude, C. S., & Longo, G. (2017). The deluge of spurious correlations in big data. Found. Sci., 22, 595–612.
Carp, J. (2012). The secret lives of experiments: Methods reporting in the fMRI literature. Neuroimage, 63(1),
289–300.
Carroll, C. M., & Macauley, S. L. (2019). The interaction between sleep and metabolism in alzheimer’s
disease: Cause or consequence of disease?
Chagnot, A., Barnes, S. R., & Montagne, A. (2021). Magnetic resonance imaging of Blood–Brain barrier
permeability in dementia. Neuroscience, 474, 14–29.
Chang Wong, E., & Chang Chui, H. (2022). Vascular cognitive impairment and dementia. Continuum, 28(3),
750–780.
Charidimou, A., & Werring, D. J. (2011). Cerebral microbleeds: Detection, mechanisms and clinical
challenges. Future Neurol., 6, 587–611.
Chavent, M., Kuentz-Simonet, V., Labenne, A., & Saracco, J. (2014). Multivariate analysis of mixed data: The
R package PCAmixdata.
Chen, G.-F., Xu, T.-H., Yan, Y., Zhou, Y.-R., Jiang, Y., Melcher, K., & Xu, H. E. (2017). Amyloid beta: Structure,
biology and structure-based therapeutic development. Acta Pharmacol. Sin., 38(9), 1205–1235.
Chen, L. Y., Ribeiro, A. L. P., Platonov, P. G., Cygankiewicz, I., Soliman, E. Z., Gorenek, B., Ikeda, T.,
Vassilikos, V. P., Steinberg, J. S., Varma, N., Bayés-de-Luna, A., & Baranchuk, A. (2022). P wave
parameters and indices: A critical appraisal of clinical utility, challenges, and future Research-A
consensus document endorsed by the international society of electrocardiology and the
international society for holter and noninvasive electrocardiology. Circ. Arrhythm. Electrophysiol.,
15(4), e010435.
Chen, Y., Kim, M., Paye, S., & Benayoun, B. A. (2022). Sex as a biological variable in nutrition research:
From human studies to animal models. Annu. Rev. Nutr., 42, 227–250.
Chow, L. S., Gerszten, R. E., Taylor, J. M., Pedersen, B. K., van Praag, H., Trappe, S., Febbraio, M. A.,
Galis, Z. S., Gao, Y., Haus, J. M., Lanza, I. R., Lavie, C. J., Lee, C.-H., Lucia, A., Moro, C., Pandey, A.,
Robbins, J. M., Stanford, K. I., Thackray, A. E., . . . Snyder, M. P. (2022). Exerkines in health,
resilience and disease. Nat. Rev. Endocrinol., 18(5), 273–289.
165

Christensen-Quick, A., Vanpouille, C., Lisco, A., & Gianella, S. (2017). Cytomegalovirus and HIV
persistence: Pouring gas on the fire. AIDS Res. Hum. Retroviruses, 33(S1), S23–S30.
Chung, Y., Addington, J., Bearden, C. E., Cadenhead, K., Cornblatt, B., Mathalon, D. H., McGlashan, T.,
Perkins, D., Seidman, L. J., Tsuang, M., Walker, E., Woods, S. W., McEwen, S., van Erp, T. G. M.,
Cannon, T. D., & North American Prodrome Longitudinal Study (NAPLS) Consortium and the
Pediatric Imaging, Neurocognition, and Genetics (PING) Study Consortium. (2018). Use of machine
learning to determine deviance in neuroanatomical maturity associated with future psychosis in
youths at clinically high risk. JAMA Psychiatry, 75, 960–968.
Cipriani, G., Nuti, A., Carlesi, C., Lucetti, C., Di Fiorino, M., & Danti, S. (2021). Categorising a problem:
Alcohol and dementia. Acta Neurol. Belg., 121(1), 1–10.
Clausen, A. N., Fercho, K. A., Monsour, M., Disner, S., Salminen, L., Haswell, C. C., Rubright, E. C.,
Watts, A. A., Buckley, M. N., Maron-Katz, A., Sierk, A., Manthey, A., Suarez-Jimenez, B.,
Olatunji, B. O., Averill, C. L., Hofmann, D., Veltman, D. J., Olson, E. A., Li, G., . . . Morey, R. A.
(2022). Assessment of brain age in posttraumatic stress disorder: Findings from the ENIGMA PTSD
and brain age working groups. Brain Behav., 12, e2413.
Clifford, D. B., Fagan, A. M., Holtzman, D. M., Morris, J. C., Teshome, M., Shah, A. R., & Kauwe, J. S. K.
(2009). CSF biomarkers of alzheimer disease in HIV-associated neurologic disease. Neurology,
73(23), 1982–1987.
Cohen, R. A., Seider, T. R., & Navia, B. (2015). HIV effects on age-associated neurocognitive dysfunction:
Premature cognitive aging or neurodegenerative disease? Alzheimers. Res. Ther., 7(1), 37.
Cole, J. H., Ritchie, S. J., Bastin, M. E., Valdés Hernández, M. C., Muñoz Maniega, S., Royle, N., Corley, J.,
Pattie, A., Harris, S. E., Zhang, Q., Wray, N. R., Redmond, P., Marioni, R. E., Starr, J. M., Cox, S. R.,
Wardlaw, J. M., Sharp, D. J., & Deary, I. J. (2018). Brain age predicts mortality. Mol. Psychiatry, 23,
1385–1392.
Cole, J. H. (2020). Multimodality neuroimaging brain-age in UK biobank: Relationship to biomedical,
lifestyle, and cognitive factors. Neurobiol. Aging, 92, 34–42.
Cole, J. H., & Franke, K. (2017). Predicting age using neuroimaging: Innovative brain ageing biomarkers.
Trends Neurosci., 40, 681–690.
Cole, J. H., Marioni, R. E., Harris, S. E., & Deary, I. J. (2019). Brain age and other bodily ’ages’: Implications
for neuropsychiatry. Mol. Psychiatry, 24, 266–281.
Cole, J. H., Poudel, R. P. K., Tsagkrasoulis, D., Caan, M. W. A., Steves, C., Spector, T. D., & Montana, G.
(2017). Predicting brain age with deep learning from raw imaging data results in a reliable and
heritable biomarker. Neuroimage, 163, 115–124.
Collins, D. L., Neelin, P., Peters, T. M., & Evans, A. C. (1994). Automatic 3D intersubject registration of MR
volumetric data in standardized talairach space. J. Comput. Assist. Tomogr., 18(2), 192–205.
166

Conen, D., Rodondi, N., Müller, A., Beer, J. H., Ammann, P., Moschovitis, G., Auricchio, A., Hayoz, D.,
Kobza, R., Shah, D., Novak, J., Schläpfer, J., Di Valentino, M., Aeschbacher, S., Blum, S., Meyre, P.,
Sticherling, C., Bonati, L. H., Ehret, G., . . . Swiss-AF Study Investigators. (2019). Relationships of
overt and silent brain lesions with cognitive function in patients with atrial fibrillation. J. Am. Coll.
Cardiol., 73(9), 989–999.
Constantinides, C., Han, L. K. M., Alloza, C., Antonucci, L., Arango, C., Ayesa-Arriola, R., Banaj, N.,
Bertolino, A., Borgwardt, S., Bruggemann, J., Bustillo, J., Bykhovski, O., Carr, V., Catts, S.,
Chung, Y.-C., Crespo-Facorro, B., Díaz-Caneja, C. M., Donohoe, G., Du Plessis, S., . . . Walton, E.
(2022). Brain ageing in schizophrenia: Evidence from 26 international cohorts via the ENIGMA
schizophrenia consortium. medRxiv, 2022.01.10.21267840.
Coupé, P., Catheline, G., Lanuza, E., Manjón, J. V., & Alzheimer’s Disease Neuroimaging Initiative. (2017).
Towards a unified analysis of brain maturation and aging across the entire lifespan: A MRI analysis.
Hum. Brain Mapp., 38(11), 5501–5518.
Coutureau, E., & Di Scala, G. (2009). Entorhinal cortex and cognition. Prog. Neuropsychopharmacol. Biol.
Psychiatry, 33(5), 753–761.
Couvy-Duchesne, B., Faouzi, J., Martin, B., Thibeau-Sutre, E., Wild, A., Ansart, M., Durrleman, S.,
Dormont, D., Burgos, N., & Colliot, O. (2020). Ensemble learning of convolutional neural network,
support vector machine, and best linear unbiased predictor for brain age prediction: ARAMIS
contribution to the predictive analytics competition 2019 challenge. Front. Psychiatry, 11, 593336.
Croteau, E., Castellano, C. A., Fortier, M., Bocti, C., Fulop, T., Paquet, N., & Cunnane, S. C. (2018). A
cross-sectional comparison of brain glucose and ketone metabolism in cognitively healthy older
adults, mild cognitive impairment and early alzheimer’s disease. Exp. Gerontol., 107, 18–26.
Dadar, M., Potvin, O., Camicioli, R., Duchesne, S., & Alzheimer’s Disease Neuroimaging Initiative. (2021).
Beware of white matter hyperintensities causing systematic errors in FreeSurfer gray matter
segmentations! Hum. Brain Mapp., 42(9), 2734–2745.
Dallaire-Théroux, C., Callahan, B. L., Potvin, O., Saikali, S., & Duchesne, S. (2017). Radiological-Pathological
correlation in alzheimer’s disease: Systematic review of antemortem magnetic resonance imaging
findings. J. Alzheimers. Dis., 57(2), 575–601.
Dallo, F. J., Kindratt, T. B., & Zahodne, L. (2021). Prevalence of self-reported cognitive impairment among
arab american immigrants in the united states. Innov Aging, 5, igaa058.
Dao, E., Barha, C. K., Zou, J., Wei, N., & Liu-Ambrose, T. (2024). Prevention of vascular contributions to
cognitive impairment and dementia: The role of physical activity and exercise. Stroke, 55(4),
812–821.
de Bruijn, R. F. A. G., & Ikram, M. A. (2014). Cardiovascular risk factors and future risk of alzheimer’s
disease. BMC Med., 12, 130.
de Godoy, L. L., Alves, C. A. P. F., Saavedra, J. S. M., Studart-Neto, A., Nitrini, R., da Costa Leite, C., &
Bisdas, S. (2021). Understanding brain resilience in superagers: A systematic review.
Neuroradiology, 63(5), 663–683.
167

de Lange, A.-M. G., Anatürk, M., Rokicki, J., Han, L. K. M., Franke, K., Alnaes, D., Ebmeier, K. P.,
Draganski, B., Kaufmann, T., Westlye, L. T., Hahn, T., & Cole, J. H. (2022). Mind the gap:
Performance metric evaluation in brain-age prediction. Hum. Brain Mapp., 43, 3113–3129.
de Lange, A.-M. G., & Cole, J. H. (2020). Commentary: Correction procedures in brain-age prediction.
Neuroimage Clin, 26, 102229.
de la Torre, J. C. (2012). Cardiovascular risk factors promote brain hypoperfusion leading to cognitive
decline and dementia. Cardiovasc. Psychiatry Neurol., 2012, 367516.
Defferrard, M., Bresson, X., & Vandergheynst, P. (2016). Convolutional neural networks on graphs with fast
localized spectral filtering. Adv. Neural Inf. Process. Syst., 29.
Desikan, R. S., Ségonne, F., Fischl, B., Quinn, B. T., Dickerson, B. C., Blacker, D., Buckner, R. L., Dale, A. M.,
Maguire, R. P., Hyman, B. T., Albert, M. S., & Killiany, R. J. (2006). An automated labeling system
for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest.
Neuroimage, 31(3), 968–980.
Dice, L. R. (1945). Measures of the amount of ecologic association between species. Ecology, 26(3), 297–302.
Dichter, G. S., Gibbs, D., & Smoski, M. J. (2015). A systematic review of relations between resting-state
functional-MRI and treatment response in major depressive disorder. J. Affect. Disord., 172, 8–17.
Dickerson, B. C., Stoub, T. R., Shah, R. C., Sperling, R. A., Killiany, R. J., Albert, M. S., Hyman, B. T.,
Blacker, D., & Detoledo-Morrell, L. (2011). Alzheimer-signature MRI biomarker predicts AD
dementia in cognitively normal adults. Neurology, 76(16), 1395–1402.
Dickerson, B. C., Bakkour, A., Salat, D. H., Feczko, E., Pacheco, J., Greve, D. N., Grodstein, F., Wright, C. I.,
Blacker, D., Rosas, H. D., Sperling, R. A., Atri, A., Growdon, J. H., Hyman, B. T., Morris, J. C.,
Fischl, B., & Buckner, R. L. (2009). The cortical signature of alzheimer’s disease: Regionally specific
cortical thinning relates to symptom severity in very mild to mild AD dementia and is detectable in
asymptomatic amyloid-positive individuals. Cereb. Cortex, 19(3), 497–510.
Dickie, E., Hodge, S., Craddock, R., Poline, J.-B., & Kennedy, D. (2017). Tools matter: Comparison of two
surface analysis tools applied to the ABIDE dataset.
Dima, D., Modabbernia, A., Papachristou, E., Doucet, G. E., Agartz, I., Aghajani, M., Akudjedu, T. N.,
Albajes-Eizagirre, A., Alnaes, D., Alpert, K. I., Andersson, M., Andreasen, N. C., Andreassen, O. A.,
Asherson, P., Banaschewski, T., Bargallo, N., Baumeister, S., Baur-Streubel, R., Bertolino, A., . . .
Karolinska Schizophrenia Project (KaSP). (2022). Subcortical volumes across the lifespan: Data
from 18,605 healthy individuals aged 3-90 years. Hum. Brain Mapp., 43(1), 452–469.
Dinga, R., Schmaal, L., Penninx, B. W. J. H., van Tol, M. J., Veltman, D. J., van Velzen, L., Mennes, M.,
van der Wee, N. J. A., & Marquand, A. F. (2019). Evaluating the evidence for biotypes of depression:
Methodological replication and extension of. Neuroimage Clin, 22, 101796.
Drouka, A., Mamalaki, E., Karavasilis, E., Scarmeas, N., & Yannakoulia, M. (2022). Dietary and nutrient
patterns and brain MRI biomarkers in dementia-free adults. Nutrients, 14.
168

Drzezga, A. (2018). The network degeneration hypothesis: Spread of neurodegenerative patterns along
neuronal brain networks. J. Nucl. Med., 59(11), 1645–1648.
Dunås, T., Wåhlin, A., Nyberg, L., & Boraxbekk, C.-J. (2021). Multimodal image analysis of apparent brain
age identifies physical fitness as predictor of brain maintenance. Cereb. Cortex, 31, 3393–3407.
Durazzo, T. C., Mattsson, N., Weiner, M. W., & Alzheimer’s Disease Neuroimaging Initiative. (2014).
Smoking and increased alzheimer’s disease risk: A review of potential mechanisms. Alzheimers.
Dement., 10(3 Suppl), S122–45.
Edgerton-Fulton, M., & Ergul, A. (2022). Vascular contributions to cognitive impairment/dementia in
diabetes: Role of endothelial cells and pericytes. Am. J. Physiol. Cell Physiol., 323(4), C1177–C1189.
Faraco, G., & Iadecola, C. (2013). Hypertension: A harbinger of stroke and dementia. Hypertension, 62(5),
810–817.
Farrall & Wardlaw. (2009). Blood–brain barrier: Ageing and microvascular disease–systematic review and
meta-analysis. Neurobiol. Aging.
Farrer, L. A., Cupples, L. A., Haines, J. L., Hyman, B., Kukull, W. A., Mayeux, R., Myers, R. H.,
Pericak-Vance, M. A., Risch, N., & van Duijn, C. M. (1997). Effects of age, sex, and ethnicity on the
association between apolipoprotein E genotype and alzheimer disease. a meta-analysis. APOE and
alzheimer disease meta analysis consortium. JAMA, 278(16), 1349–1356.
Ferrucci, L., & Fabbri, E. (2018). Inflammageing: Chronic inflammation in ageing, cardiovascular disease,
and frailty. Nat. Rev. Cardiol., 15(9), 505–522.
Filip, P., Bednarik, P., Eberly, L. E., Moheet, A., Svatkova, A., Grohn, H., Kumar, A. F., Seaquist, E. R., &
Mangia, S. (2022). Different FreeSurfer versions might generate different statistical outcomes in
case-control comparison studies. Neuroradiology, 64(4), 765–773.
Finch, C. E., & Sapolsky, R. M. (1999). The evolution of alzheimer disease, the reproductive schedule, and
apoe isoforms. Neurobiol. Aging, 20(4), 407–428.
Finch, C. E., & Stanford, C. B. (2004). Meat-adaptive genes and the evolution of slower aging in humans. Q.
Rev. Biol., 79(1), 3–50.
Finger, C. E., Moreno-Gonzalez, I., Gutierrez, A., Moruno-Manchon, J. F., & McCullough, L. D. (2022).
Age-related immune alterations and cerebrovascular inflammation. Mol. Psychiatry, 27(2), 803–818.
Fischl, B. (2012). FreeSurfer. Neuroimage, 62(2), 774–781.
Fonseka, T. M., MacQueen, G. M., & Kennedy, S. H. (2018). Neuroimaging biomarkers as predictors of
treatment outcome in major depressive disorder. J. Affect. Disord., 233, 21–35.
Forloni, G., Demicheli, F., Giorgi, S., Bendotti, C., & Angeretti, N. (1992). Expression of amyloid precursor
protein mRNAs in endothelial, neuronal and glial cells: Modulation by interleukin-1. Brain Res. Mol.
Brain Res., 16(1-2), 128–134.
169

Fortin, J.-P., Cullen, N., Sheline, Y. I., Taylor, W. D., Aselcioglu, I., Cook, P. A., Adams, P., Cooper, C.,
Fava, M., McGrath, P. J., McInnis, M., Phillips, M. L., Trivedi, M. H., Weissman, M. M., &
Shinohara, R. T. (2018). Harmonization of cortical thickness measurements across scanners and
sites. Neuroimage, 167, 104–120.
Franceschi, C., Garagnani, P., Parini, P., Giuliani, C., & Santoro, A. (2018). Inflammaging: A new
immune–metabolic viewpoint for age-related diseases. Nat. Rev. Endocrinol., 14(10), 576–590.
Frangou, S., Modabbernia, A., Williams, S. C. R., Papachristou, E., Doucet, G. E., Agartz, I., Aghajani, M.,
Akudjedu, T. N., Albajes-Eizagirre, A., Alnaes, D., Alpert, K. I., Andersson, M., Andreasen, N. C.,
Andreassen, O. A., Asherson, P., Banaschewski, T., Bargallo, N., Baumeister, S., Baur-Streubel, R., . . .
Dima, D. (2022). Cortical thickness across the lifespan: Data from 17,075 healthy individuals aged
3-90 years. Hum. Brain Mapp., 43(1), 431–451.
Franke, K., & Gaser, C. (2012). Longitudinal changes in individual BrainAGE in healthy aging, mild
cognitive impairment, and alzheimer’s disease. GeroPsych, 25, 235–245.
Franke, K., & Gaser, C. (2019). Ten years of BrainAGE as a neuroimaging biomarker of brain aging: What
insights have we gained? Front. Neurol., 10, 789.
Friedman, J. I., Tang, C. Y., de Haas, H. J., Changchien, L., Goliasch, G., Dabas, P., Wang, V., Fayad, Z. A.,
Fuster, V., & Narula, J. (2014). Brain imaging changes associated with risk factors for cardiovascular
and cerebrovascular disease in asymptomatic patients. JACC Cardiovasc. Imaging, 7(10), 1039–1053.
Fukumoto, N., Fujii, T., Combarros, O., Ilyas Kamboh, M., Tsai, S.-J., Matsushita, S., Nacmias, B.,
Comings, D. E., Arboleda, H., Ingelsson, M., Hyman, B. T., Akatsu, H., Grupe, A., Nishimura, A. L.,
Zatz, M., Mattila, K. M., Rinne, J., Goto, Y.-I., Asada, T., . . . Kunugi, H. (2009). Sexually dimorphic
effect of the Val66Met polymorphism ofBDNFon susceptibility to alzheimer’s disease: New data
and meta-analysis.
Gannicliffe, A., Sutton, R. N., & Itzhaki, R. F. (1986). Viruses, brain and immunosuppression. Psychol. Med.,
16(2), 247–249.
Garcia, A. D., & Buffalo, E. A. (2020). Anatomy and function of the primate entorhinal cortex. Annu Rev Vis
Sci, 6, 411–432.
Gardner, S. E., Humphry, M., Bennett, M. R., & Clarke, M. C. H. (2015). Senescent vascular smooth muscle
cells drive inflammation through an Interleukin-1α-Dependent Senescence-Associated secretory
phenotype. Arterioscler. Thromb. Vasc. Biol., 35(9), 1963–1974.
Gaser, C., Franke, K., Klöppel, S., Koutsouleris, N., Sauer, H., & Alzheimer’s Disease Neuroimaging
Initiative. (2013). BrainAGE in mild cognitive impaired patients: Predicting the conversion to
alzheimer’s disease. PLoS One, 8, e67346.
GBD 2019 North Africa and the Middle East Neurology Collaborators. (2024). The burden of neurological
conditions in north africa and the middle east, 1990-2019: A systematic analysis of the global
burden of disease study 2019. Lancet Glob Health.
170

Gefen, T., Peterson, M., Papastefan, S. T., Martersteck, A., Whitney, K., Rademaker, A., Bigio, E. H.,
Weintraub, S., Rogalski, E., Mesulam, M.-M., & Geula, C. (2015). Morphometric and histologic
substrates of cingulate integrity in elders with exceptional memory capacity. J. Neurosci., 35(4),
1781–1791.
Ghosh, A. K. (2004). Anaerobic threshold: Its concept and role in endurance sport. Malays. J. Med. Sci., 11(1),
24–36.
Glasser, M. F., Coalson, T. S., Robinson, E. C., Hacker, C. D., Harwell, J., Yacoub, E., Ugurbil, K.,
Andersson, J., Beckmann, C. F., Jenkinson, M., Smith, S. M., & Van Essen, D. C. (2016). A
multi-modal parcellation of human cerebral cortex. Nature, 536(7615), 171–178.
Glatard, T., Lewis, L. B., Ferreira da Silva, R., Adalat, R., Beck, N., Lepage, C., Rioux, P., Rousseau, M.-E.,
Sherif, T., Deelman, E., Khalili-Mahani, N., & Evans, A. C. (2015). Reproducibility of neuroimaging
analyses across operating systems. Front. Neuroinform., 9, 12.
Gorelick, P. B., Furie, K. L., Iadecola, C., Smith, E. E., Waddy, S. P., Lloyd-Jones, D. M., Bae, H.-J.,
Bauman, M. A., Dichgans, M., Duncan, P. W., Girgus, M., Howard, V. J., Lazar, R. M., Seshadri, S.,
Testai, F. D., van Gaal, S., Yaffe, K., Wasiak, H., Zerna, C., & American Heart Association/American
Stroke Association. (2017). Defining optimal brain health in adults: A presidential advisory from
the american heart Association/American stroke association. Stroke, 48(10), e284–e303.
Gorgolewski, K. J., & Poldrack, R. A. (2016). A practical guide for improving transparency and
reproducibility in neuroimaging research. PLoS Biol., 14(7), e1002506.
Gorgolewski, K. J., Varoquaux, G., Rivera, G., Schwarz, Y., Ghosh, S. S., Maumet, C., Sochat, V. V.,
Nichols, T. E., Poldrack, R. A., Poline, J.-B., Yarkoni, T., & Margulies, D. S. (2015). NeuroVault.org: A
web-based repository for collecting and sharing unthresholded statistical maps of the human brain.
Front. Neuroinform., 9, 8.
Grasby, K. L., Jahanshad, N., Painter, J. N., Colodro-Conde, L., Bralten, J., Hibar, D. P., Lind, P. A.,
Pizzagalli, F., Ching, C. R. K., McMahon, M. A. B., Shatokhina, N., Zsembik, L. C. P.,
Thomopoulos, S. I., Zhu, A. H., Strike, L. T., Agartz, I., Alhusaini, S., Almeida, M. A. A., Alnæs, D.,
. . . Enhancing NeuroImaging Genetics through Meta-Analysis Consortium (ENIGMA)—Genetics
working group. (2020). The genetic architecture of the human cerebral cortex. Science, 367(6484).
Green, D. A., Masliah, E., Vinters, H. V., Beizai, P., Moore, D. J., & Achim, C. L. (2005). Brain deposition of
beta-amyloid is a common pathologic feature in HIV positive patients. AIDS, 19(4), 407–411.
Gregoire, S. M., Chaudhary, U. J., Brown, M. M., Yousry, T. A., Kallis, C., Jäger, H. R., & Werring, D. J. (2009).
The microbleed anatomical rating scale (MARS): Reliability of a tool to map brain microbleeds.
Neurology, 73, 1759–1766.
Grinberg, L. T., & Thal, D. R. (2010). Vascular pathology in the aged human brain. Acta Neuropathol., 119(3),
277–290.
171

Griswold, A. J., Celis, K., Bussies, P. L., Rajabli, F., Whitehead, P. L., Hamilton-Nelson, K. L., Beecham, G. W.,
Dykxhoorn, D. M., Nuytemans, K., Wang, L., Gardner, O. K., Dorfsman, D. A., Bigio, E. H.,
Mesulam, M. M., Weintraub, S., Geula, C., Gearing, M., McGrath-Martinez, E., Dalgard, C. L., . . .
Vance, J. M. (2021). Increased APOE ϵ4 expression is associated with the difference in alzheimer’s
disease risk from diverse ancestral backgrounds. Alzheimers. Dement., 17(7), 1179–1188.
Gronenschild, E. H. B. M., Habets, P., Jacobs, H. I. L., Mengelers, R., Rozendaal, N., van Os, J., & Marcelis, M.
(2012). The effects of FreeSurfer version, workstation type, and macintosh operating system
version on anatomical volume and cortical thickness measurements. PLoS One, 7(6), e38234.
Gureje, O., Ogunniyi, A., Baiyewu, O., Price, B., Unverzagt, F. W., Evans, R. M., Smith-Gamble, V.,
Lane, K. A., Gao, S., Hall, K. S., Hendrie, H. C., & Murrell, J. R. (2006). APOE 4 is not associated with
alzheimer’s disease in elderly nigerians. Ann. Neurol., 59(1), 182–185.
Gustafson, D. R., Karlsson, C., Skoog, I., Rosengren, L., Lissner, L., & Blennow, K. (2007). Mid-life adiposity
factors relate to blood-brain barrier integrity in late life. J. Intern. Med., 262(6), 643–650.
Gutierrez, A., Norby, F. L., Maheshwari, A., Rooney, M. R., Gottesman, R. F., Mosley, T. H., Lutsey, P. L.,
Oldenburg, N., Soliman, E. Z., Alonso, A., & Chen, L. Y. (2019). Association of abnormal P-Wave
indices with dementia and cognitive decline over 25 years: ARIC-NCS (the atherosclerosis risk in
communities neurocognitive study). J. Am. Heart Assoc., 8(24), e014553.
Haacke, E. M., Xu, Y., Cheng, Y.-C. N., & Reichenbach, J. R. (2004). Susceptibility weighted imaging (SWI).
Magn. Reson. Med., 52, 612–618.
Haddad, E., Pizzagalli, F., Zhu, A. H., Bhatt, R. R., Islam, T., Ba Gari, I., Dixon, D., Thomopoulos, S. I.,
Thompson, P. M., & Jahanshad, N. (2023). Multisite test–retest reliability and compatibility of brain
metrics derived from FreeSurfer versions 7.1, 6.0, and 5.3. Hum. Brain Mapp., 44, 1515–1532.
Hampel, H., Hardy, J., Blennow, K., Chen, C., Perry, G., Kim, S. H., Villemagne, V. L., Aisen, P.,
Vendruscolo, M., Iwatsubo, T., Masters, C. L., Cho, M., Lannfelt, L., Cummings, J. L., & Vergallo, A.
(2021). The amyloid-β pathway in alzheimer’s disease. Mol. Psychiatry, 26(10), 5481–5503.
Han, L. K. M., Dinga, R., Hahn, T., Ching, C. R. K., Eyler, L. T., Aftanas, L., Aghajani, M., Aleman, A.,
Baune, B. T., Berger, K., Brak, I., Filho, G. B., Carballedo, A., Connolly, C. G., Couvy-Duchesne, B.,
Cullen, K. R., Dannlowski, U., Davey, C. G., Dima, D., . . . Schmaal, L. (2021). Brain aging in major
depressive disorder: Results from the ENIGMA major depressive disorder working group. Mol.
Psychiatry, 26, 5124–5139.
Han, X., Jovicich, J., Salat, D., van der Kouwe, A., Quinn, B., Czanner, S., Busa, E., Pacheco, J., Albert, M.,
Killiany, R., Maguire, P., Rosas, D., Makris, N., Dale, A., Dickerson, B., & Fischl, B. (2006). Reliability
of MRI-derived measurements of human cerebral cortical thickness: The effects of field strength,
scanner upgrade and manufacturer. Neuroimage, 32(1), 180–194.
172

Hancock, E. W., Deal, B. J., Mirvis, D. M., Okin, P., Kligfield, P., Gettes, L. S., Bailey, J. J., Childers, R.,
Gorgels, A., Josephson, M., Kors, J. A., Macfarlane, P., Mason, J. W., Pahlm, O., Rautaharju, P. M.,
Surawicz, B., van Herpen, G., Wagner, G. S., Wellens, H., . . . Heart Rhythm Society. (2009).
AHA/ACCF/HRS recommendations for the standardization and interpretation of the
electrocardiogram: Part v: Electrocardiogram changes associated with cardiac chamber
hypertrophy: A scientific statement from the american heart association electrocardiography and
arrhythmias committee, council on clinical cardiology; the american college of cardiology
foundation; and the heart rhythm society. endorsed by the international society for computerized
electrocardiology. J. Am. Coll. Cardiol., 53(11), 992–1002.
Harada, T., Kudo, K., Fujima, N., Yoshikawa, M., Ikebe, Y., Sato, R., Shirai, T., Bito, Y., Uwano, I., &
Miyata, M. (2022). Quantitative susceptibility mapping: Basic methods and clinical applications.
Radiographics, 42(4), 1161–1176.
Hari, K. J., Nguyen, T. P., & Soliman, E. Z. (2018). Relationship between p-wave duration and the risk of
atrial fibrillation. Expert Rev. Cardiovasc. Ther., 16(11), 837–843.
Harrison, T. M., Maass, A., Baker, S. L., & Jagust, W. J. (2018). Brain morphology, cognition, and -amyloid in
older adults with superior memory performance. Neurobiol. Aging, 67, 162–170.
Harrison, T. M., Weintraub, S., Mesulam, M.-M., & Rogalski, E. (2012). Superior memory and higher cortical
volumes in unusually successful cognitive aging. J. Int. Neuropsychol. Soc., 18(6), 1081–1085.
Hase, Y., Horsburgh, K., Ihara, M., & Kalaria, R. N. (2018). White matter degeneration in vascular and other
ageing-related dementias. J. Neurochem., 144(5), 617–633.
He, J., Tse, G., Korantzopoulos, P., Letsas, K. P., Ali-Hasan-Al-Saegh, S., Kamel, H., Li, G., Lip, G. Y. H., &
Liu, T. (2017). P-Wave indices and risk of ischemic stroke: A systematic review and Meta-Analysis.
Stroke, 48(8), 2066–2072.
Henderson, K. D., Bernstein, L., Henderson, B., Kolonel, L., & Pike, M. C. (2008). Predictors of the timing of
natural menopause in the multiethnic cohort study. Am. J. Epidemiol., 167(11), 1287–1294.
Hendrickx, J. O., Martinet, W., Van Dam, D., & De Meyer, G. R. Y. (2021). Inflammation, nitro-oxidative
stress, impaired autophagy, and insulin resistance as a mechanistic convergence between arterial
stiffness and alzheimer’s disease. Front. Mol. Biosci., 8, 651215.
Hendrie, H. C., Murrell, J., Baiyewu, O., Lane, K. A., Purnell, C., Ogunniyi, A., Unverzagt, F. W., Hall, K.,
Callahan, C. M., Saykin, A. J., Gureje, O., Hake, A., Foroud, T., & Gao, S. (2014). APOE ϵ4 and the
risk for alzheimer disease and cognitive decline in african americans and yoruba. Int. Psychogeriatr.,
26(6), 977–985.
Heye, A. K., Thrippleton, M. J., Chappell, F. M., Hernández, M. d. C. V., Armitage, P. A., Makin, S. D.,
Maniega, S. M., Sakka, E., Flatman, P. W., Dennis, M. S., & Wardlaw, J. M. (2016). Blood pressure
and sodium: Association with MRI markers in cerebral small vessel disease. J. Cereb. Blood Flow
Metab., 36, 264–274.
Higgins-Chen, A. T., Thrush, K. L., & Levine, M. E. (2021). Aging biomarkers and the brain. Semin. Cell Dev.
Biol., 116, 180–193.
173

Hinojosa, C. A., Kaur, N., VanElzakker, M. B., & Shin, L. M. (2019). Cingulate subregions in posttraumatic
stress disorder, chronic stress, and treatment. Handb. Clin. Neurol., 166, 355–370.
Hodge, S. M., Haselgrove, C., Honor, L., Kennedy, D. N., & Frazier, J. A. (2020). An assessment of the autism
neuroimaging literature for the prospects of re-executability. F1000Res., 9, 1031.
Hong, J., Yun, H. J., Park, G., Kim, S., Ou, Y., Vasung, L., Rollins, C. K., Ortinau, C. M., Takeoka, E.,
Akiyama, S., Tarui, T., Estroff, J. A., Grant, P. E., Lee, J.-M., & Im, K. (2021). Optimal method for
fetal brain age prediction using multiplanar slices from structural magnetic resonance imaging.
Front. Neurosci., 15, 714252.
Horien, C., Noble, S., Greene, A. S., Lee, K., Barron, D. S., Gao, S., O’Connor, D., Salehi, M., Dadashkarimi, J.,
Shen, X., Lake, E. M. R., Constable, R. T., & Scheinost, D. (2021). A hitchhiker’s guide to working
with large, open-source neuroimaging datasets. Nat Hum Behav, 5(2), 185–193.
Horton, W. B., & Barrett, E. J. (2021). Microvascular dysfunction in diabetes mellitus and cardiometabolic
disease. Endocr. Rev., 42(1), 29–55.
Hu, J.-R., Abdullah, A., Nanna, M. G., & Soufer, R. (2023). The Brain-Heart axis: Neuroinflammatory
interactions in cardiovascular disease. Curr. Cardiol. Rep., 25(12), 1745–1758.
Huebbe, P., & Rimbach, G. (2017). Evolution of human apolipoprotein E (APOE) isoforms: Gene structure,
protein function and interaction with dietary factors. Ageing Res. Rev., 37, 146–161.
Huuha, A. M., Norevik, C. S., Moreira, J. B. N., Kobro-Flatmoen, A., Scrimgeour, N., Kivipelto, M.,
Van Praag, H., Ziaei, M., Sando, S. B., Wisløff, U., & Tari, A. R. (2022). Can exercise training teach us
how to treat alzheimer’s disease?
Iadecola, C., & Gottesman, R. F. (2019). Neurovascular and cognitive dysfunction in hypertension. Circ. Res.,
124(7), 1025–1044.
Inoue, Y., Shue, F., Bu, G., & Kanekiyo, T. (2023). Pathophysiology and probable etiology of cerebral small
vessel disease in vascular dementia and alzheimer’s disease. Mol. Neurodegener., 18(1), 46.
Jack, C. R., Jr. (2011). Alliance for aging research AD biomarkers work group: Structural MRI. Neurobiol.
Aging, 32 Suppl 1, S48–57.
Jack, C. R., Jr, Knopman, D. S., Jagust, W. J., Petersen, R. C., Weiner, M. W., Aisen, P. S., Shaw, L. M.,
Vemuri, P., Wiste, H. J., Weigand, S. D., Lesnick, T. G., Pankratz, V. S., Donohue, M. C., &
Trojanowski, J. Q. (2013). Tracking pathophysiological processes in alzheimer’s disease: An
updated hypothetical model of dynamic biomarkers. Lancet Neurol., 12(2), 207–216.
Jensen, D. E. A., Leoni, V., Klein-Flügge, M. C., Ebmeier, K. P., & Suri, S. (2021). Associations of dietary
markers with brain volume and connectivity: A systematic review of MRI studies. Ageing Res. Rev.,
70, 101360.
Jian, B., Hu, M., Cai, W., Zhang, B., & Lu, Z. (2020). Update of immunosenescence in cerebral small vessel
disease. Front. Immunol., 11, 585655.
174

Jiang, L., Zhang, J., Monticone, R. E., Telljohann, R., Wu, J., Wang, M., & Lakatta, E. G. (2012). Calpain-1
regulation of matrix metalloproteinase 2 activity in vascular smooth muscle cells facilitates
age-associated aortic wall calcification and fibrosis. Hypertension, 60(5), 1192–1199.
Jonsson, B. A., Bjornsdottir, G., Thorgeirsson, T. E., Ellingsen, L. M., Walters, G. B., Gudbjartsson, D. F.,
Stefansson, H., Stefansson, K., & Ulfarsson, M. O. (2019). Brain age prediction using deep learning
uncovers associated sequence variants. Nat. Commun., 10, 5409.
Jovicich, J., Czanner, S., Han, X., Salat, D., van der Kouwe, A., Quinn, B., Pacheco, J., Albert, M., Killiany, R.,
Blacker, D., Maguire, P., Rosas, D., Makris, N., Gollub, R., Dale, A., Dickerson, B. C., & Fischl, B.
(2009). MRI-derived measurements of human subcortical, ventricular and intracranial brain
volumes: Reliability effects of scan sessions, acquisition sequences, data analyses, scanner upgrade,
scanner vendors and field strengths. Neuroimage, 46(1), 177–192.
Kang, S.-G., & Cho, S.-E. (2020). Neuroimaging biomarkers for predicting treatment response and
recurrence of major depressive disorder. Int. J. Mol. Sci., 21(6).
Kapasi, A., DeCarli, C., & Schneider, J. A. (2017). Impact of multiple pathologies on the threshold for
clinically overt dementia. Acta Neuropathol., 134(2), 171–186.
Kennedy, D. N., Abraham, S. A., Bates, J. F., Crowley, A., Ghosh, S., Gillespie, T., Goncalves, M., Grethe, J. S.,
Halchenko, Y. O., Hanke, M., Haselgrove, C., Hodge, S. M., Jarecka, D., Kaczmarzyk, J., Keator, D. B.,
Meyer, K., Martone, M. E., Padhy, S., Poline, J.-B., . . . Travers, M. (2019). Everything matters: The
ReproNim perspective on reproducible neuroimaging. Front. Neuroinform., 13, 1.
Khera, A., Vega, G. L., Das, S. R., Ayers, C., McGuire, D. K., Grundy, S. M., & de Lemos, J. A. (2009). Sex
differences in the relationship between C-Reactive protein and body fat.
Khurshid, S., Choi, S. H., Weng, L.-C., Wang, E. Y., Trinquart, L., Benjamin, E. J., Ellinor, P. T., & Lubitz, S. A.
(2018). Frequency of cardiac rhythm abnormalities in a half million adults. Circ. Arrhythm.
Electrophysiol., 11(7), e006273.
Kim, J. S., Singh, V., Lee, J. K., Lerch, J., Ad-Dab’bagh, Y., MacDonald, D., Lee, J. M., Kim, S. I., & Evans, A. C.
(2005). Automated 3-D extraction and evaluation of the inner and outer cortical surfaces using a
laplacian map and partial volume effect classification. Neuroimage, 27(1), 210–221.
Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv [cs.LG].
Knopman, D. S., Amieva, H., Petersen, R. C., Chételat, G., Holtzman, D. M., Hyman, B. T., Nixon, R. A., &
Jones, D. T. (2021). Alzheimer disease. Nat. Rev. Dis. Primers, 7(1), 33.
Kolbeinsson, A., Filippi, S., Panagakis, Y., Matthews, P. M., Elliott, P., Dehghan, A., & Tzoulaki, I. (2020).
Accelerated MRI-predicted brain ageing and its associations with cardiometabolic and brain
disorders. Sci. Rep., 10, 19940.
Koo, T. K., & Li, M. Y. (2016). A guideline of selecting and reporting intraclass correlation coefficients for
reliability research. J. Chiropr. Med., 15(2), 155–163.
175

Kosar, F., Aksoy, Y., Ari, F., Keskin, L., & Sahin, I. (2008). P-wave duration and dispersion in obese subjects.
Ann. Noninvasive Electrocardiol., 13(1), 3–7.
Kress, B. T., Iliff, J. J., Xia, M., Wang, M., Wei, H. S., Zeppenfeld, D., Xie, L., Kang, H., Xu, Q., Liew, J. A.,
Plog, B. A., Ding, F., Deane, R., & Nedergaard, M. (2014). Impairment of paravascular clearance
pathways in the aging brain. Ann. Neurol., 76(6), 845–861.
Kristen, H., Santana, S., Sastre, I., Recuero, M., Bullido, M. J., & Aldudo, J. (2015). Herpes simplex virus type
2 infection induces AD-like neurodegeneration markers in human neuroblastoma cells. Neurobiol.
Aging, 36(10), 2737–2747.
Lambert, J. C., Ibrahim-Verbaas, C. A., Harold, D., Naj, A. C., Sims, R., Bellenguez, C., DeStafano, A. L.,
Bis, J. C., Beecham, G. W., Grenier-Boley, B., Russo, G., Thorton-Wells, T. A., Jones, N., Smith, A. V.,
Chouraki, V., Thomas, C., Ikram, M. A., Zelenika, D., Vardarajan, B. N., . . . Amouyel, P. (2013).
Meta-analysis of 74,046 individuals identifies 11 new susceptibility loci for alzheimer’s disease. Nat.
Genet., 45, 1452–1458.
Landman, B. A., Huang, A. J., Gifford, A., Vikram, D. S., Lim, I. A. L., Farrell, J. A. D., Bogovic, J. A., Hua, J.,
Chen, M., Jarso, S., Smith, S. A., Joel, S., Mori, S., Pekar, J. J., Barker, P. B., Prince, J. L., &
van Zijl, P. C. M. (2010). Multi-parametric neuroimaging reproducibility: A 3-T resource study.
Neuroimage, 54(4), 2854–2866.
Le, T. T., Kuplicki, R. T., McKinney, B. A., Yeh, H.-W., Thompson, W. K., Paulus, M. P., & Tulsa 1000
Investigators. (2018). A nonlinear simulation framework supports adjusting for age when
analyzing BrainAGE. Front. Aging Neurosci., 10, 317.
LeBlanc, E. S., Janowsky, J., Chan, B. K., & Nelson, H. D. (2001). Hormone replacement therapy and
cognition: Systematic review and meta-analysis. JAMA, 285(11), 1489–1499.
Lee, S., Choi, Y., Jeong, E., Park, J., Kim, J., Tanaka, M., & Choi, J. (2023). Physiological significance of
elevated levels of lactate by exercise training in the brain and body. J. Biosci. Bioeng., 135(3),
167–175.
Lee, W. H., Antoniades, M., Schnack, H. G., Kahn, R. S., & Frangou, S. (2021). Brain age prediction in
schizophrenia: Does the choice of machine learning algorithm matter? Psychiatry Res
Neuroimaging, 310, 111270.
Leigh, S.-J., & Morris, M. J. (2020). Diet, inflammation and the gut microbiome: Mechanisms for
obesity-associated cognitive impairment. Biochim. Biophys. Acta Mol. Basis Dis., 1866(6), 165767.
Levakov, G., Rosenthal, G., Shelef, I., Raviv, T. R., & Avidan, G. (2020). From a deep learning model back to
the brain-identifying regional predictors and their relation to aging. Hum. Brain Mapp., 41,
3235–3252.
Lewis, J. D., Evans, A. C., Tohka, J., Brain Development Cooperative Group, & Pediatric Imaging,
Neurocognition, and Genetics Study. (2018). T1 white/gray contrast as a predictor of chronological
age, and an index of cognitive performance. Neuroimage, 173, 341–350.
176

Li, T., Huang, Y., Cai, W., Chen, X., Men, X., Lu, T., Wu, A., & Lu, Z. (2020). Age-related cerebral small vessel
disease and inflammaging. Cell Death Dis., 11(10), 932.
Liang, H., Zhang, F., & Niu, X. (2019). Investigating systematic bias in brain age estimation with application
to post-traumatic stress disorders. Hum. Brain Mapp., 40, 3143–3152.
Liao, Y.-F., Wang, B.-J., Cheng, H.-T., Kuo, L.-H., & Wolfe, M. S. (2004). Tumor necrosis factor-alpha,
interleukin-1beta, and interferon-gamma stimulate gamma-secretase-mediated cleavage of amyloid
precursor protein through a JNK-dependent MAPK pathway. J. Biol. Chem., 279(47), 49523–49532.
Libby, P., Buring, J. E., Badimon, L., Hansson, G. K., Deanfield, J., Bittencourt, M. S., Tokgözoğlu, L., &
Lewis, E. F. (2019). Atherosclerosis. Nat Rev Dis Primers, 5, 56.
Liem, F., Varoquaux, G., Kynast, J., Beyer, F., Kharabian Masouleh, S., Huntenburg, J. M., Lampe, L.,
Rahim, M., Abraham, A., Craddock, R. C., Riedel-Heller, S., Luck, T., Loeffler, M., Schroeter, M. L.,
Witte, A. V., Villringer, A., & Margulies, D. S. (2017). Predicting brain-age from multimodal imaging
data captures cognitive impairment. Neuroimage, 148, 179–188.
Lindquist, M. (2020). Neuroimaging results altered by varying analysis pipelines. Nature, 582(7810), 36–37.
Lip, G. Y. H., Fauchier, L., Freedman, S. B., Van Gelder, I., Natale, A., Gianni, C., Nattel, S., Potpara, T.,
Rienstra, M., Tse, H.-F., & Lane, D. A. (2016). Atrial fibrillation. Nat Rev Dis Primers, 2, 16016.
Littlejohns, T. J., Holliday, J., Gibson, L. M., Garratt, S., Oesingmann, N., Alfaro-Almagro, F., Bell, J. D.,
Boultwood, C., Collins, R., Conroy, M. C., Crabtree, N., Doherty, N., Frangi, A. F., Harvey, N. C.,
Leeson, P., Miller, K. L., Neubauer, S., Petersen, S. E., Sellors, J., . . . Allen, N. E. (2020). The UK
biobank imaging enhancement of 100,000 participants: Rationale, data collection, management and
future directions. Nat. Commun., 11(1), 2624.
Liu, K., Yao, S., Chen, K., Zhang, J., Yao, L., Li, K., Jin, Z., & Guo, X. (2017). Structural brain network changes
across the adult lifespan. Front. Aging Neurosci., 9, 275.
Liu, S., Buch, S., Chen, Y., Choi, H.-S., Dai, Y., Habib, C., Hu, J., Jung, J.-Y., Luo, Y., Utriainen, D., Wang, M.,
Wu, D., Xia, S., & Haacke, E. M. (2017). Susceptibility-weighted imaging: Current status and future
directions. NMR Biomed., 30.
Livingston, G., Huntley, J., Liu, K. Y., Costafreda, S. G., Selbæk, G., Alladi, S., Ames, D., Banerjee, S.,
Burns, A., Brayne, C., Fox, N. C., Ferri, C. P., Gitlin, L. N., Howard, R., Kales, H. C., Kivimäki, M.,
Larson, E. B., Nakasujja, N., Rockwood, K., . . . Mukadam, N. (2024). Dementia prevention,
intervention, and care: 2024 report of the lancet standing commission. Lancet, 0(0).
Livingston, G., Huntley, J., Sommerlad, A., Ames, D., Ballard, C., Banerjee, S., Brayne, C., Burns, A.,
Cohen-Mansfield, J., Cooper, C., Costafreda, S. G., Dias, A., Fox, N., Gitlin, L. N., Howard, R.,
Kales, H. C., Kivimäki, M., Larson, E. B., Ogunniyi, A., . . . Mukadam, N. (2020). Dementia
prevention, intervention, and care: 2020 report of the lancet commission. Lancet, 396, 413–446.
177

Lopatko Lindman, K., Weidung, B., Olsson, J., Josefsson, M., Kok, E., Johansson, A., Eriksson, S.,
Hallmans, G., Elgh, F., & Lövheim, H. (2019). A genetic signature including apolipoprotein Eϵ4
potentiates the risk of herpes simplex–associated alzheimer’s disease. Alzheimer’s & Dementia:
Translational Research & Clinical Interventions, 5(1), 697–704.
López-Otín, C., Blasco, M. A., Partridge, L., Serrano, M., & Kroemer, G. (2013). The hallmarks of aging. Cell,
153(6), 1194–1217.
López-Otín, C., Blasco, M. A., Partridge, L., Serrano, M., & Kroemer, G. (2023). Hallmarks of aging: An
expanding universe. Cell, 186(2), 243–278.
Löwe, L. C., Gaser, C., Franke, K., & Alzheimer’s Disease Neuroimaging Initiative. (2016). The effect of the
APOE genotype on individual BrainAGE in normal aging, mild cognitive impairment, and
alzheimer’s disease. PLoS One, 11, e0157514.
Loy, C. T., Schofield, P. R., Turner, A. M., & Kwok, J. B. J. (2014). Genetics of dementia. Lancet, 383(9919),
828–840.
Lyttelton, O., Boucher, M., Robbins, S., & Evans, A. (2007). An unbiased iterative group registration
template for cortical surface analysis. Neuroimage, 34(4), 1535–1544.
MacDonald, D., Kabani, N., Avis, D., & Evans, A. C. (2000). Automated 3-D extraction of inner and outer
surfaces of cerebral cortex from MRI. Neuroimage, 12(3), 340–356.
MacDonald, M. E., & Pike, G. B. (2021). MRI of healthy brain aging: A review. NMR Biomed., 34(9), e4564.
Madhavan, M., Graff-Radford, J., Piccini, J. P., & Gersh, B. J. (2018). Cognitive dysfunction in atrial
fibrillation. Nat. Rev. Cardiol., 15(12), 744–756.
Magnani, J. W., Williamson, M. A., Ellinor, P. T., Monahan, K. M., & Benjamin, E. J. (2009). P wave indices:
Current status and future directions in epidemiology, clinical, and research applications. Circ.
Arrhythm. Electrophysiol., 2(1), 72–79.
Magnani, J. W., Zhu, L., Lopez, F., Pencina, M. J., Agarwal, S. K., Soliman, E. Z., Benjamin, E. J., & Alonso, A.
(2015). P-wave indices and atrial fibrillation: Cross-cohort assessments from the framingham heart
study (FHS) and atherosclerosis risk in communities (ARIC) study. Am. Heart J., 169(1), 53–61.e1.
Maheshwari, U., Huang, S.-F., Sridhar, S., & Keller, A. (2022). The interplay between brain vascular
calcification and microglia. Front. Aging Neurosci., 14, 848495.
Mahley, R. W. (2016). Central nervous system lipoproteins: ApoE and regulation of cholesterol metabolism:
ApoE and regulation of cholesterol metabolism. Arterioscler. Thromb. Vasc. Biol., 36(7), 1305–1315.
Makowski, C., Béland, S., Kostopoulos, P., Bhagwat, N., Devenyi, G. A., Malla, A. K., Joober, R., Lepage, M.,
& Chakravarty, M. M. (2018). Evaluating accuracy of striatal, pallidal, and thalamic segmentation
methods: Comparing automated approaches to manual delineation. Neuroimage, 170, 182–198.
Malakar, A. K., Choudhury, D., Halder, B., Paul, P., Uddin, A., & Chakraborty, S. (2019). A review on
coronary artery disease, its risk factors, and therapeutics. J. Cell. Physiol., 234(10), 16812–16823.
178

Marcus, D. S., Wang, T. H., Parker, J., Csernansky, J. G., Morris, J. C., & Buckner, R. L. (2007). Open access
series of imaging studies (OASIS): Cross-sectional MRI data in young, middle aged, nondemented,
and demented older adults. J. Cogn. Neurosci., 19(9), 1498–1507.
Marmarelis, M. G., Haddad, E., Jesson, A., Jahanshad, N., Galstyan, A., & Ver Steeg, G. (2023). Partial
identification of dose responses with hidden confounders. The 39th Conference on Uncertainty in
Artificial Intelligence.
Matelsky, J., Kiar, G., Johnson, E., Rivera, C., Toma, M., & Gray-Roncal, W. (2018). Container-based clinical
solutions for portable and reproducible image analysis. J. Digit. Imaging, 31(3), 315–320.
Mattis, J., & Sehgal, A. (2016). Circadian rhythms, sleep, and disorders of aging. Trends Endocrinol. Metab.,
27(4), 192–203.
Mattson, M. P., & Arumugam, T. V. (2018). Hallmarks of brain aging: Adaptive and pathological
modification by metabolic states. Cell Metab., 27(6), 1176–1199.
McLarnon, J. G. (2021). A leaky Blood-Brain barrier to fibrinogen contributes to oxidative damage in
alzheimer’s disease. Antioxidants (Basel), 11(1).
Medland, S. E., Jahanshad, N., Neale, B. M., & Thompson, P. M. (2014). Whole-genome analyses of
whole-brain data: Working within an expanded search space. Nat. Neurosci., 17, 791–800.
Meijerman, A., Amiri, H., Steenwijk, M. D., Jonker, M. A., van Schijndel, R. A., Cover, K. S., Vrenken, H., &
Alzheimer’s Disease Neuroimaging Initiative. (2018). Reproducibility of deep gray matter atrophy
rate measurement in a large multicenter dataset. AJNR Am. J. Neuroradiol., 39(1), 46–53.
Miletić, S., Bazin, P.-L., Isherwood, S. J. S., Keuken, M. C., Alkemade, A., & Forstmann, B. U. (2022).
Charting human subcortical maturation across the adult lifespan with in vivo 7 T MRI. Neuroimage,
249, 118872.
Miller, K. L., Alfaro-Almagro, F., Bangerter, N. K., Thomas, D. L., Yacoub, E., Xu, J., Bartsch, A. J., Jbabdi, S.,
Sotiropoulos, S. N., Andersson, J. L. R., Griffanti, L., Douaud, G., Okell, T. W., Weale, P., Dragonu, I.,
Garratt, S., Hudson, S., Collins, R., Jenkinson, M., . . . Smith, S. M. (2016a). Multimodal population
brain imaging in the UK biobank prospective epidemiological study. Nat. Neurosci., 19(11),
1523–1536.
Miller, K. L., Alfaro-Almagro, F., Bangerter, N. K., Thomas, D. L., Yacoub, E., Xu, J., Bartsch, A. J., Jbabdi, S.,
Sotiropoulos, S. N., Andersson, J. L. R., Griffanti, L., Douaud, G., Okell, T. W., Weale, P., Dragonu, I.,
Garratt, S., Hudson, S., Collins, R., Jenkinson, M., . . . Smith, S. M. (2016b). Multimodal population
brain imaging in the UK biobank prospective epidemiological study. Nat. Neurosci., 19(11),
1523–1536.
Moore, E. E., & Jefferson, A. L. (2021). Impact of cardiovascular hemodynamics on cognitive aging.
Arterioscler. Thromb. Vasc. Biol., 41(4), 1255–1264.
Morinaga, A., Hirohata, M., Ono, K., & Yamada, M. (2007). Estrogen has anti-amyloidogenic effects on
alzheimer’s beta-amyloid fibrils in vitro. Biochem. Biophys. Res. Commun., 359(3), 697–702.
179

Moser, V. A., & Pike, C. J. (2016). Obesity and sex interact in the regulation of alzheimer’s disease. Neurosci.
Biobehav. Rev., 67, 102–118.
Mufford, M. S., Stein, D. J., Dalvie, S., Groenewold, N. A., Thompson, P. M., & Jahanshad, N. (2017).
Neuroimaging genomics in psychiatry-a translational approach. Genome Med., 9, 102.
Müller, V. I., Cieslik, E. C., Serbanescu, I., Laird, A. R., Fox, P. T., & Eickhoff, S. B. (2017). Altered brain
activity in unipolar depression revisited: Meta-analyses of neuroimaging studies. JAMA Psychiatry,
74(1), 47–55.
Muoio, V., Persson, P. B., & Sendeski, M. M. (2014). The neurovascular unit - concept review. Acta Physiol.,
210(4), 790–798.
Narvacan, K., Treit, S., Camicioli, R., Martin, W., & Beaulieu, C. (2017). Evolution of deep gray matter
volume across the human lifespan. Hum. Brain Mapp., 38(8), 3771–3790.
Nebuloni, M., Pellegrinelli, A., Ferri, A., Bonetto, S., Boldorini, R., Vago, L., Grassi, M. P., & Costanzi, G.
(2001). Beta amyloid precursor protein and patterns of HIV p24 immunohistochemistry in different
brain areas of AIDS patients. AIDS, 15(5), 571–575.
Neu, S. C., Pa, J., Kukull, W., Beekly, D., Kuzma, A., Gangadharan, P., Wang, L.-S., Romero, K., Arneric, S. P.,
Redolfi, A., Orlandi, D., Frisoni, G. B., Au, R., Devine, S., Auerbach, S., Espinosa, A., Boada, M.,
Ruiz, A., Johnson, S. C., . . . Toga, A. W. (2017). Apolipoprotein E genotype and sex risk factors for
alzheimer disease: A meta-analysis. JAMA Neurol., 74(10), 1178–1189.
Nianogo, R. A., Rosenwohl-Mack, A., Yaffe, K., Carrasco, A., Hoffmann, C. M., & Barnes, D. E. (2022). Risk
factors associated with alzheimer disease and related dementias by sex and race and ethnicity in
the US. JAMA Neurol., 79, 584–591.
Nichols, E., Steinmetz, J. D., Vollset, S. E., Fukutaki, K., Chalek, J., Abd-Allah, F., Abdoli, A., Abualhasan, A.,
Abu-Gharbieh, E., Akram, T. T., Al Hamad, H., Alahdab, F., Alanezi, F. M., Alipour, V.,
Almustanyir, S., Amu, H., Ansari, I., Arabloo, J., Ashraf, T., . . . Vos, T. (2022). Estimation of the
global prevalence of dementia in 2019 and forecasted prevalence in 2050: An analysis for the global
burden of disease study 2019. The Lancet Public Health, 7(2), e105–e125.
Nichols, T. E., Das, S., Eickhoff, S. B., Evans, A. C., Glatard, T., Hanke, M., Kriegeskorte, N., Milham, M. P.,
Poldrack, R. A., Poline, J.-B., Proal, E., Thirion, B., Van Essen, D. C., White, T., & Yeo, B. T. T. (2017).
Best practices in data analysis and sharing in neuroimaging using MRI. Nat. Neurosci., 20(3),
299–303.
Nielsen, J. B., Kühl, J. T., Pietersen, A., Graff, C., Lind, B., Struijk, J. J., Olesen, M. S., Sinner, M. F.,
Bachmann, T. N., Haunsø, S., Nordestgaard, B. G., Ellinor, P. T., Svendsen, J. H., Kofoed, K. F.,
Køber, L., & Holst, A. G. (2015). P-wave duration and the risk of atrial fibrillation: Results from the
copenhagen ECG study. Heart Rhythm, 12(9), 1887–1895.
Ochs, A. L., Ross, D. E., Zannoni, M. D., Abildskov, T. J., Bigler, E. D., & Alzheimer’s Disease Neuroimaging
Initiative. (2015). Comparison of automated brain volume measures obtained with NeuroQuant and
FreeSurfer. J. Neuroimaging, 25(5), 721–727.
180

Ohi, K., Shimada, T., Kataoka, Y., Yasuyama, T., Kawasaki, Y., Shioiri, T., & Thompson, P. M. (2020). Genetic
correlations between subcortical brain volumes and psychiatric disorders. Br. J. Psychiatry, 216(5),
280–283.
Ostrowska, B., Lind, L., Sciaraffia, E., & Blomström-Lundqvist, C. (2022a). Short P-Wave duration is
associated with incident atrial fibrillation. Int. Heart J., 63(4), 700–707.
Ostrowska, B., Lind, L., Sciaraffia, E., & Blomström-Lundqvist, C. (2022b). A short p-wave duration is
associated with incident heart failure in the elderly: A 15 years follow-up cohort study. J. Geriatr.
Cardiol., 19(9), 643–650.
Oveisgharan, S., Arvanitakis, Z., Yu, L., Farfel, J., Schneider, J. A., & Bennett, D. A. (2018). Sex differences in
alzheimer’s disease and common neuropathologies of aging.
Patel, M. R., Green, M., Tariq, M., Alhawli, A., Syed, N., Fleming, P. J., Ali, A., Bacon, E., Goodall, S.,
Smith, A., & Resnicow, K. (2022). A snapshot of social risk factors and associations with health
outcomes in a community sample of middle eastern and north african (MENA) people in the U.S. J.
Immigr. Minor. Health, 24, 376–384.
Peng, H., Gong, W., Beckmann, C. F., Vedaldi, A., & Smith, S. M. (2021). Accurate brain age prediction with
lightweight deep neural networks. Med. Image Anal., 68, 101871.
Perlaki, G., Horvath, R., Nagy, S. A., Bogner, P., Doczi, T., Janszky, J., & Orsi, G. (2017). Comparison of
accuracy between FSL’s FIRST and freesurfer for caudate nucleus and putamen segmentation. Sci.
Rep., 7(1), 2418.
Petersen, S. E., & Posner, M. I. (2012). The attention system of the human brain: 20 years after. Annu. Rev.
Neurosci., 35, 73–89.
Phillimore, P., Beattie, A., & Townsend, P. (1988). Health and deprivation. inequality and the north. london:
Croom helm. Health Policy.
Podcasy, J. L., & Epperson, C. N. (2016). Considering sex and gender in alzheimer disease and other
dementias. Dialogues Clin. Neurosci., 18, 437–446.
Poldrack, R. A., & Gorgolewski, K. J. (2017). OpenfMRI: Open sharing of task fMRI data. Neuroimage, 144(Pt
B), 259–261.
Poldrack, R. A., Whitaker, K., & Kennedy, D. (2020). Introduction to the special issue on reproducibility in
neuroimaging. Neuroimage, 218, 116357.
Popescu, S. G., Glocker, B., Sharp, D. J., & Cole, J. H. (2021). Local brain-age: A U-net model. Front. Aging
Neurosci., 13, 761954.
Potvin, O., Mouiha, A., Dieumegarde, L., Duchesne, S., & Alzheimer’s Disease Neuroimaging Initiative.
(2016). Normative data for subcortical regional volumes over the lifetime of the adult human brain.
Neuroimage, 137, 9–20.
181

Power, D. A., Lampert, J., Camaj, A., Bienstock, S. W., Kocovic, N., Bayes-Genis, A., Miller, M. A.,
Bayés-de-Luna, A., & Fuster, V. (2022). Cardiovascular complications of interatrial conduction
block: JACC State-of-the-Art review. J. Am. Coll. Cardiol., 79(12), 1199–1211.
Putcha, D., Katsumi, Y., Brickhouse, M., Flaherty, R., Salat, D. H., Touroutoglou, A., & Dickerson, B. C.
(2023). Gray to white matter signal ratio as a novel biomarker of neurodegeneration in alzheimer’s
disease. Neuroimage Clin, 37, 103303.
Qiu, C., & Fratiglioni, L. (2015). A major role for cardiovascular burden in age-related cognitive decline. Nat.
Rev. Cardiol., 12(5), 267–277.
Raichle, M. E., & Gusnard, D. A. (2002). Appraising the brain’s energy budget. Proc. Natl. Acad. Sci. U. S. A.,
99(16), 10237–10239.
Ramachandran, P., Zoph, B., & Le, Q. V. (2017). Searching for activation functions. arXiv [cs.NE].
Raschka, S. (2018). MLxtend: Providing machine learning and data science utilities and extensions to
python’s scientific computing stack. J. Open Source Softw., 3, 638.
Rasmussen, M. K., Mestre, H., & Nedergaard, M. (2018). The glymphatic pathway in neurological disorders.
Lancet Neurol., 17(11), 1016–1024.
Razek, A. A. K. A., & Elsebaie, N. A. (2021). Imaging of vascular cognitive impairment. Clin. Imaging, 74,
45–54.
Reyes, J. L., Norby, F. L., Ji, Y., Wang, W., Parikh, R., Zhang, M. J., Oldenburg, N. C., Lutsey, P. L.,
Jack, C. R., Jr, Johansen, M., Gottesman, R. F., Coresh, J., Mosley, T., Soliman, E. Z., Alonso, A., &
Chen, L. Y. (2023). Association of abnormal p-wave parameters with brain MRI morphology: The
atherosclerosis risk in communities neurocognitive study (ARIC-NCS). Pacing Clin. Electrophysiol.
Ribeiro-Silva, J. C., Nolasco, P., Krieger, J. E., & Miyakawa, A. A. (2021). Dynamic crosstalk between
vascular smooth muscle cells and the aged extracellular matrix. Int. J. Mol. Sci., 22(18).
Ringstad, G., Vatnehol, S. A. S., & Eide, P. K. (2017). Glymphatic MRI in idiopathic normal pressure
hydrocephalus. Brain, 140(10), 2691–2705.
Rocca, W. A., Mielke, M. M., Vemuri, P., & Miller, V. M. (2014). Sex and gender differences in the causes of
dementia: A narrative review. Maturitas, 79, 196–201.
Rosendo-Silva, D., Viana, S., Carvalho, E., Reis, F., & Matafome, P. (2023). Are gut dysbiosis, barrier
disruption, and endotoxemia related to adipose tissue dysfunction in metabolic disorders?
overview of the mechanisms involved. Intern. Emerg. Med., 18(5), 1287–1302.
Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. J.
Educ. Psychol., 66, 688–701.
Saeed, A., Lopez, O., Cohen, A., & Reis, S. E. (2023). Cardiovascular disease and alzheimer’s disease: The
Heart-Brain axis. J. Am. Heart Assoc., 12(21), e030780.
182

Said, M. A., Verweij, N., & van der Harst, P. (2018). Associations of combined genetic and lifestyle risks with
incident cardiovascular disease and diabetes in the UK biobank study. JAMA Cardiol, 3, 693–702.
Salat, D. H., Buckner, R. L., Snyder, A. Z., Greve, D. N., Desikan, R. S. R., Busa, E., Morris, J. C., Dale, A. M.,
& Fischl, B. (2004). Thinning of the cerebral cortex in aging. Cereb. Cortex, 14(7), 721–730.
Santisteban, M. M., & Iadecola, C. (2018). Hypertension, dietary salt and cognitive impairment. J. Cereb.
Blood Flow Metab., 38, 2112–2128.
Saranathan, M., Worters, P. W., Rettmann, D. W., Winegar, B., & Becker, J. (2017). Physics for clinicians:
Fluid-attenuated inversion recovery (FLAIR) and double inversion recovery (DIR) imaging. J.
Magn. Reson. Imaging, 46(6), 1590–1600.
Satizabal, C. L., Adams, H. H. H., Hibar, D. P., White, C. C., Knol, M. J., Stein, J. L., Scholz, M.,
Sargurupremraj, M., Jahanshad, N., Roshchupkin, G. V., Smith, A. V., Bis, J. C., Jian, X., Luciano, M.,
Hofer, E., Teumer, A., van der Lee, S. J., Yang, J., Yanek, L. R., . . . Ikram, M. A. (2019). Genetic
architecture of subcortical brain structures in 38,851 individuals. Nat. Genet., 51(11), 1624–1636.
Satizabal, C. L., Beiser, A. S., Fletcher, E., Seshadri, S., & DeCarli, C. (2024). A novel neuroimaging signature
for ADRD risk stratification in the community. Alzheimers. Dement., 20(3), 1881–1893.
Schaefer, A., Kong, R., Gordon, E. M., Laumann, T. O., Zuo, X.-N., Holmes, A. J., Eickhoff, S. B., &
Yeo, B. T. T. (2018). Local-global parcellation of the human cerebral cortex from intrinsic functional
connectivity MRI. Cereb. Cortex, 28(9), 3095–3114.
Schaeffer, S., & Iadecola, C. (2021). Revisiting the neurovascular unit. Nat. Neurosci., 24(9), 1198–1209.
Seeley, W. W., Crawford, R. K., Zhou, J., Miller, B. L., & Greicius, M. D. (2009). Neurodegenerative diseases
target large-scale human brain networks. Neuron, 62(1), 42–52.
Segall, J. M., Allen, E. A., Jung, R. E., Erhardt, E. B., Arja, S. K., Kiehl, K., & Calhoun, V. D. (2012).
Correspondence between structure and function in the human brain at rest. Front. Neuroinform., 6,
10.
Seitzman, B. A., Snyder, A. Z., Leuthardt, E. C., & Shimony, J. S. (2019). The state of resting state networks.
Top. Magn. Reson. Imaging, 28(4), 189–196.
Shokri-Kojori, E., Wang, G.-J., Wiers, C. E., Demiral, S. B., Guo, M., Kim, S. W., Lindgren, E., Ramirez, V.,
Zehra, A., Freeman, C., Miller, G., Manza, P., Srivastava, T., De Santi, S., Tomasi, D., Benveniste, H.,
& Volkow, N. D. (2018). β-Amyloid accumulation in the human brain after one night of sleep
deprivation. Proc. Natl. Acad. Sci. U. S. A., 115(17), 4483–4488.
Shumaker, S. A., Legault, C., Kuller, L., Rapp, S. R., Thal, L., Lane, D. S., Fillit, H., Stefanick, M. L.,
Hendrix, S. L., Lewis, C. E., Masaki, K., Coker, L. H., & Women’s Health Initiative Memory Study.
(2004). Conjugated equine estrogens and incidence of probable dementia and mild cognitive
impairment in postmenopausal women: Women’s health initiative memory study. JAMA, 291(24),
2947–2958.
183

Shumaker, S. A., Legault, C., Rapp, S. R., Thal, L., Wallace, R. B., Ockene, J. K., Hendrix, S. L.,
Jones, B. N., 3rd, Assaf, A. R., Jackson, R. D., Kotchen, J. M., Wassertheil-Smoller, S.,
Wactawski-Wende, J., & WHIMS Investigators. (2003). Estrogen plus progestin and the incidence of
dementia and mild cognitive impairment in postmenopausal women: The women’s health
initiative memory study: A randomized controlled trial. JAMA, 289(20), 2651–2662.
Shuman, D. I., Narang, S. K., Frossard, P., Ortega, A., & Vandergheynst, P. (2013). The emerging field of
signal processing on graphs: Extending high-dimensional data analysis to networks and other
irregular domains. IEEE Signal Process. Mag., 30(3), 83–98.
Sirugo, G., Williams, S. M., & Tishkoff, S. A. (2019). The missing diversity in human genetic studies. Cell,
177, 26–31.
Sled, J. G., Zijdenbos, A. P., & Evans, A. C. (1998). A nonparametric method for automatic correction of
intensity nonuniformity in MRI data. IEEE Trans. Med. Imaging, 17(1), 87–97.
Smith, S. M. (2002). Fast robust automated brain extraction. Hum. Brain Mapp., 17(3), 143–155.
Smith, S. M., Vidaurre, D., Alfaro-Almagro, F., Nichols, T. E., & Miller, K. L. (2019). Estimation of brain age
delta from brain imaging. Neuroimage, 200, 528–539.
Sochocka, M., Zwolińska, K., & Leszek, J. (2017). The infectious etiology of alzheimer’s disease. CN, 15(7).
Sohrabji, F., & Lewis, D. K. (2006). Estrogen–BDNF interactions: Implications for neurodegenerative
diseases.
Soliman, E. Z., Alonso, A., Misialek, J. R., Jain, A., Watson, K. E., Lloyd-Jones, D. M., Lima, J., Shea, S.,
Burke, G. L., & Heckbert, S. R. (2013). Reference ranges of PR duration and p-wave indices in
individuals free of cardiovascular disease: The Multi-Ethnic study of atherosclerosis (MESA). J.
Electrocardiol., 46(6), 702–706.
Sperling, R. A., Jack, C. R., Jr, Black, S. E., Frosch, M. P., Greenberg, S. M., Hyman, B. T., Scheltens, P.,
Carrillo, M. C., Thies, W., Bednar, M. M., Black, R. S., Brashear, H. R., Grundman, M., Siemers, E. R.,
Feldman, H. H., & Schindler, R. J. (2011). Amyloid-related imaging abnormalities in
amyloid-modifying therapeutic trials: Recommendations from the alzheimer’s association research
roundtable workgroup. Alzheimers. Dement., 7, 367–385.
Spiegelhalder, K., Regen, W., Baglioni, C., Nissen, C., Riemann, D., & Kyle, S. D. (2015). Neuroimaging
insights into insomnia. Curr. Neurol. Neurosci. Rep., 15(3), 9.
Stanley, L. C., Mrak, R. E., Woody, R. C., Perrot, L. J., Zhang, S., Marshak, D. R., Nelson, S. J., & Griffin, W. S.
(1994). Glial cytokines as neuropathogenic factors in HIV infection: Pathogenic similarities to
alzheimer’s disease. J. Neuropathol. Exp. Neurol., 53(3), 231–238.
Steffener, J., Habeck, C., O’Shea, D., Razlighi, Q., Bherer, L., & Stern, Y. (2016). Differences between
chronological and brain age are related to education and self-reported physical activity. Neurobiol.
Aging, 40, 138–144.
184

Stigler, S. M. (1997). Regression towards the mean, historically considered. Stat. Methods Med. Res., 6,
103–114.
Strazzullo, P., D’Elia, L., Kandala, N.-B., & Cappuccio, F. P. (2009). Salt intake, stroke, and cardiovascular
disease: Meta-analysis of prospective studies. BMJ, 339, b4567.
Strittmatter, W. J. (2012). Medicine. old drug, new hope for alzheimer’s disease. Science, 335(6075),
1447–1448.
Stuhrmann, A., Suslow, T., & Dannlowski, U. (2011). Facial emotion processing in major depression: A
systematic review of neuroimaging findings. Biol. Mood Anxiety Disord., 1(1), 10.
Sui, J., Huster, R., Yu, Q., Segall, J. M., & Calhoun, V. D. (2014). Function-structure associations of the brain:
Evidence from multimodal connectivity and covariance studies. Neuroimage, 102 Pt 1, 11–23.
Sun, F. W., Stepanovic, M. R., Andreano, J., Barrett, L. F., Touroutoglou, A., & Dickerson, B. C. (2016).
Youthful brains in older adults: Preserved neuroanatomy in the default mode and salience
networks contributes to youthful memory in superaging. J. Neurosci., 36(37), 9659–9668.
Sweeney, M. D., Sagare, A. P., & Zlokovic, B. V. (2018). Blood–brain barrier breakdown in alzheimer disease
and other neurodegenerative disorders. Nat. Rev. Neurol., 14(3), 133–150.
Tan, P.-N., Steinbach, M., & Kumar, V. (2006). Data mining introduction. People’s Posts and
Telecommunications Publishing House, Beijing.
Tang, M. X., Stern, Y., Marder, K., Bell, K., Gurland, B., Lantigua, R., Andrews, H., Feng, L., Tycko, B., &
Mayeux, R. (1998). The APOE-epsilon4 allele and the risk of alzheimer disease among african
americans, whites, and hispanics. JAMA, 279(10), 751–755.
Ter Telgte, A., van Leijsen, E. M. C., Wiegertjes, K., Klijn, C. J. M., Tuladhar, A. M., & de Leeuw, F.-E. (2018).
Cerebral small vessel disease: From a focal to a global perspective. Nat. Rev. Neurol., 14(7), 387–398.
Thaker, A. A., Weinberg, B. D., Dillon, W. P., Hess, C. P., Cabral, H. J., Fleischman, D. A., Leurgans, S. E.,
Bennett, D. A., Hyman, B. T., Albert, M. S., Killiany, R. J., Fischl, B., Dale, A. M., & Desikan, R. S.
(2017). Entorhinal cortex: Antemortem cortical thickness and postmortem neurofibrillary tangles
and amyloid pathology. AJNR Am. J. Neuroradiol., 38(5), 961–965.
Thambisetty, M., Wan, J., Carass, A., An, Y., Prince, J. L., & Resnick, S. M. (2010). Longitudinal changes in
cortical thickness associated with normal aging. Neuroimage, 52(4), 1215–1223.
Thompson, P. M., Jahanshad, N., Ching, C. R. K., Salminen, L. E., Thomopoulos, S. I., Bright, J., Baune, B. T.,
Bertolín, S., Bralten, J., Bruin, W. B., Bülow, R., Chen, J., Chye, Y., Dannlowski, U., de Kovel, C. G. F.,
Donohoe, G., Eyler, L. T., Faraone, S. V., Favre, P., . . . ENIGMA Consortium. (2020). ENIGMA and
global neuroscience: A decade of large-scale studies of the brain in health and disease across more
than 40 countries. Transl. Psychiatry, 10, 100.
Thorand, B., Baumert, J., Kolb, H., Meisinger, C., Chambless, L., Koenig, W., & Herder, C. (2007). Sex
differences in the prediction of type 2 diabetes by inflammatory markers: Results from the
MONICA/KORA augsburg case-cohort study, 1984-2002. Diabetes Care, 30(4), 854–860.
185

Toro, C. A., Zhang, L., Cao, J., & Cai, D. (2019). Sex differences in alzheimer’s disease: Understanding the
molecular impact.
Touroutoglou, A., & Dickerson, B. C. (2019). Cingulate-centered large-scale networks: Normal functions,
aging, and neurodegenerative disease. Handb. Clin. Neurol., 166, 113–127.
Townsend, R. F., Woodside, J. V., Prinelli, F., O’Neill, R. F., & McEvoy, C. T. (2022). Associations between
dietary patterns and neuroimaging markers: A systematic review. Front Nutr, 9, 806006.
Tublin, J. M., Adelstein, J. M., Del Monte, F., Combs, C. K., & Wold, L. E. (2019). Getting to the heart of
alzheimer disease. Circ. Res., 124(1), 142–149.
Tustison, N. J., Cook, P. A., Klein, A., Song, G., Das, S. R., Duda, J. T., Kandel, B. M., van Strien, N.,
Stone, J. R., Gee, J. C., & Avants, B. B. (2014). Large-scale evaluation of ANTs and FreeSurfer
cortical thickness measurements. Neuroimage, 99, 166–179.
Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D., Crivello, F., Etard, O., Delcroix, N., Mazoyer, B., &
Joliot, M. (2002). Automated anatomical labeling of activations in SPM using a macroscopic
anatomical parcellation of the MNI MRI single-subject brain. Neuroimage, 15(1), 273–289.
Udeh-Momoh, C., Watermeyer, T., & Female Brain Health and Endocrine Research (FEMBER) consortium.
(2021). Female specific risk factors for the development of alzheimer’s disease neuropathology and
cognitive impairment: Call for a precision medicine approach. Ageing Res. Rev., 71, 101459.
Ungvari, Z., Tarantini, S., Donato, A. J., Galvan, V., & Csiszar, A. (2018). Mechanisms of vascular aging. Circ.
Res., 123(7), 849–867.
Ungvari, Z., Tarantini, S., Sorond, F., Merkely, B., & Csiszar, A. (2020). Mechanisms of vascular aging, a
geroscience perspective: JACC focus seminar. J. Am. Coll. Cardiol., 75(8), 931–941.
Uyarel, H., Ozdöl, C., Karabulut, A., Okmen, E., & Cam, N. (2005). Acute alcohol intake and p-wave
dispersion in healthy men. Anadolu Kardiyol. Derg., 5(4), 289–293.
van der Eerden, A. W., van den Heuvel, T. L. A., Maas, M. C., Vart, P., Vos, P. E., Platel, B., Góraj, B. M., &
Manniesing, R. (2022). The radiological interpretation of possible microbleeds after moderate or
severe traumatic brain injury: A longitudinal study. Neuroradiology, 64, 1145–1156.
van der Velpen, I. F., Yancy, C. W., Sorond, F. A., & Sabayan, B. (2017). Impaired cardiac function and
cognitive brain aging. Can. J. Cardiol., 33(12), 1587–1596.
Van Essen, D. C., Smith, S. M., Barch, D. M., Behrens, T. E. J., Yacoub, E., Ugurbil, K., & WU-Minn HCP
Consortium. (2013). The WU-minn human connectome project: An overview. Neuroimage, 80,
62–79.
Van Gestel, H., Franke, K., Petite, J., Slaney, C., Garnham, J., Helmick, C., Johnson, K., Uher, R., Alda, M., &
Hajek, T. (2019). Brain age in bipolar disorders: Effects of lithium treatment. Aust. N. Z. J.
Psychiatry, 53, 1179–1188.
186

Vemuri, P., Decarli, C., & Duering, M. (2022). Imaging markers of vascular brain health: Quantification,
clinical implications, and future directions. Stroke, 53(2), 416–426.
Vicente-Saez, R., & Martinez-Fuentes, C. (2018). Open science now: A systematic literature review for an
integrated definition. J. Bus. Res., 88, 428–436.
Visser, E., Keuken, M. C., Douaud, G., Gaura, V., Bachoud-Levi, A.-C., Remy, P., Forstmann, B. U., &
Jenkinson, M. (2016). Automatic segmentation of the striatum and globus pallidus using MIST:
Multimodal image segmentation tool. Neuroimage, 125, 479–497.
Vogt, B. A. (2019). Cingulate impairments in ADHD: Comorbidities, connections, and treatment. Handb.
Clin. Neurol., 166, 297–314.
Walker, K. A., Basisty, N., Wilson, D. M., 3rd, & Ferrucci, L. (2022). Connecting aging biology and
inflammation in the omics era. J. Clin. Invest., 132(14).
Wang, M., Jiang, L., Monticone, R. E., & Lakatta, E. G. (2014). Proinflammation: The key to arterial aging.
Trends Endocrinol. Metab., 25(2), 72–79.
Ward, S. A., & Pase, M. P. (2020). Advances in pathophysiology and neuroimaging: Implications for sleep
and dementia. Respirology, 25(6), 580–592.
Wardlaw, J. M., Benveniste, H., Nedergaard, M., Zlokovic, B. V., Mestre, H., Lee, H., Doubal, F. N., Brown, R.,
Ramirez, J., MacIntosh, B. J., Tannenbaum, A., Ballerini, L., Rungta, R. L., Boido, D., Sweeney, M.,
Montagne, A., Charpak, S., Joutel, A., Smith, K. J., . . . colleagues from the Fondation Leducq
Transatlantic Network of Excellence on the Role of the Perivascular Space in Cerebral Small Vessel
Disease. (2020). Perivascular spaces in the brain: Anatomy, physiology and pathology. Nat. Rev.
Neurol., 16(3), 137–153.
Wardlaw, J. M., Smith, E. E., Biessels, G. J., Cordonnier, C., Fazekas, F., Frayne, R., Lindley, R. I., O’Brien, J. T.,
Barkhof, F., Benavente, O. R., Black, S. E., Brayne, C., Breteler, M., Chabriat, H., Decarli, C.,
de Leeuw, F.-E., Doubal, F., Duering, M., Fox, N. C., . . . STandards for ReportIng Vascular changes
on nEuroimaging (STRIVE v1). (2013). Neuroimaging standards for research into small vessel
disease and its contribution to ageing and neurodegeneration. Lancet Neurol., 12(8), 822–838.
Wardlaw, J. M., Valdés Hernández, M. C., & Muñoz-Maniega, S. (2015). What are white matter
hyperintensities made of? relevance to vascular cognitive impairment. J. Am. Heart Assoc., 4(6),
001140.
Weiner, M. W., Veitch, D. P., Aisen, P. S., Beckett, L. A., Cairns, N. J., Cedarbaum, J., Donohue, M. C.,
Green, R. C., Harvey, D., Jack, C. R., Jr, Jagust, W., Morris, J. C., Petersen, R. C., Saykin, A. J.,
Shaw, L., Thompson, P. M., Toga, A. W., Trojanowski, J. Q., & Alzheimer’s Disease Neuroimaging
Initiative. (2015). Impact of the alzheimer’s disease neuroimaging initiative, 2004 to 2014.
Alzheimers. Dement., 11(7), 865–884.
187

Weiner, M. W., Veitch, D. P., Aisen, P. S., Beckett, L. A., Cairns, N. J., Green, R. C., Harvey, D., Jack, C. R., Jr,
Jagust, W., Morris, J. C., Petersen, R. C., Salazar, J., Saykin, A. J., Shaw, L. M., Toga, A. W.,
Trojanowski, J. Q., & Alzheimer’s Disease Neuroimaging Initiative. (2017). The alzheimer’s disease
neuroimaging initiative 3: Continued innovation for clinical trial improvement. Alzheimers.
Dement., 13(5), 561–571.
Weinstein, G., Beiser, A. S., Choi, S. H., Preis, S. R., Chen, T. C., Vorgas, D., Au, R., Pikula, A., Wolf, P. A.,
DeStefano, A. L., Vasan, R. S., & Seshadri, S. (2014). Serum brain-derived neurotrophic factor and
the risk for dementia: The framingham heart study. JAMA Neurol., 71(1), 55–61.
Whitley, R., Kimberlin, D. W., & Prober, C. G. (2011, February). Pathogenesis and disease. In A. Arvin,
G. Campadelli-Fiume, E. Mocarski, P. S. Moore, B. Roizman, R. Whitley, & K. Yamanishi (Eds.),
Human herpesviruses: Biology, therapy, and immunoprophylaxis. Cambridge University Press.
Williams, L. R., & Leggett, R. W. (1989). Reference values for resting blood flow to organs of man. Clin. Phys.
Physiol. Meas., 10(3), 187–217.
Windham, I. A., & Cohen, S. (2024). The cell biology of APOE in the brain. Trends Cell Biol., 34(4), 338–348.
Winkler, A. M., Kochunov, P., Blangero, J., Almasy, L., Zilles, K., Fox, P. T., Duggirala, R., & Glahn, D. C.
(2010). Cortical thickness or grey matter volume? the importance of selecting the phenotype for
imaging genetics studies. Neuroimage, 53(3), 1135–1146.
World Health Organization. (2024). Global status report on alcohol and health and treatment of substance
use disorders [Accessed: 2024-8-1].
Xia, C., Vonder, M., Sidorenkov, G., Oudkerk, M., de Groot, J. C., van der Harst, P., de Bock, G. H.,
De Deyn, P. P., & Vliegenthart, R. (2020). The relationship of coronary artery calcium and clinical
coronary artery disease with cognitive function: A systematic review and Meta-Analysis. J.
Atheroscler. Thromb., 27(9), 934–958.
Xie, L., Kang, H., Xu, Q., Chen, M. J., Liao, Y., Thiyagarajan, M., O’Donnell, J., Christensen, D. J.,
Nicholson, C., Iliff, J. J., Takano, T., Deane, R., & Nedergaard, M. (2013). Sleep drives metabolite
clearance from the adult brain. Science, 342, 373–377.
Yan, C., Zhou, Y., Chen, Q., Luo, Y., Zhang, J. H., Huang, H., & Shao, A. (2020). Dysfunction of the
neurovascular unit in diabetes-related neurodegeneration. Biomed. Pharmacother., 131, 110656.
Yan, S., Qian, T., Maréchal, B., Kober, T., Zhang, X., Zhu, J., Lei, J., Li, M., & Jin, Z. (2020). Test-retest
variability of brain morphometry analysis: An investigation of sequence and coil effects. Ann
Transl Med, 8(1), 12.
Yang, J., Lunde, L. K., Nuntagij, P., Oguchi, T., Camassa, L. M. A., Nilsson, L. N. G., Lannfelt, L., Xu, Y.,
Amiry-Moghaddam, M., Ottersen, O. P., & Torp, R. (2011). Loss of astrocyte polarization in the
tg-ArcSwe mouse model of alzheimer’s disease. J. Alzheimers. Dis., 27(4), 711–722.
Yang, M., Sun, D., Wang, Y., Yan, M., Zheng, J., & Ren, J. (2021). Cognitive impairment in heart failure:
Landscape, challenges, and future directions. Front Cardiovasc Med, 8, 831734.
188

Yang, T., Sun, Y., Lu, Z., Leak, R. K., & Zhang, F. (2017). The impact of cerebrovascular aging on vascular
cognitive impairment and dementia. Ageing Res. Rev., 34, 15–29.
Yassine, H. N., Samieri, C., Livingston, G., Glass, K., Wagner, M., Tangney, C., Plassman, B. L., Ikram, M. A.,
Voigt, R. M., Gu, Y., O’Bryant, S., Minihane, A. M., Craft, S., Fink, H. A., Judd, S., Andrieu, S.,
Bowman, G. L., Richard, E., Albensi, B., . . . Schneider, L. S. (2022). Nutrition state of science and
dementia prevention: Recommendations of the nutrition for dementia prevention working group.
Lancet Healthy Longev, 3, e501–e512.
Yeo, B. T. T., Krienen, F. M., Sepulcre, J., Sabuncu, M. R., Lashkari, D., Hollinshead, M., Roffman, J. L.,
Smoller, J. W., Zöllei, L., Polimeni, J. R., Fischl, B., Liu, H., & Buckner, R. L. (2011). The organization
of the human cerebral cortex estimated by intrinsic functional connectivity. J. Neurophysiol., 106(3),
1125–1165.
You, R., Ho, Y.-S., & Chang, R. C.-C. (2022). The pathogenic effects of particulate matter on
neurodegeneration: A review. J. Biomed. Sci., 29(1), 15.
Zarb, Y., Sridhar, S., Nassiri, S., Utz, S. G., Schaffenrath, J., Maheshwari, U., Rushing, E. J., Nilsson, K. P. R.,
Delorenzi, M., Colonna, M., Greter, M., & Keller, A. (2021). Microglia control small vessel
calcification via TREM2. Sci Adv, 7(9).
Zavaliangos-Petropulu, A., Tubi, M. A., Haddad, E., Zhu, A., Braskie, M. N., Jahanshad, N., Thompson, P. M.,
& Liew, S.-L. (2022). Testing a convolutional neural network-based hippocampal segmentation
method in a stroke population. Hum. Brain Mapp., 43(1), 234–243.
Zeppenfeld, D. M., Simon, M., Haswell, J. D., D’Abreo, D., Murchison, C., Quinn, J. F., Grafe, M. R.,
Woltjer, R. L., Kaye, J., & Iliff, J. J. (2017). Association of perivascular localization of aquaporin-4
with cognition and alzheimer disease in aging brains. JAMA Neurol., 74(1), 91–99.
Zhao, M., Chu, J., Feng, S., Guo, C., Xue, B., He, K., & Li, L. (2023). Immunological mechanisms of
inflammatory diseases caused by gut microbiota dysbiosis: A review. Biomed. Pharmacother.,
164(114985), 114985.
Zhou, J., Li, A., Tan, M., Lam, M. C. Y., Hung, L. T., Siu, R. W. H., Lee, S., Lakhani, I., Chan, J. S. K.,
Bin Waleed, K., Liu, T., Jeevaratnam, K., Zhang, Q., & Tse, G. (2023). P-wave durations from
automated electrocardiogram analysis to predict atrial fibrillation and mortality in heart failure.
ESC Heart Fail, 10(2), 872–883.
Zhuang, P., Liu, X., Li, Y., Wan, X., Wu, Y., Wu, F., Zhang, Y., & Jiao, J. (2021). Effect of diet quality and
genetic predisposition on hemoglobin A1c and type 2 diabetes risk: Gene-Diet interaction analysis
of 357,419 individuals. Diabetes Care, 44, 2470–2479.
Zijdenbos, A., Forghani, R., & Evans, A. (1998). Automatic quantification of MS lesions in 3D MRI brain
data sets: Validation of INSECT. Medical Image Computing and Computer-Assisted Intervention —
MICCAI’98, 439–448.
Zlokovic, B. V. (2011). Neurovascular pathways to neurodegeneration in alzheimer’s disease and other
disorders. Nat. Rev. Neurosci., 12(12), 723–738.
189

Zuo, X.-N., Anderson, J. S., Bellec, P., Birn, R. M., Biswal, B. B., Blautzik, J., Breitner, J. C. S., Buckner, R. L.,
Calhoun, V. D., Castellanos, F. X., Chen, A., Chen, B., Chen, J., Chen, X., Colcombe, S. J.,
Courtney, W., Craddock, R. C., Di Martino, A., Dong, H.-M., . . . Milham, M. P. (2014). An open
science resource for establishing reliability and reproducibility in functional connectomics. Sci
Data, 1, 140049.
Zuo, X.-N., Xu, T., & Milham, M. P. (2019). Harnessing reliability for neuroscience research.
190

Abstract (if available)

Abstract Whether or not an individual develops late onset dementia is highly variable, as it depends on a wide range of factors, many of which are modifiable. Quantifying the rate at which the brain ages and the factors affecting it has the potential to inform treatments and recommendations, which may shape individual outcomes. In this dissertation, I explore neuroimaging markers of risk and resilience to brain aging. Chapter 1 provides background on the factors which influence brain aging and dementia risk. I discuss ways in which researchers can leverage in-vivo MRI to quantify markers of brain aging. Chapter 2 focuses on the importance of the reliability of such metrics, as derived biomarkers must be robust in order to be replicable and translatable to the clinic. Chapter 3 focuses on assessing risk for brain aging by exploring the impact that subclinical atrial abnormalities may have on the structure of brain aging networks. Chapter 4.1 focuses on brain aging resilience, where we investigate which lifestyle factors contribute most to brain structural resiliency. Finally, I focus on the generalizability of imaging associations in those underrepresented in research studies. Chapter 4.2 explores the causal role that diet has on sex-specific brain structure and chapter 5 focuses on disease prevalence and microhemorrhage associations in individuals from the Middle East and North Africa, an uncharacterized population in brain imaging studies. Collectively, I hope to highlight the utility that MRI biomarkers have on characterizing aging and disease mechanisms and informing recommendations for those at risk for neurodegeneration.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Vascular contributions to brain aging along the Alzheimer's disease continuum

PDF

Using neuroinformatics to identify genomic and proteomic markers of suboptimal aging and Alzheimer's disease

PDF

Characterizing brain aging with neuroimaging, health, and genetic data

PDF

Associations between longitudinal loneliness, epigenetic age, and dementia risk

PDF

Blood-brain barrier pathophysiology in cognitive impairment and injury

PDF

The association of cerebrovascular disease risk factors with brain structure and its modification by genetic variation

PDF

Novel multi-site brain imaging approaches to map HIV-related neuropathology

PDF

Characterization of lenticulostriate arteries using high-resolution black blood MRI as an early imaging biomarker for vascular cognitive impairment and dementia

PDF

The role of the locus coeruleus in Alzheimer’s disease and cerebrovascular function: insights from neuroimaging, neuropsychology, and biofluid markers

PDF

Independent and interactive effects of depression genetic risk and household socioeconomic status on emotional behavior and brain development

PDF

Relationships between lifetime chronic stress exposure, vascular risk, cognition, and brain structure in HIV

PDF

Characterization of the brain in early childhood

PDF

Alzheimer’s disease: dysregulated genes, ethno-racial disparities, and environmental pollution

PDF

Pattern detection in medical imaging: pathology specific imaging contrast, features, and statistical models

PDF

Neuroimaging in complex polygenic disorders

PDF

The role of inflammation in mediating effects of obesity on Alzheimer's disease

PDF

Age differences in diffusivity in the locus coeruleus and its ascending noradrenergic tract

PDF

Crosscultural differences in dementia diagnosis and care-seeking in Hispanic and non-Hispanic white outpatients

PDF

Structure and function of the locus coeruleus across the lifespan

PDF

A multi-site neuroimaging approach to studying hippocampal damage in chronic stroke

Asset Metadata

Creator Haddad, Elizabeth (author)

Core Title Neuroimaging markers of risk & resilience to brain aging and dementia

School Keck School of Medicine

Degree Doctor of Philosophy

Degree Program Neuroscience

Degree Conferral Date 2024-08

Publication Date 10/10/2024

Defense Date 08/14/2024

Publisher Los Angeles, California (original), University of Southern California (original), University of Southern California. Libraries (digital)

Tag Alzheimer's disease,biomarkers,brain aging,dementia,generalizability,lifestyle,magnetic resonance imaging,modifiable factors,MRI,neurodegeneration,neuroimaging,OAI-PMH Harvest,reliability,resilience,vascular risk factors

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Thompson, Paul M (committee chair), Irimia, Andrei (committee member), Jahanshad, Neda (committee member), Kim, Hosung (committee member), Shiroishi, Mark (committee member)

Creator Email haddad.elizabeth@gmail.com,haddade@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC11399C88P

Unique identifier UC11399C88P

Identifier etd-HaddadEliz-13591.pdf (filename)

Legacy Identifier etd-HaddadEliz-13591

Document Type Dissertation

Format theses (aat)

Rights Haddad, Elizabeth

Internet Media Type application/pdf

Type texts

Source 20241011-usctheses-batch-1218 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu