University of Southern California Dissertations and Theses

Estimation of cognitive brain activity in sickle cell disease using functional near-infrared spectroscopy and dynamic systems modeling

(USC Thesis Other)

Estimation of cognitive brain activity in sickle cell disease using functional near-infrared spectroscopy and dynamic systems modeling

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

ESTIMATION OF COGNITIVE BRAIN ACTIVITY IN SICKLE CELL
DISEASE USING FUNCTIONAL NEAR-INFRARED SPECTROSCOPY
AND DYNAMIC SYSTEMS MODELING

John Sunwoo
Department of Biomedical Engineering, University of Southern California

A Dissertation
Submitted to the
FACULTY OF THE USC GRADUATE SCHOOL
University of Southern California
in Partial Fulfillment of the
Requirements for the
Degree of
Doctor of Philosophy

Los Angeles, California
May 28, 2019

Date of degree conferral (Doctor of Philosophy):
August 15
th
, 2019
ii

Committee members:
Thomas D. Coates, MD, Professor, Children’s Hospital Los Angeles, Keck School of Medicine,
John C. Wood, MD/PhD, Professor, Children’s Hospital Los Angeles, Keck School of Medicine, and
Michael C.K. Khoo, PhD, Dissertation Chair, Professor, Department of Biomedical Engineering,
University of Southern California

Date of defense:
May 1
st
, 2019

iii

VITA
John Sunwoo was born on November 14th, 1980 in Gwangju, Korea. He graduated from Chosun
University High School, Korea, in 1999. He earned a Bachelor of Engineering in Electrical and Computer
Engineering and a Master of Science in Electrical Engineering at Auburn University, AL, in 2003 and 2005,
respectively. While in pursuit of his Master of Science degree at Auburn University, he worked under the
guidance of Dr. Charles E. Stroud as a graduate student research assistant in the Auburn University Built-
In Self-Test (AUBIST) laboratory. After obtaining a master’s degree, he worked at Electronics and
Telecommunications Research Institute from 2005 to 2010, developed wearable computing systems
including accelerometer-based gesture recognition and Field-Programmable Logic-Array (FPGA) based
Body-Area-Network (BAN) communication controller for e-textiles. He also earned a Master of Engineering
in Biomedical Engineering at Cornell University, NY, in May 2011. Finally, this led him to pursue a Doctor
of Philosophy degree in Biomedical Engineering at University of Southern California, CA, since 2011. While
in pursuit of his PhD degree, he worked under the guidance of Dr. Michael C.K. Khoo as a graduate student
research assistant since 2013. His area of interest includes functional near infrared spectroscopy (fNIRS),
non-invasive recording of physiological responses to stimuli in human, bio-signal processing, wearable
computing, software/hardware neuro-engineering platforms for entertainment and fashion industries.
iv

DISSERTATION ABSTRACT
ESTIMATION OF COGNITIVE BRAIN ACTIVITY IN SICKLE CELL DISEASE USING FUNCTIONAL
NEAR-INFRARED SPECTROSCOPY AND DYNAMIC SYSTEMS MODELING
John Sunwoo
Doctor of Philosophy, 2019
(MEng, Cornell University, New York, 2011; BS/MS, Auburn University, Alabama, 2003/2005)
73 Typed Pages
Directed by Dr. Michael C.K. Khoo
Vaso-occlusive pain episodes are a distinguishing symptom in sickle cell disease (SCD). Rigid and sickle-
shaped red blood cells occlude the blood flow in the microvasculature, causing damage to the surrounding
tissue accompanied by severe pain. SCD vaso-occlusive crises are believed to be linked with silent and
overt stroke, as well as cognitive decline. Recent MRI studies have demonstrated shrinkage of white matter
in SCD, which, in turn, could be a major factor of cognitive decline in this population. However, little is
known about the cause of a stroke and its exact linkage to cognitive decline in SCD. Analyses of functional
brain imaging data that is obtained when a subject is engaged in a cognitive task can help one understand
underlying brain mechanisms between cognitive decline and disease conditions. Therefore, the goal of this
research study is to use functional near-infrared spectroscopy (fNIRS), which is also one of the popular
functional brain imaging techniques, to study abnormal cognitive brain activity seen in SCD, as well as to
propose a signal processing method for filtering out non-neuronal physiological confounds in the original
fNIRS measurement.
We used a linear dynamic systems modeling approach (Model-based filtering, MBF) for identifying and
subtracting two major physiological confounders in the original fNIRS measurement. Using the filtered
fNIRS signal, we showed inefficiency and possible abnormal cognitive activity in SCD, especially in patients
with stroke conditions. This result was also supported by observations of consistently slower response time
to conducted mental tasks in SCD compared to healthy controls, while task accuracy was not different
between the groups. This finding is consistent with recent reports of white matter loss in SCD and suggest
the use of fNIRS-based neurocognitive responses as a potential biomarker of cognitive decline reported in
SCD.
v

ACKNOWLEDGEMENTS
The journey of my PhD study took almost a decade long - This includes finding a new advisor, switching
to a new lab, and constantly adjusting my intellectual ambitions. One may feel discouraged from the time it
took; however, looking back, I am happy and feel accomplished that I have built academic, engineering,
and soft skills but also met many inspiring scholars and lifelong friends.
First of all, I thank all my colleagues in Khoo lab. Memories of encouragement, helping each other, helpful
discussion, constructive debate and most importantly, our friendship, are the invaluable assets that I have
obtained throughout my time at USC. For this, I thank Alison, Leo, Kelby, Sadaf, Sang, Toey, Winston, and
Yunhua. At Children’s Hospital at Los Angeles (CHLA), for the very same reasons, I greatly thank Roberta,
Maha, Payal, Dr. Coates, Dr. Detterich, Dr. Wood, Anne, Ana, Obdulio, Saranya, Chris, Rose, and Dr.
Meiselman.
I am indebted to my advisor and mentor, Dr. Michael Khoo, for sharing his insights in engineering and
research, as well as giving me this precious opportunity to work with him and his colleagues at USC and
CHLA. I really and sincerely appreciate him for providing me with expert PhD training that greatly helped
me become a better and successful engineer who values knowledge, critical thinking, diligence, confidence,
perseverance, human skills, communication skills, and patience. Again, I thank you, Dr. Khoo.
I also thank you who are outside of my academics. You have been supportive throughout my PhD student
life. I want to shout out the Kappa Kappa Sigma brothers – Matthew, Mitchell, and Clark, as well as the life-
long pledge members Enrique, Leo (again), and Tommy. And I greatly thank you Gene, Uldric, Ale, Ana,
Bryce, Francesca, Gunce, Jannie, Mischal, Rachelle, and everyone in ‘Level-7’ society. I specially
acknowledge Mischal again who provided countless amount of help throughout my PhD study. I also want
to give a special thanks to Rachelle for her encouragement and continuing love. I thank you guys for
believing in me throughout this journey and being amongst the most interesting and influential people that
I have ever known.
vi

Lastly but not the least, I deeply thank my family for their sacrificing support and love in me. I thank you,
my brothers Jin and Nelson, and mom and dad. I wish the best to my brothers. I wish the most happiness
to my mom. I want to tell my dad that I admire you the most as a man and a fellow PhD himself. Finally, let
us all live long, healthy, and happy.

vii

Style manual or journal used: NeuroImage Journal style

viii

TABLE OF CONTENTS
Vita ................................................................................................................................................................ iii
Dissertation abstract .................................................................................................................................... iv
Acknowledgements ....................................................................................................................................... v
Table of Contents ........................................................................................................................................ viii
1. Introduction & dissertation statement .................................................................................................... 1
1.1 Can we estimate neuronal hemodynamics responses from original fNIRS measurements using
other indirect measurements of the confounding background physiological influences? ......................... 2
1.2 Can we detect abnormal cognitive brain activities in SCD? ......................................................... 2
2. Background ............................................................................................................................................ 4
2.1. Sickle cell disease and stroke ....................................................................................................... 4
2.2. Brain hemodynamics: Neurovascular coupling, CO2-dependent cerebral blood flow change,
autoregulation of blood flow to the arterial pressure change and autonomic nervous system ................. 5
2.3. The fNIRS principle, history, and confounding factors .................................................................. 9
2.4. Signal processing and modeling techniques for fNIRS signal refinement .................................. 14
2.4.1 Motion artifact removal and linear filtering .............................................................................. 14
2.4.2 Filtering out skin (scalp) blood flow influences ........................................................................ 14
2.4.3 Hemodynamics models for fNIRS ........................................................................................... 16
2.5. Previous applications of fNIRS and functional experiments in SCD ........................................... 20
3. Mental task induced prefrontal cortex (PFC) responses ..................................................................... 21
3.1. “N-back” standardized short-term memory and “Stroop” attention control tasks ........................ 21
3.1.1 N-back ..................................................................................................................................... 22
3.1.2 Stroop ...................................................................................................................................... 23
ix

3.1.3 Participants and physiological measurements ........................................................................ 24
3.2. Input-output dynamic systems modeling for removing non-neuronal confounders and other pre-
processing techniques ............................................................................................................................. 26
3.2.1 Signal pre-processing via linear filtering and motion artifact removal ..................................... 26
3.2.2 Model-based filtering (MBF) for removing non-brain influences ............................................. 27
3.2.3 Quantification of fNIRS signal ................................................................................................. 29
3.3. fNIRS signal improvement after model-based filtering ............................................................... 30
3.3.1 Non-neuronal contribution in fNIRS measurements ............................................................... 30
3.3.2 N-back ..................................................................................................................................... 31
3.3.3 Stroop ...................................................................................................................................... 34
3.4. Brain responses to mental tasks in SCD vs normal control subjects .......................................... 36
3.4.1 N-back – Hyper-/hypo-responses and slower processing time in SCD .................................. 37
3.4.2 Stroop – Slower processing time as seen in N-back .............................................................. 41
3.4.3 Hemoglobin level as a potential confounding/interaction factor enhanced the contrast between
stroke vs no-stroke in SCD .................................................................................................................. 43
3.5. Discussion ................................................................................................................................... 47
4. Conclusions and suggestions for future work ...................................................................................... 50
4.1. Summary and significance of new findings ................................................................................. 50
4.2. Main contribution ......................................................................................................................... 51
4.3. Limitations ................................................................................................................................... 52
4.4. Suggested future work ................................................................................................................ 53
Epilogue ...................................................................................................................................................... 54
References .................................................................................................................................................. 55
x

1

1. INTRODUCTION & DISSERTATION STATEMENT

Sickle disease (SCD) is an orphan disease due to several reasons. The patient population, mostly African
American, is small in the United States. The patients’ lower socio-economic status is also a factor for not
getting enough attention and clinical research funding opportunities (Judy Stone, 2015). However, the cost
of care in SCD is high because the disease is genetically inherited, meaning that the fight against the
disease is a lifelong process that begins at birth (Gravitz and Pincock, 2014). Moreover, frequent vaso-
occlusive pain crises can result in permanent damage to critical organs, including the brain. Silent and
acute stroke are possible medical complications in SCD, which can cause cognitive decline that can greatly
impact patients’ quality of life (QoL) and lead to lower intelligence quotient (IQ) score, academic difficulty,
as well as low employment rate (Figure 2-1) (DeBaun et al., 2012; Sanger et al., 2016; Teng et al., 2012).
Current advances in social awareness and biomedical research have helped bring about more treatments
and research opportunities in SCD in relation to cognitive decline and QoL; however, the disease processes
are still poorly understood.
The prefrontal cortex (PFC) may be considered the control tower for cognitive and executive functions,
such as reasoning and memorizing. A high prevalence of silent stroke has been reported in SCD even in
patients who received treatments on a regular basis. Functional near-infrared spectroscopy (fNIRS) is a
relatively new, non-invasive brain monitoring technique that emerged in the 1990s and has shown feasibility
in experiments that require naturalistic settings, which are often difficult to do using other brain imaging
modalities, like magnetic resonance imaging (MRI). The fNIRS can be used while a subject is engaged in
a cognitive task, and this can help one understand the underlying brain mechanisms underlying cognitive
decline in SCD. Only recently, however, has the fNIRS community come to recognize the importance of
accounting for confounding factors in fNIRS signal due to extracerebral components, such as scalp blood
flow and arterial CO2 concentration. Therefore, the goal of this dissertation is to develop and implement a
method for estimating the ‘true’ neuronal activity from the original fNIRS measurements and to use the
findings to objectively quantify the degree of cognitive decline accompanying SCD.
2

This dissertation addresses two major questions:
1.1 Can we estimate neuronal hemodynamics responses from original fNIRS
measurements using other indirect measurements of the confounding
background physiological influences?
Various signal processing techniques have been proposed for filtering out the non-neuronal influences
originated from the respiration and scalp blood flow. However, many are based on theoretical assumptions
or are challenging to implement. For example, principal component analysis (PCA) or independent
component analysis (ICA) methods can be used to separate the hidden (brain or non-brain) signals but only
under an assumption that the source signals are independent to each other. This assumption does not
always hold in practice. One can use biophysical models, for example, the balloon model which is based
on prior knowledge of the underlying cerebral hemodynamics, but this model includes many unknown
parameters that have to be estimated, which is challenging (Caldwell et al., 2016; Mildner et al., 2001).
Input-output dynamic systems modeling has proven its effectiveness in elucidating function in various
physiological systems (Li et al., 2014; Marmarelis et al., 2011). In our study, we have measured a rich set
of physiological measurements that can serve as input and output of the model, which can further help
distinguish the signal sources from many different origins. One focus of this dissertation is to determine
whether we can use the systems modeling approach for separating non-neuronal confounders in fNIRS
measurements by using other physiological measurements that are surrogate measures of the skin blood
flow and arterial CO2.
1.2 Can we detect abnormal cognitive brain activities in SCD?
After accounting for confounding influences, we can estimate the task-dependent cognitive responses
during the various interventions that are standardized and being widely used in other research communities.
We will assess the robustness and effectiveness of our proposed method by comparing our results against
the previously published results. Ultimately, we will use the refined neuronal responses obtained from fNIRS
measurements to find a contrast between SCD and controls, which can reveal abnormal PFC functions in
3

different PFC regions due to disease conditions such as SCD with stroke. Our working hypothesis is that
prefrontal cortex responses to standardized cognitive tests are altered in sickle cell disease.

Figure 1-1 My research goal is to characterize the altered brain response in sickle cell disease (SCD) by
quantifying the oxygen perfusion in prefrontal cortex using functional infrared spectroscopy (fNIRS). However,
fNIRS signals also contain non-neuronal fluctuations originating from scalp blood flow and other systemic changes,
and this can lead to false conclusions. Estimating the true brain activity from the fNIRS signal has become an active
research topic. We propose the Laguerre-Volterra modeling to filter out the confounding influences. Carbon dioxide
(CO2) and local blood flow were measured and used as the model input to characterize their influences to fNIRS
signal. The final estimation of the neuronal activity will be used to characterize the cognitive processing in SCD.
4

2. BACKGROUND

2.1. Sickle cell disease and stroke
Sickle cell disease (SCD) is characterized by acute vaso-occlusive pain episodes. Rigid sickled red blood
cells occlude the blood flow in the microvasculature, causing pain and damage to the surrounding tissue.
Frequent experiences of such vaso-occlusive crises can alter various functionalities in the central nervous
system responsible for processing pain and cognitive information (Sangkatumvong et al., 2011). However,
many previous studies have relied on subjective measurements (O’Leary et al., 2013). Choi et al. showed
a white matter shrinkage in SCD, which implied that the shortage of brain resource could be a cause of
cognitive impairment (Choi et al., 2016). Nevertheless, there is a need for improved understanding of this
process using more objective measurements and assessments of the possible cognitive decline in SCD.
Running task-driven stimulus-response experiments while monitoring brain activity can help us discover
different characteristics of mental or physical conditions. For example, the N-back test is one of the

Figure 2-1 Sickle cell disease is a genetic blood disease causes vaso-occlusive pain crises that can
lead to stroke and cognitive decline. Stroke is believed to be the cause of cognitive decline reported in
SCD, but we don’t have direct evidences to support it. We propose using functional brain imaging to study
the underlying mechanism of cognitive decline in SCD.
5

standardized protocols for assessing mental workload and has been used to show increases in brain
oxygenation during demanding memory tasks (Ayaz et al., 2012; Izzetoglu, 2008). In subjects with
amyotrophic lateral sclerosis (ALS), the n-back assessment revealed profoundly different oxygen
consumption trends along with various n-back difficulties (Kuruvilla et al., 2013). The study found that the
oxygenation consumption in ALS was not normal because there was no significant increase in brain
oxygenation during the more demanding memory task, while the ALS and control groups performed the
same. Also, patients with mild cognitive impairment while doing a work fluency task showed diminished
PFC lateralization (Yeung et al., 2016). These studies collectively show the alteration in cognitive
processing.

2.2. Brain hemodynamics: Neurovascular coupling, CO2-dependent cerebral
blood flow change, autoregulation of blood flow to the arterial pressure
change and autonomic nervous system
Cerebral hemodynamics response to various cognitive or sensory stimulation is the focus of study for this
dissertation. The key concept is the ‘neurovascular coupling’, which refers to the increase in the blood flow
as the local neuronal activity increases. Neurons consume most of the energy to set the ionic gradients (of
Na+ and K+) between the inside and outside of the neuron membrane, as it requires active pumping of the
gates on the membrane. This provides the ‘resting potential’, which is required for a neuron to evoke another
action potential. This active pumping mechanism needs the glucose, and the glucose can be metabolized
by the oxygen supplied from the blood. The byproducts of the neuronal activation, such as potassium and
nitro oxide, act as a vasodilator that can increase the blood flow, as summarized in Figure 2-2-a (Hall and
Guyton, 2011). Thus, the change in cerebral blood flow reflects the local neuronal activity (Figure 2-2-b).

6

Figure 2-2 Brain vasculature and neurovascular coupling. (a) Top: The cerebral vascular system obtained
by plastic injection to the vessels and by dissolving other non-vascular tissues. Bottom: Physiology of
neurovascular coupling shows the biproducts of firing neurons lead to blood flow increase. (b) Neurovascular
coupling of increase inflow and oxygenated red blood cells. (c) Hemodynamics upon neural activation.
Brian N. Pasley and Ralph D. Freeman (2008), Scholarpedia, 3(3):5340
(a)
https://psychcentral.com/lib/what-is-functional-magnetic-resonance-imaging-fmri/
(b)
Hall, J.E., Guyton, A.C., 2011. Guyton and Hall Textbook of Medical
Physiology, Journal of Chemical Information and Modeling.
doi:10.1017/CBO9781107415324.004
(c)

Figure 2-3 Oxy and deoxy hemoglobin change during neuronal activation and the sources of the fNIRS
measurement.
• Neurons consume oxygen (to metabolize the glucose
for energy) when firing.
Neural
activity
O2 metabolism
Cerebral blood
flow
Cerebral blood
volume
HbO
HbR
HbO
HbR
HbO
HbR
HbO
HbR
Δ concentration [Δ uM]
[Kohl 2000]
Time [sec]
HbO
HbR
fNIRS signal
Neural
activity
7

The flow change due to neurovascular coupling, which leads to changes in concentration of oxygenated
and de-oxygenated red blood cells (or Hbo and Hbr concentration) provides the key marker for brain activity.
As the brain activity increases, the Hbo and Hbr typically change in the opposite directions as shown in
Figure 2-3-a, which indicates the overly compensated blood flow that provides enough oxygen supply (rapid
increase in Hbo). At the same time, the momentarily increased de-oxygenated red blood cell content (due
to the oxygen consumption of the neurons) flushes out with the new oxygenated blood supply, resulting in
a slower and smaller amplitude decrease in Hbr after the onset of the neuronal activation (Figure 2-3-b)
(Leff et al., 2011). This activation pattern/template provides the basis for functional brain imaging techniques
such as fNIRS and blood-oxygen-level-dependent (BOLD) fMRI (Cui et al., 2010; Partington and Farmery,
2014).
There are also non-neuronal factors influencing the cerebral blood flow and (thus) cerebral blood volume
(CBV) which are highly correlated with one another. The three significant contributors to cerebral blood flow
are shown in Figure 2-4. Carbon dioxide (CO2) in the cerebral bloodstream is one of the significant
contributors in cerebral blow flow regulation because as the CO2 concentration increases by 70%, cerebral
flow can be doubled by a vasodilation mechanism (Hall and Guyton, 2011). It is believed that the
hypercapnia causes an increase in the acidity that triggers the production of extracellular H+ on vascular

Figure 2-4 Cerebral blood flow due to non-neuronal factors (Partington and Farmery, 2014). Changes in Carbon
dioxide, oxygen, and autonomic nervous system induced blood pressure in brain are major factors. The
autoregulation of the flow under variant cerebral perfusion pressure is done by active vasoaction adjustments.
8

smooth muscle, which causes the vasodilation of the cerebral vessels (Cipolla, 2009). Blood pressure
denoted as CPP in Figure 2-4 is another significant contributor to the flow because, unlike the effect of the
CO2 change, the blood flow is tightly regulated (meaning it stays constant) over quite a wide range of blood
pressure change. This ‘autoregulation’ mechanism helps prevent sudden vascular damages due to a rapid
increase in the blood pressure, which can happen during heavy exercise, for instance. Although the exact
mechanism of this autoregulation is not well understood, many reports showed that the reductions in CBF
could stimulate the release of substances (such as H+, K+, O2) from the brain that causes vasodilation of
the cerebral arterioles (Cipolla, 2009; Partington and Farmery, 2014). It is interesting to note that this
autoregulation process can increase or decrease both the oxy and de-oxy blood counts regardless of the
change in cerebral blood flow, meaning that there can be discrepancies in the correlation between changes
in blood volume and blood flow. This phenomenon, the vessel diameter can change (thus vasodilation or
vasoconstriction occur) without the change of the flow (Figure 2-4), is not due to neurocognitive causes,
which can confound the analysis and lead to false conclusions. This will be discussed more in the next
section.

Figure 2-5 The fNIRS principle uses red light that is sensitive to blood oxygen content. Our body
tissue and bones are relatively transparent to red light. The change of the measured intensity would
depend on oxy- and deoxy- hemoglobin which correlate with brain activity.
Sample at 2hz
2.5cm
separation
(deoxy)
(oxy)
(a)
(c)
(b)
9

2.3. The fNIRS principle, history, and confounding factors
Our scalp and skull are relatively transparent to the red light, which enables us to beam the lights through
them and monitor how much lights are being scattered across the brain cortex and reflected out into the
sensor located on the scalp surface (Figure 2-5-a). The fNIRS utilizes this characteristic of the light. The
fNIRS band or a cap is equipped with light sources and light detectors as shown in Figure 2-5-b. The FNIRS
we used is in a band form factor which has sixteen transmitter-receiver pairs. Each of the light sources
emits two different wavelengths of lights such that one light is more sensitive to the oxygenated blood
content while another light is more sensitive to the de-oxygenated blood content. In more detail, the
attenuated light after traveling through the tissue can be expressed in terms of the amount of concentration
of the oxy- and deoxy- hemoglobin and their associated light extinction coefficients (Figure 2-5-c and
Equation 1). This can be written into an equation using modified Beer’s law, which characterizes the
relationship between light intensity and the medium that absorbs the light (Equation 1).

For one wavelength, 𝜆:𝑂𝐷
𝜆 =𝜀𝜆 𝐻𝑏𝑅 ∙𝐶 𝐻𝑏𝑅 ∙𝑙+ 𝜀𝜆 𝐻𝑏𝑂 ∙𝐶 𝐻𝑏𝑂 ∙𝑙+𝐺 ... (1)
𝐺 : Constant attenuation factor, accounts for the measurement geometry
,where
Beer’s law in liquids: 𝑂𝐷 =𝐴 = 𝜀𝑐𝑙 =𝛼𝑙 =−log(𝑇)=log (
𝐼𝑖𝑛
𝐼𝑜𝑢𝑡 ⁄ )
𝑂𝐷 : Optical density, 𝐴 : Absorbance
𝜀 : extinction coefficient, malar absorptivity of the absorber, [𝐿 ∙𝑚𝑜𝑙 −1
∙𝑐𝑚
−1
]
𝑐 : molar concentration of the absorber, [𝑚𝑜𝑙 ∙𝐿 −1
],
𝑙 : path length, cuvette,
𝛼 : absorption coefficient, absorption factor

,and
𝑇 =
𝐼𝑜𝑢𝑡 𝐼𝑖𝑛
⁄ =𝑒 −𝛼𝑙

𝑇 : Transmissivity, 𝐼 : light intensity

By measuring the light attenuation at two wavelengths at two different time points, we can solve for the
changes in oxygenated and de-oxygenated hemoglobin concentration. This assumes that the other
unknown factors (G in Equation 1) remain constant over the two short time instances. As a result, we can
deduce additional brain hemodynamics such as:
10

Oxygenation= ∆HbO− ∆HbR…(2)
Blood volume= ∆HbO+ ∆HbR….(3)
The fNIRS has a relatively short history compared to other imaging modalities such as PET or MRI. Due
to its advantages in mobility and temporal/spatial resolution over other popular neuroimaging modalities,
the research field has grown exponentially over the past twenty-five years (Figure 2-6). We chose fNIRS
for measuring PFC activity because of the potential for large-scale studies in a daily-life environment and
its cost-effectiveness.
However, it has been only ten years since the fNIRS community started to incrementally and actively
develop signal refinement techniques for fNIRS. For example, the importance of filtering out the non-
neuronal physiological confounds from the fNIRS signal was addressed not too long ago (~10 years). This
is one of the intrinsic problems of the fNIRS technique because the transmitted light has to pass through
the scalp tissue (Figure 2-7-c]). Experimental protocols that require tasks that also trigger sympathetic
activation or changes in breathing pattern can produce substantial changes in the blood flow through the
scalp, which will confound and dominate over the true neuronal signal (F. Scholkmann et al., 2013). A
graphic in Figure 2-7 illustrates the confounding influences in fNIRS. Many make the mistake that the fNIRS
measurements are mostly based on functional activation of the brain seen as a change in cerebral
hemodynamics, but, in reality, these cerebral hemodynamics are also greatly influenced by other factors
(Figure 2-7-b) such as systemic and extra-cerebral regulation of blood flow. One of the significant known
confounding influences is from the scalp blood flow as it contains a vast amount of blood vessel network
as well (Figure 2-7-c).
11

Figure 2-6 Number of fNIRS publications per year. This trend shows the advances of fNIRS and its
applications over 25 years. It is only been around 10 years since more incremental and active development
of signal refinement techniques from other physiological confounding factors.

Yücel MA, Selb JJ, Huppert TJ, Franceschini MA, Boas DA. Functional Near Infrared Spectroscopy: Enabling routine functional brain imaging.
Curr Opin Biomed Eng. 2017;4:78-86. doi:10.1016/J.COBME.2017.09.011.

Figure 2-7 Sources of hemodynamics. (a) An erroneous assumption that the fNIRS measurement only
represents the brain neural activation. (b) There are other confounders that are mixed into the fNIRS signal
due to systemic and extra-cerebral influences. (c) Extra-cerebral layers monitored by fNIRS.
Ilias Tachtsidis, Felix Scholkmann, "False positives and false negatives in functional near-infrared spectroscopy: issues, challenges, and the way forward," 3(3) 031405 (9 March 2016)
12

Changes in peripheral blood flow due to the systemic activation was also found to be a potential
confounder in the study reported by Sunwoo et al. (Sunwoo et al., 2018). In the study, an experimental
thermal stimulus was delievered to the ventral aspect of the right forearm of participants while the subject’s
brain hemodynamics responses were recorded using fNIRS. Recorded signals from one particpant and an
estimation of non-neuronal contirbution in the original fNIRS measurements are illustrated in Figure 2-8. A
set of heat pulses was applied as shown in panel (a) while the peripheral blood flow and the fNIRS were

Figure 2-8 An example of a large contribution from peripheral blood flow in fNIRS measurement. (a) is a
heat stimulus applied and recorded while (b) peripheral blood flow and (d) fNIRS were recorded simultaneously.
There is a clear stimulus-response correlation between (a)-(b) as well as between (a)-(d) where showed
vasoconstriction responses to each of heat pulses. (c) is an estimate of dynamic relationship (i.e. an impulse
response) between the peripheral blood flow and the fNIRS measurement. The shape indicates that a unit
increase in peripheral blood flow also increases the fNIRS measurement after a short delay in time. The pink line
in panel (d) shows an estimated contribution (~50%) from the peripheral blood flow in the fNIRS measurement,
and a blue line in panel (e) is the filtered (residual) fNIRS signal after subtracting the non-neuronal contribution
from the peripheral blood flow. The upper arrows in panel (e) emphasize the amount of peripheral
vasoconstriction responses to pain stimulus that are not originated from the evoked brain activity of the interest.
Note: (b) Photo-plethysmogram amplitude (PPGa) can be a surrogate measure of peripheral blood flow. (d)
Normalized mean square error (NMSE) was converted into % contribution by (1-NMSE)*100%. More details can
be found in section 3.2.
a.u. a.u. ΔuM
ΔuM
(a)
(b) (c)
(d)
(e)
13

recorded simultaneously. A clear stimulus-response correlation indicated a vasocontriction response to
heat pain in both the peripheral blood flow and brain hemodynamics (Figure 2-8-a,b,c,d). The pink line in
panel (d) shows an estimated contribution (~50%) from the peripheral blood flow in the fNIRS measurement,
and a blue line in panel (e) is the filtered fNIRS signal after subtracting the non-neuronal contribution from
the original fNIRS measurement. The upper arrows in Figure 2-8-e emphasize the amount of peripheral
vasoconstriction response to pain stimulus that were not originated from the evoked brain activity of the
interest. A detailed methodology in model-based signal filtering is discussed in the next chapter.

14

2.4. Signal processing and modeling techniques for fNIRS signal refinement

2.4.1 Motion artifact removal and linear filtering
Common signal pre-processing involves motion artifact removal followed by conventional band-pass
filtering to remove unwanted physiological fluctuation. The first step is to examine the raw light
measurements for any motion artifacts that can significantly influence the analysis (Cui et al., 2011, 2010).
Subject movements that uncouple the light sensors and the surface of the scalp can produce signal artifacts
of high-frequency spikes or baseline shifts. Many have proposed the method for removing the spike
patterns: signal variance method (Ayaz et al., 2010), wavelet-based filtering (Dadgostar et al., 2013; Molavi
and Dumont, 2012), oxygenated and de-oxygenated hemoglobin signal correlation based (Cui et al., 2010),
principal component analysis (PCA) and independent component analysis (ICA) signal
extraction/decomposition based (Kohno et al., 2007; Santosa et al., 2013; Virtanen et al., 2009). However,
up to the current date, not many have reported methods that reliably remove abrupt step-shaped baseline
shifts (Terrighena et al., 2017; Virtanen et al., 2011). After the removal of motion artifacts, the light signal
can be converted into concentrations of oxygenated and de-oxygenated hemoglobin using the modified
Beer’s law. For removal of slow fluctuation of the signal baseline, the frequency outside of 0.01~0.1 Hz
range is filtered out using a band-pass filter in order to smooth the signal and to remove cardio-respiratory
and other physiological trends (e.g., changes due to anxiety) (Stefanovska, 2006)(Bumstead et al., 2017).

2.4.2 Filtering out skin (scalp) blood flow influences
Several techniques have been reported for improving fNIRS neuronal content over the signal
contamination originated from scalp blood flow. Reports have revealed that using short-distance source-
detector channels can significantly improve the fNIRS contrast-to-noise ratio and signal-to-noise ratio. The
short-distance channels usually have a source-to-detector distance of less than 2cm, which would only
measure the blood flow in the superficial layers of the head. The short-distance signals were fitted into the
long-distance channels using the least-squares minimization method (Saager et al., 2011), adaptive filtering
and Kalman filtering (Gagnon et al., 2012; Zhang et al., 2009), and modeling method with Kalman filter
15

estimation (Gagnon et al., 2011). It is worth noting that Gagnon et al. have significantly improved the
neuronal component of fNIRS signal by using a model of a linear combination of short-detector channel
(yss), long-distance channel (yls), and neuronal activity-related response (yb), which is shown in Eq. 4. The
superficial contribution (ass) was found using Kalman filtering followed by Rauch-Tung-Striebel smoother,
which enabled an estimate of the neuronal activity yb (Gagnon et al., 2011). However, such utilization of
the short-distance channels was only introduced recently (the late 2000s).

y
LS
[𝑛]= y
b
[𝑛]+a
SS
y
SS
, where … (4)
y
b
[𝑛]= ∑ ℎ[𝑘]𝜇[𝑛 −𝑘]
∞
−∞
, ℎ[𝑛]=∑ 𝑤 𝑖 𝑁 𝑤 𝑖=1
𝑏 𝑖[𝑛], … (5)
a
SS
: Weights for the superficial contribution.
ℎ[𝑛]: Hemodynamic response modeled by a linear combination of normalized Gaussian functions 𝑏 𝑖[𝑛].
𝑁 𝑤 : Number of Gaussian functions.
𝑤 𝑖 : Weights for the Gaussian functions.

PCA and ICA based techniques are also popular choices for removing confounding factors due to scalp
blood flow. These data-driven methods do not use additional measurements of the scalp component. The
methods assume that the influence and the contamination are more global and uniform, which also showed
a significant correlation between identified skin component and the direct skin blood flow measurement
using laser Doppler (Kohno et al., 2007; Virtanen et al., 2009; Zhang et al., 2005). However, depending on
the experiment protocol, the required assumptions for PCA and ICA (orthogonality and statistical
independence between original components) may not hold and may lead to incorrect results (Hyvärinen
and Oja, 2000). Also, due to the blind nature of PCA and ICA methods, it is difficult to evaluate whether the
extracted components are physiologically correct (Diamond et al., 2006). Instead, utilizing other physiologic
and systemic measurements such as systemic blood pressure or heart rate when available would
potentially be more accurate.
16

2.4.3 Hemodynamics models for fNIRS
One variant of the hemodynamics model was based on knowledge of underlying physiology and
biophysics. In its simplest form, many have proposed neurovascular coupling models based on the balloon
model (Buxton et al., 2004, 1998; Caldwell et al., 2016; Cui et al., 2010; Friston et al., 2000) with additional
refinements using the windkessel model (Boas et al., 2003; Mandeville et al., 1999; Olufsen et al., 2002).
Recent efforts (since the late 2000s) have been made in modeling the hemodynamics in more
comprehensive ways by adding chemical, metabolic, and physical interactions happening in the neuronal
signal pathways. One proposed model has input variables of CO2, arterial pressure (Pa), SaO2, and
extracerebral scalp blood flow (flux) all being fed into a complex structure of neurovascular and
neurometabolic coupling and systemic blood flow compartments, which are roughly sketched in Figure 2-9-
a and Figure 2-9-c (Caldwell et al., 2016, 2015). Most importantly, using this model’s various input
combinations allows simulations of ‘false-positive’ and ‘false-negative’ scenarios in which the oxy and deoxy
measurements falsely represent the actual brain state. This is shown in Figure 2-9-b. In these cases, the
changes in Pa, PaCO2, and scalp flow caused false-brain activation (left column), and also contaminated
the brain activation (right). This modeling work strongly suggests how such influences can easily confound
the neuronal activation pattern measure in fNIRS, especially in experimental settings that can trigger
systemic and respiratory changes that can contaminate underlying event-related brain responses. The
authors suggested measuring the scalp flow (for example using laser Doppler) and ETCO2 for fNIRS
studies (Caldwell et al., 2016; F. Scholkmann et al., 2013). However, the complexity of the hemodynamics
model and its a large number of unknown parameters that need to be optimized are one of the limitations
of using the biologically-plausible parametric models (Diamond, 2015; Diamond et al., 2006).
17

In contrast, approaches using non-parametric models are based on a network of simplified input-output
signals and systems paradigm that can provide a meaningful relationship between the measured signals.
The measured signals are differently weighted and combined using various optimization and regression
techniques during the model fitting process. It is advantageous because the model is data-driven, and can
be flexibly re-arranged and used to evaluate complex systems by using only a small number of free
variables (Mitsis and Marmarelis, 2002). Moreover, systemic confounders such as scalp blood flow (if the
measurements are available) can be added as additional regressors (Caldwell et al., 2016).

Figure 2-9 Recently published model of fNIRS hemodynamics based on physiology.
(a)
(b)
(c)
Caldwell M, Scholkmann F, Wolf U, Wolf M, Elwell C, Tachtsidis I. Modelling confounding effects from extracerebral contamination and systemic factors on
functional near-infrared spectroscopy. Neuroimage. 2016;143:91-105. doi:10.1016/j.neuroimage.2016.08.058.
18

Generalized linear modeling (GLM) is one of the most widely used techniques for characterizing neuronal
hemodynamics responses. Linear combinations of various measurements are used, as equations shown
in Eq. (4-6), which are examples of GLM with a regressor. We can also find the application of this modeling
technique in a recent study where the Gaussian basis functions were used to estimate the true neuronal
activation from the skin blood flow confounds (Figure 2-11) (Yücel et al., 2015). Skin blood flow was
measured using short separation channels that have a source-receiver pair spaced about 8 mm in distance.
This type of GLM was first applied in fNIRS signal processing in the year of 2004 followed by many others
(Schroeter et al., 2004)(Tak and Ye, 2014).

y
𝐿𝑜𝑛𝑔 [𝑛]=∑ ℎ[𝑛]𝑢[𝑛 −𝑘]
∞
𝑘=−∞
+∑ 𝑎 𝑖𝑦 𝑆ℎ𝑜𝑟𝑡 [𝑛 +1−𝑖]
𝑁 𝑎 𝑖=1
… (6)
, where h[n] is a sum of weighted basis functions of Gaussians as in Figure 2-11, u[n] is the stimulus
onset vector that is 1 at the stimulus onset and 0 otherwise, N is the number memory.

We can also model the non-linear behavior of biological systems using Volterra expansion technique,
which is illustrated in Figure 2-10. The Volterra expansion technique is widely applied in non-parametric
models (Carassale et al., 2010), and was also used in (Friston et al., 2000) in order to model non-linear
characteristics of brain dynamics, where authors validated whether the balloon model could accommodate
the non-linearity seen in the real data. The signals from the balloon model and 1
st
and 2
nd
order kernels
were compared and showed both methods were able to accommodate the non-linear nature of brain
hemodynamics in similar ways.
Finally, one disadvantage is that we can over-fit the data, meaning that the estimation can be largely
biased to the specific training dataset (that can also consist of irrelevant noise) such that the results do not
represent the general truth. Nevertheless, this problem can be mitigated by penalizing the excess usage of
model orders (Shiavi, 2007) or by using statistical approaches such as the k-fold cross-validation method
(“Cross-validation (statistics),” n.d.).
19

Figure 2-10 Volterra expansion technique is widely used to model non-parametric models. The diagram
and the analogous equation models a system with sum of linear and non-linear input and output relationships.
[1] L. Carassale, M. Asce, A. Kareem, and D. M. Asce, “Modeling Nonlinear
Systems by Volterra Series,” J. E ng. Mech., no. June, pp. 801–818, 2010.

Figure 2-11 An example of using Gaussian profile as the basis functions and their weighted sum to estimate
neuronal hemodynamics. (a) Gaussian basis fitted into (b) hypothetical fNIRS signals due to neuronal activation
(Yucel et al., 2015). There are more different forms of basis functions such as in Fourier basis or Laguerre basis
(c), although they haven’t been used in fNIRS modeling, yet.
Gaussian Laguerre
c
20

2.5. Previous applications of fNIRS and functional experiments in SCD
Many review papers discussing the neurocognitive aspects of SCD stated that a lack of published studies
and interventions made it difficult to address and explain the problem of possible cognitive function and
learning deficit as seen in SCD children (Day and Chismark, 2006; Yerys et al., 2003). The work by Schatz
and Roberts (Schatz and Roberts, 2005) showed that children with SCD showed difficulties in performing
specific memory task; however, the primary outcome measurement was the error rate, which lacked more
compelling evidence. One recent electroencephalography (EEG)-based study by Colombatti and
colleagues (Colombatti et al., 2015) reported that SCD showed delayed and spatially altered activation to
an auditory stimulus. Nonetheless, it is difficult to find other studies that employed functional brain imaging
techniques for assessing cognitive functions in SCD. In order to better understand the underlying
mechanism of cognitive decline in SCD, there needs to be further investment in functional neuroimaging
studies.

21

3. MENTAL TASK INDUCED PREFRONTAL CORTEX (PFC) RESPONSES

This chapter introduces two types of studies that induced mental workload and signal processing method
to identify and filter out non-neuronal influences in the original fNIRS measurement. We show the
improvement of the neuronal component of the fNIRS signal using the mental tasks as a reference. We
anticipated an increased brain response as the task was becoming difficult. And we used this pattern of
brain responses to evaluate the effectiveness of our proposed signal filtering method. Using the filtered
fNIRS signal, our interest was to look at any abnormal brain activities during the “N-back” working memory
task and “Stroop” attention control task in sickle cell disease.

3.1. “N-back” standardized short-term memory and “Stroop” attention control
tasks
The N-back and Stroop tasks were conducted in one day in one subject. They are known to activate
different parts of prefrontal cortex (PFC) (Ayaz et al., 2012; Yanagisawa et al., 2010; Yennu et al., 2016).
The first task was given randomly so that the ordering effect was minimized. State-Trait Anxiety Inventory
(STAI) anxiety questionnaires were also collected at the beginning of the study and in between the two
tasks (Spielberger et al., 1983; Yennu et al., 2016). There were 5min pre- and 3min post-baseline period
where the subject was quietly sat down while the physiological measurements were made. Pain anticipating
was a part of the study where we instructed there would be a pain stimulus but no simulation was given. In
this dissertation, STAI data and pain anticipation periods were not considered because they did not show
Figure 3-1 Mental sequence. Time sequence of the study protocol. Subjects were randomly assigned to have
the N-back or Stroop test. Blocks marked with * were not analyzed in this dissertation. Adapted from [Shah
2019]
*
* *
22

significant influence on the results of this dissertation work. Response time and accuracy rate were also
recorded for each key responses of the subject. The whole study took about an hour to complete. This is
shown in Figure 3-1. More detail procedures can be found Shah et al. (Shah et al., 2019).

3.1.1 N-back
The N-back memory load task can serve as a standardized measure of brain oxygenation and behavioral
effort for processing short-term memory. Studies have shown that neurovascular related complication can
alter the amount of oxygenation during memory tasks. We also hypothesized that the cognitive decline in
SCD could change the amount of oxygenation in the brain during a memory task.

We have designed four levels of N-back tasks (0, 1, 2, and 3 backs). Each n-back task presented a series
of alphabetic letters (500 milliseconds, one letter at a time) followed by 2500 milliseconds of a blank screen.
The subject was instructed to press a button on the keyboard when the presenting letter was the same as
the one shown n-turns ago as in Figure 3-2. For example, 2-back requires memorizing the letter shown 2-
turns ago. There were 4~6 answers in a single n-back task. Each n-back task lasted for 42.5 seconds. Each
level of difficulty was given three times (thus 4 levels x 3 repeats), and they were randomly shuffled with
other levels of n-back tasks. Between all the tasks, there were 25-second break periods that included 5
seconds of score feedback, 10 seconds of relaxing, and 10 seconds of getting ready for the next task.
Previous studies have shown that reaction time increase while accuracy rate drop as the N-back task
time
time
Figure 3-2 N-back memory task. Left: The n-back test requires a subject to remember a sequence of the
letters presented on the screen, one letter at a time (Developed by Wayne Kirchner, 1958.) Right: A subject
sits comfortably in a chair and presses the Enter key to record a response.
23

becomes more difficult as well as a greater increase in oxygenation change as task become more difficult
(Figure 3-3).

3.1.2 Stroop
Stroop test was originally named after John Ridley
Stroop, who first published the effect of verbal
responses under incongruent (or conflicting) conditions
(Stroop, 1935). The test includes naming the color of a
printed word while the word itself has a conflicting
meaning. For example, it takes a longer time to read out
the font color of the words on the top rows in (Figure
3-4) than the bottom rows.
We designed three levels of difficulty (Stroop 1, 2, and 3). The subject was instructed to click an answer
on a computer monitor using a mouse. There are 30 total questions for one block followed by 20 seconds
of a break period which included 5 seconds of score feedback, 10 seconds of relaxing, and 5 seconds of
getting ready for the next stoop block. The difficulty was controlled by the number of choices and time limit
to respond. There were four multiple choice at Stroop 1 and 2.4 seconds of the response time limit. The
number of choices was increased to 6 and 8 and the response time limit was shortened to 2.1 and 1.8
seconds at Stroop 2 and 3, respectively.

Figure 3-4 Stroop word-color incongruent test. The
test includes naming the color of a printed word while
the word itself has a conflicting meaning
Purple Red Purple
Mouse Top Face

Figure 3-3 Previously reported N-back responses. We anticipated the increase in the reaction time while the
decrease in the accuracy rate. In brain, there are greater change in oxygenation as N-back become difficult.
0- 1- 2- 3-back
Oxygenation changes (μM)
0- 1- 2- 3-back
Accuracy rate
Reaction time (ms)
0- 1- 2- 3-back
24

This test also involves controlling behavior because the participant has to overcome the urge to read out
the word itself (Kellogg, 2013). Studies also show that older adults perform the task more quickly than
younger ones, which suggests that cognitive processing becomes more efficient as we age (Demetriou et
al., 2002). This test also activated the dorsal lateral prefrontal cortex, which is also known as the region of
activation during the N-back memory task (Milham et al., 2003). Unlike the N-back memory task, this
intervention is known to induce a fair amount of sympathetic activation, as it influences galvanic skin
response (GSR) and heart rate changes (Hernando-Gallego and Artés-Rodríguez, 2015), which means it
is a stronger mental stressor. This may activate parts of functional areas in the brain; however, there will
be more unwanted influences from the scalp blood flow that can confound the analysis. E-Prime software
was used to design and present the graphical stimuli (Psychology Software Tools, Pittsburgh, PA, USA).

3.1.3 Participants and physiological measurements
The study took place at the Children’s Hospital Los Angeles (CHLA) after the approval of IRB. We
recruited a total of 23 SCD and 18 controls, who are patients received care at CHLA either on chronic
transfusion, hydroxyurea, or not on any treatment, and their age and race-matched close relatives and
friends. All participants had to be older than 11 years old,
free of vaso-occlusive crises or hospitalization in the past
10 days, and free of anxiety disorder. Each subject signed
a written consent before the study. The study was
conducted in the morning. The first hour was spent giving
instructions and preparing for data acquisition prior to
starting the mental task. As shown in Figure 3-5, we
simultaneously recorded their changes in prefrontal cortex
hemodynamics (∆HbO and ∆HbR) using fNIRS, as well as
autonomic measurements including finger blood volume
change using photo-plethysmograph (Nonin Medical Inc.)
and CO2 analyzer through nose cannula (Vacumetrics

Figure 3-5 PFC and peripheral measurements
were made using fNIRS, capnography, and
photo-plethysmography (PPG).
fNIRS
CO2
Photo-plethysmography (PPG)
25

Inc.). We used Biopac MP150 as a main acquisition system. In this dissertation, a total of 15 SCD patients
and 15 controls were analyzed. Other subjects were dropped because of missing signals (N=5), noisy
fNIRS signals (N=3), and missing a marker for the signal synchronization between fNIRS and Biopac (N=3).
Table 1 shows the detailed demographics. A more detailed description of subject demographics and
physiological measurements can be found in Shah et al. (Shah et al., 2019).

Table 1. Participants demographic.
Group Subjects Age Female/male Stroke Treatment
SCD 15 21 ± 6 6/9 4 (silent), 3 (overt) 6 (transfusion),
7 (Hydroxyurea)
Control 15 20 ± 5 8/7 n/a n/a

26

3.2. Input-output dynamic systems modeling for removing non-neuronal
confounders and other pre-processing techniques

3.2.1 Signal pre-processing via linear filtering and motion artifact removal
The measured fNIRS signal was refined using combinations of sliding window motion artifact rejection
(SMAR) (Ayaz et al., 2010), correlation-based signal improvement (CBSI) (Cui et al., 2010), and linear
filtering methods. This was followed by dynamic systems modeling for removing systemic influences as
summarized in Figure 3-6.

SMAR was applied to the fNIRS light measurement for identifying and rejecting saturated channels. After
rejecting saturated channels, two wavelengths of light were converted into ΔHbO and ΔHbR using modified
Beer’s law (Cope and Delpy, 1988). The ΔHbO and ΔHbR signals were refined and combined using CBSI
for removing both spike- and step- types of signal artifacts. The correlation-based signal improvement
(CBSI) method is based on the observation that the oxy- and deoxy- hemoglobin signals are negatively
correlated when there is a neural activity, while the correlation becomes positive when movement noise

Figure 3-6 Signal pre-processing steps. Signal contamination due to movement and non-neuronal influences
in original fNIRS measurements were removed before quantifying as brain responses.
Dynamic systems modeling:
A filter between finger blood flow
and scalp blood flow
Actively ongoing research
in fNIRS community.
Brain
oxygenation
SMAR: Sliding window motion artifact
rejection & CBSI: Correlation-based
signal improvement
Linear filtering for smoothing and
removing fast/slow physiological noise
Removal of systemic influences
Quantification of the brain response
and statistical analysis
End-tidal CO2 (ETCO2),
Photo-plethysmograph
amplitude (PPGa)
fNIRS Peripheral
motion artifacts
27

occurs. With an assumption that the noise is not correlated with the true HbO signal, enhanced HbO signal
can be computed conveniently (Cui et al., 2010). It is advantageous because the abrupt signal spikes as
well as step changes can be removed. We applied this method prior to the modeling step to 1) reduce the
signal dimensionality (from HbO and HbR to CBSI-HbO or oxygenation because CBSI-HbR will be a mirror
image of CBSI-HbO) and 2) remove/correct signal artifacts effectively (Cui et al., 2010). Therefore, in this
dissertation, fNIRS signal of interest is CBSI-HbO as an ‘HbO’ or ‘oxygenation’ response. A bandpass filter
(0.003-0.5Hz, zero-phase finite impulse response low-pass filter followed by moving average removal of
300-second Hanning window) was applied to remove signal drift and noise. All the recorded signals were
processed and analyzed using customized scripts programmed using MATLAB (Mathworks, 2015).

3.2.2 Model-based filtering (MBF) for removing non-brain influences
We identified and subtracted two major physiological confounders in the fNIRS signal by using a linear
dynamic systems modeling approach (Model-based filtering, MBF). The first physiological confounder was
the influence from skin blood flow. We used the amplitude of a finger PPG (PPGa) as a surrogate of the
skin blood flow influence (Chalacheva et al., 2017). The second physiological confounder was the change
in hemodynamics due to fluctuations in CO2 concentration in the brain, resulting from changes in breath-
to-breath ventilation. We used end-tidal CO2 (ETCO2) as the surrogate measure of cerebral CO2
concentration.
Prior to the modeling process, we normalized the PPGa signal to the individual’s 95th percentile because
each subject had different gain in the signal. All signals were downsampled to 2Hz using ‘Resample’
function in MATLAB to avoid using overly sampled time series while preventing signal aliasing. A low-pass
filter with a cut-off at 0.5Hz was applied to PPGa and ETCO2 input signals, and all input and output signals
were demeaned.
28

For each channel, we searched for the best 2-input-1-output system as suggested in Figure 3-7. We used
linear combinations of Laguerre basis functions for estimating impulse responses, which helped minimize
the number of unknown parameters to be found (2-6 basis functions and up to 30 seconds memory)
(Marmarelis, 2013). Laguerre basis function expansion technique had shown its value in modeling many
variants of biological systems, including neuron models (Li et al., 2014). However, the usage of the Laguerre
expansion technique in estimating neurovascular hemodynamics measured in fNIRS has not been
reported. We assumed linearity between the two inputs and the fNIRS output. To prevent overfitting, we
employed message description length (MDL) (Wasserman, n.d.) techniques such that using too many
Laguerre basis functions would be penalized while minimizing the variance of the residual (Nava-Guerra,
2017; Nava-Guerra et al., 2016). As a result, we estimated the amount of CO2 and blood flow influence in
the original fNIRS measurement using the model.
The model-based filtering was done in two steps, meaning that the impulse responses of PPGa and
ETCO2 input (h1_baselne and h2_baseline, respectively, Figure 3-8) were found first using the study
baseline period, then we applied the h1_baseline and h2_baseline for the mental task period by scaling
them so that the PPGa and ETCO2 input signals would provide the best model-predicted fits to fNIRS signal
during the Nback period as well as Stroop periods. We assumed the study baseline period would not contain

Figure 3-7 Two-input linear dynamic modeling for explaining CO2 and Blood flow influences in fNIRS. Impulse
responses (h1 and h2) are found using Laguerre-basis function expansion and least squares methods.
y = ℎ
1
(𝑚 1
)
1
(𝑘 −𝑚 1
)

=1
+ ℎ

(𝑚
)

(𝑘 −𝑚
)

=1
+
ℎ
1
ℎ

1

Neuronal,
𝑦 0 100 200 300 400 500 600 700 800 900 1000 Sec
fNIRS PPGa ETCO2
IR
Photo
detector
0 100 200 300 400 500 600 700 800 900 1000 Sec
0 100 200 300 400 500 600 700 800 900 1000 Sec
CO2,
1
Blood flow (Surrogate),

fNIRS, 𝑦 +Constant mean is removed.
+Low pass filter for smoothing and physiological noise removal (cutoff 0.5hz)
29

evoked brain responses of interest, and this would allow the impulse responses to be more accurate and
helped avoid overfitting the ETCO2 and PPGa input into the brain component of the fNIRS during the
cognitive tasks. The residual signal from this step was saved and analyzed as filtered fNIRS signal (v,
Figure 3-8).

3.2.3 Quantification of fNIRS signal
For numerical quantification of the response, we calculated 1) the slope of the fitted line and 2) the
average (median) changes in oxygenation responses during each trial of N-back or Stroop. We considered
the first 10-second period from each trial onset as transient or unstable and excluded it from our calculation.
The 10-second offset was adapted from the report from Herff et al. which showed that using that offset
achieved the best recognition of different levels of N-back via a machine learning model (Herff et al., 2013).
An example from a subject is plotted in Figure 3-9, which shows the oxygenation change in one fNIRS
channel during different N-back tasks and how the three trials of 0-back periods were quantified as the
slopes of the responses. Finally, each subject had quantified responses in 16 fNIRS channels over PFC x
4 (N-back) or 3 (Stroop) levels of difficulty x 3 (N-back) or 4 (Stroop) trials, totaling 192 responses per
subject.

Figure 3-8 Two-step approach for estimating evoked brain activity.
30

3.3. fNIRS signal improvement after model-based filtering

3.3.1 Non-neuronal contribution in fNIRS measurements
The proposed model-based filtering (MBF) method was applied while normalized mean square error
(NMSE) was computed. The NMSE was converted into an index of non-neuronal influences in original
fNIRS measurement by [(1-NMSE)*100%] and was consolidated. Their median and inter-quartile range
values obtained from all 30 subjects x 16 channels are shown in Figure 3-10. The fNIRS signal contained
up to 70% of non-neuronal components, but in average it showed that 24% of the original fNIRS signal was
explained by the non-neuronal inputs (i.e., ETCO2 and PPGa) during the baseline period (Figure 3-10-a).
The contribution from the PPGa was greater (11%) than that of from the ETCO2 (6%), indicating that the
skin blood flow was generally more influential confounder than the breathing pattern. The systemic
contribution in fNIRS measurement during the tasks (Stroop or N-back) was reduced compared to that of
the baseline period, which indicated newly evoked brain activities due to the mental tasks. The fNIRS

Figure 3-9 Numerical quantification using a slope of fitted line on fNIRS signal measured during a N-back
study. Top: An example of oxygenation changes due to different N-back tasks of one fNIRS channel. Different
levels of N-back are highlighted with their respective levels. Bottom: Illustration of quantifying slopes of fitted
line over the oxygenation change of three trials plotted over time.
3 3 3
2 2 2
1 1 1
0-back 0 0
sec Oxygenation (uM)
Trial1
Δ Oxygenation (uM)
0-back
Trial2
Trial3
Trial1
Trial2
Trial3 Slope (fitted line)
3-back …
31

measurements made during the Stroop contained greater confounding systemic influences than during the
N-back, which indicated the Stroop task might be more demanding than N-back (Figure 3-10-b).

3.3.2 N-back
Initial visual examination of fNIRS measurement to N-back was made to determine the general trends of
fNIRS measurement. We observed that the signal change becomes more positive as the task becomes
difficult, and there was more variability between different locations of PFC (Figure 3-11).

Figure 3-10 Up to 70% fNIRS signal contained non-neuronal components. Each of the input contributions to
fNIRS measurement were computed using normalized mean square error [(1-NMSE)*100%] after the model
estimation (a) The fNIRS signal measured during the baseline period showed that non-neuronal contribution from
peripheral blood flow (measured by PPGa) was greater than that of from breathing pattern (ETCO2). (b,c,d)
Systemic contribution in fNIRS during the tasks (Stroop or N-back) was reduced compared to that of the baseline
period, indicating that there were induced brain activities due to mental tasks. (b,c) fNIRS measurements made
during the Stroop contained greater confounding systemic influences than during the N-back, which indicated the
Stroop task might be more demanding than N-back.
Note: (d) When the two-step modeling approach was not used, (which means the task period was modeled from
scratch,) the model explained more systemic confounding effect from the same two input signals, which may indicate
the two-step modeling approach may help avoiding the over-fitting problem. Several visual comparisons made on
how the input signals were fit into the fNIRS between the case (c, two-step) and (d, TASK only) showed similar
shapes of estimated signal but in a slightly differently magnitude and smoothness (not shown). This indicated that
there were many redundant solutions to solve the model. And the slight difference in fitted results was reflected on
the difference between the percentage of non-neuronal components (7 vs 12%).
32

One example of the optimal fit is shown in Figure 3-12, showing the impulse responses of the two input
signals in the first two rows. The first row, right panel (in pink) shows the contribution of CO2 in the original
fNIRS signal, especially large when there were sigh events. The second row, right panel slows background
fluctuation in blood flow resided in the original fNIRS signal. All the fitted signal (non-neuronal influences)
were subtracted so that they were mitigated in the final residual fNIRS signal. This process was done for
each of the channels of each subject.
One subject’s brain oxygenation (HbO) responses to N-back before and after applying the model-based
filtering is shown in Figure 3-13. One of the 16 channels is plotted over 2-back and 3-back levels. There
are clear responses such as the first trial of the 3-back (in blue line), and it showed the typical hyperemia
response as the mental demand increased. However, some of the HbO dynamics were highly correlated
with the finger blood flow or ETCO2, suggesting the signal contamination (indicated by the arrows in Figure
3-13). After we filter out the CO2 and peripheral blood flow influences from the original fNIRS measurement,
we were able to see the improvements in such that the Hbo signal changes that were coincided with the
finger blood flow were filtered out (indicated by the arrows in the bottom of Figure 3-13). This improvement
in signal trends also improved bar plots such that the average oxygenation as N-back task becomes difficult

Figure 3-11 Visual examination of fNIRS measurement to N-back (a) Example of oxygenation changes in one
fNIRS channel to N-back tasks from one subject showed more positive increase in fNIRS signal to difficult N-back
task. (b) Average oxygenation of all three trials plotted in each channel over time showed some channels were more
responsive than other channels. (c) The summary bar plot for this particular subject for this particular channel
showed an average oxygenation increase as N-back task becomes difficult.
Trial1
Trial2
Trial3
Individual trials
Channel #10
Average of all trials
Channel #1~16
0-back 1-back
2-back 3-back
Oxygenation
Ch16 Ch1
. . .
0 40sec 0 40sec
0 40sec 0 40sec
0-back 1-back
2-back 3-back
(b) (a)
Oxygenation
(c)
0 1 2 3-back
Average of all trials
Channel #10
Oxygenation
33

was more consistently positive and larger in the refined HbO signal (Figure 3-13-right panel and Figure
3-14).

Figure 3-12 Modeling example of optimized N-back impulse responses and the model fit. About 30% of the total
fNIRS signal was contributed from non-neuronal components from breathing (ETCO2) and skin blood flow (PPGa).
fNIRS dynamics
to CO2
fNIRS dynamics
to skin flow
~Neuronal (Residual, )
= fNIRS – CO2 – skin blood flow
3 3 3
0 0 0
1 1 1
2 2 2
ETCO2 fit
Residual,
Linear dynamic modeling
• Structure: Laguerre-Volterra expansion
• Optimization criteria: MDL indices
PPGa fit
Hbo data
Hbo data
Hbo data
Endothelial (NO) -> myogenic (cGMP)
Finger <-> scalp -> fNIRS
(L=5,a=0.6)
(L=4,a=0.65)
Transfer function
Transfer function
Respiration

Figure 3-13 N-back fNIRS signal plotted with finger blood flow and ETCO2, before and after applying the
model-based filtering. The model-based filtering helped made fNIRS signal independent from blood flow and CO2.
0- 1- 2- 3-back
0- 1- 2- 3-back
Before
CO2 (%) Flow
sec
After
Became
independent
from skin blood
flow and CO2
Before filtering
After filtering
More positive
oxygenation
3-back
34

3.3.3 Stroop
Signal improvement seen in the Stroop task was similar to N-back. Since our version of Stroop task can
be demanding and is known to induce more autonomic nervous system responses, we could see that the
1st level of Stroop already showed a greater amount of skin blood flow vasoconstriction (Figure 3-15, third
row) and had a short time to rest between two Stroop questions (see 3.1.2). Therefore, we did not anticipate
any particular trend of brain responses to different levels of Stroop task. Nonetheless, the effect of model-
based filtering was seen such that the correlation from the PPGa and ETCO2 was much less, resulting with
more positive signal mean and slope for all Stroop tasks as shown in Figure 3-15, second panel.

Figure 3-14 Improvement on N-back responses in all subjects. (a →b) CBSI rejected motion artifact effectively
(not shown, see Cui 2010 for details) but did not influence the signal shape adversely. (b →c) The proposed model-
based filtering helped to reveal more positive and greater brain oxygenation to difficult N-back tasks, although the
effect was not as visually strong as in Stroop (See Figure 3-15).
MBF
ΔOxygenation ΔOxygenation
0-back 1-back 2-back 3-back
ΔOxygenation
CBSI
Peripheral
blood flow
(c) MBF(CBSI(Raw)))
(a) Raw
(b) CBSI(Raw)
(d)
35

Figure 3-15 Improvement on Stroop responses in all subjects. More positive and greater brain oxygenation to all
Stroop tasks (each column) are revealed after the model-based filtering method. First row: plot of all available (16
channels x 4 trials) fNIRS signal for each Stroop difficulty. Second row: average of the first row. Third & fourth row:
average PPGa and ETCO2 responses during each Stroop, showing strong correlation with original fNIRS
measurement.
Greater vasoconstriction
than in Nback
Greater response improvement than in Nback
CBSI(HbO)
→ AFTER MBF(CBSI(HBO&HBR))
BEFORE
PPGa
Stroop level 1 (1-S) 2-S 3-S
36

3.4. Brain responses to mental tasks in SCD vs normal control subjects
Quantified responses of the signal slope and average change (introduced in 3.2.3) were not necessarily
correlated all the time. Further examination suggested that the signal slope responses were more
consistently showing increasing trends of brain oxygenation for both N-back and Stroop tasks. Therefore,
we decided to use the slope of the fitted line as the main metric that represented the brain oxygenation
response to mental tasks.
Statistical analyses were made using a mixed model approach, as seen in Table 2. We tested for the
effect of disease (Diagnosis, Dx), task difficulty (level), and their interaction for different PFC locations (in 8
octants, 4 quadrants, or medial/lateral PFC, Figure 3-16). Normality assumption was checked by examining
QQ-plots and the residual of the model for each level of task difficulties and diagnosis groups. Other
possible interactions and covariates (Age and Sex) were also tested. If no statistically significant
effects/interactions were found, the covariates were excluded from the final model. A post-hoc comparison
was conducted as needed with Tukey’s, Bonferroni, or FDR correction, while accounting for effects and
interaction of the covariates.

Table 2. Two-way repeated measures mixed model
Responses (Y) Within-subject factors* Between-subject factors
Co-
variates
1. Slope of fitted line
2. Average signal change
• Task difficulty (level)
• PFC location (-group in 8 octants,
4 quadrants, or lateral/medial PFC)
• Diagnosis (SCD, SCD
with/without stroke)
• Sex
• Age
• Hb
level**
1. Accuracy rate
2. Response time
Task difficulty (level)
* Each factor and associated interactions were set as ‘random’ effects to account for the repeated
measures within a subject.
** Effect of hemoglobin level ([g/dl]) as potential confounder/interaction was tested separately by being
added to the existing model found in the subsequent sections (See Section 3.4.3).

37

3.4.1 N-back – Hyper-/hypo-responses and slower processing time in SCD
In the mixed model of testing for brain oxygenation, age and sex were dropped from the final model
because they did not show significant effect nor significant interaction with other effects. The overall
response of the brain oxygenation to N-back showed an increasing trend as the N-back became more
difficult (Figure 3-17, left panel). But the signal contrast between the easy (0-back or 1-back) and difficult
(2-back or 3-back) tasks were not statistically significant, which might indicate that the signal to noise ratio
in fNIRS as well as from the design of the study protocol were not so convincing. This may have improved
if there were more numbers of trials for the same N-back task. Similarly, there was no significant difference
between the response in SCD compared with normal control subjects, which is shown in Figure 3-17. Post-

Figure 3-17 Brain oxygenation was increased as N-back task became difficult (a). (b) statistical comparison
between SCD and control subjects did not show strong contrast, leaving only suggestive indication of hyper-
activation to 2-0back in SCD. (c) Hyper-activation in SCD became more stronger when patients with stroke
history were taken out. Each N-back was normalized by N-0back.
P-values shown are before adjusting for multiple comparisons
Slope
P=0.10
0 1 2 3 (-back)
All subjects (SCD + Controls),
all channels
(a) (b) (c)
Slope
SCD vs Controls,
all channels
1 2 3 (-0back)
Slope
SCD(No Stroke) vs Controls,
all channels
1 2 3 (-0back)
P=0.05

Figure 3-16 Different PFC location groups used for statistical testing. Original total of 16 fNIRS channels
could be grouped into 8 octants, 4 quadrants, or medial/lateral PFC.
R L R L Quad4 Quad3 Quad2 Quad1
Medial Lateral Lateral
38

hoc analyses did not find any PFC location that showed significant contrast between SCD and controls, but
with a marginal indication of hyper-activation to 2-0back in octant 2, 5, 6, and 8, suggesting that additional
recruitment was needed in SCD than controls when engaged in the same task (Figure 3-18). The same
trend of hyper-responses in SCD became more consistent when patients with stroke histories were
removed (Figure 3-18-c). Note that octant 2 is spatially close to left inferior frontal gyrus region, which was
previously reported as being sensitive to N-back task via PET and fMRI (Badre and Wagner, 2007; Owen

Figure 3-18 The strongest hyper-activation to N-back was in PFC location octant 2. Although the contrast
is not statistically significant after correcting for multiple comparisons, this location is close to left inferior frontal
gyrus, region which was previously reported as being sensitive to N-back task via PET and fMRI, suggesting that
SCD had to use more brain resources to complete the N-back task.
P=0.058
Octa6
1 2 3-0 back
Slope
P=0.015
+Marginal:
-Octa5,2-0back:P0.095
-Octa8,2-0back:P0.094
Octa2
P-values shown are before adjusting for multiple comparisons
+
+
1 2 3-0 back

Figure 3-19 Marginally slower response time to N-back in SCD than controls. Meanwhile, the correction rate
was not different between the two groups, indicating more and longer brain work was needed in SCD to complete
a task as competently as normal control subjects.
P-values are before adjusting for multiple comparisons
Nback effect: P<0.0001
P = 0.09
Response time (ms)
P=0.04
Nback effect: P<0.0001
• No difference in correction rate
Correction rate (%)
39

et al., 2005; Smith and Schatz, 2016). This may indicate that more processing power was needed in the
PFC region responsible for N-back tasks in SCD.
Both the task accuracy and response time showed significant change due to N-back difficulty levels for
all subjects (P <0.0001, Figure 3-19). Post hoc comparison showed SCD showed longer response time
than controls especially in 3-back (p= 0.04 without correcting for multiple comparisons), while there was no
difference in correction rate between the two groups.
When stroke was added as a factor, a strong interaction was found between stroke, level, and PFC
locations. Post hoc analysis found that quadrant 2 indicated that SCD with stroke history (of either overt
and silent stroke) exhibited significantly reduced PFC response to 3-0back and significantly longer
processing time than SCD without a history of strokes (Figure 3-20). Again, the task accuracy was not
different between the two groups. This indicated that SCD with strokes was able to keep up until 2-back
task with an assistance of hyper-activity in medial PFC (which is not the lateral PFC as we expect to be the

Figure 3-20 Stroke SCD patients showed significantly reduced PFC responses to 3-0back and longer
processing time than other SCD patients.
Q1 Q2 Q3 Q4
+ No difference in correction rate (P=0.96)
Nback effect: P<0.0001
P = 0.01
Response time (ms)
P=0.03
P=0.03
Q2
Slope
1 2 3 (-0back)
P=0.006*
• Suppressed PFC activity in SCD:Stroke
P-values shown are before adjusting for multiple comparisons
medial lateral lateral
Nback
* Significant at α=0.1 considering Bonferroni correction
for 12 comparisons (4quads x 3levels, < 0.008 = 0.1/12)
• Other noticeable locations at 3-0back
(Octant3&4,P=0.009; MedialPFC,P=0.018)
40

primary region for N-back) but failed to recruit the medial PFC resources to 3-back as was in SCD without
stroke. This was also supported by negative and significantly decreased signal change responses to all N-
backs in SCD with stroke histories compared to SCD without stroke histories (not shown, P=0.02).
Meanwhile, interestingly, the brain activity was higher in SCD with stroke histories than normal controls at
PFC octant 2, which is known to be sensitive to N-back tasks (Figure 3-21). No other PFC regions showed
significantly different hyper- or hypo-activation patterns between the groups (of SCD with stroke, SCD
without stroke, and normal control subjects).

Figure 3-21 Stroke as a factor linked to N-back difficulty levels and different (left ventrolateral or medial)
PFC locations. SCD with stroke was able to keep up until 2-back task with an assistance of hyper-activity in medial
PFC but failed to recruit the medial PFC resources to 3-back as was in SCD without stroke. Meanwhile, interestingly,
the brain activity was higher in SCD with stroke at PFC octant 2, where is known to be sensitive to N-back tasks.
• SCD with Stroke hypo-activate Octa3,4
(similar in Quad2, or MedialPFC)
++ longer response time
• SCD hyper-activate Octa2
+ longer response time
2back 3back 1back
Nback
(2-back)
Nback
(3-back)
• No differences
SCD
(P=0.015)
Stroke+SCD
(P=0.009)
All PFC locations
P-values shown are before adjusting for multiple comparisons
41

3.4.2 Stroop – Slower processing time as seen in N-back
In Stroop, the main metric used was the slope of the fitted line as shown in Figure 3-22-a. Similar to N-
back, the contrast between the average PFC responses of any two Stroop levels were not strong enough
to be statistically meaningful. Average signal change responses were not used because they did not show
a monotonic increase as the Stroop difficulty increased and there was no indication of different brain
activation using the metric.

Further comparison between SCD and normal control subjects did not reveal any statistically different
PFC responses. Task accuracy across all groups was not significantly different. However, consistently
longer processing time was found (Figure 3-22-b,c and Figure 3-23), indicating that SCD showed a degree
of inefficiency in completing the Stroop task, which is in line with the results seen in the N-back. Stroop
level 3 responses were not very different between the disease group, and normal control group while other
Stroop difficulty levels showed some group contrast. This might indicate that our version of the Stroop test
was too difficult a challenge, and failed to reveal the differences in responses between the groups. It is also

Figure 3-22 Brain oxygenation response to Stroop. (a) Not very strong but monotonic increase as Stroop
became difficult. (b) There were no statistical differences found between SCD (or SCD without stroke) vs normal
controls. The fact that the Stroop level 3 responses were more similar than 1 between the disease group and
normal control group could mean that our version of Stroop task may have be too difficult to differentiate the
responses between the groups, and there was possible signal contamination from systemic influences that was
not able to be accounted for using the proposed fNIRS signal refining method.
Slope
Slope
P-values shown are before adjusting for multiple comparisons
All subjects (SCD + Controls),
all channels
(a) (b)
(c)
SCD vs Controls,
all channels
SCD(No Stroke) vs Controls,
all channels
P=0.07
Slope
P=0.19
42

possible that strong systemic influence had contaminated the fNIRS measurement to a degree that the
systemic influence was not fully removed by our proposed signal refining method.

The comparison between SCD with and without stroke conditions did not show differences in response
time or accuracy, which again could be evidence that our Stroop task was too difficult for both subject
groups, therefore making it difficult for differential brain activity responses to be detected between the
groups (Figure 3-24). When multiple comparisons were not considered, there was a consistent trend of
hyper-activation of PFC to Stroop level 1 in SCD with stroke than SCD without stroke condition. However,
it is difficult to explain why having a stroke condition required more brain recruitment in Stroop, while the
opposite was seen in N-back. Although the neurobiology of our Stroop design is not clear, it is based on
the attention control, which is different than a N-back working memory task; therefore, it is possible that
SCD with stroke needs to recruit more brain sources to achieve the same Stroop task. This difference
became less apparent for Stroop 2 and 3 because the tasks were too difficult and induced the maximum
level of brain activity which cannot be differentiated between groups.

Figure 3-23 Significantly slower processing time to Stroop tasks in SCD as in N-back.
Age was significant factor (p=0.04, order: faster), interact with stroop level(p=0.03)
Stroop effect: P=0.0001
P = 0.019
P = 0.05
Longer processing time
in SCD, as in Nback
Response time (ms)
Age was significant factor(p=0.008), interact with stroop(p=0.005)
Older more accurate. No other interactions in Gender.
Stroop effect: P<0.0001
• No difference in correction rate
Correction rate (%)
P-values shown are before adjusting for multiple comparisons
43

3.4.3 Hemoglobin level as a potential confounding/interaction factor enhanced the
contrast between stroke vs no-stroke in SCD
Hemoglobin (Hb) level may be an important factor to adjust when comparing the PFC responses within
SCD because the brain hemodynamics compensatory mechanism triggered by individuals’ low or high
levels of hemoglobin can affect fNIRS measurements. It is not reasonable, though, to make adjustments to
Hb level when comparing the brain responses between normal controls and SCD because there should a
high correlation between the Hb level and diagnosis (Dx, Table 2) in that case, which results in an unstable
model. As seen in Figure 3-25, there was a significant correlation/difference between the normal vs SCD
but not in SCD without stroke (SCDsx0) and with stroke conditions (SCDsx1).

Figure 3-24 No different PFC responses nor response time to Stroop in SCD+stroke. Correction rate was
not different. The only weak indicative hyperactivation to Stroop level 1 in SCD with stroke was seen in most of
the PFC locations.
P=0.03
• Trending in SCD:Stroke
• Hyper PFC activity, and longer processing
• Octa5 (P<0.05), Octa1,4,6,7,8(P<0.1)
P-values shown are before adjusting for multiple comparisons
Slope
1 2 3Stroop
• No difference in correction rate (P=0.4)
Stroop effect: P<0.0003
Response time (ms)
Age interacted with Stroop level(p=0.06); No other Age, Gender significance/interaction
P=0.07
But neuro-biology of our Stroop is not still clear
P=0.16
44

As a result of this adjustment, we found more significantly different brain responses to 3-back between
stroke and non-stroke SCD patients (Figure 3-26-b). The improved contrast was shown in the medial
regions of PFC (Figure 3-26-d) and supported the finding from the previous section (See 3.4.1). Further
analysis showed a significant interaction between Hb level and brain responses to N-back level (nLevel,
Figure 3-26-c) which showed a negative relationship to 3-back, meaning that an increased Hb level caused
less brain blood flow for a given task. It is possible that marginally greater and/or smaller Hb levels in SCD
without stroke and with stroke, respectively (Figure 3-25), enhanced the separation of the brain oxygenation
response.

Figure 3-25 Distribution of hemoglobin (Hb) level across the groups of normal control, SCD without
stroke (SCDsx0), and SCD with stroke (SCDsx1). An ANOVA test followed by post-hoc showed Hb levels
between SCD without stroke and with stroke were not different and they could be accounted for in the
statistical model.
The distribution of the Control group was not normal. P-values are from Wilcoxon tests,
but parametric tests followed by Tukey-Kramer post-hoc test also showed similar P-values.
M = Major stroke, S = Silent stroke
R
2
=0.59
P < 0.0001
P < 0.0001
P = 0.0032
P = 0.0002
P < 0.09
45

In Stroop, also between stroke SCD vs non-stroke SCD patients, adding the Hb level as an additional
effect did not produce any significant differences in brain oxygenation responses. However, interestingly,
statistically different brain responses were found between all SCD subjects vs normal control subjects. In
the previous section when Hb level was not considered, SCD vs controls responses to Stroop were not
differentiated (see 3.4.2, Figure 3-27). While including the Hb level as another independent variable to
diagnosis (Dx) is risky because the two variables have a strong correlation (Figure 3-25), it is worth noting
for future exploration. The effect of Hb alone was significant (P=0.0012, Figure 3-27-c) without other
significant interactions. This caused further separation of the trends between SCD and controls as
illustrated in Figure 3-27-a&b. As a result, PFC location quad 3 (as well as at the medial PFC group, not
shown) showed significantly decreased brain oxygenation to Stroop level 2 in SCD than normal controls.

Figure 3-26 Enhanced contrast in brain responses to 3-back between SCD+stroke vs SCD. (a) shows the
comparison of all PFC locations and (b) shows pronounced separation in PFC responses to 3-back due to (c)
significant interaction between Hb level and brain responses to N-back (nLevel, especially to 3-back where the slope
is negative). (d) PFC location quad 2 (as well as medial PFC, not shown) showed significant decreased brain
oxygenation to 3-back in SCD with stroke than SCD without stroke condition.
P=0.1
P=0.002* P=0.02 P=0.062
Error bar: 95% confidence interval of the
mean
Adjusted for
Hb level
P=0.067 P=0.018
P-values shown are before
adjusting for multiple comparisons
Quad4 Quad3
Quad2 Quad1
All locations All locations
Before adjusted for
Hb level (Figure 3.19)
* Significant at α=0.05 considering Bonferroni correction for 12
comparisons (4quads x 3levels, < 0.004 = 0.05/12)
• Other significant location at 3-0back: MedialPFC,P=0.005
before adjusting for multiple comparisons
• Hb,P=0.9; Hb*nLevel,P=0.05
Brain oxygenation (Slope)
Brain oxygenation (Slope)
Brain oxygenation (Slope)
Significant: Hb*nLevel
(c) (a) (b)
(d)
46

Figure 3-27 Significant effect of Hb level in the SCD vs normal controls model helped to reveal significantly
decreased brain oxygenation to Stroop (level 2) at the medial PFC regions. (a) shows the comparison of all
PFC locations from the previous model where Hb level was not considered as a factor and (b) shows pronounced
separation in PFC responses to Stroop (level 2) due to (c) significant relationship between the Hb level and brain
responses to Stroop. (d) PFC location quad 3 (as well as medial PFC, not shown) showed significant decreased
brain oxygenation to Stroop level 2 in SCD than normal controls.
P=0.029
Error bar: 95% confidence interval of the
mean
Adjusted for
Hb level
P=0.009*
P-values shown are before
adjusting for multiple comparisons
Quad4 Quad3
Quad2 Quad1
All locations All locations
Before adjusted for
Hb level (Figure 3.21)
* Significant at α=0.05 considering Bonferroni correction for 12
comparisons (4quads x 3levels, < 0.004 = 0.05/12)
• Other significant locations to Stroop level 2: MedialPFC,P=0.005,
LateralPFC,P=0.02 before adjusting for multiple comparisons
• Dx,P = 0.840 (before) → 0.025 (after adjusting for Hb level)
Brain oxygenation (Slope)
Brain oxygenation (Slope)
(c) (a) (b)
(d)
Brain oxygenation (Slope)
P=0.042
P=0.011
P=0.0036*
P=0.025
Significant: Hb,P=0.0012
(No significant interactions)
47

3.5. Discussion
We have applied fNIRS in SCD while subjects were engaged in two mental tasks. Overall PFC responses
showed a monotonic increase as the task became difficult, and this was seen consistently in the slope of
the fitted line as quantified brain response. Although the contrast of the brain responses between different
tasks or diagnostic groups were not significant, we found several indications that in SCD and SCD with
stroke, the brain activity was abnormally decreased or hyper-activated and response time was longer, while
the task accuracy was the same compared to the normal control group. Slower processing time seen in our
study is also in line with other previous studies in SCD (Smith and Schatz, 2016). This strongly suggests
an inefficiency in PFC processes and functions, which is one possible explanation for the cognitive decline
seen in SCD.
N-back test is one of the standardized protocols for assessing mental workload and has been used to
show a correlation between the brain oxygenation and mental workload (Ayaz et al., 2012). Inefficient PFC
functions in SCD with stroke were found using the N-back task by showing statistically significant hypo-
activation in medial PFC (quadrant 2, using signal slope as a response, Figure 3-20) with longer processing
time required to complete the same task than SCD without stroke condition. This was also supported by
the negative and significantly decreased signal change to all N-backs in SCD with stroke compared to SCD
without stroke. This finding suggests an alteration of PFC brain function due to stroke. Similar reports were
found in amyotrophic lateral sclerosis (ALS) and bipolar disorder, where the N-back assessment revealed
profoundly attenuated oxygenation to difficult N-back tasks at medial PFC locations, suggesting prefrontal
neuronal loss (Jogia et al., 2012; Kuruvilla et al., 2013). This is also in line with the decreased white matter
volume found in Choi et al. (Choi et al., 2017) because the white matter acts as a signal conduit between
different cognitive centers across the brain, which can also affect the processing power and speed when a
subject is engaged in an N-back task.
Furthermore, indicative hyper-activation of the left lateral PFC (octant 2) to 2 and 3back tasks (SCD
without stroke) or 2back (SCD with stroke) showed additional recruitment was made at the PFC region that
is known to be sensitive to N-back task (Ayaz et al., 2012; Badre and Wagner, 2007). Combined with the
fact that additional recruitment was also seen at medial PFC in SCD without stroke, while SCD with stroke
48

failed to recruit the same medial PFC, it is reasonable to say SCD recruited wider PFC regions (lateral +
medial PFC) and stronger activity to complete the same task than the controls. However, in 3-back, which
was the most difficult task in our study, SCD with stroke could not continue with such recruitment added at
the medial PFC, showing significantly decreased brain responses compared to SCD without stroke. This
was more evident when each individual’s hemoglobin level was taken into account in the model. SCD
required to put more brain power for a longer time than the normal controls while having a stroke condition
made it difficult to utilize the brain at the most difficult task, resulted in an even longer time than others to
complete the same task as in normal control and SCD without stroke subjects.
There was no statistical difference in PFC responses to Stroop but the SCD subjects showed consistently
longer processing time, while the correction rates were not significantly different. Only indicative PFC
responses were hyper-activation patterns to Stroop level 1 in SCD with stroke compare to SCD without
stroke. Such elevated brain activation and longer processing time to Stroop seen in SCD can also indicate
a degree of inefficiency in line with the results seen in N-back. However, results suggest that our version of
Stroop was too difficult to differentiate the responses between the groups, and there was a possibility that
strong systemic influences had contaminated the fNIRS measurement to the degree that our proposed
signal refining method was less effective. Accounting for the subject’s hemoglobin level resulted in showing
significantly decreased brain oxygenation responses to Stroop; however, this result requires further
validation. This is because the hemoglobin level was significantly correlated with another independent
variable in the model (i.e., diagnosis) which could result in an instability of the model and lead to a biased
conclusion.
There are a few limitations of this study. The responses of fNIRS signal to the cognitive tasks were not
strong because there were no statistical differences seen between the easiest tasks and the most difficult
tasks for both N-back and Stroop studies. This may be due to our small sample size. A total of 30 subjects
may not be enough to see clear brain responses. It may have improved if there were more trials for the
same N-back or Stroop task. In terms of the model-based filtering, there could be more factors to be
considered. We have used a surrogate measure of the peripheral/skin blood flow and ETCO2, but other
systemic influences such as blood pressure and arterial O2 saturation can also “contaminate” the fNIRS
49

signal (Caldwell et al., 2016). Also, the model-based filtering would have been more effective if we could
have recorded and used the blood flow measurement from the scalp instead of from the finger, since it has
been shown that peripheral blood flow is not always global, and peripheral blood flow measurements close
to the site of fNIRS sensors are more correlated with the corresponding fNIRS measurements (Gagnon et
al., 2014, 2012). It may also be plausible to use a non-linear model with 2
nd
order kernels/impulse responses
because we cannot guarantee the stationary of the (body) system due to brain activation, especially if the
task is highly demanding (as in our version of Stroop test). Using more complex model can fit the data
better, but this approach should be applied with caution because it can result in overfitting the data. The
overfitting can result in the removal of the brain component we are interested in. Finally, it seems that the
average signal change as the metric of brain response was not very useful in Stroop because the overall
trend did not show a monotonic increase as the Stroop became more difficult. This can be seen in two
ways. First, compared to the slope response, average signal change response is sensitive to many factors,
such as systemic influences and possible signal drift and motion artifact that generally occurs in fNIRS.
This resulted in false negative and positive responses when the fNIRS signal shifted its signal baseline
abruptly at the beginning part of the trial block. For this reason, we believe it is more reasonable and reliable
to use a slope of a fitted line after the 10 seconds from the trial onset, which can capture the direction of
the brain activation over a course of time (Herff et al., 2013). Secondly, we do not yet understand the
physiology underlying our version of Stroop test which was too demanding; therefore, further studies are
warranted.

50

4. CONCLUSIONS AND SUGGESTIONS FOR FUTURE WORK

4.1. Summary and significance of new findings
Based on fNIRS measurements, we found signs of possible abnormal cognitive activities in SCD,
particularly in patients who had a history of stroke (Figure 4-1). To the best of our knowledge, this study is
the first in which fNIRS has been applied to SCD patients to assess neurocognitive responses. Our fNIRS
findings are supported by our observations of consistently slower response time during mental tasks (both
in N-back and Stroop) in SCD than healthy controls, while the task accuracy was not different between the
groups. Our results suggest that SCD subjects with previous history of stroke were not able to work with
their full brain resources and, therefore, had to work for a longer time with inefficiency. This may also be
linked to the white matter loss reported in Choi et al. (Choi et al., 2017) because the white matter acts as a
signal conduit between different cognitive centers across the brain, and this can also affect the processing
speed when a subject is engaged in a task.

Figure 4-1 Stroke SCD patients showed significantly decreased PFC responses to 3-0back and longer
processing time than other SCD patients.
P=0.002*
Q1 Q2 Q3 Q4
+ No difference in correction rate (P=0.96)
Nback effect: P<0.0001
P = 0.01
Response time (ms)
P=0.03
P=0.03
Q2
Slope
1 2 3 (-0back)
• Decreased PFC activity in SCD:Stroke
P-values shown are before adjusting for multiple comparisons
medial lateral lateral
Nback
* Significant at α=0.05 considering Bonferroni correction for
12 comparisons (4quads x 3levels, < 0.004 = 0.05/12)
• Other significant locations at 3-0back:
MedialPFC,P=0.005
Error bar: 95% confidence interval of the mean
51

The challenges we encountered in this study included the need for careful signal processing for removal
of non-neuronal physiological influences that can significantly confound the results. First, unwanted motion
artifact was corrected by correlation-base signal improvement (CBSI) technique which greatly improved
signal quality. This was followed by model-based filtering where we found dynamic relationships between
the non-brain influences and fNIRS measurement, such that we could subtract the non-brain components
from the original fNIRS measurement. As a result, we demonstrated an augmentation of cognitive-task-
induced fNIRS responses after applying the proposed model-based filtering on the original fNIRS
measurements.

4.2. Main contribution
We used finger blood flow (inferred from PPGa) as a surrogate measure of scalp blood flow that can
greatly influence the original fNIRS measurement. PPGa, along with ETCO2, were used as inputs to a
dynamic systems model. This was advantageous in two ways. First, any discrepancy between the finger
blood flow (measured) and scalp blood (actual information we wanted but not available) would be mitigated
by the impulse response dynamics that we optimized for the model. Second, the use of physiological
measurements helped avoid making blind assumptions that other suggested methods such as PCA and
ICA. Finger PPGa was a reasonable surrogate of scalp blood flow, and ETCO2 which reflected the CO2
concentration in the brain were reported to be a critical confounder in fNIRS signal (Felix Scholkmann et
al., 2013; Yücel et al., 2015). The PPGa showed good correlation with a scalp blood volume measured
using a short-distanced PPG on the forehead (r>0.5, shown in qualifying exam proposal). Another
contribution related to modeling was the use of two-step modeling. First, the dynamic relationship (i.e.,
impulse responses) between the non-neuronal confounders and the fNIRS measurement was found using
the quiet baseline period where we assumed there was no task-related brain activity. Then these baseline
dynamics were subsequently used to find the neuronal component during the mental task period, averting
the problem of overfitting and underestimating the original brain component of interest.

52

4.3. Limitations
One limitation which also served as the main motivation of the research was that the fNIRS device used
for this dissertation did not have short separation channels that could directly measure the scalp blood flow.
If we had such direct blood flow measurement of the scalp, a more precise estimate for the scalp influence
confounding the brain component in the original fNIRS measurement would be possible. One solution is to
use the laser Doppler flowmetry on the forehead; however, the laser Doppler sensor we used only
measured the blood flow in a subcutaneous layer, which might not be enough to represent the blood flow
in the deeper scalp tissue. Furthermore, it is difficult to place additional sensors around the fNIRS band as
it covers most of the forehead.
Moreover, validating or assessing the performance of our model was difficult and limited, partially due to
the lack of short-separation fNIRS channels and other reasons, as 1) we did not have isolated
measurements of the brain activity alone (via fMRI or microelectrodes) so that we could compare our results
with and 2) there was no established method that we could compare with.
The small sample size was another limitation. This not only includes the number of subjects who
participated in the study but also the number of trials made for the same difficulty of N-back and Stroop
tasks. There were three repeated trials and four repeated trials in N-back and Stroop, respectively, which
was comparably a lower number than other similar studies (e.g., more than 7 repeats and 10 repeats of
each difficulty of N-back in Ayaz et al. and Herff et al, respectively (Ayaz et al., 2012; Herff et al., 2013).).
One previous report showed that utilizing both HbO and HbR signals was beneficial compared to using
only one of them (see Ch 2) (Herff et al., 2013). One of the benefits of fNIRS compared to fMRI is that it
can measure both the HbO and HbR activity and therefore much richer information than BOLD-fMRI (mostly
dependent on HbR). In this dissertation, our primary signal of interest was CBSI-HbO, which was computed
using both original HbO and HbR signal. However, CBSI technique has its own limitations that may
contaminate or dilute the original brain activity of interest.
There are newer versions of fNIRS optode setups that come in different form factors, such as a cap that
can snuggly fit onto the head. Some of the newer fNIRS are designed to avoid or mitigate the problem of
53

hair interference by using the cap-type fNIRS sensor with an opening so that the hair under the optical
sensors can be brushed away using a small stick. Similarly, initial and periodic training of the experimental
team in operating fNIRS sensors should be considered for obtaining good quality signals. Other limitations
have been discussed at the end of Chapter 3.

4.4. Suggested future work
As future work, we can re-investigate the mental and pain study using a new fNIRS device 1) with the short-
separation capability, 2) with a wider measuring area over the head, and 3) with a larger sample size and
repeated trials as mentioned in the Limitations section.
It is also important to further validate the model-based filtering. The best way is to repeat several studies
done here but with isolated measurements of the brain activity via fMRI or other invasive recordings (e.g.,
using microelectrodes), so that the refined fNIRS signal after the application of model-based filtering method
can be compared to the isolated brain activity measurement that is free of non-brain components.
Finally, in context of neuroscience, it will also be meaningful to investigate the functional role of different
parts of PFC (e.g., medial vs lateral, ventral vs dorsal PFC) in relation to sickle cell disease and the mental
studies done in this dissertation. What was the significance of seeing the decreased PFC activity in the
medial part of the PFC, as opposed to other parts of PFC? This is an ongoing question. Also, investigation
of functional connectivity of the PFC is another potentially interesting direction, particularly since fNIRS
generally has better time resolution than fMRI and measurements can be made in parallel using multiple
channels. On the other hand, the use of simultaneous multiple channel measurements brings about other
issues, such as the statistical treatment of multiple comparisons – therefore, the application of more
sophisticated but robust statistical tests deserves careful consideration as well.

54

EPILOGUE
Among many emerging and popular functional brain imaging techniques, fNIRS has a potential in clinical
research where the need for portable, large-scale, and daily-life-situation monitoring is required. Recent
studies have also shown applications to human-human and human-computer interaction, which are difficult
to apply using other popular functional brain imaging techniques such as fMRI. Along with its cost-
effectiveness, ease-of-use, and ongoing efforts in miniaturization, fNIRS can aid future research trying to
understand the possible alterations in brain functions in different types of disease in various environmental
contexts.

55

REFERENCES
Ayaz, H., Izzetoglu, M., Shewokis, P. a., Onaral, B., 2010. Sliding-window motion artifact rejection for
Functional Near-Infrared Spectroscopy. Conf. Proc. ... Annu. Int. Conf. IEEE Eng. Med. Biol. Soc.
IEEE Eng. Med. Biol. Soc. Annu. Conf. 2010, 6567–70. doi:10.1109/IEMBS.2010.5627113
Ayaz, H., Shewokis, P.A., Bunce, S., Izzetoglu, K., Willems, B., Onaral, B., 2012. Optical brain monitoring
for operator training and mental workload assessment. Neuroimage 59, 36–47.
doi:10.1016/j.neuroimage.2011.06.023
Badre, D., Wagner, A.D., 2007. Left ventrolateral prefrontal cortex and the cognitive control of memory.
Neuropsychologia 45, 2883–2901. doi:10.1016/j.neuropsychologia.2007.06.015
Boas, D. a, Strangman, G., Culver, J.P., Hoge, R.D., Jasdzewski, G., Poldrack, R. a, Rosen, B.R.,
Mandeville, J.B., 2003. Can the cerebral metabolic rate of oxygen be estimated with near-infrared
spectroscopy? Phys. Med. Biol. 48, 2405–2418. doi:10.1088/0031-9155/48/15/311
Bumstead, J.R., Bauer, A.Q., Wright, P.W., Culver, J.P., 2017. Cerebral functional connectivity and Mayer
waves in mice: Phenomena and separability. J. Cereb. Blood Flow Metab. 37, 471–484.
doi:10.1177/0271678X16629977
Buxton, R.B., Uludağ, K., Dubowitz, D.J., Liu, T.T., 2004. Modeling the hemodynamic response to brain
activation. Neuroimage 23, S220–S233. doi:10.1016/j.neuroimage.2004.07.013
Buxton, R.B., Wong, E.C., Frank, L.R., 1998. Dynamics of blood flow and oxygenation changes during brain
activation: the ballon model. Magn Reson Med 39, 855–864. doi:10.1002/mrm.1910390602
Caldwell, M., Hapuarachchi, T., Highton, D., Elwell, C., Smith, M., Tachtsidis, I., 2015. BrainSignals
revisited: Simplifying a computational model of cerebral physiology. PLoS One 10, e0126695.
doi:10.1371/journal.pone.0126695
Caldwell, M., Scholkmann, F., Wolf, U., Wolf, M., Elwell, C., Tachtsidis, I., 2016. Modelling confounding
effects from extracerebral contamination and systemic factors on functional near-infrared
spectroscopy. Neuroimage 143, 91–105. doi:10.1016/j.neuroimage.2016.08.058
56

Carassale, L., Asce, M., Kareem, A., Asce, D.M., 2010. Modeling Nonlinear Systems by Volterra Series. J.
Eng. Mech. 801–818. doi:10.1061/(ASCE)EM.1943-7889.0000113
Chalacheva, P., Khaleel, M., Sunwoo, J., Shah, P., Detterich, J.A., Kato, R.M., Thuptimdang, W.,
Meiselman, H.J., Sposto, R., Tsao, J., Wood, J.C., Zeltzer, L., Coates, T.D., Khoo, M.C.K., 2017.
Biophysical markers of the peripheral vasoconstriction response to pain in sickle cell disease. PLoS
One 12, e0178353. doi:10.1371/journal.pone.0178353
Choi, S., Bush, A.M., Borzage, M., Joshi, A., Coates, T.D., Leahy, R., C., W.J., 2016. Hemoglobin Level
and Platelet Size Predicts Grey and White Matter Volume Loss Measured By Tensor Based
Morphology in Sickle Cell Disease. Blood 128, 2481.
Choi, S., Bush, A.M., Borzage, M.T., Joshi, A.A., Mack, W.J., Coates, T.D., Leahy, R.M., Wood, J.C., 2017.
Hemoglobin and mean platelet volume predicts diffuse T1-MRI white matter volume decrease in sickle
cell disease patients. NeuroImage. Clin. 15, 239–246. doi:10.1016/j.nicl.2017.04.023
Cipolla, M.J., 2009. The Cerebral Circulation, The Cerebral Circulation.
Colombatti, R., Ermani, M., Rampazzo, P., Manara, R., Montanaro, M., Basso, G., Meneghetti, G., Sainati,
L., 2015. Cognitive evoked potentials and neural networks are abnormal in children with sickle cell
disease and not related to the degree of anaemia, pain and silent infarcts. Br. J. Haematol. 169, 597–
600. doi:10.1111/bjh.13232
Cope, M., Delpy, D.T., 1988. System for long-term measurement of cerebral blood and tissue oxygenation
on newborn infants by near infra-red transillumination. Med. Biol. Eng. Comput. 26, 289–94.
doi:10.1007/BF02447083
Cross-validation (statistics) [WWW Document], n.d. . Wikipedia. URL https://en.wikipedia.org/wiki/Cross-
validation_(statistics)#cite_note-McLachlan-7 (accessed 1.1.17).
Cui, X., Bray, S., Bryant, D.M., Glover, G.H., Reiss, A.L., 2011. A quantitative comparison of NIRS and
fMRI across multiple cognitive tasks. Neuroimage 54, 2808–21.
doi:10.1016/j.neuroimage.2010.10.069
Cui, X., Bray, S., Reiss, A.L., 2010. Functional near infrared spectroscopy (NIRS) signal improvement
57

based on negative correlation between oxygenated and deoxygenated hemoglobin dynamics.
Neuroimage 49, 3039–3046. doi:10.1016/j.neuroimage.2009.11.050
Dadgostar, M., Setarehdan, S.K., Akin, A., 2013. Detection of Motion Artifacts in fNIRS via the Continuous
Wavelet Transform. ICBME 243–246. doi:10.1109/ICBME.2013.6782227
Day, S., Chismark, E., 2006. The Cognitive and Academic Impact Of Sickle Cell Disease. J. Sch. Nurs. 22,
330–335. doi:10.1177/10598405060220060401
DeBaun, M.R., Armstrong, F.D., McKinstry, R.C., Ware, R.E., Vichinsky, E., Kirkham, F.J., 2012. Silent
cerebral infarcts: A review on a prevalent and progressive cause of neurologic injury in sickle cell
anemia. Blood 119, 4587–4596. doi:10.1182/blood-2011-02-272682
Demetriou, A., Christou, C., Spanoudis, G., Platsidou, M., Fischer, K.W., Dawson, T.L., 2002. The
Development of Mental Processing: Efficiency, Working Memory, and Thinking. Source Monogr. Soc.
Res. Child Dev. 67, 1–167. doi:10.2307/3181583
Diamond, S., 2015. Multimodal Neuroimaging with Functional Near-Infrared Spectroscopy - YouTube
[WWW Document]. Dartmouth - Youtube. URL
https://www.youtube.com/watch?v=L0_g4mAsPpk&t=2077s (accessed 12.28.17).
Diamond, S.G., Huppert, T.J., Kolehmainen, V., Franceschini, M.A., Kaipio, J.P., Arridge, S.R., Boas, D.A.,
2006. Dynamic physiological modeling for functional diffuse optical tomography. Neuroimage 30, 88–
101. doi:10.1016/j.neuroimage.2005.09.016
Friston, K.J.J., Mechelli, A., Turner, R., Price, C.J.J., 2000. Nonlinear responses in fMRI: the Balloon model,
Volterra kernels, and other hemodynamics. Neuroimage 12, 466–77. doi:10.1006/nimg.2000.0630
Gagnon, D., Matthew Brothers, R., Ganio, M.S., Hastings, J.L., Crandall, C.G., 2014. Forehead versus
forearm skin vascular responses at presyncope in humans. Am. J. Physiol. Regul. Integr. Comp.
Physiol. 307, R908-13. doi:10.1152/ajpregu.00204.2014
Gagnon, L., Cooper, R.J., Yücel, M.A., Perdue, K.L., Greve, D.N., Boas, D.A., 2012. Short separation
channel location impacts the performance of short channel regression in NIRS. Neuroimage 59,
2518–2528. doi:10.1016/j.neuroimage.2011.08.095
58

Gagnon, L., Perdue, K., Greve, D.N., Goldenholz, D., Kaskhedikar, G., Boas, D.A., 2011. Improved
recovery of the hemodynamic response in diffuse optical imaging using short optode separations and
state-space modeling. Neuroimage 56, 1362–1371. doi:10.1016/j.neuroimage.2011.03.001
Gravitz, L., Pincock, S., 2014. Sickle-cell disease 515, S1–S1. doi:10.1038/515S1a
Hall, J.E., Guyton, A.C., 2011. Guyton and Hall Textbook of Medical Physiology, Journal of Chemical
Information and Modeling. doi:10.1017/CBO9781107415324.004
Herff, C., Heger, D., Fortmann, O., Hennrich, J., Putze, F., Schultz, T., 2013. Mental workload during n-
back task-quantified in the prefrontal cortex using fNIRS. Front. Hum. Neurosci. 7, 935.
doi:10.3389/fnhum.2013.00935
Hernando-Gallego, F., Artés-Rodríguez, A., 2015. Individual performance calibration using physiological
stress signals, in: Workshop of Shimmer Sensors IEEE Body Sensor Networks Conference.
Hyvärinen, A., Oja, E., 2000. Independent Component Analysis: Algorithms and Applications. Neural
Networks 13, 411–430. doi:10.1016/S0893-6080(00)00026-5
Izzetoglu, K., 2008. Neural correlates of cognitive workload and anesthetic depth: fNIR spectroscopy
investigation in humans. Drexel Univ.
Jogia, J., Dima, D., Kumari, V., Frangou, S., 2012. Frontopolar cortical inefficiency may underpin reward
and working memory dysfunction in bipolar disorder 2975. doi:10.3109/15622975.2011.585662
Judy Stone, 2015. Sickle Cell Disease Highlights Racial Disparities In Healthcare [WWW Document].
Forbes. URL https://www.forbes.com/sites/judystone/2015/06/19/sickle-cell-disease-highlights-racial-
disparities-in-healthcare/#2b66f66c3b75 (accessed 1.11.18).
Kellogg, R.T., 2013. The making of the mind: The neuroscience of human nature., The making of the mind:
The neuroscience of human nature.
Kohno, S., Miyai, I., Seiyama, A., Oda, I., Ishikawa, A., Tsuneishi, S., Amita, T., Shimizu, K., 2007. Removal
of the skin blood flow artifact in functional near-infrared spectroscopic imaging data through
independent component analysis. J. Biomed. Opt. 12, 062111. doi:10.1117/1.2814249
59

Kuruvilla, M.S., Green, J.R., Ayaz, H., Murman, D.L., 2013. Neural correlates of cognitive decline in ALS:
an fNIRS study of the prefrontal cortex. Cogn. Neurosci. 4, 115–21.
doi:10.1080/17588928.2013.797889
Leff, D.R., Orihuela-Espina, F., Elwell, C.E., Athanasiou, T., Delpy, D.T., Darzi, A.W., Yang, G.-Z.Z., 2011.
Assessment of the cerebral cortex during motor task behaviours in adults: A systematic review of
functional near infrared spectroscopy (fNIRS) studies. Neuroimage 54, 2922–2936.
doi:10.1016/j.neuroimage.2010.10.058
Li, W.X.Y., Xin, Y., Chan, R.H.M., Song, D., Berger, T.W., Cheung, R.C.C., 2014. Laguerre-Volterra model
and architecture for MIMO system identification and output prediction. Conf. Proc. ... Annu. Int. Conf.
IEEE Eng. Med. Biol. Soc. IEEE Eng. Med. Biol. Soc. Annu. Conf. 2014, 4539–4542.
doi:10.1109/EMBC.2014.6944633
Mandeville, J.B., Marota, J.J.A., Ayata, C., Zaharchuk, G., Moskowitz, M.A., Rosen, B.R., Weisskoff, R.M.,
1999. Evidence of a cerebrovascular postarteriole windkessel with delayed compliance. J. Cereb.
Blood Flow Metab. 19, 679–89. doi:10.1097/00004647-199906000-00012
Marmarelis, V.., 2013. Nonlinear Dynamic Modeling of Physiological Systems, Journal of Chemical
Information and Modeling. doi:10.1017/CBO9781107415324.004
Marmarelis, V.Z., Shin, D.C., Diaz-Arrastia, R., Zhang, R., 2011. Diagnostic biomarkers for Alzheimer’s
disease using dynamic nonlinear models based on Principal Dynamic Modes. Conf. Proc. ... Annu.
Int. Conf. IEEE Eng. Med. Biol. Soc. IEEE Eng. Med. Biol. Soc. Annu. Conf. 2011, 7119–22.
doi:10.1109/IEMBS.2011.6091799
Mathworks, C., 2015. Statistics and Machine Learning Toolbox
TM
User ’ s Guide R 2015 a.
Mildner, T., Norris, D.G., Schwarzbauer, C., Wiggins, C.J., 2001. A qualitative test of the balloon model for
BOLD-based MR signal changes at 3T. Magn. Reson. Med. 46, 891–9. doi:10.1002/mrm.1274
Milham, M.. P., Banich, M.. T., Claus, E.. D., Cohen, N.. J., 2003. Practice-related effects demonstrate
complementary roles of anterior cingulate and prefrontal cortices in attentional control. Neuroimage
18, 483–493. doi:10.1016/S1053-8119(02)00050-2
60

Mitsis, G.D., Marmarelis, V.Z., 2002. Modeling of nonlinear physiological systems with fast and slow
dynamics. I. Methodology. Ann. Biomed. Eng. 30, 272–281. doi:10.1114/1.1458591
Molavi, B., Dumont, G.A., 2012. Wavelet-based motion artifact removal for functional near-infrared
spectroscopy. Physiol. Meas. 33, 259–270. doi:10.1088/0967-3334/33/2/259
O’Leary, J.D., Crawford, M.W., Odame, I., Shorten, G.D., McGrath, P. a, 2013. Thermal Pain and Sensory
Processing in Children With Sickle Cell Disease. Clin. J. Pain 30, 244–250.
doi:10.1097/AJP.0b013e318292a38e
Olufsen, M.S., Nadim, A., Lipsitz, L.A., 2002. Dynamics of cerebral blood flow regulation explained using a
lumped parameter model. Am. J. Physiol. - Regul. Integr. Comp. Physiol. 282, R611–R622.
doi:10.1152/ajpregu.00285.2001
Owen, A.M., McMillan, K.M., Laird, A.R., Bullmore, E., 2005. N-back working memory paradigm: A meta-
analysis of normative functional neuroimaging studies. Hum. Brain Mapp. 25, 46–59.
doi:10.1002/hbm.20131
Partington, T., Farmery, A., 2014. Intracranial pressure and cerebral blood flow. Anaesth. Intensive Care
Med. 15, 189–194. doi:10.1016/j.mpaic.2014.02.002
Saager, R.B., Telleri, N.L., Berger, A.J., 2011. Two-detector Corrected Near Infrared Spectroscopy (C-
NIRS) detects hemodynamic activation responses more robustly than single-detector NIRS.
Neuroimage 55, 1679–1685. doi:10.1016/j.neuroimage.2011.01.043
Sanger, M., Jordan, L., Pruthi, S., Day, M., Covert, B., Merriweather, B., Rodeghier, M., DeBaun, M.,
Kassim, A., 2016. Cognitive deficits are associated with unemployment in adults with sickle cell
anemia. J. Clin. Exp. Neuropsychol. 38, 661–671. doi:10.1080/13803395.2016.1149153
Sangkatumvong, S., Khoo, M.C.K., Kato, R., Detterich, J. a., Bush, A., Keens, T.G., Meiselman, H.J., Wood,
J.C., Coates, T.D., 2011. Peripheral vasoconstriction and abnormal parasympathetic response to
sighs and transient hypoxia in sickle cell disease. Am. J. Respir. Crit. Care Med. 184, 474–481.
doi:10.1164/rccm.201103-0537OC
Santosa, H., Hong, M.J., Kim, S.-P., Hong, K.-S., 2013. Noise reduction in functional near-infrared
61

spectroscopy signals by independent component analysis. Rev. Sci. Instrum. 84, 073106.
doi:10.1063/1.4812785
Schatz, J., Roberts, C.W., 2005. Short-term memory in children with sickle cell disease: executive versus
modality-specific processing deficits. Arch. Clin. Neuropsychol. 20, 1073–1085.
doi:10.1016/j.acn.2005.06.008
Scholkmann, F., Gerber, U., Wolf, M., Wolf, U., 2013. End-tidal CO2: An important parameter for a correct
interpretation in functional brain studies using speech tasks. Neuroimage 66, 71–79.
doi:10.1016/j.neuroimage.2012.10.025
Scholkmann, Felix, Wolf, M., Wolf, U., 2013. The Effect of Inner Speech on Arterial CO2 and Cerebral
Hemodynamics and Oxygenation: A Functional NIRS Study. pp. 81–87. doi:10.1007/978-1-4614-
7411-1_12
Schroeter, M.L., Bucheler, M.M., Muller, K., Uludag, K., Obrig, H., Lohmann, G., Tittgemeyer, M., Villringer,
A., von Cramon, D.Y., Bücheler, M.M., Müller, K., Uludağ, K., Obrig, H., Lohmann, G., Tittgemeyer,
M., Villringer, A., von Cramon, D.Y., Bucheler, M.M., Muller, K., Uludag, K., Obrig, H., Lohmann, G.,
Tittgemeyer, M., Villringer, A., von Cramon, D.Y., 2004. Towards a standard analysis for functional
near-infrared imaging. Neuroimage 21, 283–290. doi:10.1016/j.neuroimage.2003.09.054
Shah, P., Khaleel, M., Thuptimdang, W., Sunwoo, J., Veluswamy, S., Chalacheva, P., Kato, R.M., Detterich,
J., Wood, J.C., Zeltzer, L., Sposto, R., Khoo, M.C.K., Coates, T.D., 2019. Mental stress causes
vasoconstriction in sickle cell disease and normal controls. Haematologica haematol.2018.211391.
doi:10.3324/haematol.2018.211391
Shiavi, R., 2007. Introduction to applied statistical signal analysis : guide to biomedical and electrical
engineering applications. Elsevier Academic Press.
Smith, K.E., Schatz, J., 2016. Working Memory in Children With Neurocognitive Effects From Sickle Cell
Disease: Contributions of the Central Executive and Processing Speed. Dev. Neuropsychol. 41, 231–
244. doi:10.1080/87565641.2016.1238474
Spielberger, C.D., Gorsuch, R.L., Lushene, R., Vagg, P.R., Jacobs, G.A., 1983. Manual for the State-Trait
62

Anxiety Inventory. Consulting Psychologists Press, Palo Alto, CA.
Stefanovska, A., 2006. Coupled oscillators. Complex but not complicated cardiovascular and brain
interactions. IEEE Eng. Med. Biol. Mag. 26, 25–9. doi:10.1109/IEMBS.2006.259557
Stroop, J.R., 1935. Studies of interference in serial verbal reactions. J. Exp. Psychol. 18, 643–662.
doi:10.1037/h0054651
Sunwoo, J., Chalacheva, P., Sposto, R., Kato, R.M., Detterich, J., Wood, J.C., Coates, T.D., Khoo, M.C.K.,
Zeltzer, L.K., Wood, J.C., Coates, T.D., Khoo, M.C.K., Sunwoo, J., Chalacheva, P., Sposto, R., Kato,
R.M., Detterich, J., Zeltzer, L.K., Wood, J.C., Coates, T.D., Khoo, M.C.K., Sunwoo, J., Chalacheva,
P., Sposto, R., Kato, R.M., Detterich, J., Zeltzer, L.K., Wood, J.C., Coates, T.D., Khoo, M.C.K., 2018.
A novel cross-correlation methodology for assessing biophysical responses associated with pain. J.
Pain Res. 2207–2219. doi:10.2147/JPR.S142582
Tak, S., Ye, J.C., 2014. Statistical analysis of fNIRS data: a comprehensive review. Neuroimage 85 Pt 1,
72–91. doi:10.1016/j.neuroimage.2013.06.016
Teng, E., Tassniyom, K., Lu, P.H., 2012. Reduced Quality-of-Life Ratings in Mild Cognitive Impairment:
Analyses of Subject and Informant Responses. Am. J. Geriatr. Psychiatry 20, 1016–1025.
doi:10.1097/JGP.0b013e31826ce640
Terrighena, E.L., Lu, G., Yuen, W.P., Lee, T.M.C., Keuper, K., 2017. Effects of vicarious pain on self-pain
perception: Investigating the role of awareness. J. Pain Res. 10, 1821–1830.
doi:10.2147/JPR.S132744
Virtanen, J., Noponen, T., Meriläinen, P., 2009. Comparison of principal and independent component
analysis in removing extracerebral interference from near-infrared spectroscopy signals. J. Biomed.
Opt. 14, 054032. doi:10.1117/1.3253323
Virtanen, Jaakko, Noponen, T., Kotilahti, K., Virtanen, Juha, Ilmoniemi, R.J., 2011. Accelerometer-based
method for correcting signal baseline changes caused by motion artifacts in medical near-infrared
spectroscopy. J. Biomed. Opt. 16, 087005. doi:10.1117/1.3606576
Yanagisawa, H., Dan, I., Tsuzuki, D., Kato, M., Okamoto, M., Kyutoku, Y., Soya, H., 2010. Acute moderate
63

exercise elicits increased dorsolateral prefrontal activation and improves cognitive performance with
Stroop test. Neuroimage 50, 1702–1710. doi:10.1016/j.neuroimage.2009.12.023
Yennu, A., Tian, F., Smith-Osborne, A., Gatchel, R.J., Woon, F.L., Liu, H., 2016. Prefrontal responses to
Stroop tasks in subjects with post-traumatic stress disorder assessed by functional near infrared
spectroscopy. Sci. Rep. 6, 30157. doi:10.1038/srep30157
Yerys, B.E., White, D.A., Salorio, C.F., McKinstry, R., Moinuddin, A., DeBaun, M., 2003. Memory strategy
training in children with cerebral infarcts related to sickle cell disease. J Pediatr Hematol Oncol 25,
495–498.
Yeung, M.K., Sze, S.L., Woo, J., Kwok, T., Shum, D.H.K., Yu, R., Chan, A.S., 2016. Altered frontal
lateralization underlies the category fluency deficits in older adults with mild cognitive impairment: A
near-infrared spectroscopy study. Front. Aging Neurosci. 8, 1–15. doi:10.3389/fnagi.2016.00059
Yücel, M.A., Selb, J., Aasted, C.M., Petkov, M.P., Becerra, L., Borsook, D., Boas, D.A., 2015. Short
separation regression improves statistical significance and better localizes the hemodynamic
response obtained by near-infrared spectroscopy for tasks with differing autonomic responses.
Neurophotonics 2, 035005. doi:10.1117/1.NPh.2.3.035005
Zhang, Q., Strangman, G.E., Ganis, G., 2009. Adaptive filtering to reduce global interference in non-
invasive NIRS measures of brain activation: How well and when does it work?, NeuroImage.
doi:10.1016/j.neuroimage.2008.12.048
Zhang, Y., Brooks, D.H., Franceschini, M.A., Boas, D. a, 2005. Eigenvector-based spatial filtering for
reduction of physiological interference in diffuse optical imaging. J. Biomed. Opt. 10, 11014.
doi:10.1117/1.1852552

Abstract (if available)

Abstract Vaso-occlusive pain episodes are a distinguishing symptom in sickle cell disease (SCD). Rigid and sickle-shaped red blood cells occlude the blood flow in the microvasculature, causing damage to the surrounding tissue accompanied by severe pain. SCD vaso-occlusive crises are believed to be linked with silent and overt stroke, as well as cognitive decline. Recent MRI studies have demonstrated shrinkage of white matter in SCD, which, in turn, could be a major factor of cognitive decline in this population. However, little is known about the cause of a stroke and its exact linkage to cognitive decline in SCD. Analyses of functional brain imaging data that is obtained when a subject is engaged in a cognitive task can help one understand underlying brain mechanisms between cognitive decline and disease conditions. Therefore, the goal of this research study is to use functional near-infrared spectroscopy (fNIRS), which is also one of the popular functional brain imaging techniques, to study abnormal cognitive brain activity seen in SCD, as well as to propose a signal processing method for filtering out non-neuronal physiological confounds in the original fNIRS measurement. ❧ We used a linear dynamic systems modeling approach (Model-based filtering, MBF) for identifying and subtracting two major physiological confounders in the original fNIRS measurement. Using the filtered fNIRS signal, we showed inefficiency and possible abnormal cognitive activity in SCD, especially in patients with stroke conditions. This result was also supported by observations of consistently slower response time to conducted mental tasks in SCD compared to healthy controls, while task accuracy was not different between the groups. This finding is consistent with recent reports of white matter loss in SCD and suggest the use of fNIRS-based neurocognitive responses as a potential biomarker of cognitive decline reported in SCD.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Susceptibility-weighted MRI for the evaluation of brain oxygenation and brain iron in sickle cell disease

PDF

Modeling of cardiovascular autonomic control in sickle cell disease

PDF

Role of the autonomic nervous system in vasoconstriction responses to mental stress in sickle cell disease: a bioengineering perspective

PDF

Photoplethysmogram-based biomarker for assessing risk of vaso-occlusive crisis in sickle cell disease: machine learning approaches

$Quantitative MRI for the measurement of cerebral oxygen extraction fraction in sickle cell disease$

PDF

Quantitative MRI for the measurement of cerebral oxygen extraction fraction in sickle cell disease

PDF

Cerebrovascular disease of white matter in patients with chronic anemia syndrome

PDF

Assessing cerebral oxygen supply and demand in sickle cell anemia using MRI

PDF

A virtual reality exergaming system to enhance brain health in older adults at risk for Alzheimer’s disease

PDF

Quantitative MRI for oxygenation imaging in cerebrovascular diseases

PDF

Clinical significance of oxyhemoglobin dissociation curves on the interpretation of pulse oximetry in Sickle Cell disease

PDF

Modeling autonomic peripheral vascular control

PDF

Cross-sectional association of blood pressure, antihypertensive medications, MRI volumetric measures and cognitive function scores in an aging population

PDF

Contributions of dynamic cerebrovascular function to cognitive decline and dementia: development and validation of a novel neuroimaging approach

PDF

Alpha globin and vascular reactivity in humans: loss of alpha globin genes in human subjects is associated with improved nitric oxide-mediated vascular perfusion

PDF

Mapping water exchange rate across the blood-brain barrier

PDF

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

PDF

Blood-brain barrier pathophysiology in cognitive impairment and injury

PDF

Long-term treadmill exercise delays the onset of motor dysfunction, cognitive impairments, and mood disturbances in the CAG 140 mouse model of Huntington's disease via restoration of dopamine neu...

PDF

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

PDF

Understanding the role of APP and DYRK1A in human brain pericytes

Asset Metadata

Creator Sunwoo, John (author)

Core Title Estimation of cognitive brain activity in sickle cell disease using functional near-infrared spectroscopy and dynamic systems modeling

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Biomedical Engineering

Publication Date 06/27/2019

Defense Date 05/01/2019

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

Tag brain oxygenation,cognitive decline,N-back,OAI-PMH Harvest,prefrontal cortex,sickle cell disease,stroke,Stroop

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Khoo, Michael C.K. (committee chair), Coates, Thomas D. (committee member), Wood, John C. (committee member)

Creator Email john@johnsunwoo.com,johnfsunwoo@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-178723

Unique identifier UC11660973

Identifier etd-SunwooJohn-7516.pdf (filename),usctheses-c89-178723 (legacy record id)

Legacy Identifier etd-SunwooJohn-7516-0.pdf

Dmrecord 178723

Document Type Dissertation

Format application/pdf (imt)

Rights Sunwoo, John

Type texts

Source 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 a...

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