Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Brain-based prediction of chronic pain progression: a longitudinal study of urologic chronic pelvic pain syndrome using baseline resting state connectivity from the periaqueductal gray
(USC Thesis Other)
Brain-based prediction of chronic pain progression: a longitudinal study of urologic chronic pelvic pain syndrome using baseline resting state connectivity from the periaqueductal gray
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
Sonja J. Fenske Brain-based prediction of chronic pain progression: A longitudinal study of Urologic Chronic Pelvic Pain Syndrome using baseline resting state connectivity from the Periaqueductal Gray by Sonja J. Fenske A Dissertation Presented to the FACULTY OF THE USC GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY Neuroscience December 2020 ii Acknowledgements The work in this thesis was undertaken in the laboratory of Dr. Jason Kutch in the Division of Biokinesiology and Physical Therapy, University of Southern California, Los Angeles, USA. Data collection was performed by the Multidisciplinary Approach to the Study of Chronic Pelvic Pain (MAPP) research network. Work was supported by funding from the Neuroscience Graduate Program at the University of Southern California and the National Institute of Health grants: R01DK110669 and U01DK082370. Special thanks and appreciation to my supervisor Dr. Jason Kutch for his guidance and support throughout my degree. His mentorship, candor, and attention to my work was invaluable throughout my studies. I thank also my committee members Dr. Sook-Lei Liew, Dr. Judy Pa, Dr. Richard Leahy, and Dr. John Van Horn for their supervision. I thank my colleagues in the MAPP research network and members of Dr. Kutch’s research group for their input throughout. I thank the administrative staff at all the institutes for their support. I would like to recognize the invaluable contribution of the participants in our studies without whose involvement this work would not be possible. I would finally like to thank my family and friends for their support throughout my studies. iii Table of Contents Acknowledgements ................................................................................................................ ii List of Figures and Tables ..................................................................................................... v Abstract ................................................................................................................................ vi Chapter 1: Introduction ......................................................................................................... 1 Participants, data collection, and methods: ...................................................................................4 Urologic Chronic Pelvic Pain Syndrome ............................................................................................................. 4 MAPP datasets...................................................................................................................................................... 6 MAPP-I prediction in UCPPS .............................................................................................................................. 9 The Periaqueductal Gray ............................................................................................................ 10 The PAG and the descending pain modulatory system ...................................................................................... 12 Analgesic effects associated with different subregions of the PAG ................................................................... 13 The PAG’s acute response to threat and inflammation ...................................................................................... 14 Involvement of the PAG and UCPPS ................................................................................................................. 15 Chapter Summaries .................................................................................................................... 18 Chapter 2: Longitudinal UCPPS symptom change as a predictive outcome measure ............. 20 Introduction................................................................................................................................ 20 Methods ...................................................................................................................................... 21 MAPP Phase II dataset overview ....................................................................................................................... 21 Participants ......................................................................................................................................................... 21 Measures for analysis ......................................................................................................................................... 22 Analysis .............................................................................................................................................................. 24 Results ........................................................................................................................................ 26 Discussion ................................................................................................................................... 34 Conclusion .................................................................................................................................. 36 Chapter 3: Sensitivity of functional connectivity to periaqueductal gray localization ............. 37 Introduction................................................................................................................................ 37 Methods ...................................................................................................................................... 39 Participants ......................................................................................................................................................... 39 Image collection ................................................................................................................................................. 42 Image processing ................................................................................................................................................ 42 PAG ROI definition ............................................................................................................................................ 43 PAG ROI signal comparison .............................................................................................................................. 46 Whole-brain connectivity ................................................................................................................................... 47 Cluster-based analysis in UCPPS patients versus healthy controls.................................................................... 49 Independent component analysis signal correction for potential physiological effects ..................................... 50 Results ........................................................................................................................................ 51 PAG ROI BOLD signal non-uniformities in healthy control participants ......................................................... 51 Whole-brain functional connectivity differences in healthy control participants .............................................. 51 Cluster-based functional connectivity differences in UPPS patients versus healthy controls ........................... 54 Head motion ....................................................................................................................................................... 58 iv Independent component analysis signal correction effect .................................................................................. 58 Discussion ................................................................................................................................... 62 Conclusion .................................................................................................................................. 70 Chapter 4: Using the PAG to whole-brain connectivity to predict UCPPS symptom change... 72 Introduction................................................................................................................................ 72 Methods ...................................................................................................................................... 74 Participants ......................................................................................................................................................... 74 Image collection ................................................................................................................................................. 74 Image processing ................................................................................................................................................ 75 Predictive measures: Functional connectivity preprocessing ............................................................................. 77 Outcome measures for prediction ....................................................................................................................... 78 Split dataset used in analysis .............................................................................................................................. 79 Machine learning prediction and biomarker evaluation ..................................................................................... 81 Results: ....................................................................................................................................... 84 Discussion ................................................................................................................................... 93 Conclusion .................................................................................................................................. 99 Chapter 5: Discussion........................................................................................................ 100 Machine learning results ........................................................................................................... 100 UCPPS symptoms ..................................................................................................................... 102 Potential biomarkers ................................................................................................................ 104 Future steps .............................................................................................................................. 106 References......................................................................................................................... 108 v List of Figures and Tables Table 1.1 MAPP-II site distribution................................................................................................ 9 Figure 1.1: Preliminary prediction data. ....................................................................................... 10 Figure 1.2: Sagittal view of a manually traced PAG in an individual participant ........................ 12 Figure 2.1: 12-month symptom progression slope ....................................................................... 25 Figure 2.2: 3-month symptom progression slope ......................................................................... 25 Figure 2.3: Baseline pain score is associated with 12-month slope values .................................. 27 Figure 2.4: Baseline pain score is associated with 3-month slope values .................................... 27 Figure 2.5: Average pain score split into three groups by 12-month slope value ........................ 29 Figure 2.6: Average pain score split into three groups by 3-month slope value .......................... 29 Figure 2.7: Residual slope value distribution ............................................................................... 30 Figure 2.8: One-way ANOVA comparison of 12-month slope values at each site in the MAPP-II dataset ........................................................................................................................................... 31 Figure 2.9: A one-way ANOVA comparison of 3-month slope values at each site in the MAPP-II dataset ........................................................................................................................................... 32 Table 3.1: Demographic characteristics of participant datasets in Chapter 3 ............................... 40 Figure 3.1: Analysis of the BOLD rs-fMRI time series signal from the three different localization techniques of the PAG .................................................................................................................. 44 Figure 3.2: Whole-brain functional connectivity analysis of three different localization techniques of the PAG .................................................................................................................. 48 Figure 3.3: Performance of the PAG ROIs on whole-brain functional connectivity .................... 53 Table 3.2: The average connectivity of significant clusters found connected to the MNI-Sphere ROI and MNI-trace ROI in UCPPS patients and healthy controls ............................................... 56 Figure 3.4: Functional connectivity differences in UPPS patients versus healthy controls ......... 58 Figure 3.5: Independent component analysis (ICA) correction for potential physiological noise effects ............................................................................................................................................ 62 Figure 4.1: Preprocessing and machine learning workflow on resting state fMRI data ............... 76 Figure 4.2: Examples of the split distributions across the three datasets...................................... 81 Figure 4.3: 12-month symptom slope generalized prediction model and significant weights based on the Power 264 atlas .................................................................................................................. 88 Table 4.1: Connectivity values in improving and worsening participants .................................... 89 Figure 4.4: 12-month symptom LOOCV prediction..................................................................... 92 vi Abstract The work of this thesis was to establish whether the prediction of long-term symptom change in urologic chronic pelvic pain syndrome (UCPPS) is achievable using baseline resting state functional magnetic resonance imaging (rs-fMRI) measures. The development of chronic pain is not well understood, and pain symptoms fluctuate over months to years. Studies from the Multidisciplinary Approach to the Study of Chronic Pelvic Pain (MAPP) research network have previously shown changes in brain function and structure between UCPPS patients and healthy controls. Previous studies in our laboratory have been able to predict UCPPS symptom change up to 3-months with baseline rs-fMRI along with longitudinal patient data from MAPP. In order to make the results more clinically relevant there is a need to validate and expand these results for longer time periods, as well as develop more definitive biomarkers representative of symptoms progression. We identified the periaqueductal gray (PAG), well-established as a component in pain modulation, as our main region of interest. We hypothesized that the PAG may be associated with reduced ability to self-regulate altered physiological responses in UCPPS and influence symptom change over time. To predict change in symptoms up to a year requires a better understanding of how symptom pain trajectories change over time. We defined our outcome variable as the slope, or the difference, in reported symptoms from baseline up to 12-months. Using data from 389 participants in MAPP’s longitudinal collection, we found that a portion of the participants improve significantly over the 12-month timeframe. This trend, and its differentiation from patients who do not improve, is what we predicted using baseline connectivity from the PAG to the rest of the brain. To do this accurately we needed to verify the location of the PAG in the brainstem with resting state connectivity measures, as previous studies have used coordinates representative of the vii PAG more anterior than the true anatomical location. We found that differing PAG localization measures impacted whole-brain connectivity in healthy participants and the detection of differences in UCPPS patients compared to healthy participants. To make a prediction on the change in symptoms over 12-months based on resting state connectivity measures from the PAG, we used a hand-traced MNI location. This region of interest was shown to be a close approximation to the “ground-truth” or participant specific anatomical location of the PAG and was used as the location to seed the PAG in our study. Our final analysis sought to predict the change in symptoms over 12-months using whole brain connectivity. We optimized brain parcellation and feature selection and used a support vector regression to develop our model. We show that our method makes a significant prediction and our model is able to generalize to independent datasets. Furthermore, we identify biomarkers, such as the PAG’s connectivity to the insula, that were found to be influential in driving the predictive model. These results further our understanding of the mechanisms behind pain chronicity and identify potential prognostic biomarkers that could be targeted in future therapeutic interventions to alleviate the symptoms of chronic pain. 1 Chapter 1: Introduction Chronic pain is one of the most prevalent medical conditions, recently being estimated to effect from 11-40% of the adult population of the United States of America (1) and costing upwards of $560 billion a year (2). This is more than the annual individual costs of diabetes, heart disease, or cancer. Despite these statistics, the definite prevalence of chronic pain and the mechanisms for the shift from acute to chronic pain remain an enigma in the medical field. For example, 70-80% of adults are affected by lower back pain sometime in their life, however, around 90% of these people recover and about 5% have symptoms that become chronic. The physiological mechanisms producing persistent symptoms are not well known. Furthermore, the term “chronic” is broadly defined and depends on the disease studied (3). Though both are termed “chronic,” the time scale at which back pain and post-herpetic neuralgia are diagnosed is around six months and three months respectively (4–7). Due to the effects that chronic pain has on daily life, it is often associated with psychological comorbidities such as depression and anxiety (8). Apkarian et al. states that “depression is ranked as one of the strongest predictors for low back pain” and studies have “argued that…chronic pain is…dependent on psychosocial and occupational factors” (5,9). The identification of specific biological mechanisms that underlie pain is difficult because of the various associated etiologies and differences in symptom progression. In this thesis we propose specific brain regions whose activity offers some insight into the neurophysiology underlying chronic pain and neural signals that provide biomarkers capable of predicting symptom progression. Such areas could be future targets for intervention aimed at treating chronic pain. While sources of chronic pain are diverse, more studies have shown its effect and functioning as a part of the central nervous system (10). The theory of the neuromatrix was defined 2 by Melzack; “pain…is produced by output of a widely distributed neural network in the brain rather than directly by injury, inflammation, or other pathology” and that the “brain…[acts] as an active system that filters, selects and modulates inputs” (11,12). In contrast with acute and inflammatory pain responses, chronic pain may involve distinct spinal cord nociceptive neurons and supraspinal projections (5). The reasoning behind this transition from acute to chronic pain could be plasticity in the brain, leading to reorganization due to pain conditions (for a review see: May et al. (13)). Previously, scientists have defined themselves as “localizationists,” – those who are focused on specific regions of brain functioning, such as only measuring chronic pain as a function of spinal cord processing and spinothalamic pathway transmission (5). This approach is in contrast to a more modern perspective that the brain is part of an interconnected network where many of the cognitive processes associated with pain have integrating and overlapping features (13,14). Both theories are excepted by the scientific community; however, due to the nature of chronic pain and the comorbidities that often accompany pain, pain is becoming more understood as an “integration of activity in distinct neuronal structures” (13). Neuroimaging measures such as resting state functional magnetic resonance imaging (rs-fMRI) and voxel-based morphometry have shown that central plasticity could be a result of pain as these measurers reveal physiological and structural changes in brain due to persistent pain. Evidence for this is the functional reorganization observed with chronic back pain (15) and grey matter alterations in the subcortical and brainstem structures in back and pelvic pain respectively (15,16). As discussed above there are many different symptoms and biological causes for the effects of pain and there is no standardized method for diagnosing chronic pain. Treatments for chronic pain have been inadequate or lacking. Back pain, for example, has been treated using a myriad of approaches, from oral and injectable medications to surgical devices, physical therapy and 3 phycological interventions (5). However, the response to drug treatments varies for different types of pain, in part due to how little is known about the individual and common pathophysiology of different pain conditions (17). Furthermore, the side-effects of current drug treatments can be dangerous, leading to organ toxicity and addiction, such as well documented cases of opioid addiction and its impact beyond pain, on cognitive function (17). A review by May and colleagues questions the cause and effect of chronic pain; are changes in the brain seen because of the disorder, or, does the disorder exist because of the changes seen in the brain (13)? As with the hypothesis towards central plasticity, a recent study shows that grey matter decreases, after correcting for age and gender, over a long period in subacute back pain patients (15). Others, as highlighted by Bushnell and colleagues, include grey matter loss in “fibromyalgia, headache, [irritable bowel syndrome], complex regional pain syndrome,[and] osteoarthritis” (18). These studies indicate that over time, the effects in the brain are result of chronic pain and not vice-versa. Though, additional research is needed to understand the relationship between pain and brain changes over longer periods of time and to uncover biomarkers that define this change for potential treatment. An important aspect in finding the proper treatment for chronic pain, is the identification and development of therapies that are targeted to specific individuals and specific pain conditions (19). In our studies, we focus on Urologic Chronic Pelvic Pain Syndrome (UCPPS). To begin, making the distinction between the patients whose symptoms will worsen or improve is the first step towards the correct treatment. Symptom prediction and prevention may be the most cost- effective measures for medical diagnosis and achieving this using neuroimaging in longitudinal clinical trials will provide a better understanding of disease progression. Neuroimaging also allows us to view altered regions of the brain or brain networks related to pain conditions and find potential targets for early treatment. Therefore, the goal of this thesis is to test whether prediction 4 is possible in UCPPS patients and to define rs-fMRI features found to be predictive of longitudinal change in UCPPS symptoms. Participants, data collection, and methods: Urologic Chronic Pelvic Pain Syndrome Urologic Chronic Pelvic Pain Syndrome (UCPPS) affects millions of men and women in the US, as high as 5-10% of the general population, with annual per person costs from ranging from $3500-7000 (20,21). UCPPS disorders include interstitial cystitis/bladder pain syndrome (IC/BPS) and chronic prostatitis/chronic pelvic pain syndrome (CP/CPPS) (22). Onset age is between 20 to 50 years with urinary control and pelvic functioning issues that can last decades (23,24). Symptoms consist of problems with pelvic functioning and urinary control and pain (20). These conditions affect different parts of the pelvic region, though there is overlap of these regions in many of the clinical features of each disorder (25). Furthermore, symptoms are often accompanied with other chronic pain symptoms such as allodynia which can also lead to the magnification of the perception of non-painful stimuli (26). Also, hyperalgesia, when stimuli are perceived more painful than expected (26), occurs in both men (27) and women (28) and there is evidence of increased hypersensitivity in bladder filling (29). Due to the persistence of these conditions without successful treatment, there is a clear need to improve the clinical care of UCPPS patients. Many of the same issues associated with chronic pain, involving symptom variability and heterogeneous etiologies, are also seen in patients who suffer from UCPPS. For example, once thought to be the source of UCPPS symptoms, the bladder or the prostate were treatment targets for alpha receptor blockers (30). These treatments were originally developed for genitourinary 5 problems due to the high density of receptors in this area, but such agents work only at specific types of pharmacological sites, which may not necessarily be effective when treating UCPPS due the range of underlying causes (31). In subsequent research there has been no identifiable pathology in these organs relating specifically to UCPPS and the effectiveness of receptor blocker treatments have varied per individual (32). As a result of the heterogeneous nature of causes underlying UCPPS, identifying effective treatment targets with known physiological mechanisms has been a challenge. Many studies have attempted to characterize UCPPS, but there has been “no consensus agreement” on the pathophysiology at the cellular level and in-vivo/ or animal model systems (22). Therefore, more research is needed to determine specific biomarkers related to the syndrome before a treatment can be provided. Recent studies point to a more systematic view of UCPPS where biomarkers in the brain may play an important factor in identifying potential treatment targets (33–37). As with chronic pain symptomology, there is no specific source for UCPPS patient’s pain symptoms. However, there is a “systemic view of the disease” where the CNS has been shown to play an important factor (22). For example, a study by Schrepf et al. focuses on toll-like receptors, TLR-4 specifically, as a strong predictor for IC/BPS (19). This study showed that there are specific biomarkers that may help discern patients from healthy controls. This is significant because “IC/BPS patients lack any discernible end organ inflammation, and do not respond to treatment of peripheral tissues” (38). Schrepf et al. also demonstrated that the dysregulation of the hypothalamic-pituitary-adrenal axis, also found in other pain conditions, is a feature of IC/BPS, showing a connection to central nervous system regulation of pain. Current research focusing on the CNS in UCPPS patients involves central processing of pain and viscerosensory signals (36). Resting state fMRI (rs-fMRI) has shown alterations in frequency power distribution and functional 6 connectivity patterns in the cortico-cerebellar network of women with IC/BPS (36) as well as altered functional connectivity compared to healthy controls in CP/CPPS (37). Though these cross- sectional studies indicate UCPPS as a condition affecting the brain, UCPPS is a chronic disorder with symptoms fluctuating over months (25,39). For the reason that UCPPS symptoms tend to fluctuate over time, there is a need for longitudinal studies assessing pain symptoms. Longitudinal data should provide more informative treatment targets by focusing on changes over time within the patient population rather than solely between healthy and patient populations. MAPP datasets New studies collecting data from UCPPS patients have taken a more interdisciplinary approach. A research group, of which our lab is a member, called the Multidisciplinary Approach to Chronic Pelvic Pain (MAPP) was formed by the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) to apply basic, translational, and clinical science studies to phenotype data associated with UCPPS and “ultimately [translate the research] into improved clinical management” (22). Experts include researchers in pain, neurobiology, neuroimaging, infectious disease, biomarker discovery, animal modeling, epidemiology, psychology, immunology, and more (22). A main goal in the clinical application of this research is to define robust and meaningful subgroups of UCPPS patients (40), such as defining patients that are likely to be resistant to treatment or, alternatively, patients who improve over time without treatment intervention. The diagnosis of UCPPS relies on patient reported symptoms and through methods of elimination, such as infectious disorders or other neurological conditions affecting the bladder or bowel (see (32,41) for further details). Questionnaires including the Genitourinary Pain Index (GUPI), the Interstitial Cystitis Symptom Index (ICSI) and Problem Index (ICPI) are designed to 7 measure baseline symptom severity of pain and urinary conditions, and are additional tools to measure patient populations (20,22,42). Under MAPP, researchers have discovered biomarkers underlying symptoms that may provide a better understanding of UCPPS (43) and lead to the development of clinical treatments (see the MAPP website: https://www.mappnetwork.org/). The data collected for our studies and used throughout this thesis were collected by MAPP Research Network and organized by the Data Coordinating Center (DCC). There are two phases of MAPP. The first phase, MAPP-I, or the Trans-MAPP Epidemiology/Phenotyping Study (EPS), was a 5-year project period with recruitment between 2009-2012 with 1039 participants enrolled, including 424 with UCPPS, 200 participants with chronic overlapping pain symptoms, and 415 healthy controls (44). All participants were collected at 5 institutional review board approved study sites across the United States with neuroimaging quality standards met. Baseline neuroimaging and questionnaire data was collected from participants. Questionnaire data included basic demographics, brief pain inventory (BPI), and anxiety and depression (HADS). UCPPS patients were followed up to a year with bi-weekly internet-based symptom assessments and on follow up visits quantitative sensory testing (QST) and biospecimens were collected (for a more comprehensive list and summarization of findings see Clemens et al. 2019 (43)). Phase 2 of MAPP, MAPP-II, or the Trans-MAPP Symptom Patterns Study (SPS), enrolled 620 participants with baseline data and followed UCPPS participants for up to 3 years between 2015 through 2019 (the number of participants in the longitudinal dataset is ongoing) (40,44). Neuroimaging was recorded again at baseline, but also during bladder filling and at follow-up visits. More extensive phenotyping was collected as well (see Clemens et al. 2020 (44) for an review of data collection). MAPP-II participants were collected at 6 study sites: Northwestern, 8 University of California Los Angeles, University of Iowa, University of Michigan, University of Washington, Washington University in St. Louis (44). Classification studies in our lab have used neuroimaging and non-neuroimaging measures to observe how changes in UCPPS symptom improvement is related to brain connectivity (39). A dataset of 52 patients’ longitudinal data from MAPP-I shows our preliminary prediction results, see below for a brief review. Furthermore, phenotypic and psychological variables were used in our preliminary results to characterize the patient population based on non-neuroimaging correlates (39). Our current work for this thesis uses MAPP-I data in Chapter 3 for our PAG analysis. The original cohort consisted of 318 participants from MAPP-I, with 30 health controls for our first analysis. The second part of the analysis reduced the 318 participants to 209, including 100 UCPPS patients and 109 healthy controls. MAPP-II data was used in Chapter 2 for our analysis of our UCPPS outcome measures and Chapter 4 for our analysis on longitudinal prediction of symptoms. We have a sample of 235 women and 129 men from MAPP-II, for a total of 389 with longitudinal data collected for up to 1 year. Table 1.1 shows the distribution of this data across the 6 collection sites and the imaging scanners used at each site. Our outcome measures were based on baseline questionnaire data and follow-up data collected at 3-months, 6-months, and 12-months. In all of these datasets from MAPP-I we only used baseline neuroimaging. In the case of MAPP- II neuroimaging, we only used images taken after voiding. 9 Table 1.1: Table 1.1 MAPP-II site distribution. Key: NW, Northwestern, UCLA, University of Southern California Los Angeles, UI, University of Iowa, Umich, University of Michigan,UW, University of Washington,WahsU, Washington University in St. Louis, GE, General Electric MAPP-I prediction in UCPPS Classification studies in our lab have used neuroimaging and non-neuroimaging measures to observe how changes in UCPPS symptom improvement is related to brain connectivity (39). A dataset of 52 patients’ baseline resting state fMRI was first collected from MAPP-I and patient questionnaire data on primary symptoms were assessed every two weeks (beginning in the first four weeks after enrollment to allow for symptom stabilization). Questionnaire data included those from the Genitourinary Pain Index (GUPI)(1) and the Interstitial Cystitis Symptom Index (ICSI)(25), combined for a total index from 0-28 (25,39,45). Two representative subjects are shown in Figure 1.1 A). Outcome variables were defined as the slope of the linear regression fit to the change in symptoms at 3 months - short-term, and 12 months - long-term (not shown). This was dichotomized for classification measures, separating those that improved (improvers) versus those that do not improve (non-improvers) by the median (Figure 1.1 B.). Baseline fMRI was used to predict these longitudinal outcome variables using a support vector machine classifier (SVM). NW UCLA UI UMich UW WashU Total # of patients 52 62 79 63 67 66 389 Sex (M/F) 16/29 14/46 19/59 12/45 43/20 24/36 129/235 Age (years) 44.2±15.3 39.7±15.3 47.8±16.3 41.4±14.8 47.6±16.0 44.1±14.7 44.4±15.7 Scanner Siemens Siemens Siemens GE Phillips Siemens 10 Overall the classification accuracy for improvers and non-improvers was 73.1% at 3 months (significantly greater than chance p=0.001)(39). Also, results of the classifier indicated that specific connections in the frontal parietal regions of the brain were associated to those categorized as “improvers” at this short-term time point (Figure 1.1 C.). While this is strong preliminary work, sensitivity was 69.2% and though promising, classification measures were not significant at the 12-month time point (p=0.08) (39). The MAPP-II dataset provides a valuable resource to properly address these trends and motivates a further analysis in longitudinal prediction of UCPPS symptoms. Figure 1.1 Figure 1.1: Preliminary prediction data. A. Two examples of UCPPS patient’s baseline rs-fMRI scan and 2-week symptom assessment for 3 months. B. Symptom change was dichotomized into two categories: improvers and non- improvers based on their median (the large dots represent the two examples in A). C. Machine learning support vector machine identified connections from baseline rs-fMRI data that were stronger in improvers. Red spheres represent frontal regions and green represent parietal regions, where the size of the sphere represents the number of connections to that region. The Periaqueductal Gray Patients with Urologic Chronic Pelvic Pain Syndrome (UCPPS) exhibit variable pain and urologic symptoms that can last years (22,41). With no currently effective treatment plan there is clearly a clinical need to develop methods, such as biomarker identification, to distinguish the 11 mechanisms for pain variability and to reveal why some patients improve over time while others do not (33,43,46). Understanding changes in pain symptoms in particular, for instance how chronicity progresses and why it is maintained, is important for targeted treatment (43). The periaqueductal gray (PAG) is a critical control center encircling the central aqueduct in the midbrain (see Figure 1.2). The PAG region has been attributed to a range of functions, from autonomic control (47–52), such as micturition and cardiovascular regulation, to behaviors generalized to fear, anxiety (47,53,54), sexual behavior (55), and pain perception (56). Neuroanatomical connections common in both primate and humans from the prefrontal cortex to the PAG (48,57) have been shown to contribute towards top down control of spinal and sensorimotor circuits, such as in pain modulation (58,59). The PAG’s functionality in the response to acute threat as well as its autonomic response to inflammatory states sets up an alarm system to potential danger. This may fail to reset in chronic pain disorders. Such maladaptive behavior of the PAG may lead to the manifestation and maintenance of UCPPS symptoms. We provide an explanation of these concepts intended to highlight the PAG as an initial region of particular interest in our research on longitudinal UCPPS symptom prediction. 12 Figure 1.2: Figure 1.2: Sagittal view of a manually traced PAG in an individual participant. This is an example trace (in green) of one of participants for our analysis covered in Chapter 3. The voxels were traced by researchers at the Medical College of Wisconsin on a T1-weighted structural image in 1mm grid space. The PAG and the descending pain modulatory system The descending pain modulatory system (DPMS) consists of a set of brain networks which regulate acute pain processes by inhibiting or exciting afferent pain signals (18). Examples of this system reducing experience to pain are seen in injured soldiers wounded in battle and mothers during childbirth (60). In these emergency situations, severe pain does not hinder acts of survival (61). The DPMS comprises regions in the forebrain that are also involved in emotional processing, which include the prefrontal cortex (PFC), anterior cingulate cortex (ACC), amygdala, thalamus, hypothalamus and insular cortex (62). Pathways from the cortex project to brainstem nuclei, one of which is the PAG. The PAG is an essential hub for many autonomic functions related to survival (47–49,63– 66) and regulates the raphe nuclei (specifically nucleus raphes magnus), critical to the DPMS. Initially discovered as a key component in downstream analgesic effects through stimulation (67– 13 70), the PAG is also known as a comparator of nociceptive signals, matching pain signals from the periphery to expectations from higher cognitive centers (61), thus further enforcing its critical role as a modulator. In neuroimaging, the PAG is shown to contribute to the “descending pathway for attentional control of pain” in humans, where activity in the PAG significantly increased during a distraction to pain condition (59,71). The PAG’s role in acute pain has been established (59,67,71– 74), however, it is not well understood how changes in regions associated with DPMS may lead to chronic pain. Studies have demonstrated connectivity changes in the PAG to the ventral medial PFC (75) and predicted pain persistence from cortical striatal connectivity (15) in chronic low backpain patients, but further research into chronic pain is needed given its various etiologies. Analgesic effects associated with different subregions of the PAG The PAG has been shown to be associated with different downstream analgesic effects (70,76). The main neurotransmitters clearly controlling descending inhibition to the spinal levels and dorsal horn laminae via the brainstem are serotonin, noradrenaline, and endogenous opioids (58,62,77–80). These neurotransmitters play various roles in excitatory and inhibitory modulation of pain (72,81). Descending pain control within the PAG is mediated by both opioid and non- opioid related receptors within the PAG ventrolateral and dorsolateral (and lateral) regions respectively (56,76). The descending control of pain via opioid receptors is associated with activity that occurs more ventrally in the PAG. Regions in the cortex, particularly the prefrontal areas, are identified in relation to the brainstem in this descending effect. Naloxone has been used to reduce placebo effects, such as placebo analgesia, associated with DPMS regions, showing that the anterior cingulate cortex to PAG connectivity (as well as activation of the rostral ventromedial medulla) was predictive of both pain rating and neural placebo effects (78). In addition to 14 opioidergic control, Huang and colleagues demonstrate that the basolateral amygdala-PFC-PAG connections reduce descending serotoninergic and noradrenergic spinal pain signals in mice in ventral regions of the PAG (82). The dorsolateral PAG (dlPAG) has been associated with non- opioid analgesia (58,76). For example, cannabinoid receptors in the dlPAG in rat models were proven to mediate stressed induced analgesia (83,84). Stressed induced analgesia may be differentially modified in the PAG possibly due to different environmental demands in both animal and human studies (76,85,86). The development of chronic pain may therefore be based on activity that changes the normal function of specific sub-regions in the PAG. The PAG’s acute response to threat and inflammation Chronic pain may be a maladaptive response tied to the PAG’s acute response to threat and inflammation. The relationship between the PFC to the PAG is a prominent factor in the directed response to threat, a critical component to escape and recovery. For example, the PFC affects activity in the PAG related to defensive behaviors during fear discrimination in mice (87). In a neuroimaging task, as threat becomes more immediate, human BOLD activity changes from the PFC to the PAG, where fast escape decisions are associated with higher PAG activity (53,88,89). These neuroimaging studies discuss fear, where there exists a tangible threat, and anxiety, where the threat is abstract and future oriented, with respect to these frontal and midbrain regions. Divergence in behavior depends on environmental demands. Active versus passive freezing is a function to a threat response in animal models, activating the ventral and dorsal regions of the PAG respectively. The opioid mediated ventrolateral (vlPAG) has been activated during passive coping or parasympathetic behavior whereas the non-opioid mediated dlPAG is associated with fight or flight sympathetic response (76). As such, medial PFC projections to the dorsal PAG have 15 been identified and demonstrate weakened functional connectivity in behavioral adaptation to social defeat, a behavioral outcome of threat (90), in mice (91). Franklin and colleagues’ result (91), underscored that inhibitory projections from the medial PFC to the brainstem have a “dual role” in “regulating neuromodulatory tone” in response to threat through more ventral regions involving the dorsal raphe and through behavioral modifications made via more dorsal regions (92). Repeated social defeat has also caused inflammation in the spinal cord and increased pain sensitivity in mice models (93). The vlPAG activity is associated with assisting recovering and healing following and injury (63,76,94), and post-encounter to a threat behavior (76). This response may tie in with the parasympathetic descending vagal influence, important in regulating metabolic homeostasis (95). Neuroanatomical studies have shown that the vagus nerve projects indirectly to the PAG by way of the solitary nucleus (96,97) and that studies stimulating ventral PAG increases heart rate variability and decrease pain and inflammatory response (98). Koenig and colleagues used heart rate variability as a proxy for vagal activity and demonstrated differences in participants with chronic pain and healthy controls (99). Responses to threat and inflammation mediation from the PAG via the vagus nerve may serve as a link to acute to chronic pain manifestation and maintenance of chronic pain disorders. Involvement of the PAG and UCPPS The PAG could play a central role in UCPPS if the acute threat and inflammatory response to injury become impaired. PAG activity may reflect this change in UCPPS as an autonomic control center. In comparing irritable bowel syndrome (IBS) and ulcerative colitis, both chronic symptoms of the pelvic region and gut, ulcerative colitis patients activated pain inhibition networks 16 from the right lateral frontal cortex to the PAG during rectal distention and anticipation of distention (100). This behavior, occurring also in healthy controls, did not appear in IBS patients, suggesting a failure to “appropriately activate an inhibitory corticolimbic system” in IBS. In primary dysmenorrhea and endometriosis, conditions that are risk factors for developing chronic pelvic pain in women, PAG grey matter volume was shown to increase compared to healthy controls (16,101). Interestingly, gray matter volume increases noted in PAG were also found in panic disorder (102). A study of women with UCPPS demonstrated decreased vagal activity in patients and, based on changes in heart rate variability, implied a shift to a more sympathetic dominant state (103). These changes were not seen in participants with myofascial pelvic pain and only existed in those with UCPPS. The rationale behind this distinction in UCPPS patients was a reduced ability to “self-regulate” physiological responses, also implicated in individual differences in interoceptive sensitivity and pain tolerance (104). In another UCPPS study, authors revealed an increase in the connectivity from the PAG to the amygdala in patients with UCPPS compared to healthy controls (62). These studies have shown alterations in brain function and structure of the PAG using neuroimaging in UCPPS patients and chronic pelvic disorders (43), but none have directly compared the whole-brain resting-state connectivity of the PAG. Evidence of alterations related to the PAG in UCPPS along with previous knowledge of differential activation in the ventral and dorsal region of the PAG through descending opioid and non-opioid analgesia, suggests a potential modulatory effect. The PAG, as a key part of the DPMS, is a site of importance in understanding and addressing the longitudinal pain symptoms of UCPPS. This therefore justifies further analysis of the PAG in the UCPPS population. The inability to improve UCPPS symptoms over time may be connected to the reduced ability to self-regulate altered physiological responses and therefore, interferes with the body’s natural healing process 17 after an injury or threat to the system. In addition to Williams and colleagues’ analysis regarding an imbalance of parasympathetic activity in UCPPS patients (103), there are other pain studies showing autonomic response changes. For example, post-intubation of balloon distention in the esophagus is associated with sympathetic activation and reduction in parasympathetic function, determined by various cardiac measures and mean blood pressure (106). For the purpose of survival, this type of noxious stimuli modulates the response in the autonomic nervous system (73) and enhances the sympathetic system response. If there is a shift from more parasympathetic to sympathetic responses as implied by these studies, we would see a decreased ability of the vlPAG to implement post-encounter healing mechanisms and changes to the dlPAG activity as a reaction to heightened stress. Patients whose symptoms improve over time may develop protective mechanisms, as suggested by As-Sanie and colleagues (16) with increased grey matter in the PAG in patients with endometriosis, but no pain. These changes may stabilize the parasympathetic response towards healing or assist in ability to balance autonomic processes, as seen in vagal stimulation with an increased parasympathetic behavior and suppression of the inflammatory response (95). However, patients with the inability to heal, due to a malfunction in the inflammatory response or changes in the cortex, may require more invasive therapies. For instance, changes in regions associated with the “pain matrix” in the cortex where nociceptive-independent pain information is altered (18) may require stimulation therapy, such as with treatment resistant depression (107). In our resting-state analysis, we tested the entire trace of the PAG region. Most functional neuroimaging studies in humans have not specified activity from the subregions in the PAG (47,59,100). Others have erroneously modeled the location, without specification to subregions, from as set of coordinates based on a heat pain task (65,66,75,108–112). This is likely because 3 18 tesla fMRI resolution, commonly used in previous studies and our own, and the voxel size sampled are not adequate to define the subregions (47). With 7 tesla, visualization of distinct subregions is possible and has been confirmed in animal studies (113), as well as differentiation between painful and innocuous stimulation (114). Even so, susceptibility effects and artifacts such as physiological noise can also be greater at higher MRI field strengths (115,116). Furthermore, MRI lacks the ability to precisely distinguish activity in nucleus subregions that can be achieved with techniques such as optogenetic modulation (92) and tracer imaging (117). However, these methods are not possible in human studies. Subregions involved with improvement or non-improvement in UCPPS symptoms is still unknown. The 3 tesla and 3mm voxel resolution of our dataset limits our main analysis to the PAG as a whole. As a potential modulator for UCPPS symptom change, we focused on the PAG as a seed in whole-brain connectivity analysis. Chapter Summaries This main research work of this thesis will be divided into three chapters. Each chapter is summarized below: Chapter 2 presents an analysis of longitudinal UCPPS symptom change which we use as the outcome measures for our prediction. This analysis seeks to identify the groups of patients whose symptoms improve, worsen or remain stable from the symptom reporting. Chapter 3 is an extensive analysis on the PAG location. We found that PAG coordinates reported in visceral pain literature appeared to be several millimeters anterior to the anatomical location of the PAG. Before moving into our prediction studies based on PAG connectivity measures in rs-fMRI, our goal was to determine whether measures of PAG functional connectivity were sensitive to the localization technique used. Chapter 4 is our final analysis and our prediction of our longitudinal symptom 19 change using baseline resting state functional magnetic imaging (rs-fMRI). In this chapter we use the outcome measures defined in Chapter 2 as the target prediction measure and the PAG location defined in Chapter 3 to whole-brain regions define our connectivity features. Chapter 5 is a discussion of the final prediction results and their future work implications. This section covers review of the machine learning techniques used and biomarker discovery and potential application. 20 Chapter 2: Longitudinal UCPPS symptom change as a predictive outcome measure Introduction UCPPS is marked by heterogenous etiologies and variability across symptom progression (22,24,33,118). A hallmark of the Multidisciplinary Approach to the Study of Chronic Pelvic Pain Research Network (MAPP) was to distinguish symptom patterns in patients and identify clinically relevant subpopulations that could be targeted with precise treatment. As a main focus of MAPP, neuroimaging measures have been used in conjunction with clinical symptoms to better understand potential subpopulations. Likewise, other studies outside of pain, such as depression (119), have attempted to combine neuroimaging measures with reported clinical symptoms to define biotypes and to characterize subpopulations. Recently, MAPP has used a consensus clustering algorithm to group participants according to characteristics defining patients symptoms (120), such as baseline urological symptoms, pain sensations (121), and body areas of reported pain (122). In our analysis of PAG activity in UCPPS (see Chapter 4), we extend the prediction of the relationship between resting state neuroimaging measures and the change in patient pain scores over time (39). In this chapter we define the symptom progression slope which is used in our analysis as the outcome measure that we aim to predict. We report these scores as a measure of change of reported symptoms up to a year. We further aim here to determine whether we can separate UCPPS symptoms into distinct categories of improving, stable, or worsening (123). These measures will be used as relative markers for subgroups in our prediction analysis. 21 Methods MAPP Phase II dataset overview The MAPP Research Network follows a systemic and longitudinal approach to the study of Interstitial Cystitis (IC)/ Bladder Pain Syndrome (BPS) in men and women, and Chronic Prostatitis (CP)/Chronic Pelvic Pain Syndrome (CPPS) in men. Established by the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) of the National Institutes of Health (NIH), data is collected across multiple sites. In the second phase of MAPP (MAPP-II), research carried in MAPP-I was continued, following UCPPS participants for up to 3 years (40). Emphasis is placed on understanding particular subgroups’ pathophysiological profiles that may be treated with clinical interventions (see the MAPP website: https://www.mappnetwork.org/). Data for our study from MAPP comprises phenotyping data from baseline, or when patients first visit, through 12 months reported symptoms. See Clemens et al. (22,43,44) for more detail on the MAPP collection. Participants The dataset for this analysis included 389 participants from the MAPP-II cohort, with 129 males and 235 females. Patients received a clinical diagnosis of UCPPS and were 18 years or older to be included in the study (for more specific inclusion and exclusion of symptoms, please see Landis et al. (32) and Clemens et al. (44)). MAPP-II participant data was collected across six sites: Northwestern University, University of California, Los Angeles, University of Iowa, University of Michigan, University of Washington, and Washington University, St. Louis. Participant data was organized by the Data Coordinating Center (DCC). All participants provided informed consent 22 according to the Declaration of Helsinki and the institutional review board approved the collection at each site, as reported in MAPP-I (22,32,124). Measures for analysis Pain intensity score Measures used in this study consisted of the Genitourinary Pain Index (GUPI), Interstitial Cystitis Symptom Index (ICSI) and Problem Index (ICPI), which provide pain and urinary symptoms as well as quality of life measures (20,125,126). The overall pain symptom score used in our study is a composite of the GUPI questionnaire (total 0-23 points from questions) and ICSI question “Have you experienced pain or burning in your bladder” (total 0-5 points from question), for an overall used in this study of 0-28 points (39). Questionnaires were collected at baseline and every three months post-baseline producing a total of 5 pain score values (baseline, 3 months, 6 months, 9 months, and 12 months) per participant across a 12-month time frame. Baseline pain score measures were measured after the first 4 weeks of enrollment, as previous symptom assessments have shown a regression to the mean upon initial recruitment (45). In this analysis we focus on the difference between baseline and 12 months pain score to address the primary goal of this study, which is novel to MAPP. We have included 3-month pain score to validate our previous results in longitudinal symptoms (seen in 3-month symptom Figures 2.2, 2.4, 2.6, 2.7, and 2.9). Furthermore, though UCPPS is marked by both pain and urinary symptoms, these measures should not be combined into one score (42). For the purpose for this study and the results from our previous prediction (39), we focused only on pain score. 23 Covariates Factors that may influence longitudinal symptom change in our prospective neuroimaging analysis are age, sex, and baseline pain. Age has been associated with decreased stress in relation to the duration of UCPPS symptoms (126), while both age and sex were baseline factors associated with lower likelihood of improvement in physical and mental health-related quality of life in UCPPS (120). In addition, the extent of widespread pain at baseline is predictive of worsening pain symptoms (127). Treatment history A goal of the second phase of MAPP was to analyze effects of treatment reported to better understand the etiology and treated natural history of UCPPS (128). Prior MAPP-I collection observed treated natural history of patients. In our study, we used longitudinally recorded treatment history at baseline, 6 months, and 12 months. Treatments reported included acupuncture, alpha adrenergic blockers, bladder installation, bladder training, Botox, chiropractic treatment, cystoscopy, dietary changes, Elmiron, heat or cold application, home exercise (including yoga), message, neuropathic pain treatment, oral opioids, pelvic floor physical therapy, ATLAS therapy, counseling and psychotherapy, generalized physical therapy, sacral neuromodulation, and tricyclic antidepressants. These treatments were reported as ever taken, recently taken (within 3 months), and actively taken (within 1 month). 24 Analysis Primary outcome measure: Slope calculation and assessment The purpose of this analysis was to determine whether we could detect meaningful outcome measures for our prediction of longitudinal symptom change. Figure 2.1 shows an example of the pain intensity score reported across 5 visits for two subjects with differing symptom outcomes up to a year. Participant A’s symptoms improve over time, while participant B’s symptoms worsen. The slope of these scores is calculated as the change of reported symptoms from baseline up to 12- months (3-month example slopes are shown in Figure 2.2). These measures were based on our previous analysis set up for prediction over 3, 6, and 12 months on MAPP-I data (39). To estimate the slope over 12-months, the DCC provided a linear mixed effects model (calculated in SAS) with a random intercept and random slope. Slope in pain score was also estimated after the first 4 weeks of enrolment due to the regression to the mean effects (45). The slope was negative for improving symptoms, or decreasing pain score, and positive for worsening symptoms, or increasing pain score. These slope values were used to split patients with improving (negative slope), stable (close to zero slope), or worsening symptoms (positive slope) by rank ordering the 12-month slope values and separating the data into three categories based on the distribution tertile. Multiple regression was used to assess the potential impact of age, sex, and baseline pain prior to splitting each category. Site effects from the 6 collection sites were assessed by observing the relationship between, improving, stable, or worsening symptoms across each site. The 3-month slope values were evaluated with the same analysis as applied to 12-month slope. 25 Figure 2.1: Figure 2.1: 12-month symptom progression slope. Example participant longitudinal pain trajectories for 12-month reported symptoms selected based on their 12-month symptom change. The slope of these pain scores is the measure we used to define whether a participant improves or does not. These pain scores were derived from a combination of the Genitourinary Pain Index (GUPI), Interstitial Cystitis Symptom Index (ICSI). Slope was calculated 4 weeks after enrollment due to regression to the mean effects. Figure 2.2: Figure 2.2: 3-month symptom progression slope. Example participant longitudinal pain trajectories for 12-month reported symptoms selected based on their 3-month symptom change. The slope of these pain scores is the measure we used to define whether a participant improves or does not. These pain scores were derived from a combination of the Genitourinary Pain Index (GUPI), Interstitial Cystitis Symptom Index (ICSI), and Interstitial Cystitis Problem Index (ICPI). Slope was calculated 4 weeks after enrollment due to regression to the mean effects. 26 Treatment effects The effect of specific treatment was an important measure second phase of MAPP. Maintenance of UCPPS symptoms may be due to centralized pain. Centrally acting treatments may mediate chronic pain symptoms and may be markers of improvement (43). We focused on consistently taken treatments, or treatments that were actively taken, over the 12-month timeframe. We were also interested in treatment effects in general and its influence on symptom trajectory. We did not place emphasis on a specific drug or type of treatment due to limitations in the number of participants taking a specific treatment. We provide a summary of the patients consistently taking the range of treatments recorded. Our main analysis was to determine whether these treatments influenced patient improvement over 12-months. Results The UCPPS outcome measures for our prediction analysis are based on the slope of the pain intensity score over time (see Chapter 2, Methods, ‘Primary outcome measure: Slope calculation and assessment’). Here we focused on 12-month measures, however parallel results from our 3-month measures can been seen in 3-month results Figures 2.4, 2.6, 2.7, and 2.9. In our multiple regression analysis, we examined the influence of age, sex and baseline pain on our 12- month measures. Age has a significant linear relationship with the slope values (p=0.022), however, this was lost when baseline pain is included in the model (p=0.155). Age and sex are significantly correlated with baseline pain (age: p=0.001; sex: p=3.0e-04), suggesting multicollinearity in the dataset. Additionally, we observed an interaction with sex and baseline pain for the 12-month slope (p=0.004). Baseline pain had the strongest effect on our slope measures (p=3.0e-06) (see Figure 2.3). The results seen with 3-month outcome measures, also 27 show an association to baseline pain, which we have accounted for in our final model for 3-months (Figure 2.4). Figure 2.3: Figure 2.3: Baseline pain score is associated with 12-month slope values. Blue dots show participants’ baseline pain score (N=387) taken at the end of the 4-week run-in period linearly related to 12-month slope values (p=3.0e-06). The red line shows the line of best fit to the data. Figure 2.4: Figure 2.4: Baseline pain score is associated with 3-month slope values. Blue dots show participants’ baseline pain score (N=387) taken at the end of the 4-week run-in period linearly related to 3-month slope values (p = 3.0e-06). The red line shows the line of best fit to the data. p = 3.0e-06 p = 8.7e-07 - - 28 To control for the potential effect from age, sex, and baseline pain, we corrected for these factors in our model of UCPPS outcome measures by including these covariates in a linear model with the slope value as the dependent variable. The corrected slope values are equivalent to the residuals of this linear model. For the purpose of our neuroimaging predictive study (Chapter 4) we decided to include these factors in our model. The number of participants was reduced to 387 because two participants did not have data at baseline, and so their slope values were incomplete. After correction, the residual slope values were split into three categories: improving, stable, and worsening symptoms. We classified more negative slope values as improving symptoms and more positive values as worsening symptoms. Figure 2.5 shows percent change in pain score across all participants from baseline to 12-months (3-months in Figure 2.6). Percent change has been used previously in longitudinal pain studies as a measure of improvement, stability, or worsening symptoms in chronic low back pain (15). For 12-month slope trajectory, participants symptom scores improve with a -46.3% change, are stable with a +8.4% change, or worsen with a +19.1% change from baseline. Figure 2.7 A and Figure 2.7 B show the distribution between these three categories at 12-month and 3-month slope values. This trend is maintained within each site and there are no significant differences between slopes per site when pooled (12-month, Figure 2.8, and 3-month, Figure 2.9). The 3-month averaging of pain scores, Figure 2.6, shows that after grouping the data into improving, stable, and worsening symptoms at the 3-month timeframe, pain- scores average towards a more stable trajectory by the 12-month timeframe. This trend is likely due to the variability in pain score as seen in Figure 2.2. 29 Figure 2.5: Figure 2.5: Average pain score split into three groups by 12-month slope value: patients that improve (blue), stay stable (green), or worsen over time (red). Those that improve, show an average percent change of 46.3% from their baseline score (at 4 weeks, shaded area) to 12-months (48 weeks) Those that stay stable or worsen show a percent change of 8.44% and 19.1%, respectively. Figure 2.6: Figure 2.6: Average pain score split into three groups by 3-month slope value: patients that improve, stay stable, or worsen over time. Those that improve, show an average percent change of -38.8% from their baseline score (at 4 weeks) to 3-months (12 weeks), whereas, those that stay stable or worsen show a percent change of +3.6% and +21.1%, respectively. Improve by 46.3% Improve by 38.8% 30 Figure 2.7: Figure 2.7: Residual slope value distribution. A comparison of A. 12-month slope values and B. 3-month slope values within participant data split into categories of improving, stable, and worsening symptoms. Box plots provide a comparison of each category, with the central mark as the median and outliers plotted as a red ‘+’ symbol. 31 Figure 2.8: Figure 2.8: One-way ANOVA comparison of 12-month slope values at each site in the MAPP-II dataset. Box plots provide a comparison of each site, with the central mark as the median and outliers plotted as a red ‘+’ symbol. Key: NW, Northwestern, UCLA, University of California Los Angeles, UI, University of Iowa, UMich, University of Michigan, UW, University of Washington, WashU, Washington University in St. Louis. 32 Figure 2.9: Figure 2.9: A one-way ANOVA comparison of 3-month slope values at each site in the MAPP-II dataset. Box plots provide a comparison of each site, with the central mark as the median and outliers plotted as a red ‘+’ symbol Key: NW, Northwestern, UCLA, University of California Los Angeles, UI, University of Iowa, UMich, University of Michigan, UW, University of Washington, WashU, Washington University in St. Louis. 33 Treatment effects were based on the 12-month timeframe. In particular, we were interested in consistent treatment and whether these treatments influenced the improvement of symptoms. Figure 2.10 A shows the treatments that patients consistently underwent during data collection. The largest number of patients underwent diet changes. Even so, the number of participants taking a specific treatment consistently is a small fraction of the total number of participants (e.g. (diet changes N=57 of 389 total participants). To understand the effect of these treatments on the improving, stable, and worsening slope trajectories, we compared the number of participants within each category that took consistent treatment with the total number of patients in each category (Figure 2.10 B). The difference between patients taking a consistent treatment with improving symptoms relative to the total number of patients with improving symptoms is not significantly different from those with stable or worsening symptoms as calculated using a Chi- square test (X2 (1, N = 258) = 0.018, p=0.893) (Figure 2.10 B). Figure 2.10: Figure 2.10: Treatment effects. A. The number of subjects who consistently reported an active treatment at all of baseline, 6 months, and 12 months reporting. The active treatments with the largest number of participants are in bold: antidepressants, dietary changes, heat/cold treatments, and home exercise, including yoga. B. The total number of participants compared to the number of participants taking consistent treatment (orange) who are separated into each category of improving, stable, or worsening symptoms. The difference between consistent treatment within improvers compared to participants that worsen is not significant (p=0.893). 34 Discussion We identify and define the potential subpopulations in our analysis of longitudinal UCPPS symptom change up to a year. Pain score was used to distinguish subgroups associated with potential predictive outcome measures. We categorized three groups, those who improve, those who remain stable and those who worsen. We use the slope of the linear fit between baseline and 12-month and 3-month pain score to establish a trajectory separating improvers from non- improvers. The main goal was to determine a possible behavioral pattern to predict using our baseline neuroimaging measures in a further analysis (see Chapter 4). The method to define these three categories, was used to distinguish the extremes in the dataset, in particular those that improved over time. We focused mainly on pain symptom trajectories because of our previous prediction MAPP study (39). Urinary severity is also highly descriptive of the GUPI, ICSI, and ICPI surveys, but it was previously shown that this was not a strong predictive measure (39,42) compared to pain. Baseline pain proved to be strongly associated with our symptom trajectories, and while this measure may provide insight on the trajectory of pain symptoms, it does not provide information on how we may treat symptoms. Therefore, the logical next step in this analysis is to use neuroimaging to understand the correlates in the neural signal, measured by fMRI, that may lead to improvement. Multicollinearity of age and sex was not shown to influence symptom trajectory and does not detract from the predictive power from our analysis. Furthermore, we control for these covariates in our final model as an extra precaution, as age and sex have been controlled for in other neuroimaging studies on pain due to their association with symptoms (25,33,37,129). The interaction effect from sex on baseline pain relative to duration of symptoms has been seen before in MAPP literature (126,127). Although we did not cover chronic overlapping 35 pain conditions or other non-urologic measures, other studies have shown their predictive power of longitudinal symptoms (127,130). Differing 12-month trajectories for patients with improving and worsening symptoms based on their 12-month and 3-month slope values are shown in Figure 2.1 and Figure 2.2. This demonstrates that a patient who improves within the 3-month time frame can also have symptoms that worsen over the 12-months. However, another patient may not improve as drastically within the 3-month timeframe, but over 12-months steadily improve. These different patients may show differing neural correlates to baseline neuroimaging for short-term and long-term improvement. Our previous MAPP-I prediction demonstrated significant prediction between whole brain neural activity and pain slope at 3-months, but not at 12-months. We were limited to 52 participants, and the slope values were dichotomized into improvers and non-improvers based on their median. We have extended these results here to tripartite the data into 3 groups of improving, stable, and worsening symptoms. However, prediction of these outcome measures may be better suited to a continuous rather than discrete model. Though we see a significant trend in the dataset, measured by percent change (a measure used in other pain and UCPPS studies (15,45,130)), the boundaries between categories are not distinct and so we may see a loss of information with our method of triparting the data, and a lowered sensitivity towards improvers. For example, due to the continuous distribution of the outcome measure, we may categorize a participant as having stable symptoms rather than improving. A recent study by Locke and colleagues has provided a more in- depth cluster analysis (120), using more measures than just pain symptoms, including urinary and body map scores, and unsupervised k-means clustering of UCPPS patients into K=2 groups. Furthermore, another MAPP study (123) has used K-class functional mixed effects clustering on 36 longitudinal symptoms into improving, stable, and worsening symptoms. Instead of triparting rank ordered data, by using a form of clustering, we may avoid potential misclassification. For the purpose of this analysis as a validation and extension of our previous MAPP-I findings, we followed similar procedures to those established in the previous MAPP-I prediction, for further use with rs-fMRI. The importance in this study is that we established significant trends via our slope values that we can use as markers for clinically relevant outcomes, i.e. what are the characteristics that define a patient who will eventually improve. In this chapter we observe potential trends in the data based on triparting the distribution of the outcome measure. This method allows us to view extremes of the distribution and categorize participants into equal groups. Though due to the continuous nature of values, prediction should be applied on continuous data and not on these discreet categories. Even so, categorization may be necessary for prospective interpretation and potential diagnostics or intervention. Conclusion Meaningful outcome measures were constructed from longitudinally recorded UCPPS symptom change. The change of symptoms up to a year, calculated as slope score compared to baseline, was split into three distinct groups based on the trajectory of pain measurements. These significant trends provide us with subgroup of UCPPS patients whose symptoms improve, stay stable or worsen. Moving forward with these subgroups, we can investigate possible neural correlates related to improvement or non-improvement in symptoms UCPPS up to a year. 37 Chapter 3: Sensitivity of functional connectivity to periaqueductal gray localization Introduction The periaqueductal gray (PAG) is considered one of the most important centers of activity in the brain, integrating and controlling behavior crucial for survival (47–49,63–66). Functions specific to the PAG in humans have been attributed to autonomic control (47–52), such as with micturition and cardiovascular regulation, but also relate to functions as diverse as fear and anxiety (47,53,54), sexual behavior (55), and antinociception (56). The PAG connects to the medial prefrontal cortical regions in Macaque monkeys (48,51) and governs descending pain modulation pathways in humans (58,59). These connections support an anatomical basis for top down control of spinal and sensorimotor circuits that could be particularly important in chronic pain disorders. Defining brain regions and their relationships noninvasively through neuroimaging, such as with functional connectivity, has gained traction in recent literature (49,131–134). This approach is fundamental to understanding the brain function that may underlie human behavior (133,135). Much work has been done to define the PAG in animals and humans invasively (48,51,56,64,67,117,136), but a benefit to neuroimaging is that it can measure human behavior in vivo, noninvasively (47) (see Linnman et al. for a review on neuroimaging PAG). Anatomically, the PAG, a small, curved, hollow partial cylinder approximately 10-14 mm long and 4-5 mm in external diameter encircling the central aqueduct in the midbrain (n.b. the nuclei ventral to the aqueduct are distinct from the PAG), can be identified visually using high-resolution structural magnetic resonance sequences (72,78,113). However, the cytoarchitectonic pattern and the behavioral role of the four longitudinal columns within the PAG, the dorsomedial, dorsolateral, 38 lateral, and ventrolateral, well-established in invasive animal studies, remains unclear in humans (47,49,72,134). Recent work has attempted to define these subregions of the PAG using data- driven rs-fMRI connectivity methods (49). Another approach activates the region through known functional stimuli, pain, fear, or autonomic regulation changes. The challenge with this type of measurement is in anatomically localizing the region of activation, optimally performed with additional landmarks and across a range of experimental conditions (72,135,137,138). At this juncture, functional magnetic resonance imaging (fMRI) still provides marginal spatial resolution and significant scanner and physiological artifacts (47,72,78,113) when identifying brainstem regions. Many functional neuroimaging studies have used PAG coordinates as a seed or region of interest (ROI) analysis based on previously established functional connectivity measures. However, a number of these studies (65,66,75,108–111,139,140) have modeled the PAG as a sphere, despite its curved cylindrical shape, around coordinates located several millimeters anterior to the anatomical location of the PAG. These coordinates were localized in a low and high heat pain fMRI contrast (141) and validated in a follow-up study analyzing the functional connectivity of the ventrolateral region of the PAG (65). Though the ventrolateral region is justified in this study, others have used these coordinates to represent the whole PAG. The discrepancy between the functional and the anatomical location of the PAG leads us to further establish benchmarks that define the true PAG and discern approaches that render functional neuroimaging more reproducible and robust. We defined three different localization techniques for the PAG as a whole, including coordinates from the literature, and compared each ROI with various resting-state neuroimaging approaches. These three ROIs were labeled the MNI-sphere, coordinates from literature 39 surrounded by a 3 mm radius sphere, the MNI-trace, a hand-traced ROI of the PAG from a MNI template, and the participant-trace, another hand-traced ROI in the structural space of individual participants (this trace was assumed to represent the optimal rendering of an individual’s true PAG region, or “gold-standard”). We measured differences in extracted time series signals of resting- state functional neuroimaging data from healthy controls, analyzing the signal directly from the PAG ROI defined in these three ways, and compared the functional connectivity throughout the brain across methods. For this study, we also demonstrated that the location of the ROI affects the differences seen in voxelwise whole-brain functional connectivity between urologic chronic pelvic pain syndrome (UCPPS) patients and healthy controls. We hypothesized that the differences that emerged between healthy control subjects and patients would differ based on the anatomic region identified as the ‘PAG’. UCPPS patients were used in previously published cross-sectional functional neuroimaging studies with involvement of the PAG (105), but this is the first study to directly compare the whole-brain resting-state connectivity of the PAG in such a population. Methods Participants Our initial cohort consisted of 318 participants from the first phase of the multi-site Multidisciplinary Approach to the Study of Chronic Pain (MAPP) Research Network study§, selected based on a clinical diagnosis of UCPPS or as a healthy control without a history of UCPPS. Participants were recruited from five collection sites: Northwestern University, University of California, Los Angeles, University of Michigan, University of Alabama at Birmingham, and Stanford University (22,33,34). All participants provided informed consent according to the Declaration of Helsinki and the institutional review board approved the collection 40 at each site (124). Selection criteria were based on whether neuroimaging data met specific quality standards as regards to motion and dataset balancing protocols. The first step in our analysis examined PAG location in healthy controls. These participants were selected with the following goals in mind: 1) equalizing number of males and females, 2) spanning a wide age range, and 3) distributing across the five sites equally to account for any differences in acquisition technique. Due to the time-intensive work of hand tracing the PAG voxel by voxel for each participant, we selected 15 healthy controls (Table 3.1, N=9 female, N=6 male) after excluding three participants (N=1 female, N=2 males) due to image preprocessing issues. Table 3.1: Demographic characteristics of participants age (y) sex (male/female) site Healthy control datasets Test (N=15) 37.6±19.1 6/9 NW, UCLA, Michigan, UAB, Stanford Validation (N=15) 37.1±11.4 7/8 UCLA Patients vs. healthy controls datasets Patients (N=100) 39.2±13.3 34/66 NW, UCLA, Michigan, UAB, Stanford Healthy controls (N=109) 36.7±12.2 34/75 NW, UCLA, Michigan, UAB, Stanford Table 3.1: Demographic characteristics of participant datasets in Chapter 3. Key: NW, Northwestern University, UCLA, University of California Los Angeles, UAB, University of Alabama at Birmingham. 41 An independent dataset (Table 3.1, N=8 female, N=7 male) with all participants from the University of California Los Angeles, was matched to our first dataset for validation. For clarification, these two datasets are labeled as test, which is our first dataset, and validation. For our validation dataset, we followed the same procedure as our test dataset, but the selection was made from only one site to verify that the results from our test dataset were driven by the localization technique and not site differences. For example, if site differences drive our findings, the test and validation datasets would differ significantly since they originated from different site groupings. Participants from the validation dataset were selected only after removing participants included in the test dataset from the University of California, Los Angeles. The validation dataset was also matched in size and balanced by the number of males and females to the test dataset. Age is known to be related to volumetric variability of the brainstem (142). Maximum deviation in age, an iterative selection of participants with the largest age difference, was calculated in the selection of participants in both test and validation datasets for males and females separately. The age distribution was not significantly different in our test and validation datasets (p=0.92). In our secondary analysis, our initial 318 MAPP cohort was reduced to 209 (Table 3.1, N=100 patients, N=109 healthy controls), based on the conditions described by Landis and colleagues (32) characterizing UCPPS patients. Furthermore, only natural history was recorded in MAPP (22) and we did not control for treatment factors within patients. Inclusion criteria included 1) UCPPS or healthy control participants and 2) participants that passed image preprocessing quality control. This resulted in the exclusion of 68 positive control patients with non-urologic associated syndromes (32) and 41 participants based on image preprocessing failures. Focus was placed only on neuroimaging measures and no symptom data were utilized in study analysis. 42 Image collection All 3D T1-weighted structural MRI and resting-state fMRI data were collected and quality controlled by MAPP on 3 tesla scanners (33). A high resolution structural T1-weighted magnetization-prepared rapid gradient-echo (MP-RAGE) pulse sequence was used to collect the data from each participant from all sites except Stanford University, which used an inversion- recovery fast spoiled gradient echo (IR-FSPGR) sequence. The University of California, Los Angeles collected slices with interleaved sequence, all other sites collected slices sequentially. Parameters for structural MRI collection included a repetition time (TR) = 2200 ms, echo time (TE) = 3.26 ms, slice thickness = 1 mm, 176 slices, 256x256 acquisition matrices, and voxel resolution = 1x1x1 mm (33,37,124). The participants closed their eyes for 10 minutes during the rs-fMRI, with 40-slice whole-brain volume, slice thickness = 4 mm, TR = 2000 ms, TE = 28 ms, flip angle = 77o, and FOV = 220. Further details are published in previous MAPP Research Network studies (33,143). Image processing Resting-state fMRI data were preprocessed with a software package developed by the Analysis of Functional Neuroimaging (AFNI, https://afni.nimh.nih.gov/, Version AFNI_18.1.18, (144)) with parameters adjusted according to Plitt and colleagues and Drysdale and colleagues (119,145,146). One of the benefits for this pipeline is its stringent correction for distance- dependent motion artifacts and use of local white matter regressor (AFNI’s ANATICOR). First, FreeSurfer (https://surfer.nmr.mgh.harvard.edu/) is used to segment the T1-weighted image and generate the white matter and ventricle masks. The rs-fMRI image was preprocessed by removing the first four volumes, despiked, slice-time corrected, co-registered to the T1-weighted image 43 output from FreeSurfer, automatically resampled to 3 mm isotropic voxels, normalized to reflect the percent signal change, non-linearly registered to the MNI template scan, and spatially smoothed with a 6 mm full width half-maximum Gaussian kernel (119,146). Nuisance variables included the average ventricular time series, local average white matter time series (AFNI’s ANATICOR), twelve parameter estimates for head motion, including their first derivative (145). The residual BOLD time series from the rs-fMRI images are simultaneously bandpass filtered at 0.01-0.1 Hz and nuisance regressed (AFNI’s 3dTproject). The time series is censored by removing volumes exceeding the Euclidean norm of the motion derivative at 0.3 mm or volumes with outliers, defined as time points a certain distance from the trend of the time series (see AFNI’s 3dToutcount), whose voxels exceed 10% of the masked voxels. The median number of volumes removed was 2 across participants, with a maximum number of 84 (119,146) . PAG ROI definition Three different localization techniques were used to define PAG ROIs. As seen in Figure 3.1 A, these ROIs are defined as MNI-sphere, MNI-trace, and participant-trace. The MNI-sphere was a 3 mm radius sphere with a location previously published as MNI coordinates for the PAG, (MNI: 4 -26 -14). The MNI coordinates for the PAG can be found in Wei et al. (66). The MNI- trace PAG ROI was created from a 1 mm MNI152 standard-space T1-weighted average structural template image and the participant-trace PAG ROI was identified on each participant’s structural T1-weighted image in 1x1x1 mm grid-space using a combination of AFNI and FreeSurfer’s Freeview. Consultation with several PAG morphology and anatomy experts helped us establish the anatomic margins of the PAG for the MNI-trace and participant-trace. The caudal margin was considered as the location where the cerebral aqueduct no longer appeared as a circle, due its 44 opening into the fourth ventricle. The rostral border of the PAG was defined as the point at which the cerebral aqueduct had definitively formed in the axial image from the third ventricle. This can be delineated using a combination of axial and sagittal images to determine the inferior border of the third ventricle. The general morphology of the PAG resembles a three-dimensional kite, which tapers as the cerebral aqueduct enters the fourth ventricle. We did not fix a set PAG length due to variations in brain size, which may affect the length of the cerebral aqueduct and the PAG, but rather allowed the border definitions to provide this information, intrinsic to the brain-shape of the individual participant. Figure 3.1 Figure 3.1: Analysis of the BOLD rs-fMRI time series signal from the three different localization techniques of the PAG. A. Top left: An 3D illustration of the brain highlighting the brainstem. Top right: The average normalized rs- fMRI across participants in the healthy control dataset. Bottom: The three different regions of interest (ROIs) for the PAG: a previously published Montreal Neurological Institute (MNI) set of coordinates surrounded by a 3mm radius sphere (MNI-sphere in green, shown top right in average image), a hand-traced ROI in a MNI template brain (MNI- trace in blue, shown top right in average image), and a hand-drawn ROI in each of the structural images from healthy control participants (participant-trace in red). B. Example rs-fMRI time series signals comparing the MNI-sphere ROI the MNI-trace ROI to the participant-trace ROI. C. The differences in connectivity across all healthy controls (test dataset, N=15, p<0.001 and validation dataset, N=15, p<0.0001). The connectivity is given by the standardized correlation of the signal extracted from the MNI-sphere ROI to the participant-trace ROI and from the MNI-trace ROI to the participant-trace ROI. The T1-weighted MPRAGE image allowed for optimal differentiation of grey and white matter to identify the PAG. Using AFNI and FreeSurfer’s Freeview to trace the PAG, the outline of the PAG ROI was identified using visual differentiation between grey and white matter, guided **** 0. 1 0.8 *** -10 0 10 -10 0 10 20 sec BOLD signal participant-trace MNI-sphere MNI-trace MNI-sphere participant-trace MNI-trace participant-trace B. Connectivity measure (z-score) C. MNI-sphere: participant-trace MNI-trace: participant-trace Test Validated MNI-sphere MNI-trace Averaged rs-fMRI A. 45 by the understanding of general PAG morphology provided by our PAG experts. We consulted several individuals who work with the PAG on a regular basis and whose work has required its identification from different perspectives. These experts included one of the foremost brainstem anatomists, Dr. Cliff Saper at Harvard, a neuroradiologist with particular expertise in the brainstem, Dr. Scott Rand at MCW, and a physiologist who entirely focuses her work on PAG function, Dr. Caron Dean-Bernhoft. We met with them together on a regular basis until a consensus was reached by all 3 of the exact anatomic boundaries of the PAG on the anatomic images. Literature which informed our anatomic localization of the PAG included Linnman et al., and Satpute et al. (47,113). In instances of uncertainty about the definitive border of the PAG, a more conservative approach was taken: voxels that could not be identified as either PAG or surrounding white matter were not included in the PAG ROI mask. The cerebral aqueduct was then outlined and the area of the PAG was filled in. After the PAG ROI mask was created, the edges were further refined on a separate day. In total, each mask took approximately 1.5 hours to complete. This process was completed for the PAG ROI of the MNI template, or the MNI-trace, as well as the 30 individual masks, or the participant-traces, from MAPP. To assess intra-rater and inter-rater reliability, two separate manual tracings of the PAG were completed by a single rater for 10 participants and by two independent raters for 4 participants in a separate dataset. All these separate sets of tracings were performed in different random orders and the raters were blinded to the order. The degree of spatial overlap between the two traced ROIs for each participant was measured using the Dice similarity coefficient (147), which ranges for 0 (no overlap) to 1 (perfect agreement). The mean intra-rater Dice coefficient (+ standard deviation) was 0.83 + 0.05 (range: 0.76 – 0.91), and the mean inter-rater Dice coefficient was 0.80 + 0.02 (range: 0.78 - 0.83), which both suggest strong agreement. 46 PAG ROI signal comparison To compare the three different PAG localization techniques, we looked for non- uniformities in the extracted rs-fMRI brainstem signals from each ROI in healthy controls separately in our test (N=15) and validation datasets (N=15). A combination of FMRIB Software Library (FSL, https://fsl.fmrib.ox.ac.uk/fsl/fslwiki) and AFNI functions were used to transform PAG ROIs into the participants’ preprocessed rs-fMRI 3x3x3 mm grid-space and then to extract signals from the residual BOLD time series. The signal from the MNI-sphere ROI was directly extracted, while the MNI-trace ROI was output to a coordinate text file based in MNI space (AFNI’s 3dmaskdump). To transform the participant-trace ROI into the correct image space for signal extraction, a similar technique to the registration of the white matter mask to standardized participant specific space was applied. With this procedure, FSL’s FLIRT was used to register the output of the anatomical image from FreeSurfer and AFNI’s 3dNWarpApply, a nonlinear 3D warping technique, into the preprocessed rs-frmi image space. Extracted signals were averaged for further analysis. We assumed the participant-trace ROI to be the “truest” measure of the anatomical location of the PAG because the ROI, a structure that is easily identifiable visually, is traced in anatomical space specific to the participant. The MNI-sphere and MNI-trace extracted signals were therefore each compared to the participant-trace signal using Pearson’s correlation (Figure 3.1 B). Differences between the resulting correlation coefficients were determined after standardization using Fisher z-transformation with a paired t-test. 47 Whole-brain connectivity Whole-brain connectivity analysis allowed us to identify quantifiable differences throughout the brain between the MNI-sphere and MNI-trace ROIs with reference to the participant-trace ROI. For this analysis, voxelwise signals were extracted and averaged in each of the Power 264 atlas regions (148) from the residual BOLD time series output file from AFNI in healthy controls. Individual voxelwise signals were also extracted without averaging from a standard template brain-mask covering the entire brain, including areas of the brain not covered in the Power 264 atlas. Connectivity error was first assessed in a test dataset (N=15) and then validated in a matched dataset (N=15). Connectivity error was defined as the standardized correlation value from each signal extracted from the Power 264 atlas regions to the participant-trace signal (z1, the “gold standard”) subtracted from the standardized correlation values from each signal extracted from the Power 264 atlas regions to the MNI-sphere signal (z2) or the MNI-trace signal (z3) (Figure 3.2 .A). The interquartile range (IQR) across participants and distance of the median (D) of the participants from zero were measures used to ascertain connectivity error. For example, in Figure 3.2 B, smaller differences (i.e. smaller IQR and D) are seen in the connectivity error of the MNI- trace as compared to the MNI-sphere for an extracted signal (MNI: x = 44, y= -53, z = 47) from the Power 264 atlas. In this case the connectivity of the MNI-trace ROI shows better performance in estimating PAG connectivity compared to the connectivity of the MNI-sphere ROI across participants for this particular region, as the MNI-trace ROI has a smaller connectivity error in relation to our reference trace, participant-trace ROI. A larger connectivity error (i.e. larger IQR and D), would indicate worse performance. In the voxelwise analysis, the same comparison was 48 applied to individual voxels using the IQR and D for each extracted voxel instead of each Power 264 atlas region. Figure 3.2: Figure 3.2: Whole-brain functional connectivity analysis of three different localization techniques of the PAG. A. A measure of functional connectivity from each PAG ROI to 264 regions throughout the brain defined by the Power 264 atlas. An example of rs-fMRI BOLD residual time series signal from an atlas region and its connectivity measure to each PAG ROI. The connectivity error of the example region and the MNI-sphere ROI to the participant-trace ROI (z2-z1) and the connectivity error of the example region and the MNI-trace ROI to the participant-trace ROI (z3-z1) are calculated per healthy control participant. B. The distribution of the connectivity error as calculated in A (N=15, test dataset). For this example region, smaller differences are seen in the MNI-trace ROI to the participant-trace ROI as compared to the MNI-sphere ROI to the participant-trace ROI. Signals from A come from the participant circled in purple. -10 0 10 -10 0 10 20 sec -10 0 10 A. B. Connectivity error relative to participant-trace (z-z 1 ) MNI-sphere: participant-trace MNI-trace: participant-trace BOLD signal BOLD signal MNI-sphere participant-trace MNI-trace participant-trace region region region Example participant z 2 -z 1 = 0.34 z 3 -z 1 = -0.04 -0.4 0 0.4 49 Cluster-based analysis in UCPPS patients versus healthy controls Due to the PAG’s association with chronic pain (37,49,75,108,149) and other chronic conditions (47,150,151), our objective was to lay the foundation for optimal localization of the PAG in UCPPS patients. Data from 100 UCPPS patients and 109 healthy controls was used for this analysis. A participant-trace is not practical for widespread use with more than a few participants. Our analysis was therefore focused on the two PAG ROIs traced in standard MNI space: the MNI-sphere ROI and the MNI-trace ROI. Connectivity was calculated by following a voxelwise correlation procedure established in AFNI (https://afni.nimh.nih.gov/SimAna). Next, to control for site related differences, the participant data was demeaned for each MAPP Research Network site and the global mean for all sites was added. We first established healthy control connectivity and defined clusters from each PAG ROI, and subsequently used this connectivity as a backdrop against which to compare our findings in patients with chronic pelvic pain. Specific clusters were defined by their significant connection to each PAG ROI across healthy control participants (AFNI’s 3dClustSim) (p<0.00001, 𝛼 = 0.01, uc = 6). The use of the most recent update of 3dClustSim for cluster analysis is taken into account after results in regard to false positives (152). Each set of clusters was binarized and saved as an image file generated in AFNI’s interactive clustering interface accessed via AFNI’s GUI and based on parameters from 3dClustSim. The image file saved contains multiple masks created from clusters in the healthy control dataset (N=109). The image file was then used to extract the average signals (AFNI’s 3dROIstats) from both patient and healthy control participants’ voxelwise whole-brain connectivity to one of the two PAG ROIs, respectively. AFNI’s CA_ML_18_MNIA atlas (144,153) was queried (AFNI’s whereami) for cluster labels by centering on the voxel peak coordinates and location was checked with FSL’s Harvard-Oxford cortical and subcortical 50 structural probabilistic atlases (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Atlases). UCPPS patients and healthy control z-scores were statistically compared with a two-sample t-test for each cluster and significance was determined after p-values were corrected for multiple comparisons (false discovery rate using the Benjamini and Hochberg procedure) across each set of clusters. Independent component analysis signal correction for potential physiological effects Finally, the impact of physiological noise on whole brain connectivity differences needs to be addressed considering the localization of our PAG traces within the brainstem. In the absence of physiological recordings, we have implemented this correction using FSL’s ICA-FIX in an attempt to auto-classify signals and remove potential noise confounds due to physiological noise at the individual subject level (154,155). FSL-FIX trains features based on either the spatial map, time series, or frequency spectrum and provides a score of the likelihood of whether the feature is a signal (156). A threshold is then used on these scores to classify which features are labeled as noise. In our analysis we followed this protocol, beginning with individually preprocessed, but not nuisance regressed rs-fMRI images. One of the key aspects in following FSL-FIX’s protocol is to train components classified as noise in the analyzed dataset. However, in order to maintain an automated preprocessing pipeline with as little experimenter input as necessary, we used the training data provided in FSL-FIX, “Standard.RData” (3.5 mm isotropic resolution, TR = 3 seconds), which was the closest match to our dataset’s parameters and has been implemented successfully in previous literature (157,158). For the reason that we did not create a unique training set specific to our dataset, we examined the impact of signal to noise threshold at a range of values suggested by the literature (5, 40, and 75) (154,158). Following independent component analysis noise correction, resulting datasets were nuisance regressed (AFNI’s 3dTproject). The same methods as above were completed by 1) comparing the three different PAG localization techniques 51 in healthy controls (test and validation) and 2) the cluster-based analysis in UCPPS patients versus healthy controls. These two analyses were chosen to evaluate potential differences in the direct PAG signal and to evaluate the sensitivity of healthy control versus patient differences, respectively. Results PAG ROI BOLD signal non-uniformities in healthy control participants Our results included a total of 30 healthy control participants in our test (N=15) and validation (N=15) datasets (Table 3.1). To test BOLD signal uniformities from each PAG ROI, the residual time series signal from the MNI-sphere and the MNI-trace ROIs was correlated to the signal from the participant-trace ROI, our best measure of the true PAG location. Figure 3.1 C shows significant differences in the standardized correlation coefficient of the BOLD signal across participants in a paired t-test (test: p<0.001, validation: p<0.0001). This relationship demonstrates that the MNI-trace signal is significantly more similar to the participant-trace signal than the MNI- sphere signal (test: MNI-sphere z-score = 0.51±0.17; MNI-trace z-score = 0.72±0.04; validated: MNI-sphere z-score = 0.50±0.13; MNI-trace z-score = 0.69±0.08 ). Whole-brain functional connectivity differences in healthy control participants We evaluated whether the MNI-sphere and the MNI-trace signals performed better or worse by their connectivity error, or the standardized connectivity difference, to the participant- trace signal. Using the Power 264 atlas regions, 73% of regions (out of 264) showed better performance, or regions with smaller IQR and D when compared to the MNI-sphere connectivity error, in the test dataset for the MNI-trace connectivity error (Figure 3.3 A.i). This was replicated 52 in the validation dataset with 69% of regions and in the overlap of the test and validation datasets with 52% of regions with better performance (Figure 3.3 A.i). Worse performance, or regions with larger IQR and D, was only seen in 3% of regions in the test dataset, 2% of regions in the validation dataset, and no regions in the overlap (Figure 3.3 B.i) for the MNI-trace compared to the MNI sphere connectivity error. No particular network or pattern of regions emerged, though clearly the MNI-sphere and the MNI-trace connectivities differ in major ways throughout the brain. 53 Figure 3.3: Figure 3.3: Performance of the PAG ROIs on whole-brain functional connectivity. A. Brain regions whose interquartile range (IQR) of the connectivity error was smaller and median closer to zero (D) in the MNI-trace ROI to participant-trace ROI compared to the MNI-sphere ROI to participant-trace ROI. B. Power 264 atlas regions whose interquartile range of the connectivity error were greater and median further to zero in their connectivity to the MNI- trace ROI to participant-trace ROI compared to the MNI-sphere ROI to participant-trace ROI. The comparison was performed at a for each Power 264 atlas region (A.i and B.i) and at each voxel within a brain-masked covering the entire brain (A.ii and B.ii). The analysis included the test dataset (N=15), validation dataset (N=15), and overlap of the two datasets. Key: SMM, somatosensory/motor (hand), SMMm, somatosensory/motor (mouth), COTC, cingulo- opercular task-control, Aud, auditory, DMN, default-mode, MR, memory retrieval, Vis, visual, FPTC, frontoparietal task-control, SN, salience, subC, subcortical, VAN, ventral attention, DAN, dorsal attention, CBL, cerebellum, Other, other regions. A. B. (i) (i) (ii) (ii) Test Validation Overlap Vis SN SSM COTC DMN FTPC subC VAN DAN CBL Aud R. R. 54 Whole-brain voxelwise analysis showed similar results. Better performance of the MNI- trace connectivity error was seen with 72% of brain masked voxels (out of 73409) in the test dataset, 68% of voxels in the validation dataset, and 49% in the overlap (Figure 3.3 A.ii). Worse performance was found in 2% of voxels in the test dataset, 2% of voxels in the validation, and less than 0.1% in the overlap (Figure 3.3 B.ii). These results indicate smaller connectivity errors from the MNI-trace signal throughout the brain when compared to the MNI-sphere signal across healthy control participants. Cluster-based functional connectivity differences in UPPS patients versus healthy controls Our final goal examined not only whole-brain connectivity differences in choice of PAG localization, but also tested whether the choice would alter connectivity in determining patient versus healthy control differences in the MNI-sphere and the MNI-trace signals. The presence of both ROI’s in MNI space removed any restriction on participant numbers. Whole-brain functional connectivity differences in healthy controls suggest that the MNI-trace ROI was more similar, though not equivalent, to our participant-trace ROI. After neuroimaging preprocessing and quality controls, 209 participants (17.3% reduction from our initial number) were included in our final dataset (N=100 patients, N=109 healthy controls). Following connectivity analysis and corrections for site effects we found significantly connected clusters in healthy controls from the MNI-sphere and the MNI-trace connectivity to the entire brain. We used these clusters to extract the average signal from each participant’s standardized functional connectivity data in both the patients and healthy controls. Table 3.2 shows all cluster labels defined using voxel peak coordinates except for the Right Posterior Cingulate for 55 each PAG ROI. Peak coordinates for this cluster were not located in any atlas region and were defined based on neighboring voxels in the cluster. Also, we did not subdivide the largest cluster for either of the ROIs (MNI-sphere: 24130 voxels, MNI-trace: 13876 voxels) because our cluster defining primary threshold was already very low (p<0.00001) (159) and we were able to acquire a number of regions throughout the brain for comparison. Collectively, there were smaller p-values from the group difference measured in patients versus healthy control clusters associated with the MNI-trace connectivity than those associated with the MNI-sphere connectivity (Table 3.2). The only clusters with significant differences after multiple corrections (p<0.01, FDR q<0.05) were those significantly connected to the MNI-trace signal, defined as the left rostral part of the inferior frontal gyrus (pars orbitalis) and the left inferior parietal lobule (Figure 3.4 A). Figure 3.4 B shows that there were no clusters significantly connected to the MNI-sphere signal with significant differences in patients and healthy controls even in the inferior frontal gyrus (pars orbitalis) (Table 3.2, p=0.42, FDR q=0.51), a coinciding cluster significantly connected to the MNI-trace signal. Table 3.2: Average connectivity of significant clusters in UCPPS patients and healthy controls X (mm) Y (mm) Z (mm) volume (3 mm2) patients (z- score) healthy controls (z- score) p-value/FDR corrected p- value Cluster name MNI-sphere PAG ROI/midline 3 -31 -10 24130 0.12±0.07 0.13±0.07 0.57/0.74 Right medial temporal pole 36 14 -37 44 0.08± 0.12 0.09±0.12 0.35/0.74 Left medial temporal pole -30 8 -37 40 0.09±0.15 0.09±0.13 0.68/0.74 Right precentral gyrus 42 -19 47 21 0.08±0.15 0.09±0.17 0.37/0.74 Right posterior cingulate 21 -43 26 20 -0.07±0.14 -0.09±0.15 0.73/0.74 Right postcentral gyrus 21 -37 65 18 0.08±0.17 0.10±0.17 0.74/0.74 Left calcarine gyrus -15 -85 17 17 0.08±0.16 0.09±0.19 0.15/0.67 Left postcentral gyrus -42 -25 44 11 0.07±0.16 0.10±0.18 0.14/0.67 56 Left inferior frontal gyrus (pars orbitalis) -30 35 -10 8 0.08±0.18 0.08±0.15 0.53/0.74 MNI-trace PAG ROI/midline 0 -37 -4 13876 0.12±0.09 0.14±0.09 0.14/0.31 Right inferior frontal gyrus (pars triangularis) 51 20 35 210 0.06±0.11 0.10±0.12 0.03*/0.13 Left inferior frontal gyrus (pars opercularis) -54 11 38 189 0.06±0.11 0.09±0.13 0.02*/0.13 Left middle frontal gyrus -24 26 53 158 0.08±0.13 0.10±0.15 0.43/0.57 Right superior temporal gyrus 51 -22 11 151 0.07±0.15 0.10±0.13 0.19/0.36 Right precentral gyrus 45 -19 47 76 0.07±0.14 0.10±0.15 0.09/0.25 Left postcentral gyrus -42 -25 47 63 0.09±0.13 0.11±0.16 0.26/0.50 Right posterior cingulate 12 -40 23 29 -0.07±0.16 -0.11±0.12 0.05*/0.17 Right superior parietal lobule 45 -52 62 26 0.05±0.16 0.09±0.16 0.07/0.23 Left precentral gyrus -42 -10 56 18 0.08±0.16 0.09±0.15 0.58/0.64 Left superior frontal gyrus -24 59 23 17 0.09±0.15 0.09±0.16 0.86/0.86 Right precentral gyrus 21 -34 68 16 0.09±0.20 0.10±0.18 0.71/0.75 Left SMA 0 14 68 13 0.09±0.18 0.11±0.20 0.57/0.64 Right SMA 6 20 71 11 0.05±0.19 0.09±0.17 0.10/0.25 Left inferior frontal gyrus (pars orbitalis) -33 29 -4 10 0.03±0.16 0.09±0.14 0.004**/0.04* Left inferior parietal lobule -54 -43 41 7 0.02±0.16 0.08±0.15 0.004**/0.04* Right superior frontal gyrus 24 59 17 6 0.07±0.20 0.09±0.16 0.40/0.57 Right superior frontal gyrus 18 17 74 6 0.07±0.17 0.09±0.17 0.41/0.57 Right SMA 6 5 80 6 0.09±0.22 0.12±0.22 0.45/0.57 Table 3.2: The average connectivity of significant clusters found connected to the MNI-Sphere ROI and MNI-trace ROI in UCPPS patients and healthy controls. Cluster labels, Montreal Neurological Institute (MNI) for the peak voxel coordinates, total number of voxels in each cluster, average UCPPS patient z-score and standard deviation, average healthy control z-score and standard deviation, and p-value, uncorrected and false discovery rate corrected. AFNI’s CA_ML_18_MNIA atlas (45,54) was used for cluster labels based on their peak voxel coordinates (except for the Right posterior cingulate, whose label was defined using neighboring voxels within the cluster). Key: PAG, periaqueductal gray, SMA, supplementary motor area. 57 Figure 3.4: 58 Figure 3.4: Functional connectivity differences in UPPS patients versus healthy controls. A. Selection of clusters based on significant voxelwise connectivity in healthy controls to the MNI-sphere ROI (HC: p<0.00001,𝛼 = 0.01 , uc = 6). B. Selection of clusters based on significant voxelwise connectivity in healthy controls to the MNI-trace ROI HC: p<0.00001,𝛼 = 0.01, uc= 6).Uncorrected significant clusters identified between patients and healthy controls are circled in purple (HC vs. UCPPS: p<0.05) and corrected significant clusters are identified with an asterisk (HC vs. UCPPS: p<0.01, FDR q<0.05). Key: HC, healthy control, UCPPS, patients. Head motion Head motion is known to be a confounding factor in the estimation of BOLD functional connectivity (119,160–162). Even small amounts of movement are known to bias functional connectivity measures, especially those regions found in the brainstem, and have been shown to increase relative to the distance between two brain regions of interest (160). Our neuroimaging preprocessing measures attempt to correct these issues limiting framewise displacement by censoring volumes related to movement greater than 0.3 mm, setting this parameter in our AFNI pipeline according to previous publications (119,146). However, in our analysis, motion differences may contribute to differences between UCPPS patients and healthy controls. We used the same time series file for censoring movement volumes in our AFNI pipeline based on the Euclidean norm of the derivatives of the motion parameters for patients (N=98, two files removed due to incomplete motion times series) and healthy control (N=109). After comparing average movement across all time series per participant, we found no significant differences between these groups in overall head motion (p=0.64). Independent component analysis signal correction effect To control for potential physiological effects from our PAG signal we implemented the auto-classification FSL-FIX at multiple signal to noise thresholds. First, we compared the three different PAG localization techniques across FIX thresholds 5, 40 and 75 in both test and validation 59 datasets (Figure 3.5 A). The higher the threshold indicates the inclusion of more potential noise components that are removed in the preprocessed rs-fMRI image. Our results show that significance was maintained after implementation between the standardized correlation coefficient of the MNI-sphere and MNI-trace ROIs to the participant-trace ROI. In fact, the p-value decreases at higher FIX thresholds (FIX Thr.40: p<0.0001, FIX Thr.75: p<0.0001). The average standardized correlation coefficient of the MNI-Sphere ROI to participant-trace ROI consistently remains lower in value compared to the MNI-trace ROI coefficient values. Likewise, in our cluster-based analysis, in which we determine patient versus healthy control connectivity differences to the MNI-trace ROI, qualitative differences are maintained in significant clusters from our previous analysis without independent component analysis application (Figure 3.5 B). These clusters, defined as the inferior frontal gyrus cluster (IF) and inferior parietal lobule cluster (IP), have consistently higher connectivity to the MNI-trace ROI in healthy controls than in patients across signal to noise thresholds. This same quality is not necessarily maintained in other clusters connected to the MNI-trace ROI. However, as more noise components are removed with higher thresholds, there is a loss in the significant differences between the groups. Based on these changes in connectivity, we were interested in key noise components that influenced connectivity from the MNI-trace PAG ROI. The contribution from these potential noise components was calculated at the MNI-trace PAG ROI, inferior frontal gyrus cluster, and inferior parietal lobule cluster (Figure 3.5 C). The maximal value at each location was used to determine the noise component averaged across participants. The voxel threshold across components is based on the normalized voxel z-scored values for the chosen maximum component (calculated by dividing the z-scores for that component by the maximum average value at the location, e.g. MNI- 60 trace PAG ROI). Figure 3.5 C shows the number of participants as a percentage (out of N=209) whose signal is above a voxel threshold of 0.1. For example, at the location of the MNI-trace PAG ROI, there is an increased number of participants with a voxel threshold of 0.1 or greater around the area of the brainstem and cerebellum. There were fewer noise components removed at a FIX threshold of 5 across all participants than at higher FIX thresholds, 40 and 75. 61 Figure 3.5: Participants above voxel threshold 0.1 (%) 0 0.8 No ICA FIX Thr.5 FIX Thr.40 FIX Thr.75 MNI-sphere: participant-trace MNI-trace: participant-trace Test Validated Test Validated Test Validated Test Validated *** **** A. Connectivity measure (z-score) FIX Thr.5 FIX Thr.40 FIX Thr.75 MNI-trace PAG (x=4) IF (x=-33) IP (x=-54) 50 100 PAG trace and corrected significant HC>UCPPS HC<UCPPS cl usters from Figure 4 Averaged FIX noise components C. *** **** **** **** **** **** Connectivity MNI-trace PAG:IF Connectivity MNI-trace PAG:IP (z-score) (z-score) B. ** ** 0 0.1 0 0.1 HC UCPPS No ICA FIX Thr.5 FIX Thr.40 FIX Thr.75 No ICA FIX Thr.5 FIX Thr.40 FIX Thr.75 ** ** *** ** 62 Figure 3.5: Independent component analysis (ICA) correction for potential physiological noise effects without ICA application and across FIX thresholds 5, 40 and 75. A. The differences in connectivity across healthy controls in test (N=15) and validation (N=15) datasets (No ICA: p<0.001, FIX Thr.5: p<0.001, FIX Thr.40: p<0.0001, FIX Thr.75: p<0.0001) B. A comparison of UCPPS patient versus healthy control connectivity differences to the MNI-trace ROI in significant clusters, left inferior frontal gyrus (IF) and left inferior parietal lobule (IP), from Figure 3.4 (IF HC vs. UCPPS: No ICA: p<0.01, FIX Thr.5: p<0.01, IP HC vs. UCPPS: p<0.01, FIX Thr.5: p<0.01, IF HC No ICA vs IF HC FIX Thr.75: p<0.01, IF HC FIX Thr.5 vs IF HC FIX Thr.75: p<0.001). C. ICA noise component contribution at the MNI-trace ROI, IF, and IP, calculated from the maximum value at each location and averaged across participants. Threshold values are based on the number of participants with a signal above a voxel threshold of 0.01. Key: HC, healthy control, UCPPS, patients, IF, left inferior frontal gyrus, IP, left inferior parietal lobule. Discussion The goal of this study was to determine whether the type of localization methodology affected PAG functional connectivity and impacted the detection of differences in chronic pain patients compared to healthy controls. Using the time-intensive, hand-drawn, individual participant-trace ROI as a reference, we found significant non-uniformities in the extracted brainstem rs-fMRI signals from our MNI-sphere ROI and MNI-trace ROI in healthy controls. ROI localization methodology impacts functional connectivity findings throughout the brain using either region-based or voxel-based analysis. When tested in a larger population of healthy controls and UCPPS patients, we found significant differences in connectivity in regions previously examined in pain patients (13,65,66,139,163,164). To the best of our knowledge, this is the first study to show that such a comparison depends on PAG localization methodology. How the PAG is localized as a whole, impacts findings of connectivity to other regions of the brain and should be considered when analyzing disease-states associated with PAG connectivity. This study establishes that the PAG localization methodology frequently reported in literature (65,66,75,108–111,139,140), as representing the entire PAG region and labeled in this study as the MNI-sphere, does not adequately represent the anatomical PAG. This finding leads us to ask what brainstem portion the MNI-sphere may represent. The MNI coordinates reported in literature (MNI: 4 -26 -14) align with another study (165) whose peak voxel coordinates from a 63 positron emission measurement relate to improvement in depression symptoms after treatment of a selective serotonin reuptake inhibitor. The dorsal raphe nucleus (DRN), a region in the brain with the largest population of serotonergic neurons (54,166,167), sits very close to the ventrolateral PAG in the brainstem (166–168). The DRN is also involved in descending pain modulation (58,169) and, although it is histologically different from the PAG (54,77), produces analgesia when stimulated (58,68,170); similarly, the ventrolateral PAG has been stimulated to evoke opioid mediated analgesia (47,56,59,65). The MNI-sphere in the literature could therefore be identifying the DRN. Nevertheless, fMRI coordinates reported in literature for the DRN region are ambiguous (171–173), often cited as part of the anatomical location of the PAG and it is unclear if these regions are functionally distinct (56,58,68). Clearly the optimal ROI would be defined both anatomically and functionally from within each participant (137,174). Though not interchangeable, we demonstrated that the participant-trace and MNI-trace ROIs were similar in their connectivity patterns. Caution should still be taken when considering the use of the MNI-trace ROI as a PAG ROI because the trace is derived from an average of 152 T1-weighted structural images. For example, if the study of interest includes variability between participants, such as size and shape variations of the PAG, there is a likelihood of mixed functional activity patterns (133). Our results showed no Power 264 atlas regions and less than 0.1% of voxels with consistent error differences of the whole-brain functional connectivity of the MNI-trace ROI compared to the MNI-sphere ROI in our test and validation datasets (Figure 3.3). This was indicative of a better estimation of the MNI-trace ROI to our participant-trace ROI using participant variation as a measure of differences (interquartile range and distance of the median from 0). Even so, the method for defining the “gold standard” remains complicated for extracting the BOLD time-series signal from the entire PAG. 64 Characterizing a region in the brainstem using fMRI has been met with difficulty (72,116,175). Recent efforts focus on defining a brain region first, before testing its functionality with different experimental paradigms (135). Traditionally brain regions were defined by stimulation of the region of interest or lesion-deficient approaches, such as with the functional segregation of the PAG in humans (47,49,67,135,136). Heat stimulation tasks are the main choice for eliciting a PAG response for measuring pain-induced activation with fMRI or positron emission tomography (47). However, one behavioral outcome may be related to multiple neuronal behaviors (135,176) and the behavior elicited to identify a region is not specific to what that region actually does. We defined the PAG by expertly hand tracing the grey matter voxels in the brainstem of a T1-weighted structural MRI in patient specific space, labelled participant-trace, or from the MNI152 1mm standard space, labelled MNI-trace. This process can be time consuming and details about distinct subregions are unidentifiable (72,175). An alternative to our method is the creation of a probabilistic grey matter map for the study group of interest encompassing the entire PAG (177). For the definition of specific functional subgroups within the PAG, a combination of imaging approaches and atlas references can be used to maximize spatial resolution and compare intrinsic functional connectivity within the region (49,131). A normalization approach optimized for the brainstem as to reduce further sources of potential error with PAG definition (178) is an additional consideration. Overall, the integration of different approaches by pooling data from multiple sources, meta-analysis (47,137,138), and established databases, such as Neurosynth (179) and BrainMaps (180), rather than focusing only on single study activations to define regions, provides a more reliable brain definition. 65 The purpose of this study was not to provide a specific method for analyzing the PAG, but to highlight the erroneous application of published PAG coordinates in chronic visceral pain. Nonetheless, the MNI-trace ROI provides a potential representation of the PAG. Anatomical PAG connections to the cortical and subcortical parts of the brain are well established. Connections to the parts of the medial prefrontal, anterior cingulate, dorsal medial, orbital, posterior cingulate, temporal, ventral insula, and amygdalar regions of the brain were revealed by anterograde and retrograde tracers in Macaques (48). Human studies have shown similar connections using deep brain stimulation (DBS) and diffusion tractography depending on the seeded region; though not all connections fully correlate to animal studies (50,134) and differences exist across species (134). It is important to keep in mind the BOLD signal provides only a surrogate marker for the actual neuronal brain responses (181) and since functional connectivity depends entirely on the BOLD signal, it cannot demonstrate actual anatomical connections (47). In addition, evaluating functional connectivity throughout the entire brain produces a high family-wise error rate causing some important connections to be overlooked (137), especially when dealing with small volume regions. By limiting our analysis between patients and healthy controls to predefined regions connected to our ROIs within our healthy control dataset, we were able to distinguish important differences in two clusters of the MNI-trace ROI associated with PAG connectivity that the MNI-sphere ROI failed to identify. While this approach may bias the connectivity differences towards healthy controls, most cluster differences do not pass multiple comparisons and only two in the MNI-trace ROI: inferior frontal and inferior parietal, were significant (assessed in with both one and two- tailed, two sample t-tests). The first cluster was in the orbitofrontal cortex, a region known for emotionally driven cognitive states and its impairment in chronic pain (13,139,164,169). Connectivity from the 66 orbitofrontal cortex to the PAG is established in humans in pain-related studies (79,182) and was shown within both the MNI-sphere and MNI-trace ROI cluster-based connections from our voxelwise analysis (Table 2); however, significant differences between UCPPS patients and healthy controls were only seen in the MNI-trace ROI. The second cluster occurred in the inferior parietal lobule. Despite this region not showing a direct connection in animal studies to the PAG (51), there is minor evidence of its linkage in rat studies (117) and direct linkage to the dorsolateral prefrontal cortex and its association with visceromotor control (51,183). In patients with chronic inflammation, the inferior parietal lobule has recently been found to be associated with the magnitude of inflammatory response (184). Additionally, a fibromyalgia study (163) between patients and healthy controls showed resting- state network connectivity differences in the secondary somatosensory cortex of the default node network (DMN), a region located very close to the parietal area reported in our study. Another study (185) found that the secondary somatosensory cortex and the temporoparietal junction were activated during an attention to pain fMRI task. Though these studies validate our patient and healthy control comparison for the MNI-trace connectivity, further studies are needed to understand its relationship as a functional measure for the PAG. It is important to note that these connectivity findings fit well with a larger view of the PAG’s role in chronic pain states as an orchestrator of the acute threat response (50,53). Prior studies comparing irritable bowel syndrome with ulcerative colitis have demonstrated that frontal connectivity with PAG may determine whether the PAG is “set to sound the alarm” (100) which would generate both an autonomic sympathetic response and a conscious pain experience. A volumetrically larger PAG primarily characterizes women with endometriosis who do not have pain, compared to those who do (16), confirming that PAG function and its trophic state may play 67 an important role in chronic pelvic pain. It is also known that vagal parasympathetic influence is reduced in both chronic pain and inflammatory states (99,103), and the vagus nerve plays a major role in the control of inflammation (186). The PAG may carry out its primary function of orchestrating a response to an acute threat by linking cortical processes focused on the readiness of bodily systems such as the inflammatory response, with descending vagal and sympathetic influences which carry out these tasks. Further, the PAG could play a central role in a chronic pelvic pain syndrome, if the systems designed to reset it after the acute threat response fail to do so, resulting in a state of “chronic threat response”, and consequent exhaustion of resources designed only for an acute need. Unlike the MNI-sphere ROI, our MNI-trace and participant-trace ROIs surround the cerebral aqueduct which carries cerebrospinal fluid (CSF) between the third and fourth ventricles (187). One of the challenges in fMRI of the brainstem is spatial resolution at 3 tesla, which shows the brainstem as a mostly homogeneous structure (175). The anatomical size of the PAG is 4-5 mm around the aqueduct (47,175), therefore the MNI-trace and participant-trace likely include artifacts from the CSF due to cardiac and respiratory effects (116,175). The PAG’s association with autonomic control suggests that these movement artifacts could also be temporally correlated to behavior (47,50,72,134,171,188,189). The significant differences in UCPPS patients and healthy controls in our results (Figure 3.4) could be due to autonomic functional differences or coupled physiological noise effects. We recognize that the p-values across a number of the clusters in MNI-trace ROI appear smaller than in the MNI-sphere ROI, which may be due to proximity of this region to the cerebral aqueduct and the effect of cardiac pulsation on the CSF (116,154,190,191). The variance in global signal, which was not corrected for in our pipeline because of distant dependent bias (145), is linked to both fluctuations in cardiac rate and respiratory 68 activity (192), and could account for the relationship in patient versus control differences in the MNI-trace ROI connectivity to the rest of the brain. Even though we have not directly controlled for physiological measures such as heart rate and breathing, we utilized an independent component analysis approach in our preprocessing pipeline to address the potential effect. At increasing FIX thresholds, we show that the signal differences from ROI placement exist regardless of ICA correction (Figure 3.5 A). However, significant differences between patients and healthy controls were lost at higher signal to noise FIX thresholds (Figure 3.5 B). This was likely due to potential noise around the aqueduct, caused by fluctuations in cardiac and respiratory activity on the CSF, and arterial blood flow from the posterior cerebral artery (154) near the PAG (Figure 3.5 C). The spread across the PAG at higher FIX thresholds in Figure 3.5 C could also be enhanced because of spatial smoothing (191). In lower spatial resolution datasets, such as in our study, signal can spread to nearby veins in brainstem and cerebellar areas (190) and should not be removed from the data. Nevertheless, we also observe a maintenance in the qualitative relationship between the connectivity in healthy controls and patients (a quality not maintained in other clusters connected to the MNI-trace ROI) and a valid significant difference in connectivity at a more conservative FIX threshold. When using automatic classification methods without the input from expert training on noise component selection, more conservative FIX thresholds are recommended (154) so as not to remove signal. We demonstrate that FSL-FIX’s independent component analysis is a logical preprocessing step in interpreting the relationship of the PAG to other areas of the brain if physiological measures have not been recorded directly. Where this automated technique may assist in determining signals that constitute potential physiological noise, more selective approaches are recommended, such as 69 creating a training set for selecting heart rate and breathing components by several trained experts as defined by Salimi-Khorshidi and colleagues (154). In addition to potential physiological noise artifact we reduced motion through our AFNI pipeline (AFNI’s ANATICOR) which accounts for local artifacts rather than using the global signal (144,145). Even so, we have found no significant differences in overall head motion in our patient and healthy control populations (p=0.64). Given that the PAG’s location is so close to such a physiologically dynamic structure affected by heart rate and breathing, and dealing with signals derived from this activity, the aqueduct must be considered carefully, and these findings show that caution should be taken when analyzing PAG functional connectivity. Many prior studies did not address these potential confounds (65,66,75,109,139). This analysis of PAG localization demonstrates the dependence of disease-state differences on ROI localization methodology in functional neuroimaging studies, particularly in an area such as the brainstem. Improved imaging modalities will likely lead to better understanding of both anatomical morphology and functional connectivity of the PAG. For example, enhanced resolution of the GE 7T MR950 MRI at the Center for Imaging Research at the Medical College of Wisconsin improved anatomic identification of the PAG borders and facilitated identification of subregions within the PAG during aversive image viewing (113). However, this particular imaging device suffers from blurring in areas of heterogeneous density weakening its capacity for functional imaging particularly in the frontal areas where significant fluid-bone-air interfaces occur. Ultimately, it will be critical to corroborate PAG localization methods anatomically when performing neuroimaging meta-analyses in order to account for activation peaks found outside the PAG region (which weaken spatial specificity), to align processing procedures, and to limit mislabeling (47). 70 In summary, this study evaluates an anatomical PAG localization method directly from structural MRI images. Future PAG studies may benefit from this foundational work. Based on this study, we recommend anatomic rather than functional localization procedures, either through hand-traced participant specific ROI, or a PAG traced in average space, depending on the research problem. As a note of caution, we have supported use of the MNI-trace ROI based on our connectivity results corroborated by prior literature and our similarity measures to the participant- trace, differences between UCPPS patients and healthy controls alone do not determine choice of ROI. Interest in the PAG will likely increase in reference to chronic pelvic pain and bladder pain, as the PAG has been shown to activate in response to bladder filling and during micturition (50,52,193). The MAPP research group has shown that painful bladder filling or urgency is associated with a possible subset of UCPPS patients (29,105). The present MNI-trace ROI can be used to compare changes in PAG activity in UCPPS patients and healthy control participants and their relationship to pelvic pain. The MNI-trace ROI may also be useful in predicting longitudinal symptom evolution in UCPPS patients, since whole-brain functional connectivity patterns of the MNI-trace ROI signal are similar to our reference participant-trace ROI across defined Power 264 atlas regions. Connectivity from the MNI-trace ROI may link to individual differences in pain symptoms in the larger UCPPS population. Given these applications, there is still a need for further study to establish the reproducibility of the functional connectivity of the PAG in humans. Conclusion To determine the optimal localization of the PAG ROI in fMRI, we compared the rs-fMRI functional connectivity in three different traces of the PAG: 1) traditional MNI-sphere, 2) MNI- trace, and 3) participant-trace. The direct fMRI time series signal from the MNI-trace ROI 71 paralleled the participant-trace ROI, our presumptive reference measure of the PAG signal, far more closely than the traditional MNI-sphere. The same finding emerged from further investigation using whole-brain functional connectivity to these ROIs and likewise, found that the connectivity to the MNI-trace ROI was more similar to the participant-trace ROI. In our second analysis, we compared whole-brain voxelwise connectivity to the MNI-sphere and MNI-trace ROI, respectively in patients versus healthy controls. Regions established to be connected to these ROIs in healthy controls only showed patient versus healthy control differences in the MNI-trace ROI and not in the traditional MNI-sphere ROI. Our results confirm that the PAG ROI, MNI-sphere, often reported in literature as the PAG is not equivalent to the anatomical trace of the PAG (both the MNI-trace and the participant-trace), and may reflect another region in the brainstem related to pain, perhaps the dorsal raphe nucleus. This study demonstrates that care and validation are required when defining the PAG, especially when probing for disease-state differences. Assumptions about localization may alter our understanding of neuronal processing and influence prospective clinical conclusions. The connectivity findings reported here support a specific hypothesis linking the PAG’s role as orchestrator of the response to threat and a potential neuronal infrastructure for the development and maintenance of a chronic pelvic pain syndrome and its autonomic co-morbidities (194). 72 Chapter 4: Using the PAG to whole-brain connectivity to predict UCPPS symptom change Introduction The lives of millions of women and men are affected by Urologic Chronic Pelvic Pain Syndrome (UCPPS). UCPPS is comprised of interstitial cystitis/bladder pain syndrome (IC/BPS) in women and chronic prostatitis/chronic pelvic pain syndrome (CP/CPPS) in men. The pathophysiology of UCPPS is not well understood, and no generally effective treatments have been identified. The relationship between UCPPS and the bladder and the prostate has been studied extensively, though without any definitive biomarkers (22,26,43). This has led researchers to focus on a more systematic view of UCPPS. Recent studies have begun to show that brain structure and function are important factors in distinguishing UCPPS patients from healthy individuals (34,41,43,124,195). Our group has shown that whole-brain connectivity measured with resting-state functional magnetic resonance imaging (rs-fMRI) at baseline may predict progression of UCPPS symptoms longitudinally (39). This work in a cohort of 52 patients was performed as part of the Multidisciplinary Approach to the Study of Chronic Pelvic Pain (MAPP) research study. We were able to distinguish which patients had strong improvement over 3 months with 73% (p=0.0012) prediction accuracy based on functional connectivity within the left frontoparietal network. 12-month prediction accuracy was 59.6% and trended toward significance (p=0.079). Though promising, there are several factors currently limiting the clinical utility of these results. First, higher accuracy is required for clinically useful tests. Second, improvement of the prediction of long-term symptoms is also desired. Third, 73 neuroimaging predictive results should be interpreted with their relationship to the clinical relevance and validity of the features defined in the longitudinal prediction. The goal of the work presented here is to overcome these limitations by optimizing an algorithm for predicting UCPPS symptom progression by focusing on resting state connectivity. As a step towards developing this algorithm, we decided to focus on the connectivity to the PAG due to its prominent role in pain modulation and potential influence on chronic pain manifestation (see Chapter 1, ‘The Periaqueductal Gray’). We optimize brain parcellation, feature selection, classification, and cross-validation. To further test generalizability, we also validate our optimized predictive algorithm on an independent dataset held out from the initial stages of analysis. To achieve our goal, we have used the neuroimaging dataset collected as part of the second phase of MAPP, MAPP-II, with a total of 389 UCPPS participants with longitudinal symptoms reported up to a year, or 12-months (See Chapter 2, Methods, ‘Measures for analysis – Pain intensity slope’). We designed our algorithm by using outcome measures based on symptom change over time (established in Chapter 2) and the knowledge that there was a significant improvement in symptoms in a subgroup of our dataset. For example, the lower third of the participants show an improvement in symptoms of approximately 46% from baseline, whereas others have stable trajectories or even worsen over a 12-month timeframe. Additionally, we focused on regions in the brain that drive our prediction and these differences in behavior. The overall aim is to test whether symptom prediction is possible in UCPPS patients at 12-months and to define the potential biomarkers that lead to this prediction. 74 Methods Participants We obtained a total cohort of 389 participants for our longitudinal analysis from the MAPP-II study, or the Trans-MAPP symptom patterns study (SPS), based on collection criteria for UCPPS setup in MAPP-I (32,44). Our cohort came from a total of 620 participants, 210 males and 410 females with UCPPS and 73 healthy controls that were enrolled in the MAPP-II study (44). This dataset included 83 participants that were re-enrolled from MAPP-I, of which the longitudinal study covered in this chapter used 58 participants, not included in our MAPP-I prediction. In this study, we focused only on longitudinal collection up to one year provided by the Data Coordinating Core (DCC), and that have passed quality control measures from the Neuroimaging scan repository and Reading Center at UCLA/USC. Participant data collected beyond one year will be used in other MAPP-II manuscripts as in this study we attempt to expand the results found in our MAPP-I prediction by focusing on 12-month symptoms. Participants were recruited from six collection sites: Northwestern, University of California Los Angeles, University of Iowa, University of Michigan, University of Washington, Washington University in St. Louis (44). At each site, participants provided informed consent according to the Declaration of Helsinki and the institutional review board approved the collection in accordance with MAPP-I (22,44). Image collection Following the same guidelines established in MAPP-I, 3D T1-weighted structural MRI and resting state fMRI were collected and quality controlled (33). In the MAPP-II data collection, two resting state scans were taken: 1) bladder filling proceeded by 2) bladder voiding. These were collected at multiple time points across the study. For the purpose of our analysis, we focused on 75 the low urgency (bladder voiding) scan taken only at the first visit. For additional details on image collection see Chapter 3, Methods, ‘Image collection’ (33,143). Image processing Resting state fMRI raw data was processed using a combination of Analysis of Functional Neuroimaging (AFNI, https://afni.nimh.nih.gov/, Version AFNI_19.3.12 'Nero', (144)) and Freesurfer (https://surfer.nmr.mgh.harvard.edu/freesurfer-i386-apple-darwin11.4.2-stable6- 20170119). Parameters within these analysis packages were adjusted according to Plitt and colleagues (146) and Drysdale and colleagues (119). These packages were selected based on the analysis performed previously on resting state fMRI prediction studies. In particular, a benefit of AFNI’s imaging pipeline is it’s use of the white matter regressor (AFNI’s ANATICOR) and stringent control of distance-dependent motion artifacts (further explained in Drysdale and colleagues (119)) used with the Python program afni_proc.py (https://afni.nimh.nih.gov/pub/dist/doc/program_help/afni_proc.py.html, see Example 11). The first step within our preprocessing pipeline was to segment the white matter and ventricle mask from the T1-weighted anatomical image using Freesurfer. The rs-fMRI image time series was then preprocessed by removing the first four volumes. Data was then despiked, slice- time corrected, co-registered to the T1 weighted image from the output of Freesurfer and automatically resampled to 3mm isotropic voxels. This was then normalized to reflect the percent signal change, non-linearly registered to the MNI template scan, and spatially smoothed with 6mm full width half maximum Gaussian kernel. The nuisance variables include in the regression were average ventricular time series, local average white matter time series (from AFNI’s ANATICOR), and twelve parameter estimates for head motion (including their first derivatives). The time series 76 was simultaneously bandpass filtered between 0.01-0.1 Hz and nuisance regressed. Censoring was accomplished by removing volumes that were greater than 0.3mm, the Euclidean norm of the motion derivative, and volumes with outlying voxels exceeding more than 10% of the masked voxels (see AFni’s 3dToutcount). After regression corrections, the output from AFNI results in a residual time series file used for further analysis. For a summary of this process see Figure 4.1 A, recreated from Jo and colleagues (145). Figure 4.1: Figure 4.1: Preprocessing and machine learning workflow on resting state fMRI data. A. The rs-fMRI is corrected using AFNI’s pipeline (diagram recreated from Jo et al. (145)). A feature of this pipeline is that nuisance regressors are simultaneously bandpass filtered and global signal is not removed. B. Further preprocessing steps followed after resting state preprocessing to calculate functional connectivity. Green box indicates preprocessing steps. C. Machine learning optimization steps for prediction, highlighted in the orange box. 77 Predictive measures: Functional connectivity preprocessing The goal of our study was to determine if baseline rs-fMRI could predict and validate future UCPPS symptom trends. In our study, we used functional connectivity measures derived from rs- fMRI to make our prediction. Using the preprocessed residual BOLD time series output from AFNI, the next steps to generate our predictive measures were brain parcellation, additional quality control of noise signals, and functional connectivity estimation and standardization (Figure 4.1 B). For brain parcellation, we focused on a region of interest (ROI) analysis based on previous neuroimaging prediction papers (119,146) using the Power 264 atlas (148). Voxelwise signals were extracted from the residual BOLD time series and averaged in each of the atlas regions. In addition to the Power 264 atlas, we attempted to replicate similar prediction findings in a voxelwise analysis by using a standard template brain-mask covering the entire brain created in AFNI’s output file. Individual voxelwise signals were extracted from each voxel in the mask without averaging. Subsequently, the temporal signal to noise measures (TSNR, given as the voxelwise mean of the time series divided by the standard deviation) were taken to address signal quality or scan coverage as explained by Drysdale and colleagues (119). In the ROI analysis, regions were first removed if the TSNR was less than 100 in 5% or more of the participants for that region (35 regions). Next, only participants with TSNR greater than 100 in each voxel were used to calculate the average BOLD signal extracted in each region. Finally, participants were removed if the SNR was less than 100 in the remaining regions after elimination. In the voxelwsie analysis, we only removed voxels with TSNR less than 100 in more than 5% of the participants for that voxel (approximately 20,000 voxels) because these values were not averaged. Regions and voxels that were removed were located in ventral regions of the frontal and temporal lobes where the SNR is typically low in rs-fMRI. 78 Due to our previous results on PAG localization (see Chapter 3), to reduce the number of predictive data points we decided to focus on connectivity solely to the PAG, established as a key region in pain modulation. Final functional connectivity estimation was calculated between the PAG MNI-trace (established in Chapter 3) to every remaining Power 264 atlas region after TSNR corrections with Pearson’s pairwise linear correlation and Fisher transformation. Site correction and additional standardization depended on the machine learning procedure utilized. To control for site related differences, datasets were corrected by demeaning participant connectivity data for each region by each MAPP Research Network site. This was accomplished by subtracting the average connectivity over each region within a specific site from that region. In the voxelwise analysis, the same procedure for functional connectivity estimation was followed for each voxel to the PAG trace. The number of functional connectivity features resulted in 229 regions in the ROI analysis and 48,573 voxels in the voxelwise analysis. Outcome measures for prediction The outcome measure for our prediction is symptom change over time. After a 4-week run- in period, baseline, 3-month, 6-month, and 12-month questionnaire data and “deep phenotyping” of symptoms were taken (44). Full details are explained in detail in Chapter 2. To mirror our analysis in MAPP-I prediction, we decided to focus on pain intensity scores measured with the Genitourinary Pain Index (GUPI), Interstitial Cystitis Symptom Index (ICSI) and Problem Index (ICPI), see Chapter 2 Methods, ‘Measures for analysis – Pain intensity score.’ The pain intensity score is converted to a slope value, or the change of reported symptoms over time. For our study, in order to mirror the analysis in our previous MAPP-I prediction (39), we focused on 3-month and 12-month slope values as outcome measures. In this analysis, our objective was firstly to find 79 comparable results in our 3-month prediction and secondly optimize the prediction for 12-months. In this study, we underscored our 12-month slope analysis in our optimization and focused on these measures in our design. A crucial aspect in determining whether baseline rs-fMRI can predict future symptom trends in UCPPS patients (39), is our ability to delineate meaningful change in symptoms over time. In our previous prediction work, the symptom trends, or slope values, were dichotomized into “improvers” or “non-improvers” by the median split of the data. Other MAPP studies have been able to cluster pain scores into improving, stable, and worsening symptoms (123) and show overall improvement in symptoms with improving quality of life scores (130). We were also able to find improving, stable, and worsening symptoms in our MAPP-II dataset (see Chapter 2). However, UCPPS does not have a clinically proven threshold for symptom change and there is no definitive split in our data that we can use for discrete classification. Therefore, we used a continuous regression approach to associate rs-fMRI predictors with the continuous symptom change. Datasets were also corrected for site related differences by demeaning slope values for each MAPP Research Network site. Split dataset used in analysis The participant data was separated into three balanced groups after image processing (Total: 364, Dataset 1: training = 124, Dataset 2: test = 120, Dataset 3: validation = 120) by using a data splitting algorithm, the Kullback Leibler (KL) Divergence. This algorithm measures the difference in an approximate probability distribution, p(x), to an original distribution, q(x), through measurement of information loss one distribution to the other. KL-Divergence was calculated by the difference in the log values of these two distributions. It is important to note that the difference 80 metric calculated is not a symmetrical quantity, e.g. KL(p(x)∥q(x)) ≢ KL(q(x)∥p(x)) ((196)). However, in our use of the KL-Divergence algorithm, our goal was to minimize the difference between the distributions of our entire dataset and that of each split. This was done by fixing the distribution of the entire dataset, q(x), and varying each split, p(x), across 10,000 iterations and comparing this total dataset distribution to one of the three splits at a time. The constraints for splitting the dataset are based on the KL-Divergence distributions for 1) participant age, 2) sex, 3) site number, 4) baseline pain, 5) slope at 3-months, 6) slope at 12-months. An example of the distribution for 12-months slope values and data acquisition site from the split data is shown in Figure 4.2. Root mean square error (RMSE) values are calculated from each group probability distribution trace (indicated by their subscript) to the entire dataset probability distribution trace. Site did not group as well as the other constraints in the dataset as seen in Figure 4.2 B. Our final datasets were sent to the DCC to pre-register our data splits and to avoid the possibility of hypothesizing after the results are known (197). 81 Figure 4.2: Figure 4.2: Examples of the split distributions across the three datasets for A. 12-month slope values and B. data acquisition sites. The distribution across all participants is shown with the dashed red line, labeled as total dataset. Root mean square error (RMSE) values are from each dataset’s probability distribution trace (indicated by their subscript) to the total dataset probability distribution trace. For 12-month slope values, the greatest deviation exists in Group 1, with a 4% difference in the proportion of people with a 12-month slope value between -0.08 and 0.24 from the total dataset. For site distribution, the greatest deviation exists in Group 3, with a 4% difference in the proportion of people at Northwestern from the total dataset. Key: NW, Northwestern, UCLA, University of Southern California Los Angeles, UI, University of Iowa, Umich, University of Michigan, UW, University of Washington, WahsU, Washington University in St. Louis. Machine learning prediction and biomarker evaluation The aim of our machine learning algorithm was to first learn a model on training data and then predict on a test set of unseen data. We wanted to answer whether prediction of future UCPPS symptom change was possible using functional connectivity measures from baseline rs-fMRI and, more importantly, whether this prediction generalized to an independent dataset. 12-month prediction was a more clinically relevant outcome measure due to the long-term effects of UCPPS. While 3-month was predicted to confirm similar findings to our previous MAPP-I outcome, these functional measures were not correlated to the PAG. We have established our method by first focusing on 12-month symptom slope values using the Power 264 atlas. After optimization of feature selection, regression, and testing (Figure 4.1 C), we could then move on to 3-month and 82 voxelwise datasets using a similar approach. We used each of the three splits of the data to 1) train, 2) test, and 3) validate these results, respectively. In addition to prediction, we were interested in the biomarkers for regions of the brain that drove potentially successful results. Initially, each dataset was processed independently from one another. We removed low TSNR signals, as explained in Methods, ‘Predictive measures: Functional connectivity preprocessing,’ and applied site corrections by demeaning by each site for both the connectivity and slope values within Dataset 1. This dataset was used to train our prediction model. Additional options in our optimization pipeline were to standardize connectivity values with a z- transformation within participants across regions (146) and a z-transformation within regions across participants, used in separate predictions. We optimized our features by focusing on the PAG connectivity, and thus, limited connections to one region in the brain. Next, we optimized our regression with a combination of methods from MATLAB’s programing (Matlab 2018b, Mathworks) including regularization options such as lasso and ridge regression (MATLAB function: ridge, lasso, fitrlinear), and support vector machines (MATLAB function: fitrsvm), which incorporates regularization parameters. Different cross validation techniques were also applied to investigate prediction accuracy and potential generalization of the data. Leave-one-out cross validation (LOOCV) was used in our initial prediction, whereas 10-fold and leave-one-site out were used to understand the expected generalizability. Accuracy was determined by evaluating adjusted R-squared values and significance was determined with permutation tests. This was done by shuffling our outcome measure 1000 times to build a null distribution, testing the true adjusted R-squared value to be greater than 95% (p<0.05) than the values achieved with permutation. The work on Dataset 1, set the first steps in our prediction and determined if we could then move on to generalization. 83 With Dataset 2 we used the same approach as on Dataset 1. To test our results from Dataset 1, we checked the LOOCV accuracy within the independently processed dataset to determine predictability. However, for generalization of the model trained on Dataset 1, we incorporated TSNR corrections and site demeaning processing steps combining Dataset1 with Dataset 2 due to better noise approximation. Moreover, we also combined Dataset 1 and Dataset 2 in a separate prediction to test LOOCV on a greater number of participants. Dataset 3 was our final validation dataset. Again, we applied our LOOCV test to the independently processed dataset to check within dataset prediction. Next, we incorporated our processing measures combining Dataset 1, 2, and 3, and predicted Dataset 3 values using the same predictive model trained on Dataset 1. However, if the model from Dataset 1 was not successful previously on Dataset 2, we used the model trained on Dataset 2 to validate generalization to Dataset 3. Using the procedure first established within each LOOCV fold and the model that generalized from one dataset to the other, we examined the top brain regions that contributed toward the prediction. LOOCV trains a model on each k-fold number of participants in the dataset and tests on the left-out participant. In each of these LOOCV folds, the weights applied to each feature are trained and fit to the model. We were interested in how many of these top weights, ranked by their absolute value, were repeated in each fold of the LOOCV and how these weights compared between split datasets. The top weights were selected to be within the top 20% of the data with the highest absolute values. The model weights for each feature derived from either Dataset 1 or Dataset 2 that generalized well to Dataset 3, were tested for significance with permutation testing. Moreover, in order to gain a better understanding of how the model weights influenced brain regions and also differed between those categorized as having improving or 84 worsening symptoms, we observed differences in the significant weights calculated against null distributions by permuting the model with shuffled outcome measures. In the voxelwise analysis we took additional measures to reduce the data, due to the large number of voxels, or features, in each dataset (N=48,573). The number of features was reduced using pairwise correlation to the top 10,000, 1,000, and 100 voxels. In addition to these univariate reductions, we further reduced features by using only the clusters connected to the PAG MNI-trace found to be significantly different between healthy controls and patients (excluding multiple comparisons) in our PAG analysis, see Chapter 4 Results. Other than these reductions, we selected a method for prediction on the voxelwise dataset with the highest prediction accuracies and generalizability measured by adjusted R-squared values and significance testing. Results: Our study was designed to predict UCPPS symptoms using baseline rs-fMRI, test and validate this prediction on a separate dataset from this original model, and determine the biomarkers that lead to the prediction. By following an exhaustive procedure for preprocessing and analysis, we have established a successful model based on 12-month symptom values using baseline rs-fMRI Fisher z-transformed connectivity measures parcellated by the Power 264 atlas. The data was site corrected by demeaning both the connectivity measures and slope values. To achieve this prediction, we performed LOOCV on each of our three datasets (leaving the last validation model to test after Datasets 1 and 2 were appropriately tested) and then generalized the model established to have successfully performed with LOOCV (i.e. significant linear dependence of predictive slope values on actual values using the adjust R-squared for accuracy measurement). The best machine learning model in these predictions was based on support vector machine 85 regression (MATLAB’s fitrsvm, with default parameters). Our results show successful predictions and generalizability using this method. The model established in our Dataset 1 generalized successfully to our test (Dataset 2), and validation (Dataset 3) datasets (Figure 4.3 A). Outliers from each dataset slope value distributions were removed after site demeaning by removing elements more than 1.5 interquartile ranges above the upper quartile (75%) or below the lower quartile (25%) (MATLAB’s rmoutliers). Our first step was to test generalization of the model trained on Dataset 1 on the connectivity data from Dataset 2. Further corrections for noise were applied by combining Dataset 1 and Dataset 2 prior to prediction (total dataset: participants N=232, features: N=229). The application of the model from Dataset 1 on the connectivity data from Dataset 2 resulted in a significant prediction (adjusted R- squared=0.049, p=0.01, ppermuted=0.01). Once this generalization was established in Dataset 1 and 2, Dataset 3 was added for noise corrections (total dataset: participants N=348, features N=229). The distribution of slope values in each dataset was not significantly different between Dataset 1 and Dataset 2 (p>0.05, two-sample Kolmogorov-Smirnov test, Dataset 1: N=115, Dataset 2: N=104) or Dataset 1 and Dataset 3 (p>0.05, two-sample Kolmogorov-Smirnov test, Dataset 1: N=115, Dataset 3: N=110) after corrections in all three datasets (Figure 4.3 A top left and bottom left). Additionally, following these corrections with all datasets combined, generalization was again successful in Dataset 2 (adjusted R-squared=0.046, p=0.02, ppermuted=0.01) and the in final validation with Dataset 3 (adjusted R-squared=0.05, p=0.01, ppermuted=0.009) as seen in Figure 4.3 A, top and bottom right. The model appears to fit both datasets similarly, with the line of best fit for Dataset 2 and 3 represented by slope: 0.24, intercept: -0.006 and slope:0.26, intercept: -0.025, respectively. Outliers seen circled in Dataset 3 Figure 4.3 A bottom right for actual versus predicted slope values, are removed in the final model calculation. To note, generalization was 86 also successful with the same approach on datasets that were uncorrected for noise, but corrections on the combined Dataset 1, Dataset 2, and Dataset 3 appeared to strengthen the prediction. 87 Figure 4.3: 88 Figure 4.3: 12-month symptom slope generalized prediction model and significant weights based on the Power 264 atlas. A. Top left: Slope value distributions for Dataset 1 and Dataset 2 after corrections (demeaning by site and removing outliers). Bottom left: Slope value distribution for Dataset 1 and Dataset 3 after corrections (demeaning by site and removing outliers). Top right: Prediction of the model trained on dataset 1 applied on dataset 2. Bottom right: Prediction of the model trained on dataset 1 applied on dataset 3. B. Regions with significant weights compared to a null distribution (p<0.05) for each weight. Colors refer to rs-fMRI connectivity in patients with improving symptoms. C. Regions with significant weights compared to a null distribution (p<0.0.05) for each weight. Negative colors refer to corrected rs-fMRI connectivity in patients with improving symptoms and Positive colors refer to corrected rs-fMRI connectivity in patients with worsening symptoms. Potential biomarkers related to the prediction model are seen in Figure 4.3 B and C. The weights from the model trained on Dataset 1 and applied to Dataset 2 and 3, were tested for significance by permuting the outcome measures for the model trained in Dataset 1 1,000 times and testing the actual weight value for the model against the null distribution (p<0.05). The resulting 13 significant regions from the Power 264 atlas are visualized with the median connectivity data across participants. We took the median because the connectivity from each region does not follow a normal distribution. This connectivity data was categorized into those with improving symptoms (Figure 4.3 B) or worsening symptoms (Figure 4.3 C), as defined previously in Chapter 2. Regions that show opposite connectivity in those with improving symptoms to those with worsening symptoms include the left postcentral gyrus, right putamen/insula, right Rolandic operculum, left fusiform gyrus, and left putamen (queried with AFNI’s ‘whereami’ using labels from the AFNI’s CA_ML_18_MNIA atlas (144,153) and checked with FSL’s Harvard-Oxford cortical and subcortical structural probabilistic atlases (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Atlases). Furthermore, the model weight values are applied to Figures 4.3 B and 4.3 C indicated by sphere size. The PAG shows negative median connectivity across participants in patients with improving symptoms (10 out of 13 regions) and positive connectivity in patients with worsening symptoms (8 out of 13 regions) (see Table 4.1). The right putamen/insula area shows the largest weight value of those that are significant in the model and is also the only region with significantly different connectivity between patients that improve and 89 patients that worsen (p=0.02, two-sided Wilcoxon rank sum test, improve: N=107, worsen: N=109). The insula is also the only region with a significant linear association to slope value (p=0.03, linear regression, total dataset: N=327). Even if the other regions do not show significant relationships, these regions significantly affect our prediction and are important in our understanding of regions that drive our results. Table 4.1: X (mm) Y (mm) Z (mm) Improving corrected median (z-score(IQR) ) Worsening corrected median (z-score(IQR)) p- value/FDR corrected p- value p-value (linear associatio n with slope) Regions Right postcentral gyrus -23 -30 72 -0.019(0.21) 0.013(0.22) 0.59/0.60 0.29 Left postcentral gyrus 51 -6 32 -0.017(0.18) -0.002(0.18) 0.16/0.60 0.08 Right putamen/insula 37 1 -4 -0.033(0.23) 0.014(0.22) 0.02*/0.32 0.03* Right Rolandic operculum 56 -5 13 -0.009(0.21) 0.003(0.22) 0.46/0.60 0.77 Left superior frontal gyrus -20 45 39 0.019(0.18) 0.006(0.16) 0.47/0.60 0.66 Right cuneus -20 45 39 -0.007(0.19) -0.007(0.18) 0.60/0.60 0.29 Left fusiform gyrus 15 -87 37 -0.020(0.26) 0.018(0.26) 0.21/0.60 0.12 Left middle frontal gyrus -33 -79 -13 -0.004(0.21) -0.029(0.16) 0.52/0.60 0.48 Left angular gyrus -42 38 21 0.004(0.18) 0.010(0.19) 0.54/0.60 0.27 Right middle cingulate cortex -42 -55 45 -0.018(0.21) -0.011(0.17) 0.50/0.60 0.16 Left putamen 11 -39 50 -0.023(0.20) 0.006(0.18) 0.37/0.60 0.40 Left cerebellum -22 7 -5 -0.024(0.19) -0.007(0.22) 0.58/0.60 0.62 Right superior parietal lobule 25 -58 60 0.001(0.20) 0.037(0.18) 0.35/0.60 1.00 Table 4.1: Connectivity values in improving and worsening participants from regions in the Power 264 atlas with significant weights derived from the model that generalizes. Region labels, Montreal Neurological Institute (MNI) for the peak voxel coordinates, median UCPPS improving z-score and interquartile range (IQR), median UCPPS worsening z-score and interquartile range (IQR), and p-value (two-sided Wilcoxon rank sum test): uncorrected and false discovery rate corrected, p-value for linear relationship of connectivity values to slope values. AFNI’s CA_ML_18_MNIA atlas (45,54) was used for region labels based on Power 264 atlas coordinates. The above results were derived from the initial training and testing with LOOCV in each independently corrected Dataset 1, 2, and 3. Figure 4.4 shows successful LOOCV at each of these 90 datasets. LOOCV with SVR was first applied on Dataset 1 (adjusted R-squared=0.424, p=3.81e- 16, ppermuted=0; participants N=121, features N=229) as a training dataset (Figure 4.4 A). This was then followed by the model’s application separately to Dataset 2 (adjusted R-squared=0.437, p=2.38e-15, ppermuted=0; participants N=110, features N=229) (Figure 4.4 B). Once both these LOOCV models showed successful predictions, the model was tested for generalizability, as explained above (see Methods, ‘Machine learning prediction and biomarker evaluation’). The same steps were applied to Dataset 3 (adjusted R-squared=0.453, p=8.06e-17, ppermuted=0; participants N=116, features N=229) (Figure 4.4 B). Contributing weights that appeared in each fold were visualized using their Power 264 Atlas (Figure 4.4). In particular, regions within the somatosensory motor, default mode, and visual areas as well as regions near the insular cortex appear in all three datasets, though their distribution is varied. 91 Figure 4.4: 92 Figure 4.4: 12-month symptom LOOCV prediction for A. Dataset 1 (participants N=121, features N=229), B. Dataset 2 (participants N=110, features N=229 ) and C. Dataset 3 (participants N=116, features N=229) and in contributing consistent weights found in each fold using the Power Atlas. The color scheme refers to the Power Atlas (see Key). Key: SMM, somatosensory/motor (hand), SMMm, somatosensory/motor (mouth), COTC, cingulo-opercular task- control, Aud, auditory, DMN, default-mode, MR, memory retrieval, Vis, visual, FPTC, frontoparietal task-control, SN, salience, subC, subcortical, VAN, ventral attention, DAN, dorsal attention, CBL, cerebellum, Other, other regions. We tested other MAPP datasets using the same methods as described above for our 12- month prediction. For 3-month symptom trajectories using the Power 264 atlas parcellation, only the model trained on Dataset 2 generalized to the values in Dataset 3. While this prediction was significant (adjusted R-squared=0.034, p=0.03, ppermuted=0.02) the model did not generalize from Dataset 1 to 2 and we found that it was not as robust as the 12-month predictions. LOOCV was not significant in Dataset 1 (p>0.05, permutation test, N=1,000) and may be a marker of why this dataset did not generalize to the other two. In the voxelwise analysis, we analyzed 12-month and 3-month slope values. However, due to the large number of voxels in all datasets the number of features were reduced using pairwise correlation to the top 10,000, 1,000, and 100 voxels (see Methods, ‘Machine learning prediction and biomarker evaluation’). We also used the clusters connected to the MNI-trace, found to be significantly different between healthy controls and patients in a two-tailed t-test (ignoring multiple comparisons) in our PAG analysis. At 10,000 and 1,000 voxels, the prediction appeared to be overfitting the dataset. Adjusted R-squared values were highly significant (p<10e-12) but when the actual versus predicted values were plotted, they did not follow the line of unity. Moreover, at 1,000 top features none of the datasets generalized. At 100 top features, LOOCV results appeared promising at 12-months (Dataset 1 and 2 were significant, p<0.05, and Dataset 3 p=0.121), but these results did not generalize well. In the last analysis, using the significant clusters connected to the MNI-trace, LOOCV results significant (p<0.05, permutation test, N=1,000) and fit the unity line at 12-months compared to 3-months, however, these datasets did not generalize well either. 93 Discussion This research sought to determine if baseline resting state connectivity could be used to predict changing UCPPS symptoms up to a year. Using statistical and machine learning approaches to preprocess and analyze our dataset, we successfully predicted symptom changes from baseline. The major finding of our work was that we were able to not only make these predictions, but also generalize our support vector machine model onto a new, untested dataset (Figure 4.3 A, bottom left panel). One of the hallmarks of a successful machine learning prediction is its ability to translate to a novel independent dataset. Typically, as in our previous MAPP-I prediction (39), many neuroimaging studies have used machine learning to make predictions and develop biomarkers (119,146,197,198). Few have looked at longitudinal outcome measures and many of these studies have also been limited by the number of participants (n<100), which in turn limits the flexibility to generalize (197). Cross validation is an essential step in prediction modeling with a smaller number of participants, with LOOCV utilized on smaller datasets. In our MAPP-II dataset we had the opportunity to split the number of participants into three groups of approximately 120 participants. This provided a strong dataset for training, testing, and validating our algorithm. We were able to train Dataset 1 using LOOCV and other cross validation methods, as well as test these methods on balanced groups. Despite this, our dataset was not without outliers. The outliers seen in our Dataset 3 (Figure 4.3 A, bottom left panel) had estimated negative slope values to be more positive than the rest of the dataset (outlier 1: actual value=-0.15, predicted value=0.24; outlier 2: actual value=-0.18, predicted value=0.08). These are corrected within the final linear calculation between actual and predicted values. It is important to note, the 94 generalization of the model trained on Dataset 1 to Dataset 2 provided the basis to consider this removal of the two outliers and make a significant estimation on Dataset 3. The bulk of our effort and analysis was delegated to the number of processing steps needed to develop the final algorithm. Ensuring that our input data was clean and accurate for predicting UCPPS symptoms. Steps included preprocessing in AFNI (Figure 4.1) to correct potential motion related artifacts and outliers as well as limiting the dataset to connectivity between the PAG and the Power 264 atlas. We used a basic support vector regression for the final prediction and did not tune our parameters because our prediction was successful in LOOCV (Dataset 1: adjusted R- squared=0.424, p=3.81e-16, ppermuted=0). This type of model fits a maximum margin hyperplane to separate two classes, however, in our algorithm we used the model to fit a regression due to the continuous nature of our outcome data. The goal of the support vector regression is to maintain a large margin while minimizing the number of errors, or interior data points within the margin. The C-parameter in support vector machines (SVM) is a metric that helps avoid misclassification in the training data and is included in soft-margin SVM, where a penalty is given for misclassified data points. Also, epsilon provides a penalization of misclassified points based on the margins. Though our prediction generalized at 12-month symptoms, 3-month and voxelwise analysis did not perform as well. By adjusting parameters such as C and epsilon with a grid search may improve these predictions in future work. For example, the model from Dataset 1 for 3-months did not generalize well and adjusting for these parameters may provide better translation to new datasets. Moreover, the prediction at 12-months may be further improved, with a stronger adjusted R- squared in our generalization (Dataset 2: adjusted R-squared=0.046, p=0.02, ppermuted=0.01; Dataset 3: adjusted R-squared=0.05, p=0.01, ppermuted=0.009). Alternatively, if the accuracy of the 95 support vector model does not change with adjusted parameters, this may tell us that the are no soft-margin errors in the dataset and that the data is fit with the best line. Another explanation to why one model may generalize, whereas others do not, may be due to the outcome distributions. Corrections were made by removing outliers in the 12-month slope value outcome measure prior to support vector prediction, such as removing outliers and demeaning by site. While our LOOCV prediction did not perform as well after removing outliers on Dataset 1 (adjusted R-squared=0.0228, p=0.05, ppermuted=0.20) as without outliers (adjusted R- squared=0.424, p=3.81e-16, ppermuted=0), Figure 4.4 A), after outlier removal it was possible to train a model that was able to generalize to Dataset 2 and 3. Differences in generalizability could also be due to differences in site split (Figure 4.2 B). Demeaning slope values helps adjust for this discrepancy as well as outlier removal from the slope values in Dataset 1, 2, and 3. The distribution of slope values in each dataset after corrections did not differ one another after testing with a two- sample Kolmogorov-Smirnov test (p>0.05, Dataset 1: N=115, Dataset 2: N=104, Dataset 3: N=110). However, analysis of the confidence interval (Dataset 1: 95% CI[ -0.0144, 0.0177], Dataset 2: 95% CI [ -0.0168, 0.0168], and Dataset 3 95% CI[-0.0053, 0.0264]) and application of skewness in MATLAB, shows that Dataset 2 is more negatively skewed (s=-0.20) and Dataset 3 is slightly positively skewed (s=0.02). Dataset 1 possibly performs better with generalizing on Datasets 2 and 3, than a model trained on Dataset 2 and generalized to 3 (p>0.05, permutation test, N=1000) because the slope values exhibit a more balanced representation of the dataset. A key component of our study was to further validate our predictive results by interpreting the features that defined our prediction and establishing these regions as acceptable biomarkers for longitudinal prediction. An important aspect of support vector regression is that feature weights from this model allow us to understand the baseline connectivity features that potentially drive a 96 patient to either improve or worsen over time. The Power 264 atlas uses a combination of meta- analytic and functional connectivity methods to define regions of the brain that are functionally relevant to brain organization (148). This atlas regulated the number of features, reducing the potential effects of overfitting and allowing for better signal to noise in the data from averaged regions. Most of the regions that were significant in our final generalized model, have been shown to be directly or indirectly connected to the PAG. In particular, regions that showed contrasting effects in patients that improve versus patients that worsen overtime may contribute to the coricostriatal and corticolimbic pathway’s influence on modulation of pain symptoms (100,199,200). We show this in the putamen, insula, and operculum. The insula (the only region that was significantly different in patients with improving versus worsening as well as linearly related to slope value, see Table 4.1) is well known to be involved in the pain experience, playing a meaningful role in sensation and affect (18). Furthermore, the insula has appeared in MAPP studies (195) showing differences between patients and healthy controls. Other neuroimaging studies have shown that the insula is more active due to pain, when a participant focuses on the pain versus distracted by a task (18,201–203) and greater activity is associated with pain catastrophizing (204). The somatosensory cortex also activates in such studies, which is another region where we observe contrasting effects between improving and worsening participants. Moreover, anatomical effects, such as grey matter reductions, are most commonly found in the insular region as a cause of chronic pain development (16,18). An interesting observation after comparing the connectivity of patients that improve over time to patients that worsen, is that the majority of regions have negative connectivity with improving symptoms and positive connectivity with worsening symptoms (10 out of 13 and 8 out of 13, respectively). The prefrontal cortex has been shown to have an inhibitory effect on the PAG, 97 via control of the amygdala and hypothalamic pathways (100,205). Mayer and colleagues found that there was a mediating inhibitory connection from the medial prefrontal cortex to the dorsal pons/PAG region only seen in healthy controls and patients with ulcerative colitis, and not in irritable bowel symptom patients (100). As stated, this inhibition of the PAG, where “less activation is seen in faciliatory regions thought to inhibit the PAG” such as dorsal medial prefrontal cortex, anterior cingulate cortex, and amygdala is corroborated by functional neuroanatomical and neuroimaging approaches (100). Another study using optogenetic approaches in mice models has shown that the circuit from the basolateral amygdala to inhibitory interneurons in the prefrontal cortex mediates feedforward inhibition to the ventrolateral PAG, to alter pain behavior (205). It is possible that the higher connectivity to the PAG in regions associated with descending control of pain, such as those in the putamen and insula, are due to inactivation or malfunctioning of the inhibitory corticolimbic system, as suggested by Mayer and colleagues for patients with IBS (100). However, negative correlation to the PAG in improving patients does not directly relate to inhibition and, furthermore, the physiological properties of negative correlation have been debated (206). Other relevant regions in the study of pain, such as the amygdala and parts of the anterior cingulate cortex, are not covered by the Power 264 atlas and so are absent from our study. Nevertheless, they may contribute toward this in inhibition of PAG activity. Alternative approaches, such as a more in depth voxelwise analysis or an addition of such regions to the Power 264 atlas, may also assist in characterizing this possible inhibitory behavior in our dataset. Future work is necessary to better understand how baseline connectivity relates to improving and worsening patients. The regions shown in this study that contribute significantly to our predictive model are promising and overlap with regions associated with the PAG and circuitry related to pain perception and modulation (18). The PAG is included in the same circuit involving 98 regions such as the insular cortex and studies have shown an anatomical connection to the insula (48,207). We have not stratified our data by gender in our study, and this step may improve the specificity of our prediction. We found an interaction effect of baseline pain and sex relative to 12-month slope values which should be addressed in future studies. In a recent study on sex differences in pain related to the functional connectivity of PAG, men were shown to have increased connectivity between the PAG and putamen, whereas this was not seen in women (111). However, this study by Linnmen and colleagues also used the coordinates of the PAG found to be more anterior than the anatomical location we use here as a seed, informed by our previous work (see Chapter 3). Our connectivity measure does not provide any information on the directionality of the connectivity, and many of the significant regions in our model are shown in both “bottom- up” and “top-down” attention to pain (18,203). Methods such as moderation and mediation are important techniques in neuroimaging in examining how a third component fits into the relationship (208), as between the PAG and insula. We can use moderation to understand how our outcome measure is changed by connectivity to the PAG, and whether patients will improve by testing the connectivity strength from the PAG to the insula. These results would corroborate our prediction that the connectivity is more negative at baseline in patients who improve. With mediation we can test the effect of the PAG on symptom change with the insula as a mediator. Alternatively, we can test for other regions, such as the prefrontal cortex’s role in PAG descending activation. By defining the relationship with the PAG more clearly through our understanding of the directionality of connectivity of significant regions to the PAG, we can then identify the best treatment to target and adjust regions in patients that do not improve over time, such as with targeted brain stimulation. 99 Conclusion To predict longitudinal changes in UCPPS symptoms using baseline functional connectivity derived from rs-fMRI, we developed and optimized an algorithm with extensive preprocessing and machine learning techniques. After splitting our dataset of 389 participants into three different groups to train, test, and validate our model, our main finding suggests we are able use our preprocessed connectivity data from baseline to significantly predict and generalize 12- month changes in UCPPS symptoms. The regions associated with our predictive model overlap with those that have been found in previous studies to contribute to pain perception and modulation. Additionally, these regions have been established in previous literature (18,49,182,207,209,210) as having a functionally direct relationship to the PAG. This study demonstrates that prediction is a useful method in understanding the differential effects that may occur in regions that lead to a UCPPS patient improving over time. Future work will develop these findings to establish treatment targets for clinical applications. 100 Chapter 5: Discussion The prognosis of chronic pain symptoms, whether a patient will get better over time, is arguably one of the most crucial issues in chronic pain management. Evidence suggests functional and anatomical changes occur in the brain due to chronic pain relating to both pain control and altered cognition (15,18,36,37,100,105,108,195,211). Resting state connectivity derived from fMRI provides an opportunity to identify underlying markers intrinsic to the individual that may tell us about future outcomes. By analyzing resting state neuroimaging measures at baseline stage in UCPPS pain patients, we established it is possible to some extent to predict future long-term changes of UCPPS pain symptoms, reported up to a year. We focused on whole-brain regions in relation to the periaqueductal gray (PAG). The PAG is one of the most prominent and extensively studied pain regions, known to be a key factor in modulation (47,59,63,64,70,78,136,210,212). Our hypothesis suggested that the maladaptive behavior of the PAG due to chronic pain, may lead to the maintenance of UCPPS symptoms. By examining whole-brain connectivity to the PAG, we examined other regions that may contribute to the symptoms of UCPPS. These regions have the potential to define biomarkers specific to UCPPS patients. Machine learning results In order to make our prediction of UCPPs symptom progression we used a combination of statistical preprocessing and machine learning techniques on our dataset of 389 participants from the Multidisciplinary Approach to the Study of Chronic Pelvic Pain (MAPP) research network. A distinguishing characteristic of this dataset is the sample size. Only until recently (119) have longitudinal neuroimaging studies included enough participant data needed to validate on a 101 completely independent dataset. Even so, as a part of the machine learning algorithm, the predictor data was set up to minimize the potential for overfitting. Overfitting takes place when the number of features is much greater than the number of outcome values predicted, as with the curse of dimensionality, and the model predicts noise rather than a meaningful pattern in the data (197). We split our data into three datasets of approximately 120 participants to train and test our model. We left the third of these independent datasets to validate our finalized model. Overfitting was also controlled by using linear support vector machines (SVM) and leave-one-out cross validation in our data. The linear SVM model has also been shown to be more appropriate for generalization and is more computationally efficient for larger datasets (213,214). SVM models are one of the most popular in neuroimaging studies (214), and are a good choice for creating explainable models (215,216) where clear biologically relevant inference can be made between the features used and the model results. An important part of the predictive algorithm developed in this thesis, was the model’s ability to generalize to our independent dataset. In neuroimaging, machine learning is an excellent method to create models that generalize well (197,198,217). Models generated using this method, tend to generalize much better than those derived through simple correlation between MRI and independent outcome measures (197). Many studies have utilized machine learning to identify biomarkers for different brain disorders (18,215). However, there are a limited number of studies which have also tested on a separate dataset (119,197), crucial for validation. Furthermore, the models developed on certain datasets should be able to generalize across individuals, methods, settings, and populations (215). Part of this generalizability includes sharing analysis workflows and methods among laboratories (215,218). To illustrate this issue, Botvinik-Nezer and colleagues showed that the same neuroimaging dataset analyzed by 70 independent teams lead to variability 102 in outcomes (218). Here, we were able to generalize our data to both our test and validation datasets based on our model from our first training dataset. Though not a strong linear effect (Dataset 2: adjusted R-squared=0.046, p=0.02, ppermuted=0.01, Dataset 3: adjusted R-squared=0.05, p=0.01, ppermuted=0.009), this significant prediction was made with data from across six different collection sites, using similar methods at each site for consistency. However, some methodological inconsistency, such as data collection on different scanners, could not be avoided. We hope generalization in our data will make our model and the biomarkers we identified potentially more transferable to new datasets and more robust to additional methodical changes. It may even show some level of robustness to different pain populations, though accuracy inevitably may decrease. UCPPS symptoms Assuming a linear relationship, we have shown by using the model generated in Chapter 4, UCPPS symptom change up to 12-months is a feasible measure for prediction. One of the issues with pain reported symptoms is that they are given as a subjective experience of the individual patient and perception of pain at one time point may change relative to the next. To generate symptom indices for UCPPS, an exploratory factor analysis was applied to multiple questionnaires known for characterizing UCPPS (42). Pain score along with urinary symptoms (but not combined) were found to describe UCPPS accurately. We were able to take the slope of these symptom changes for pain and use these for our prediction. Slope is a measure calculated relative to the individual patient’s baseline, which is a method to standardize subjective change based on the patient’s own measurement. Combining self-reporting with fMRI strengthens the validity of such self-reported metrics and underscores their use as prognostic measures (215). In our previous MAPP-I prediction we were only able to predict using the UCPPS pain score (39). We used this 103 MAPP-I result to structure our predictive model presented in this thesis on the MAPP-II dataset. Though the outcome measures are subjective, our predictors allow for the estimation of a more objective measurement. UCPPS is marked by symptom variability and this is likely to be a source of noise in the prediction of change a year in advance. Our model could be strengthened by stratifying the data into subcategories of the population (i.e. stratification by sex) to improve precision. However, this in turn would also decrease the sample size. Furthermore, we focus our predictive results on the extremes in the data, patients that improve or worsen. We do not attend to stable change, as unchanging symptoms over time provide less information on potential treatment markers. However, as the slope is an average over the 12-month timeframe, stable participants likely have fluctuating slope trajectories in the short-term. Prediction at the 3-month timeframe may take into account changes not seen in these stable slope trajectories at 12-months. In this case different connectivity profiles and biomarkers leading to 3-month prediction will provide potential treatment targets for these shorter changes in symptoms from baseline. Beyond only the pain score, comorbidities and overlapping pain conditions may be related to our predictive outcome measure as potential modulators. There are multiple systems that are involved in pain relief. For example, emotional state and attention to pain both influence descending modulatory effects on pain (18). Negative emotion associated with the anterior cingulate cortex-fronto-PAG circuitry correlated with pain-evoked activity in the anterior cingulate cortex related to negative mood (219). Naliboff and colleagues showed that pain catastrophizing and self-reported stress in UCPPS patients’ predicted pain outcome, however, this did not apply to anxiety and depression (127). For further specificity, biomarkers that are generated from our prediction should be mainly representative of pain itself if they are to reflect the mechanism of pain (215,220). Nonetheless, it is likely that our results overlap with other factors 104 that influence the pain experience. It is important to note, that we were able to see a significant trend in the pain scores towards improvement over time, but it is not clear if change in UCPPS pain score is mediated by other comorbidities. Potential biomarkers Of the significant regions in our model, improving symptoms represented a decreasing slope, which is related to more negatively correlated regions with the PAG (Chapter 4, Figure 4.3 B). On the other hand, worsening symptoms represented an increasing slope, which is related to more positively connected regions with the PAG (Chapter 4, Figure 4.3 C). It is possible that this behavior is partially driven by the linear fit in the prediction, but these significant regions all impact pain perception via descending pathways, such as emotion and placebo analgesia activating the prefrontal to PAG regions and attention activating projections from the superior parietal to the somatosensory and insula to the amygdala and PAG (18). While these differences in connectivity are not significant across all regions (see Chapter 4, Table 4.1), this behavior provides insight to how patients that improve naturally over the course of a year do so relative to patients with worsening symptoms over the same period. The PAG to insula connectivity had the strongest weight value in the regions that are significant from our predictive model. It is also the only significantly different region between patients that improve compared to patients that worsen (see Chapter 4, Table 4.1). The insula is well documented in pain studies to be associated with pain sensation and affect (18). The ventral insula has been shown in primates to project directly to the lateral and dorsolateral columns of the PAG from agranular areas (48,207) and may receive inhibitory feedback from the insula and others (i.e. the lateral hypothalamus and basal amygdala). Furthermore, the lateral and dorsolateral 105 regions of the PAG were shown to be positively and negatively correlated to the insula, respectively, in a resting-state fMRI analysis of PAG subdivisions (49). In our study, the negative correlation of the insula to the PAG may be a marker of this inhibitory effect, whereas, patients that worsen may have lost this control. A recent study by Huang and colleagues has also shown direct influence of the medial prefrontal cortex on the ventrolateral PAG, altering neuropathic pain behavior in mice models (205). This study showed that afferent signaling from nerve injury strengthened synaptic input of the basolateral amygdala to the medial prefrontal cortex which consequently, decreased signal to the ventrolateral PAG, increasing the pain experience. Optogenetic inhibition of the basolateral amygdala to medial prefrontal cortex resulted in pain relief (205). As we do not subdivide the PAG, the mixture of signals potentially coming from different areas of the PAG may influence our results. In the future our predictive accuracy may improve by seeding based on these distinct regions (49,210). Even so, neuroimaging studies have continued to use the PAG location as a whole (209) (as well as erroneous locations of the PAG, see Chapter 3) due to the resolution achievable with 3 tesla scanners. Notably, we were still able to make a prediction with the whole PAG signal. While the insula and the PAG contribute towards are prediction and are both involved in the descending modulation of pain, further analysis is needed to establish these as prognostic biomarkers. First, resting state connectivity is an important starting point for our prediction, but does not give any information regarding the directionality of the signal. Many of the regions involved in pain modulation both incorporate afferent and efferent signaling, enhancing or reducing the incoming pain signal in multiple regions. Furthermore, other studies have tailored the Power 264 atlas to define additional regions in their analysis. Key regions in pain modulation (18,209) that show anatomical connectivity to the PAG (48), such as the subgenual anterior 106 cingulate, hypothalamus, amygdala, and orbitofrontal regions, should be included in the brain parcellation. Once predictive regions are established, but before they are used as prognostic biomarkers, they need to be validated with further testing on new datasets that increasingly show better differentiation between patients that are improving and worsening. Future steps There are few prognostic biomarkers that have been validated in chronic pain. Establishing biologically meaningful as well as interpretable features for prediction of symptoms is imperative to our understanding of pain mechanisms and treatment of pain (46,215). The first step in further defining such biomarkers established in the predictive model in this thesis, is to 1) test the performance of the predictive model in UCPPS patients to new chronic pain population datasets 2) determine whether biomarkers originally defined in UCPPS predictions generalize to other conditions, and 3) determine if and how these regions are influenced following treatment response or intervention. Prognostic biomarkers defined in the longitudinal progression of symptoms can be used to tailor personalized treatment paradigms to individuals susceptible of developing or currently suffering from chronic pain. In UCPPS pain symptoms alone we have seen almost a 50% decrease in reported pain in about a third of the study group over a period of 12-months, whereas other participants’ symptoms have either remained steady or worsened (Chapter 2). Based on our prediction and its validity across other populations, we can define specific biomarkers for those whose symptoms improved compared to those who do not and then test for predictive biomarkers in patient subpopulations that may benefit from treatment. For instance, is there a treatment (e.g. transcranial magnetic stimulation, cognitive behavioral therapy, pharmacotherapy, deep brain 107 stimulation, etc.) that will alter activity in a specific brain area, such as the insula, of those who’s symptoms do not improve over time so that the activity becomes closer to that of those that do improve? In other words, can we manipulate patients into becoming improvers? Additionally, do these biomarkers indicate a population of patients who are more or less likely to respond to treatment (i.e. a subpopulation that responds differently to certain treatments compared to improvers)? There may also be a discrepancy between pain populations, for example, our method for prediction may translate to new datasets, but the model may be driven by alternative biomarkers. In this case, we can then examine why distinct features are unique to different conditions of chronic pain. This distinction could then be used for tailoring specific types of treatment uniquely to each chronic pain condition. Future research should assess and validate what underlying mechanisms drive pain progression across various chronic pain conditions and elucidate the brain regions that relate to pain mechanisms specifically in order to make a proper treatment assessment. 108 References 1. Dahlhamer J. Prevalence of Chronic Pain and High-Impact Chronic Pain Among Adults — United States, 2016. MMWR Morb Mortal Wkly Rep [Internet]. 2018 [cited 2020 Jun 5];67. Available from: https://www.cdc.gov/mmwr/volumes/67/wr/mm6736a2.htm 2. Gaskin DJ, Richard P. The Economic Costs of Pain in the United States. J Pain. 2012 Aug 1;13(8):715–24. 3. Treede R-D, Rief W, Barke A, Aziz Q, Bennett MI, Benoliel R, et al. A classification of chronic pain for ICD-11. Pain. 2015 Jun;156(6):1003–7. 4. Frank A. Low back pain. BMJ. 1993 Apr 3;306(6882):901–9. 5. Apkarian AV, Baliki MN, Geha PY. Towards a theory of chronic pain. Prog Neurobiol. 2009 Feb;87(2):81–97. 6. Argoff CE. Review of current guidelines on the care of postherpetic neuralgia. Postgrad Med. 2011 Sep;123(5):134–42. 7. Nalamachu S, Morley-Forster P. Diagnosing and Managing Postherpetic Neuralgia. Drugs Aging. 2012 Nov;29(11):863–9. 8. Gureje O, Von Korff M, Simon GE, Gater R. Persistent pain and well-being: a World Health Organization Study in Primary Care. JAMA. 1998 Jul 8;280(2):147–51. 9. Young Casey C, Greenberg MA, Nicassio PM, Harpin RE, Hubbard D. Transition from acute to chronic pain and disability: a model including cognitive, affective, and trauma factors. Pain. 2008 Jan;134(1–2):69–79. 10. Tracey I, Bushnell MC. How Neuroimaging Studies Have Challenged Us to Rethink: Is Chronic Pain a Disease? J Pain. 2009 Nov 1;10(11):1113–20. 11. Melzack R. Pain and the neuromatrix in the brain. J Dent Educ. 2001 Dec;65(12):1378–82. 12. Melzack R. From the gate to the neuromatrix. Pain. 1999 Aug;Suppl 6:S121-126. 13. May A. Chronic pain may change the structure of the brain. Pain. 2008 Jul;137(1):7–15. 14. Simons L, Elman I, Borsook D. Psychological Processing in Chronic Pain: A Neural Systems Approach. Neurosci Biobehav Rev. 2014 Feb;0:61–78. 15. Baliki MN, Petre B, Torbey S, Herrmann KM, Huang L, Schnitzer TJ, et al. Corticostriatal functional connectivity predicts transition to chronic back pain. Nat Neurosci. 2012 Aug;15(8):1117–9. 109 16. As-Sanie S, Harris RE, Napadow V, Kim J, Neshewat G, Kairys A, et al. Changes in regional gray matter volume in women with chronic pelvic pain: a voxel-based morphometry study. Pain. 2012 May;153(5):1006–14. 17. Plow EB, Pascual-Leone A, Machado A. Brain stimulation in the treatment of chronic neuropathic and non-cancerous pain. J Pain Off J Am Pain Soc. 2012 May;13(5):411–24. 18. Bushnell MC, Čeko M, Low LA. Cognitive and emotional control of pain and its disruption in chronic pain. Nat Rev Neurosci. 2013 Jul;14(7):502–11. 19. Institute of Medicine (US) Committee on Advancing Pain Research, Care, and Education. Relieving Pain in America: A Blueprint for Transforming Prevention, Care, Education, and Research [Internet]. Washington (DC): National Academies Press (US); 2011 [cited 2016 Aug 1]. (The National Academies Collection: Reports funded by National Institutes of Health). Available from: http://www.ncbi.nlm.nih.gov/books/NBK91497/ 20. Clemens JQ, Calhoun EA, Litwin MS, McNaughton-Collins M, Kusek JW, Crowley EM, et al. Validation of a modified National Institutes of Health chronic prostatitis symptom index to assess genitourinary pain in both men and women. Urology. 2009 Nov;74(5):983– 7, quiz 987.e1-3. 21. Clemens JQ, Markossian T, Calhoun EA. Comparison of economic impact of chronic prostatitis/chronic pelvic pain syndrome and interstitial cystitis/painful bladder syndrome. Urology. 2009 Apr;73(4):743–6. 22. Clemens JQ, Mullins C, Kusek JW, Kirkali Z, Mayer EA, Rodríguez LV, et al. The MAPP research network: a novel study of urologic chronic pelvic pain syndromes. BMC Urol. 2014;14:57. 23. Michael YL, Kawachi I, Stampfer MJ, Colditz GA, Curhan GC. Quality of life among women with interstitial cystitis. J Urol. 2000 Aug;164(2):423–7. 24. Allsop SA, Erstad DJ, Brook K, Bhai SF, Cohen JM, Dimitrakoff JD. The DABBEC Phenotyping System: towards a mechanistic understanding of CP/CPPS. Nat Rev Urol. 2011; 25. Clemens JQ, Clauw DJ, Kreder K, Krieger JN, Kusek JW, Lai HH, et al. COMPARISON OF BASELINE UROLOGIC SYMPTOMS IN MEN AND WOMEN IN THE MULTIDISCIPLINARY APPROACH TO THE STUDY OF CHRONIC PELVIC PAIN (MAPP) RESEARCH COHORT. J Urol. 2015 May;193(5):1554–8. 26. Fall M, Baranowski AP, Elneil S, Engeler D, Hughes J, Messelink EJ, et al. EAU Guidelines on Chronic Pelvic Pain. Eur Urol. 2010 Jan;57(1):35–48. 27. Yang CC, Lee JC, Kromm BG, Ciol MA, Berger RE. Pain Sensitization in Male Chronic Pelvic Pain Syndrome: Why Are Symptoms so Difficult to Treat? J Urol. 2003 Sep 1;170(3):823–7. 110 28. Lai HH, Gardner V, Ness TJ, Gereau RW. Segmental hyperalgesia to mechanical stimulus in interstitial cystitis/bladder pain syndrome: evidence of central sensitization. J Urol. 2014 May;191(5):1294–9. 29. Lai H. Henry, Krieger John N., Pontari Michel A., Buchwald Dedra, Hou Xiaoling, Landis J. Richard, et al. Painful Bladder Filling and Painful Urgency are Distinct Characteristics in Men and Women with Urological Chronic Pelvic Pain Syndromes: A MAPP Research Network Study. J Urol. 2015 Dec 1;194(6):1634–41. 30. Potts JM, O’Dougherty E. Pelvic floor physical therapy for patients with prostatitis. Curr Urol Rep. 2000 Jun;1(2):155–8. 31. Lee SWH, Liong ML, Yuen KH, Liong YV, Krieger JN. Chronic Prostatitis/Chronic Pelvic Pain Syndrome: Role of Alpha Blocker Therapy. Urol Int. 2007;78(2):97–105. 32. Landis JR, Williams DA, Lucia MS, Clauw DJ, Naliboff BD, Robinson NA, et al. The MAPP research network: design, patient characterization and operations. BMC Urol. 2014;14:58. 33. Alger JR, Ellingson BM, Ashe-McNalley C, Woodworth DC, Labus JS, Farmer M, et al. Multisite, multimodal neuroimaging of chronic urological pelvic pain: Methodology of the MAPP Research Network. NeuroImage Clin. 2016 Jan 6;12:65–77. 34. Bagarinao E, Johnson KA, Martucci KT, Ichesco E, Farmer MA, Labus J, et al. Preliminary structural MRI based brain classification of chronic pelvic pain: A MAPP network study. PAIN®. 2014 Dec;155(12):2502–9. 35. Farmer MA, Huang L, Martucci K, Yang CC, Maravilla KR, Harris RE, et al. Brain White Matter Abnormalities in Female Interstitial Cystitis/Bladder Pain Syndrome: A MAPP Network Neuroimaging Study. J Urol. 2015 Jul;194(1):118–26. 36. Kilpatrick LA, Kutch JJ, Tillisch K, Naliboff BD, Labus JS, Jiang Z, et al. Alterations in resting state oscillations and connectivity in sensory and motor networks in women with interstitial cystitis/painful bladder syndrome. J Urol. 2014 Sep;192(3):947–55. 37. Kutch JJ, Yani MS, Asavasopon S, Kirages DJ, Rana M, Cosand L, et al. Altered resting state neuromotor connectivity in men with chronic prostatitis/chronic pelvic pain syndrome: A MAPP: Research Network Neuroimaging Study. NeuroImage Clin. 2015;8:493–502. 38. Schrepf A, O’Donnell M, Luo Y, Bradley CS, Kreder K, Lutgendorf S, et al. Inflammation and inflammatory control in interstitial cystitis/bladder pain syndrome: Associations with painful symptoms. Pain. 2014 Sep;155(9):1755–61. 39. Kutch JJ, Labus JS, Harris RE, Martucci KT, Farmer MA, Fenske S, et al. Resting-state functional connectivity predicts longitudinal pain symptom change in urologic chronic pelvic pain syndrome: A MAPP Network Study. Pain. 2017 Jun;158(6):1069–82. 111 40. Mayer E. Project Information - NIH RePORTER - NIH Research Portfolio Online Reporting Tools Expenditures and Results [Internet]. [cited 2020 Jun 4]. Available from: https://projectreporter.nih.gov/project_info_description.cfm?aid=9820886&icde=5029283 2&ddparam=&ddvalue=&ddsub=&cr=2&csb=default&cs=ASC&pball= 41. Woodworth D, Mayer E, Leu K, Ashe-McNalley C, Naliboff BD, Labus JS, et al. Unique Microstructural Changes in the Brain Associated with Urological Chronic Pelvic Pain Syndrome (UCPPS) Revealed by Diffusion Tensor MRI, Super-Resolution Track Density Imaging, and Statistical Parameter Mapping: A MAPP Network Neuroimaging Study. PLOS ONE. 2015 Oct 13;10(10):e0140250. 42. Griffith JW, Stephens-Shields AJ, Hou X, Naliboff B, Pontari M, Edwards TC, et al. Pain and urinary symptoms should not be combined into one score: Psychometric findings from the Multi-disciplinary Approach to the Study of Chronic Pelvic Pain (MAPP) Research Network. J Urol. 2016 Apr;195(4 Pt 1):949–54. 43. Clemens JQ, Mullins C, Ackerman AL, Bavendam T, van Bokhoven A, Ellingson BM, et al. Urologic chronic pelvic pain syndrome: insights from the MAPP Research Network. Nat Rev Urol. 2019 Mar;16(3):187–200. 44. Clemens JQ, Kutch JJ, Mayer EA, Naliboff BD, Rodriguez LV, Klumpp DJ, et al. The Multidisciplinary Approach to The Study of Chronic Pelvic Pain (MAPP) Research Network*: Design and implementation of the Symptom Patterns Study (SPS). Neurourol Urodyn [Internet]. [cited 2020 Jun 28];n/a(n/a). Available from: https://onlinelibrary.wiley.com/doi/abs/10.1002/nau.24423 45. Stephens-Shields AJ, Clemens JQ, Jemielita T, Farrar J, Sutcliffe S, Hou X, et al. Symptom Variability and Early Symptom Regression in the MAPP Study, a Prospective Study of Urological Chronic Pelvic Pain Syndrome. J Urol [Internet]. [cited 2016 Jul 31]; Available from: http://www.sciencedirect.com/science/article/pii/S002253471630310X 46. Mackey S, Greely HT, Martucci KT. Neuroimaging-based pain biomarkers: definitions, clinical and research applications, and evaluation frameworks to achieve personalized pain medicine. PAIN Rep. 2019 Aug;4(4):e762. 47. Linnman C, Moulton EA, Barmettler G, Becerra L, Borsook D. Neuroimaging of the Periaqueductal Gray: State of the Field. Neuroimage. 2012 Mar;60(1):505–22. 48. An X, Bandler R, Ongür D, Price JL. Prefrontal cortical projections to longitudinal columns in the midbrain periaqueductal gray in macaque monkeys. J Comp Neurol. 1998 Nov 30;401(4):455–79. 49. Coulombe M-A, Erpelding N, Kucyi A, Davis KD. Intrinsic functional connectivity of periaqueductal gray subregions in humans. Hum Brain Mapp. 2016;37(4):1514–30. 50. Benarroch EE. Periaqueductal gray: An interface for behavioral control. Neurology. 2012 Jan 17;78(3):210–7. 112 51. Öngür D, Price JL. The Organization of Networks within the Orbital and Medial Prefrontal Cortex of Rats, Monkeys and Humans. Cereb Cortex. 2000 Mar 1;10(3):206–19. 52. Fowler CJ, Griffiths D, de Groat WC. The neural control of micturition. Nat Rev Neurosci. 2008 Jun;9(6):453–66. 53. Mobbs D, Petrovic P, Marchant JL, Hassabis D, Weiskopf N, Seymour B, et al. When Fear Is Near: Threat Imminence Elicits Prefrontal-Periaqueductal Gray Shifts in Humans. Science. 2007 Aug 24;317(5841):1079–83. 54. Silva C, McNaughton N. Are periaqueductal gray and dorsal raphe the foundation of appetitive and aversive control? A comprehensive review. Prog Neurobiol. 2019 Jun 1;177:33–72. 55. Ferris CF, Snowdon CT, King JA, Sullivan JM, Ziegler TE, Olson DP, et al. Activation of neural pathways associated with sexual arousal in non-human primates. J Magn Reson Imaging. 2004;19(2):168–75. 56. Bandler R, Shipley MT. Columnar organization in the midbrain periaqueductal gray: modules for emotional expression? Trends Neurosci. 1994 Jan 1;17(9):379–89. 57. Ongür D, Price JL. The organization of networks within the orbital and medial prefrontal cortex of rats, monkeys and humans. Cereb Cortex N Y N 1991. 2000 Mar;10(3):206–19. 58. Stamford JA. Descending control of pain. Br J Anaesth. 1995 Aug 1;75(2):217–27. 59. Tracey I, Ploghaus A, Gati JS, Clare S, Smith S, Menon RS, et al. Imaging Attentional Modulation of Pain in the Periaqueductal Gray in Humans. J Neurosci. 2002 Apr 1;22(7):2748–52. 60. Vanderah T, Gould D. Nolte’s The Human Brain 7th Edition: an Introduction to Its Functional Anatomy [Internet]. eBook : Document : English : Seventh edition. Philadelphia, Pennsylvania: Elsevier; 2015 [cited 2020 Jun 1]. Available from: https://www.us.elsevierhealth.com/noltes-the-human-brain- 9780323653985.html?campid=BIN4E22502&utm_source=google&utm_medium=cpc&ut m_term=&device=c&matchtype=b&adgroup=88512439832&geo_interest=&geo=903111 1&gclid=EAIaIQobChMIn7jWs_7h6QIV9f3jBx3DIQDeEAAYASAAEgLcpfD_BwE 61. Chapman CR, Vierck CJ. The Transition of Acute Postoperative Pain to Chronic Pain: An Integrative Overview of Research on Mechanisms. J Pain. 2017 Apr 1;18(4):359.e1- 359.e38. 62. Martin J. Neuroanatomy Text and Atlas, Fourth Edition. Blacklick: McGraw-Hill Publishing; 2012. 63. Koutsikou S, Apps R, Lumb BM. Top down control of spinal sensorimotor circuits essential for survival. J Physiol. 2017 Jul 1;595(13):4151–8. 113 64. Koutsikou S, Watson TC, Crook JJ, Leith JL, Lawrenson CL, Apps R, et al. The Periaqueductal Gray Orchestrates Sensory and Motor Circuits at Multiple Levels of the Neuraxis. J Neurosci. 2015 Oct 21;35(42):14132–47. 65. Kong J, Tu P, Zyloney C, Su T. Intrinsic functional connectivity of the periaqueductal gray, a resting fMRI study. Behav Brain Res. 2010 Aug 25;211(2):215–9. 66. Wei S-Y, Chao H-T, Tu C-H, Li W-C, Low I, Chuang C-Y, et al. Changes in functional connectivity of pain modulatory systems in women with primary dysmenorrhea. Pain. 2016 Jan;157(1):92–102. 67. Richardson DE, Akil H. Pain reduction by electrical brain stimulation in man: Part 1: Acute administration in periaqueductal and periventricular sites. J Neurosurg. 1977 Aug 1;47(2):178–83. 68. Liebeskind JC, Guilbaud G, Besson J-M, Oliveras J-L. Analgesia from electrical stimulation of the periaqueductal gray matter in the cat: behavioral observations and inhibitory effects on spinal cord interneurons. Brain Res. 1973 Feb 28;50(2):441–6. 69. Reynolds DV. Surgery in the Rat during Electrical Analgesia Induced by Focal Brain Stimulation. Science. 1969 Apr 25;164(3878):444–5. 70. Besson J-M, Fardin V, Olivéras J-L. Analgesia Produced by Stimulation of the Periaqueductal Gray Matter: True Antinoceptive Effects Versus Stress Effects. In: Depaulis A, Bandler R, editors. The Midbrain Periaqueductal Gray Matter: Functional, Anatomical, and Neurochemical Organization [Internet]. Boston, MA: Springer US; 1991 [cited 2020 Jun 2]. p. 121–38. (NATO ASI Series). Available from: https://doi.org/10.1007/978-1- 4615-3302-3_8 71. Tracey I, Mantyh PW. The Cerebral Signature for Pain Perception and Its Modulation. Neuron. 2007 Aug 2;55(3):377–91. 72. Tracey I, Iannetti GD. Chapter 6 Brainstem functional imaging in humans. In: Cruccu G, Hallett M, editors. Supplements to Clinical Neurophysiology [Internet]. Elsevier; 2006 [cited 2019 Jul 11]. p. 52–67. (Brainstem Function and Dysfunction; vol. 58). Available from: http://www.sciencedirect.com/science/article/pii/S1567424X09700595 73. Lee I-S, Necka EA, Atlas LY. Distinguishing pain from nociception, salience, and arousal: How autonomic nervous system activity can improve neuroimaging tests of specificity. NeuroImage. 2020 Jan 1;204:116254. 74. Seminowicz DA, Moayedi M. The Dorsolateral Prefrontal Cortex in Acute and Chronic Pain. J Pain. 2017 Sep 1;18(9):1027–35. 75. Yu R, Gollub RL, Spaeth R, Napadow V, Wasan A, Kong J. Disrupted functional connectivity of the periaqueductal gray in chronic low back pain. NeuroImage Clin. 2014 Aug 23;6:100–8. 114 76. Bandler R, Keay KA, Floyd N, Price J. Central circuits mediating patterned autonomic activity during active vs. passive emotional coping. Brain Res Bull. 2000 Sep 1;53(1):95– 104. 77. Descarries L, Watkins KC, Garcia S, Beaudet A. The serotonin neurons in nucleus raphe dorsalis of adult rat: A light and electron microscope radioautographic study. J Comp Neurol. 1982;207(3):239–54. 78. Eippert F, Bingel U, Schoell ED, Yacubian J, Klinger R, Lorenz J, et al. Activation of the Opioidergic Descending Pain Control System Underlies Placebo Analgesia. Neuron. 2009 Aug;63(4):533–43. 79. Wager TD, Scott DJ, Zubieta J-K. Placebo effects on human -opioid activity during pain. Proc Natl Acad Sci. 2007 Jun 26;104(26):11056–61. 80. Wey H-Y, Catana C, Hooker JM, Dougherty DD, Knudsen GM, Wang DJJ, et al. Simultaneous fMRI-PET of the Opioidergic Pain System in Human Brain. NeuroImage. 2014 Nov 15;102(0 2):275–82. 81. Fields HL, Heinricher MM. Anatomy and Physiology of a Nociceptive Modulatory System. Philos Trans R Soc Lond B Biol Sci. 1985;308(1136):361–74. 82. Huang KW, Ochandarena NE, Philson AC, Hyun M, Birnbaum JE, Cicconet M, et al. Molecular and anatomical organization of the dorsal raphe nucleus. Kauer JA, Marder E, Harkany T, Quintana-Urzainqui I, editors. eLife. 2019 Aug 14;8:e46464. 83. Hohmann AG, Suplita RL, Bolton NM, Neely MH, Fegley D, Mangieri R, et al. An endocannabinoid mechanism for stress-induced analgesia. Nature. 2005 Jun;435(7045):1108–12. 84. Suplita RL, Farthing JN, Gutierrez T, Hohmann AG. Inhibition of fatty-acid amide hydrolase enhances cannabinoid stress-induced analgesia: Sites of action in the dorsolateral periaqueductal gray and rostral ventromedial medulla. Neuropharmacology. 2005 Dec 1;49(8):1201–9. 85. Woodhams SG, Chapman V, Finn DP, Hohmann AG, Neugebauer V. The cannabinoid system and pain. Neuropharmacology. 2017 Sep 15;124:105–20. 86. Butler RK, Finn DP. Stress-induced analgesia. Prog Neurobiol. 2009 Jul 1;88(3):184–202. 87. Rozeske RR, Jercog D, Karalis N, Chaudun F, Khoder S, Girard D, et al. Prefrontal- Periaqueductal Gray-Projecting Neurons Mediate Context Fear Discrimination. Neuron. 2018 Feb 21;97(4):898-910.e6. 88. Fung BJ, Qi S, Hassabis D, Daw N, Mobbs D. Slow escape decisions are swayed by trait anxiety. Nat Hum Behav. 2019 Jul;3(7):702–8. 115 89. Mobbs D, Marchant JL, Hassabis D, Seymour B, Tan G, Gray M, et al. From Threat to Fear: The Neural Organization of Defensive Fear Systems in Humans. J Neurosci. 2009 Sep 30;29(39):12236–43. 90. Sakai RR, Tamashiro KLK. Chapter 1.6 - Social hierarchy and stress. In: Steckler T, Kalin NH, Reul JMHM, editors. Techniques in the Behavioral and Neural Sciences [Internet]. Elsevier; 2005 [cited 2020 Jun 1]. p. 113–32. (Handbook of Stress and the Brain; vol. 15). Available from: http://www.sciencedirect.com/science/article/pii/S0921070905800525 91. Franklin TB, Silva BA, Perova Z, Marrone L, Masferrer ME, Zhan Y, et al. Prefrontal cortical control of a brainstem social behavior circuit. Nat Neurosci. 2017 Feb;20(2):260– 70. 92. Challis C, Beck SG, Berton O. Optogenetic modulation of descending prefrontocortical inputs to the dorsal raphe bidirectionally bias socioaffective choices after social defeat. Front Behav Neurosci [Internet]. 2014 Feb 17 [cited 2020 May 29];8. Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3925846/ 93. Sawicki CM, Kim JK, Weber MD, Faw TD, McKim DB, Madalena KM, et al. Microglia Promote Increased Pain Behavior through Enhanced Inflammation in the Spinal Cord during Repeated Social Defeat Stress. J Neurosci. 2019 Feb 13;39(7):1139–49. 94. Walker P, Carrive P. Role of ventrolateral periaqueductal gray neurons in the behavioral and cardiovascular responses to contextual conditioned fear and poststress recovery. Neuroscience. 2003 Feb 17;116(3):897–912. 95. Pavlov VA, Tracey KJ. The vagus nerve and the inflammatory reflex—linking immunity and metabolism. Nat Rev Endocrinol. 2012 Dec;8(12):743–54. 96. Frangos E, Ellrich J, Komisaruk BR. Non-invasive Access to the Vagus Nerve Central Projections via Electrical Stimulation of the External Ear: fMRI Evidence in Humans. Brain Stimulat. 2015 May 1;8(3):624–36. 97. Berthoud H-R, Neuhuber WL. Functional and chemical anatomy of the afferent vagal system. Auton Neurosci. 2000 Dec 20;85(1):1–17. 98. Ea P, G L, S W, Pm S, Ja H, Jf S, et al. Ventral periaqueductal grey stimulation alters heart rate variability in humans with chronic pain. Exp Neurol. 2010 Feb 21;223(2):574–81. 99. Koenig J, Loerbroks A, Jarczok MN, Fischer JE, Thayer JF. Chronic Pain and Heart Rate Variability in a Cross-Sectional Occupational Sample: Evidence for Impaired Vagal Control. Clin J Pain. 2016 Mar;32(3):218–25. 100. Mayer E, Berman S, Suyenobu B, Labus J, Mandelkern M, Naliboff B, et al. Differences in brain responses to visceral pain between patients with irritable bowel syndrome and ulcerative colitis. Pain. 2005 Jun;115(3):398–409. 116 101. Tu C-H, Niddam DM, Chao H-T, Liu R-S, Hwang R-J, Yeh T-C, et al. Abnormal cerebral metabolism during menstrual pain in primary dysmenorrhea. NeuroImage. 2009 Aug 1;47(1):28–35. 102. Protopopescu X, Pan H, Tuescher O, Cloitre M, Goldstein M, Engelien A, et al. Increased brainstem volume in panic disorder: a voxel-based morphometric study. Neuroreport. 2006 Mar 20;17(4):361–3. 103. Williams DP, Chelimsky G, McCabe NP, Koenig J, Singh P, Janata J, et al. Effects of Chronic Pelvic Pain on Heart Rate Variability in Women. J Urol. 2015 Nov;194(5):1289– 94. 104. Pollatos O, Füstös J, Critchley HD. On the generalised embodiment of pain: How interoceptive sensitivity modulates cutaneous pain perception. PAIN®. 2012 Aug 1;153(8):1680–6. 105. Kleinhans Natalia M., Yang Claire C., Strachan Eric D., Buchwald Dedra S., Maravilla Kenneth R. Alterations in Connectivity on Functional Magnetic Resonance Imaging with Provocation of Lower Urinary Tract Symptoms: A MAPP Research Network Feasibility Study of Urological Chronic Pelvic Pain Syndromes. J Urol. 2016 Mar 1;195(3):639–45. 106. Paine P, Kishor J, Worthen SF, Gregory LJ, Aziz Q. Exploring relationships for visceral and somatic pain with autonomic control and personality. PAIN®. 2009 Aug 1;144(3):236– 44. 107. Mayberg HS, Lozano AM, Voon V, McNeely HE, Seminowicz D, Hamani C, et al. Deep Brain Stimulation for Treatment-Resistant Depression. Neuron. 2005 Mar 3;45(5):651–60. 108. Harper DE, Ichesco E, Schrepf A, Hampson JP, Clauw DJ, Schmidt-Wilcke T, et al. Resting functional connectivity of the periaqueductal gray is associated with normal inhibition and pathological facilitation in conditioned pain modulation. J Pain Off J Am Pain Soc. 2018 Jun;19(6):635.e1-635.e15. 109. Wei S-Y, Chen L-F, Lin M-W, Li W-C, Low I, Yang C-J, et al. The OPRM1 A118G polymorphism modulates the descending pain modulatory system for individual pain experience in young women with primary dysmenorrhea. Sci Rep [Internet]. 2017 Jan 6 [cited 2019 Jul 11];7. Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5216367/ 110. Liu P, Wang G, Liu Y, Zeng F, Lin D, Yang X, et al. Disrupted intrinsic connectivity of the periaqueductal gray in patients with functional dyspepsia: A resting-state fMRI study. Neurogastroenterol Motil. 2017 Aug;29(8):e13060. 111. Linnman C, Beucke J-C, Jensen KB, Gollub RL, Kong J. Sex similarities and differences in pain-related periaqueductal gray connectivity: Pain. 2012 Feb;153(2):444–54. 117 112. Niu X, Bai L, Sun Y, Wang S, Cao J, Sun C, et al. Disruption of periaqueductal grey-default mode network functional connectivity predicts persistent post-traumatic headache in mild traumatic brain injury. J Neurol Neurosurg Psychiatry. 2019 Mar 1;90(3):326–32. 113. Satpute AB, Wager TD, Cohen-Adad J, Bianciardi M, Choi J-K, Buhle JT, et al. Identification of discrete functional subregions of the human periaqueductal gray. Proc Natl Acad Sci. 2013 Oct 15;110(42):17101–6. 114. Faull OK, Jenkinson M, Clare S, Pattinson KTS. Functional subdivision of the human periaqueductal grey in respiratory control using 7tesla fMRI. NeuroImage. 2015 Jun 1;113:356–64. 115. Pinto J, Nunes S, Bianciardi M, Dias A, Silveira LM, Wald LL, et al. Improved 7 Tesla resting-state fMRI connectivity measurements by cluster-based modeling of respiratory volume and heart rate effects. NeuroImage. 2017 Jun;153:262–72. 116. Brooks JCWP, Faull OK, Pattinson KTSDpF, Jenkinson MP. Physiological Noise in Brainstem fMRI. Front Hum Neurosci [Internet]. 2013 [cited 2019 Jul 23];7. Available from: https://www.frontiersin.org/articles/10.3389/fnhum.2013.00623/full 117. Herrero MT, Insausti R, Gonzalo LM. Cortically projecting cells in the periaqueductal gray matter of the rat. A retrograde fluorescent tracer study. Brain Res. 1991 Mar 15;543(2):201– 12. 118. Nickel JC, Shoskes D, Irvine-Bird K. Clinical Phenotyping of Women With Interstitial Cystitis/Painful Bladder Syndrome: A Key to Classification and Potentially Improved Management. J Urol. 2009 Jul;182(1):155–60. 119. Drysdale AT, Grosenick L, Downar J, Dunlop K, Mansouri F, Meng Y, et al. Resting-state connectivity biomarkers define neurophysiological subtypes of depression. Nat Med. 2017 Jan;23(1):28–38. 120. Locke Kenneth, Lai H. Henry, Pontari Michel A., Clemens J. Quentin, Kreder Karl J., Krieger John N., et al. Mp07-20 discovery, validation and novel visualization of subgroups in urologic chronic pelvic pain syndrome (ucpps): consensus clustering findings from the mapp research network. J Urol. 2020 Apr 1;203(Supplement 4):e104–e104. 121. Vollert J, Kramer M, Barroso A, Freynhagen R, Haanpää M, Hansson P, et al. Symptom profiles in the painDETECT Questionnaire in patients with peripheral neuropathic pain stratified according to sensory loss in quantitative sensory testing. Pain. 2016;157(8):1810– 8. 122. Cleeland CS, Ryan KM. Pain assessment: Global use of the Brief Pain Inventory. Ann Acad Med Singap. 1994;23(2):129–38. 123. Guo W, Yi J, Landis JR. K-class functional mixed effects clustering with application to longitudinal urological chronic pelvic pain syndrome symptom data. :1. 118 124. Kutch JJ, Ichesco E, Hampson JP, Labus JS, Farmer MA, Martucci KT, et al. Brain signature and functional impact of centralized pain: a multidisciplinary approach to the study of chronic pelvic pain (MAPP) network study. Pain. 2017 Oct;158(10):1979–91. 125. O’Leary MP, Sant GR, Fowler FJ, Whitmore KE, Spolarich-Kroll J. The interstitial cystitis symptom index and problem index. Urology. 1997 May;49(5A Suppl):58–63. 126. Rodríguez L, Stephens A, Clemens J, Buchwald D, Yang C, Lai H, et al. SYMPTOM DURATION IN PATIENTS WITH UROLOGIC CHRONIC PELVIC PAIN SYNDROME (UCPPS) IS NOT ASSOCIATED WITH PAIN SEVERITY, NON-UROLOGIC SYNDROMES AND MENTAL HEALTH SYMPTOMS: A MULTIDISCIPLINARY APPROACH TO THE STUDY OF CHRONIC PELVIC PAIN (MAPP) NETWORK STUDY. Urology. 2019 Feb;124:14–22. 127. Naliboff BD, Stephens AJ, Lai HH, Griffith JW, Clemens JQ, Lutgendorf S, et al. Clinical and Psychosocial Predictors of Urologic Chronic Pelvic Pain Symptom Change Over One Year: A Prospective Study from the MAPP Research Network. J Urol. 2017 Oct;198(4):848–57. 128. Trans-MAPP Symptom Patterns Study (SPS) - Full Text View - ClinicalTrials.gov [Internet]. [cited 2020 Jun 18]. Available from: https://clinicaltrials.gov/ct2/show/NCT02514265 129. Shpaner M, Kelly C, Lieberman G, Perelman H, Davis M, Keefe FJ, et al. Unlearning chronic pain: A randomized controlled trial to investigate changes in intrinsic brain connectivity following Cognitive Behavioral Therapy. NeuroImage Clin. 2014 Jan 1;5:365– 76. 130. Clemens JQ, Stephens-Shields AJ, Newcomb C, Rodriguez LV, Lai HH, Bradley CS, et al. Correlates of 1-Year Change in Quality of Life in Patients with Urologic Chronic Pelvic Pain Syndrome: Findings from the Multidisciplinary Approach to the Study of Chronic Pelvic Pain (MAPP) Research Network. J Urol. 0(0):10.1097/JU.0000000000001080. 131. Tillman RM, Stockbridge MD, Nacewicz BM, Torrisi S, Fox AS, Smith JF, et al. Intrinsic functional connectivity of the central extended amygdala. Hum Brain Mapp. 2018;39(3):1291–312. 132. Cohen AL, Fair DA, Dosenbach NUF, Miezin FM, Dierker D, Van Essen DC, et al. Defining functional areas in individual human brains using resting functional connectivity MRI. NeuroImage. 2008 May 15;41(1):45–57. 133. Eickhoff SB, Yeo BTT, Genon S. Imaging-based parcellations of the human brain. Nat Rev Neurosci. 2018 Nov;19(11):672. 134. Ezra M, Faull OK, Jbabdi S, Pattinson KT. Connectivity-based segmentation of the periaqueductal gray matter in human with brainstem optimized diffusion MRI. Hum Brain Mapp. 2015;36(9):3459–71. 119 135. Genon S, Reid A, Langner R, Amunts K, Eickhoff SB. How to Characterize the Function of a Brain Region. Trends Cogn Sci. 2018 Apr 1;22(4):350–64. 136. Bittar RG, Kar-Purkayastha I, Owen SL, Bear RE, Green A, Wang S, et al. Deep brain stimulation for pain relief: A meta-analysis. J Clin Neurosci. 2005 Jun 1;12(5):515–9. 137. Poldrack RA, Mumford JA, Nichols TE. Visualizing, localizing, and reporting fMRI data [Internet]. Handbook of Functional MRI Data Analysis. 2011 [cited 2019 Jul 30]. Available from: /core/books/handbook-of-functional-mri-data-analysis/visualizing-localizing-and- reporting-fmri-data/AAEE56293C17DA35104E17A65E4F4C1B 138. Costafreda SG. Pooling fMRI Data: Meta-Analysis, Mega-Analysis and Multi-Center Studies. Front Neuroinformatics [Internet]. 2009 Sep 30 [cited 2019 Jul 23];3. Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2759345/ 139. Chen Z, Chen X, Liu M, Liu S, Ma L, Yu S. Disrupted functional connectivity of periaqueductal gray subregions in episodic migraine. J Headache Pain [Internet]. 2017 Mar 21 [cited 2019 Jul 11];18(1). Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5359195/ 140. Li Z, Liu M, Lan L, Zeng F, Makris N, Liang Y, et al. Altered periaqueductal gray resting state functional connectivity in migraine and the modulation effect of treatment. Sci Rep [Internet]. 2016 Apr [cited 2019 Aug 20];6(1). Available from: http://www.nature.com/articles/srep20298 141. Kong J, Loggia ML, Zyloney C, Tu P, LaViolette P, Gollub RL. Exploring the brain in pain: activations, deactivations and their relation. Pain. 2010 Feb;148(2):257. 142. Lambert C, Chowdhury R, FitzGerald THB, Fleming SM, Lutti A, Hutton C, et al. Characterizing Aging in the Human Brainstem Using Quantitative Multimodal MRI Analysis. Front Hum Neurosci [Internet]. 2013 Aug 20 [cited 2020 May 20];7. Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3747448/ 143. Glover GH, Mueller BA, Turner JA, van Erp TGM, Liu TT, Greve DN, et al. Function biomedical informatics research network recommendations for prospective multicenter functional MRI studies. J Magn Reson Imaging. 2012 Jul 1;36(1):39–54. 144. Cox RW. AFNI: Software for Analysis and Visualization of Functional Magnetic Resonance Neuroimages. Comput Biomed Res. 1996 Jun;29(3):162–73. 145. Jo HJ, Gotts SJ, Reynolds RC, Bandettini PA, Martin A, Cox RW, et al. Effective Preprocessing Procedures Virtually Eliminate Distance-Dependent Motion Artifacts in Resting State FMRI. J Appl Math. 2013 Jun 10;2013:e935154. 146. Plitt M, Barnes KA, Wallace GL, Kenworthy L, Martin A. Resting-state functional connectivity predicts longitudinal change in autistic traits and adaptive functioning in autism. Proc Natl Acad Sci. 2015 Dec 1;112(48):E6699–706. 120 147. Dice LR. Measures of the Amount of Ecologic Association Between Species. Ecology. 1945;26(3):297–302. 148. Power JD, Cohen AL, Nelson SM, Wig GS, Barnes KA, Church JA, et al. Functional network organization of the human brain. Neuron. 2011 Nov 17;72(4):665–78. 149. Segerdahl AR, Themistocleous AC, Fido D, Bennett DL, Tracey I. A brain-based pain facilitation mechanism contributes to painful diabetic polyneuropathy. Brain. 2018 Feb 1;141(2):357–64. 150. Hamani C, Mayberg H, Stone S, Laxton A, Haber S, Lozano AM. The Subcallosal Cingulate Gyrus in the Context of Major Depression. Biol Psychiatry. 2011 Feb 15;69(4):301–8. 151. Harricharan S, Rabellino D, Frewen PA, Densmore M, Théberge J, McKinnon MC, et al. fMRI functional connectivity of the periaqueductal gray in PTSD and its dissociative subtype. Brain Behav. 2016;6(12):e00579. 152. Eklund A, Nichols TE, Knutsson H. Cluster failure: Why fMRI inferences for spatial extent have inflated false-positive rates. Proc Natl Acad Sci. 2016 Jul 12;113(28):7900–5. 153. Eickhoff SB, Stephan KE, Mohlberg H, Grefkes C, Fink GR, Amunts K, et al. A new SPM toolbox for combining probabilistic cytoarchitectonic maps and functional imaging data. NeuroImage. 2005 May 1;25(4):1325–35. 154. Salimi-Khorshidi G, Douaud G, Beckmann CF, Glasser MF, Griffanti L, Smith SM. Automatic denoising of functional MRI data: Combining independent component analysis and hierarchical fusion of classifiers. NeuroImage. 2014 Apr 15;90:449–68. 155. Griffanti L, Salimi-Khorshidi G, Beckmann CF, Auerbach EJ, Douaud G, Sexton CE, et al. ICA-based artefact removal and accelerated fMRI acquisition for improved resting state network imaging. NeuroImage. 2014 Jul 15;95:232–47. 156. Bijsterbosch J, Smith S, Beckmann C. An Introduction to Resting State fMRI Functional Connectivity. Oxford, New York: Oxford University Press; 2017. 160 p. (Oxford Neuroimaging Primers). 157. Boyacioğlu R, Schulz J, Koopmans PJ, Barth M, Norris DG. Improved sensitivity and specificity for resting state and task fMRI with multiband multi-echo EPI compared to multi-echo EPI at 7 T. NeuroImage. 2015 Oct 1;119:352–61. 158. Marchitelli R, Minati L, Marizzoni M, Bosch B, Bartrés‐Faz D, Müller BW, et al. Test- retest reliability of the default mode network in a multi-centric fMRI study of healthy elderly: Effects of data-driven physiological noise correction techniques. Hum Brain Mapp. 2016;37(6):2114–32. 159. Woo C-W, Krishnan A, Wager TD. Cluster-extent based thresholding in fMRI analyses: Pitfalls and recommendations. NeuroImage. 2014 May 1;91:412–9. 121 160. Power JD, Barnes KA, Snyder AZ, Schlaggar BL, Petersen SE. Spurious but systematic correlations in functional connectivity MRI networks arise from subject motion. NeuroImage. 2012 Feb 1;59(3):2142–54. 161. Van Dijk KRA, Sabuncu MR, Buckner RL. The influence of head motion on intrinsic functional connectivity MRI. NeuroImage. 2012 Jan 2;59(1):431–8. 162. Satterthwaite TD, Wolf DH, Loughead J, Ruparel K, Elliott MA, Hakonarson H, et al. Impact of in-scanner head motion on multiple measures of functional connectivity: relevance for studies of neurodevelopment in youth. NeuroImage. 2012 Mar;60(1):623–32. 163. Napadow V, LaCount L, Park K, As-Sanie S, Clauw DJ, Harris RE. Intrinsic brain connectivity in fibromyalgia is associated with chronic pain intensity. Arthritis Rheum. 2010 Aug;62(8):2545–55. 164. Apkarian VA, Sosa Y, Krauss BR, Thomas SP, Fredrickson BE, Levy RE, et al. Chronic pain patients are impaired on an emotional decision-making task: Pain. 2004 Mar;108(1):129–36. 165. Mayberg HS, Brannan SK, Tekell JL, Silva JA, Mahurin RK, McGinnis S, et al. Regional metabolic effects of fluoxetine in major depression: serial changes and relationship to clinical response. Biol Psychiatry. 2000 Oct 15;48(8):830–43. 166. Hornung J-P. The human raphe nuclei and the serotonergic system. J Chem Neuroanat. 2003 Dec 1;26(4):331–43. 167. Michelsen KA, Prickaerts J, Steinbusch HWM. The dorsal raphe nucleus and serotonin: implications for neuroplasticity linked to major depression and Alzheimer’s disease. In: Di Giovann G, Di Matteo V, Esposito E, editors. Progress in Brain Research [Internet]. Elsevier; 2008 [cited 2019 Jul 24]. p. 233–64. (Serotonin–Dopamine Interaction: Experimental Evidence and Therapeutic Relevance; vol. 172). Available from: http://www.sciencedirect.com/science/article/pii/S0079612308009126 168. Naidich TP, Duvernoy HM, Delman BN, Sorensen AG, Kollias SS, Haacke EM, editors. Internal Architecture of the Brain Stem with Key Axial Section. In: Duvernoy’s Atlas of the Human Brain Stem and Cerebellum: High-Field MRI: Surface Anatomy, Internal Structure, Vascularization and 3D Sectional Anatomy [Internet]. Vienna: Springer Vienna; 2009 [cited 2019 Jul 30]. p. 53–93. Available from: https://doi.org/10.1007/978-3-211-73971-6_4 169. Babalian A, Eichenberger S, Bilella A, Girard F, Szabolcsi V, Roccaro D, et al. The orbitofrontal cortex projects to the parvafox nucleus of the ventrolateral hypothalamus and to its targets in the ventromedial periaqueductal grey matter. Brain Struct Funct. 2019;224(1):293–314. 170. Oliveras JL, Besson JM, Guilbaud G, Liebeskind JC. Behavioral and electrophysiological evidence of pain inhibition from midbrain stimulation in the cat. Exp Brain Res. 1974 Apr 1;20(1):32–44. 122 171. Napadow V, Dhond RP, Purdon P, Kettner N, Makris N, Kwong KK, et al. Correlating Acupuncture fMRI in the Human Brainstem with Heart Rate Variability. In: 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference. 2005. p. 4496–9. 172. Pedroni A, Koeneke S, Velickaite A, Jäncke L. Differential magnitude coding of gains and omitted rewards in the ventral striatum. Brain Res. 2011 Sep 9;1411:76–86. 173. Beliveau V, Svarer C, Frokjaer VG, Knudsen GM, Greve DN, Fisher PM. Functional connectivity of the dorsal and median raphe nuclei at rest. NeuroImage. 2015 Aug 1;116:187–95. 174. Nieto-Castañón A, Fedorenko E. Subject-specific functional localizers increase sensitivity and functional resolution of multi-subject analyses. NeuroImage. 2012 Nov 15;63(3):1646– 69. 175. Sclocco R, Beissner F, Bianciardi M, Polimeni JR, Napadow V. Challenges and opportunities for brainstem neuroimaging with ultrahigh field MRI. NeuroImage. 2018 Mar;168:412–26. 176. Price CJ, Friston KJ. Degeneracy and cognitive anatomy. Trends Cogn Sci. 2002 Oct 1;6(10):416–21. 177. Brooks JCW, Davies W-E, Pickering AE. Resolving the Brainstem Contributions to Attentional Analgesia. J Neurosci. 2017 Mar 1;37(9):2279–91. 178. Diedrichsen J, Maderwald S, Küper M, Thürling M, Rabe K, Gizewski ER, et al. Imaging the deep cerebellar nuclei: A probabilistic atlas and normalization procedure. NeuroImage. 2011 Feb 1;54(3):1786–94. 179. Yarkoni T, Poldrack RA, Nichols TE, Van Essen DC, Wager TD. Large-scale automated synthesis of human functional neuroimaging data. Nat Methods. 2011 Jun 26;8(8):665–70. 180. Mikula S, Trotts I, Stone JM, Jones EG. Internet-Enabled High-Resolution Brain Mapping and Virtual Microscopy. NeuroImage. 2007 Mar;35(1):9–15. 181. Logothetis NK. What we can do and what we cannot do with fMRI. Nature. 2008 Jun 12;453(7197):869–78. 182. Coulombe M-A, Lawrence KSt, Moulin DE, Morley-Forster P, Shokouhi M, Nielson WR, et al. Lower Functional Connectivity of the Periaqueductal Gray Is Related to Negative Affect and Clinical Manifestations of Fibromyalgia. Front Neuroanat [Internet]. 2017 Jun 8 [cited 2019 Aug 20];11. Available from: http://journal.frontiersin.org/article/10.3389/fnana.2017.00047/full 183. Quintana J, Fuster JM. From Perception to Action: Temporal Integrative Functions of Prefrontal and Parietal Neurons. Cereb Cortex. 1999 Apr 1;9(3):213–21. 123 184. Schrepf A, Kaplan CM, Ichesco E, Larkin T, Harte SE, Harris RE, et al. A multi-modal MRI study of the central response to inflammation in rheumatoid arthritis. Nat Commun. 2018 Jun 8;9(1):1–11. 185. Kucyi A, Salomons TV, Davis KD. Mind wandering away from pain dynamically engages antinociceptive and default mode brain networks. Proc Natl Acad Sci. 2013 Nov 12;110(46):18692–7. 186. Pavlov VA, Tracey KJ. Controlling inflammation: the cholinergic anti-inflammatory pathway. Biochem Soc Trans. 2006 Dec 1;34(6):1037–40. 187. Bandler R, Carrive P, Depaulis A. Emerging Principles of Organization of the Midbrain Periaqueductal Gray Matter. In: Depaulis A, Bandler R, editors. The Midbrain Periaqueductal Gray Matter: Functional, Anatomical, and Neurochemical Organization [Internet]. Boston, MA: Springer US; 1991 [cited 2019 Jul 30]. p. 1–8. (NATO ASI Series). Available from: https://doi.org/10.1007/978-1-4615-3302-3_1 188. Sclocco Roberta, Beissner Florian, Desbordes Gaelle, Polimeni Jonathan R., Wald Lawrence L., Kettner Norman W., et al. Neuroimaging brainstem circuitry supporting cardiovagal response to pain: a combined heart rate variability/ultrahigh-field (7 T) functional magnetic resonance imaging study. Philos Trans R Soc Math Phys Eng Sci. 2016 May 13;374(2067):20150189. 189. Chang C, Metzger CD, Glover GH, Duyn JH, Heinze H-J, Walter M. Association between heart rate variability and fluctuations in resting-state functional connectivity. NeuroImage. 2013 Mar;68:93–104. 190. Griffanti L, Douaud G, Bijsterbosch J, Evangelisti S, Alfaro-Almagro F, Glasser MF, et al. Hand classification of fMRI ICA noise components. Neuroimage. 2017 Jul 1;154:188–205. 191. Beissner F. Functional MRI of the Brainstem: Common Problems and their Solutions. Clin Neuroradiol. 2015 Oct 1;25(2):251–7. 192. Liu TT, Nalci A, Falahpour M. The Global Signal in fMRI: Nuisance or Information? NeuroImage. 2017 Apr 15;150:213–29. 193. Blok BF, Willemsen AT, Holstege G. A PET study on brain control of micturition in humans. Brain. 1997 Jan 1;120(1):111–21. 194. Chelimsky GG, Yang S, Sanses T, Tatsuoka C, Buffington CAT, Janata J, et al. Autonomic neurophysiologic implications of disorders comorbid with bladder pain syndrome vs myofascial pelvic pain. Neurourol Urodyn. 2019 Jun 1;38(5):1370–7. 195. Martucci KT, Shirer WR, Bagarinao E, Johnson KA, Farmer MA, Labus JS, et al. The Posterior Medial Cortex in Urologic Chronic Pelvic Pain Syndrome: Detachment from Default Mode Network. A Resting-State Study from the MAPP Research Network. Pain. 2015 Sep;156(9):1755–64. 124 196. Bishop CM. Pattern Recognition and Machine Learning. Springer: New York; 2006. 197. Dubois J, Adolphs R. Building a Science of Individual Differences from fMRI. Trends Cogn Sci. 2016 Jun 1;20(6):425–43. 198. Arbabshirani MR, Plis S, Sui J, Calhoun VD. Single subject prediction of brain disorders in neuroimaging: Promises and pitfalls. NeuroImage [Internet]. [cited 2016 Jul 30]; Available from: http://www.sciencedirect.com/science/article/pii/S105381191600210X 199. Thompson JM, Neugebauer V. Cortico-limbic pain mechanisms. Neurosci Lett. 2019 May 29;702:15–23. 200. Shipp S. The functional logic of corticostriatal connections. Brain Struct Funct. 2017;222(2):669–706. 201. Brooks JCW, Nurmikko TJ, Bimson WE, Singh KD, Roberts N. fMRI of Thermal Pain: Effects of Stimulus Laterality and Attention. NeuroImage. 2002 Feb 1;15(2):293–301. 202. Touroutoglou A, Hollenbeck M, Dickerson BC, Feldman Barrett L. Dissociable large-scale networks anchored in the right anterior insula subserve affective experience and attention. NeuroImage. 2012 May 1;60(4):1947–58. 203. Legrain V, Damme SV, Eccleston C, Davis KD, Seminowicz DA, Crombez G. A neurocognitive model of attention to pain: Behavioral and neuroimaging evidence. PAIN. 2009 Aug;144(3):230–232. 204. Seminowicz DA, Davis KD. Cortical responses to pain in healthy individuals depends on pain catastrophizing. Pain. 2006 Feb;120(3):297–306. 205. Huang J, Gadotti VM, Chen L, Souza IA, Huang S, Wang D, et al. A neuronal circuit for activating descending modulation of neuropathic pain. Nat Neurosci. 2019 Oct;22(10):1659–68. 206. Chen G, Chen G, Xie C, Li S-J. Negative Functional Connectivity and Its Dependence on the Shortest Path Length of Positive Network in the Resting-State Human Brain. Brain Connect. 2011 Sep;1(3):195–206. 207. Evrard HC, Forro T, Logothetis NK. Von Economo neurons in the anterior insula of the macaque monkey. Neuron. 2012 May 10;74(3):482–9. 208. Baron RM, Kenny DA. The moderator–mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. J Pers Soc Psychol. 1986;51(6):1173–82. 209. Wiech K, Jbabdi S, Lin CS, Andersson J, Tracey I. Differential structural and resting state connectivity between insular subdivisions and other pain-related brain regions. Pain. 2014 Oct;155(10):2047–55. 125 210. Kragel PA, Bianciardi M, Hartley L, Matthewson G, Choi J-K, Quigley KS, et al. Functional Involvement of Human Periaqueductal Gray and Other Midbrain Nuclei in Cognitive Control. J Neurosci. 2019 Jul 31;39(31):6180–9. 211. Martucci KT, Borg N, MacNiven KH, Knutson B, Mackey SC. Altered prefrontal correlates of monetary anticipation and outcome in chronic pain. Pain. 2018 Aug;159(8):1494–507. 212. Hemington KS, Coulombe M-A. The periaqueductal gray and descending pain modulation: why should we study them and what role do they play in chronic pain? J Neurophysiol. 2015 Feb 11;114(4):2080–3. 213. Song S, Zhan Z, Long Z, Zhang J, Yao L. Comparative Study of SVM Methods Combined with Voxel Selection for Object Category Classification on fMRI Data. PLOS ONE. 2011 Feb 16;6(2):e17191. 214. Orrù G, Pettersson-Yeo W, Marquand AF, Sartori G, Mechelli A. Using Support Vector Machine to identify imaging biomarkers of neurological and psychiatric disease: A critical review. Neurosci Biobehav Rev. 2012 Apr 1;36(4):1140–52. 215. Davis KD, Aghaeepour N, Ahn AH, Angst MS, Borsook D, Brenton A, et al. Discovery and validation of biomarkers to aid the development of safe and effective pain therapeutics: challenges and opportunities. Nat Rev Neurol. 2020 Jun 15;1–20. 216. Fellous J-M, Sapiro G, Rossi A, Mayberg H, Ferrante M. Explainable Artificial Intelligence for Neuroscience: Behavioral Neurostimulation. Front Neurosci [Internet]. 2019 [cited 2020 Jan 13];13. Available from: https://www.frontiersin.org/articles/10.3389/fnins.2019.01346/full 217. Kassraian-Fard P, Matthis C, Balsters JH, Maathuis MH, Wenderoth N. Promises, Pitfalls, and Basic Guidelines for Applying Machine Learning Classifiers to Psychiatric Imaging data, with Autism as an example. Front Psychiatry. 2016 Dec 1;7:177. 218. Botvinik-Nezer R, Holzmeister F, Camerer CF, Dreber A, Huber J, Johannesson M, et al. Variability in the analysis of a single neuroimaging dataset by many teams. Nature. 2020 Jun;582(7810):84–8. 219. Villemure C, Bushnell MC. Mood influences supraspinal pain processing separately from attention. J Neurosci Off J Soc Neurosci. 2009 Jan 21;29(3):705–15. 220. Wager TD, Atlas LY, Lindquist MA, Roy M, Woo C-W, Kross E. An fMRI-Based Neurologic Signature of Physical Pain. N Engl J Med. 2013 Apr 11;368(15):1388–97.
Abstract (if available)
Abstract
The work of this thesis was to establish whether the prediction of long-term symptom change in urologic chronic pelvic pain syndrome (UCPPS) is achievable using baseline resting state functional magnetic resonance imaging (rs-fMRI) measures. The development of chronic pain is not well understood, and pain symptoms fluctuate over months to years. Studies from the Multidisciplinary Approach to the Study of Chronic Pelvic Pain (MAPP) research network have previously shown changes in brain function and structure between UCPPS patients and healthy controls. Previous studies in our laboratory have been able to predict UCPPS symptom change up to 3-months with baseline rs-fMRI along with longitudinal patient data from MAPP. In order to make the results more clinically relevant there is a need to validate and expand these results for longer time periods, as well as develop more definitive biomarkers representative of symptoms progression. We identified the periaqueductal gray (PAG), well-established as a component in pain modulation, as our main region of interest. We hypothesized that the PAG may be associated with reduced ability to self-regulate altered physiological responses in UCPPS and influence symptom change over time. ❧ To predict change in symptoms up to a year requires a better understanding of how symptom pain trajectories change over time. We defined our outcome variable as the slope, or the difference, in reported symptoms from baseline up to 12-months. Using data from 389 participants in MAPP’s longitudinal collection, we found that a portion of the participants improve significantly over the 12-month timeframe. This trend, and its differentiation from patients who do not improve, is what we predicted using baseline connectivity from the PAG to the rest of the brain. To do this accurately we needed to verify the location of the PAG in the brainstem with resting state connectivity measures, as previous studies have used coordinates representative of the PAG more anterior than the true anatomical location. We found that differing PAG localization measures impacted whole-brain connectivity in healthy participants and the detection of differences in UCPPS patients compared to healthy participants. To make a prediction on the change in symptoms over 12-months based on resting state connectivity measures from the PAG, we used a hand-traced MNI location. This region of interest was shown to be a close approximation to the “ground-truth” or participant specific anatomical location of the PAG and was used as the location to seed the PAG in our study. Our final analysis sought to predict the change in symptoms over 12-months using whole brain connectivity. We optimized brain parcellation and feature selection and used a support vector regression to develop our model. We show that our method makes a significant prediction and our model is able to generalize to independent datasets. Furthermore, we identify biomarkers, such as the PAG’s connectivity to the insula, that were found to be influential in driving the predictive model. These results further our understanding of the mechanisms behind pain chronicity and identify potential prognostic biomarkers that could be targeted in future therapeutic interventions to alleviate the symptoms of chronic pain.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Identifying neural markers of centralized pain
PDF
Decoding the neurological and genetic underpinnings of chronic pain
PDF
Cerebrovascular mechanisms of cognitive enhancement after periodized resistance training in older adults
PDF
The footprint of pain: investigating persistence of altered trunk control in recurrent low back pain
PDF
Chronic eye disease epidemiology in the multiethnic ophthalmology cohorts of California study
PDF
A series of longitudinal analyses of patient reported outcomes to further the understanding of care-management of comorbid diabetes and depression in a safety-net healthcare system
PDF
Exploring the effects of mindfulness on psychosocial factors for patients receiving hand therapy
Asset Metadata
Creator
Fenske, Sonja Joan
(author)
Core Title
Brain-based prediction of chronic pain progression: a longitudinal study of urologic chronic pelvic pain syndrome using baseline resting state connectivity from the periaqueductal gray
School
College of Letters, Arts and Sciences
Degree
Doctor of Philosophy
Degree Program
Neuroscience
Publication Date
09/11/2020
Defense Date
07/22/2020
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
chronic pelvic pain,functional magnetic resonance imaging,OAI-PMH Harvest,periaqueductal gray,resting state
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Liew, Sook-Lei (
committee chair
), Kutch, Jason (
committee member
), Leahy, Richard (
committee member
), Pa, Judy (
committee member
)
Creator Email
sfenske@usc.edu,sjfxctrack@gmail.com
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c89-369019
Unique identifier
UC11666186
Identifier
etd-FenskeSonj-8952.pdf (filename),usctheses-c89-369019 (legacy record id)
Legacy Identifier
etd-FenskeSonj-8952.pdf
Dmrecord
369019
Document Type
Dissertation
Rights
Fenske, Sonja Joan
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
Tags
chronic pelvic pain
functional magnetic resonance imaging
periaqueductal gray
resting state