Opinion | Cognition and Behavior Pushing the boundaries of psychiatric neuroimaging to ground diagnosis in biology https://doi.org/10.1523/ENEURO.0384-19.2019

Cite as: eNeuro 2019; 10.1523/ENEURO.0384-19.2019 Received: 20 September 2019 Revised: 22 October 2019 Accepted: 23 October 2019

This Early Release article has been peer-reviewed and accepted, but has not been through the composition and copyediting processes. The final version may differ slightly in style or formatting and will contain links to any extended data.

Alerts: Sign up at www.eneuro.org/alerts to receive customized email alerts when the fully formatted version of this article is published.

Copyright © 2019 Saggar and Uddin This is an open-access article distributed under the terms of the Creative Commons Attribution 4.0 International license, which permits unrestricted use, distribution and reproduction in any medium provided that the original work is properly attributed.

Manuscript Title Pushing the boundaries of psychiatric neuroimaging to ground diagnosis in biology

Abbreviated Title Pushing the boundaries of psychiatric neuroimaging

List all Author Names and Affiliations in order as they would appear in the published article Manish Saggar1, Ph.D. and Lucina Q. Uddin2, Ph.D. 1Department of Psychiatry & Behavioral Sciences, Stanford University 2Department of Psychology, University of Miami

Author Contributions: Both authors contributed to literature survey, discussion and writing of the manuscript

Correspondence should be addressed to Manish Saggar 401 Quarry Rd, Stanford CA 94305 Email: [email protected]

Number of figures: 2 Number of tables: 0 Number of words for Abstract: 47 Number of words for Significance Statement: 79 Number of words for Introduction: 590 Number of words for Discussion: 3,134

Acknowledgements: We thank the anonymous reviewers for their valuable feedback.

Conflict of Interest: No, Authors report no conflict of interest

Funding Sources: This work was supported in part by a Career Development Award (K99/R00 MH104605) and a New Innovator Award (DP2 MH119735) from the NIMH to M.S. and by an R01 Award (R01 MH107549) from the NIMH to L.U.

1

1 Abstract 2 To accurately detect, track progression of, and develop novel treatments for mental illnesses, a 3 diagnostic framework is needed that is grounded in biological features. Here we present the case 4 for utilizing personalized neuroimaging, computational modeling, standardized computing, and 5 ecologically valid neuroimaging to anchor psychiatric nosology in biology. 6 7 Significance Statement 8 There is a growing recognition that the boundaries of human neuroimaging data acquisition and 9 analysis must be pushed to ground psychiatric diagnosis in biology. For successful clinical 10 translations, we outline several proposals across the four identified domains of human 11 neuroimaging, namely, (a) reliability of findings; (b) effective clinical translation at the 12 individual subject level; (c) capturing mechanistic insights; and (d) enhancing ecological validity 13 of lab findings. Advances across these domains will be necessary for further progress in 14 psychiatric neuroimaging. 15 16 1. Introduction 17 Mental illness affects a large proportion of the global population and has a significant economic 18 impact due to treatment costs and lost productivity. In the United States alone, nearly one in five 19 adults lives with a mental illness (44.7 million in 2016; retrieved from 20 https://www.samhsa.gov/data/). Given this widespread prevalence and societal cost, there is a 21 crucial need for finding ways to prevent and treat mental illness. Despite the challenges of 22 understanding the complexity of the human brain in health and disease, researchers have made 23 large strides in developing tools and methodologies that have allowed us to get a sneak peek into 24 brain functioning over the last two decades. It is, however, shocking that even after such an 25 accelerated pace of discovery in neuroscience and bioengineering, the pace of development for 26 treatment of mental illness has not only been slow but has almost stagnated. This slow progress 27 in the development of psychiatric treatments could be largely attributed to the lack of an accurate 28 and neurobiologically-grounded diagnostic nosology (Redish and Gordon 2016). 29 30 The diagnostic nosology, or disorder classification system, most commonly used in psychiatry 31 (the Diagnostic and Statistical Manual of Mental Disorders; DSM-V) is built entirely upon 32 assessment of symptoms by clinicians. The Research Domain Criteria (RDoC) initiative put forth 33 by the National Institute of Mental Health (Cuthbert and Insel 2013) aims to address the 34 challenge of revising this diagnostic nosology by creating a framework integrating multiple 35 levels information from genomics to neural circuits and behavior to explore basic dimensions of 36 function across clinical and non-clinical populations. Grounding of a psychiatric diagnosis in 37 biological features can not only potentially provide reliable and valid diagnosis, but can also 38 reveal specific biomarkers to track the course of illness and test the efficacy of new treatments. 39 To borrow an example from (Redish and Gordon 2016), type II diabetes has a complex etiology 40 and pathophysiology; however, unlike psychiatric illnesses, the diagnosis for type II diabetes is 41 largely based on the biological feature of the amount of hemoglobin A1c in the blood. This

2

42 biological feature not only helps in precise diagnosis, but also in proper management of blood 43 glucose as well as in testing the efficacy of treatments over time. 44 45 Different disciplines of science are actively engaged to illuminate biological features that 46 contribute to major psychiatric illnesses. Specifically, several studies from genetics and 47 neuroimaging are at the frontier. Large-scale genomic investigations have linked molecular 48 genetic variations to major psychiatric illnesses. These studies not only present evidence of 49 heterogeneity and polygenicity of psychiatric disorders, but also reveal that connecting multiple 50 levels of molecular, cellular, and circuit functions to complex human behavior is immensely 51 challenging (Geschwind and Flint 2015). Further, we are just beginning to understand how 52 observed genetic variants give rise to changes in brain function and behavior (Gandal et al. 53 2018). Along the same lines, large-scale neuroimaging investigations of both brain structure and 54 function have seemingly failed to pinpoint differences across phenotypically distinct psychiatric 55 diagnosis (Etkin 2019). Interestingly, instead of finding disorder-specific differences, many 56 large-scale studies have either shown converging evidence towards common circuit dysfunction 57 in brain structure (Goodkind et al. 2015) and/or provided evidence for non-specificity in brain 58 function (Baker et al. 2019; (Uddin 2015)). 59 60 In this review article, we first outline some of the pan-disciplinary issues that broadly hamper 61 progress in the search for disorder-specific biological features in psychiatry. We then focus 62 specifically on issues unique to neuroimaging research, and lastly present novel avenues that are 63 capitalizing on recent advances in machine learning and biophysical network modeling to push 64 the boundaries of personalized neuroimaging. 65 66 2. Pan-disciplinary issues in anchoring psychiatric nosology in biology 67 One of the major challenges in finding disorder-specific biomarkers is the fact that clinical 68 symptoms, on which diagnoses are based, may not have a one-to-one mapping to the underlying 69 biological mechanisms. In other words, different biological mechanisms may have led to the 70 same cluster of clinical symptoms. This lack of one-to-one relation between clinical symptoms 71 and biology goes against the traditionally held assumption that symptoms-based stratification 72 could help us discover disorder-specific biomarkers, which in turn would largely explain the 73 observed heterogeneity and comorbidity repeatedly observed in psychiatric disorders (Bijl, 74 Ravelli, and van Zessen 1998; Uddin et al. 2017). Thus, several researchers now argue that 75 instead of performing small scale case-controlled studies, where stratification of the sample is 76 based on symptoms, large-scale transdiagnostic/dimensional studies are needed (Etkin 2019). 77 78 The second major and related challenge for biomarker studies is how to validate the observed 79 biotypes. The lack of one-to-one mapping between clinical symptoms and biology postulates that 80 clinical symptoms alone cannot be used to validate the observed biotypes. Thus, newer

3

81 approaches for validation should be explored, including treatment outcomes and performance on 82 tasks assessing different dimensions of functioning (e.g., RDoC). 83 84 The third challenge pertains to lack of group-to-individual translation of findings. Although most 85 studies are conducted at the group level, for the best translational outcomes, such group-level 86 findings need to be reliable even at the single-patient level. However, it has been previously 87 argued that due to the non-ergodicity in human social and psychological processes, the 88 inferences based on group-level data are challenging to generalize to an individual experience or 89 behavior (Fisher, Medaglia, and Jeronimus 2018). Thus, novel methods that can provide 90 statistically significant as well as behaviorally relevant information at both group and single- 91 participant levels are needed (e.g., Saggar et al. 2018). 92 93 3. Issues specific to neuroimaging 94 To characterize neural substrates of psychiatric disorders, modern neuroimaging tools and 95 sophisticated data analytics have been developed. These neuroimaging methods have already 96 been used to measure a range of neural differences across psychiatric disorders, including, (a) 97 volumetric and morphological differences (Goodkind et al. 2015); (b) structural connectomics 98 (van den Heuvel, Bullmore, and Sporns 2016); (c) static and dynamic functional connectivity at 99 rest and during task performance (Calhoun et al. 2014); and (d) activation differences across a 100 variety of psychological paradigms (Picó-Pérez et al. 2019) (McTeague et al. 2017). 101 102 Yet, neuroimaging as a diagnostic tool still struggles with several specific issues. Here, we 103 broadly classified these issues into four domains: (a) reliability of findings; (b) lack of group-to- 104 individual translation; (c) lack of mechanistic insights; and (d) absence of ecological validity in 105 lab environments and experiments. 106 107 The psychological sciences have been reportedly going through a major replication crisis (Open 108 Science Collaboration 2015; De Boeck and Jeon 2018), and replicability of previously reported 109 brain-behavior relations in neuroimaging studies are also under heavy scrutiny (Button et al. 110 2013; Kharabian Masouleh et al. 2019). A variety of explanations have been put forth to account 111 for this lack of replication in neuroimaging findings. First, and perhaps the most straightforward, 112 is the need for large-scale samples (n ~ several hundred of participants) to find reliable and 113 replicable findings (Kharabian Masouleh et al. 2019). Fortunately, there is now an ever- 114 increasing number of consortia that provide such big datasets, including the Human Connectome 115 Project (HCP; (Van Essen et al. 2013)), UK-Biobank (Smith et al. 2019), and the Adolescent 116 Brain Cognitive Development (ABCD) study (Bjork et al. 2017). Other explanations for the lack 117 of replicability in neuroimaging studies include high corruptibility of data due to head 118 movement and other related artifacts (Power et al. 2012); lack of validation data to show 119 robustness of effects (Davis and Poldrack 2013); lack of reporting of null findings (Kharabian 120 Masouleh et al. 2019); lack of sharing of raw data or unthresholded statistical parametric maps

4

121 across labs (Poldrack et al. 2008); and lack of standardized neuroimaging data pre-and post- 122 processing (Eklund, Nichols, and Knutsson 2016). 123 124 Although large-scale studies and data sharing are putatively successful in addressing the 125 replicability issue, see a sample list of available datasets at 126 https://sites.google.com/site/publicdatadatabase/, several cautions regarding such studies are 127 worth noting. First, due to larger sample sizes, more focus should be devoted to the clinical 128 significance or effect size as opposed to the statistical significance of a given effect (Etkin 2019). 129 Second, a large proportion of currently available large-scale data banks are skewed towards 130 “healthy” participants, and not necessarily treatment-seeking participants or patients. Third, most 131 of the available datasets are cross-sectional in nature, as opposed to longitudinal; the assumption 132 being that the effects observed in large-scale cross-sectional studies could putatively translate to 133 individual participants and hence could have clinical significance. However, we also know that 134 the statistical conditions (i.e., ergodicity) needed for such translation to be legitimate are very 135 unlikely to hold in the case of human social and psychological processes (Hamaker 2012). This 136 raises a major concern for translating insights from studies conducted at the group level to 137 ultimately helping an individual patient in the clinic. 138 139 Another, and perhaps deeper issue with neuroimaging-based brain-behavior associations in 140 psychiatric disorders is that even when we find replicable relationships, we are still far from a 141 causal and mechanistic understanding of neural processes underlying psychopathology (Etkin 142 2018). Neuroimaging-based brain-behavior associations at best can only provide descriptive 143 information about what is disrupted in the brain due to psychopathology, but not why or how 144 such disruptions occur in the first place (i.e., mechanistic insights). 145 146 Lastly, psychiatric neuroimaging, and the field of psychological sciences as a whole, suffers 147 from a lack of connection between highly controlled laboratory environments and the real-world. 148 Ecological validity refers to the extent to which research findings can be generalized to real-life 149 settings. In the case of psychiatric neuroimaging, it is unclear how reliably the research findings 150 obtained in a highly controlled laboratory environment with high resolution equipment, 151 sophisticated methods, and typically unmedicated participants could be applicable to real-world 152 clinical care settings with hospital-grade equipment and treatment-seeking patients (Etkin 2019). 153 154 4. Pushing the boundaries of psychiatric neuroimaging 155 To address some of the issues with the current state of psychiatric neuroimaging, we provide 156 suggestions for moving the field forward in four main directions (Figure 1). 157 158 4.1 Improving replication and reliability 159 In the domain of reliability and replication of findings, psychiatry can emulate the momentum 160 that has already developed in the broader fields of neuroimaging and cognitive neuroscience

5

161 (Poldrack and Gorgolewski 2014). Specifically, data sharing should be encouraged across labs 162 irrespective of the sample size of the study. It has been previously argued that instead of 163 prioritizing large-scale targeted data collection and sharing (e.g. HCP (Van Essen et al. 2013)), it 164 is equally important to also share small-scale data collected by small labs and individual 165 investigators. This approach advocates for ‘data bazaars’ instead of large-scale ‘data factories’ 166 (Poldrack and Gorgolewski 2014; Poldrack and Gorgolewski 2016; Gorgolewski et al. 2017). 167 The Autism Brain Imaging Data Exchange (Di Martino et al. 2014) represents one such grass- 168 roots data sharing initiative. High impact psychiatric neuroimaging may be accomplished by 169 stitching together data/findings from smaller-scale studies, as large-scale studies may be 170 prohibitively costly. A recent case study by Milham and colleagues provides clear evidence for 171 not only accelerated science, but also massive cost benefits of data reuse as opposed to de novo 172 data generation (Milham et al. 2018). Interestingly, during the 2010-2016 period, the cost-benefit 173 analysis of the data shared via the International Neuroimaging Data-sharing Initiative (INDI) 174 consortia alone saved funding agencies north of hundreds of millions of dollars in data 175 collection. Further, availability of such data bazaars could also help with failing faster, 176 developing innovative methods, improving statistical power for future studies (Mumford 2012), 177 as well as validating older results on newer datasets. 178 179 In the area of improving reliability and replicability of findings, there is also a vital need to move 180 away from ad hoc data processing workflows towards standardized neuroimaging platforms that 181 provide analysis-agnostic tools with minimal subjective input from the users. Fortunately, 182 several such platforms have been recently developed and are actively curated for most up-to-date 183 processing workflows (Esteban et al. 2019; Glasser et al. 2013). These standardized platforms 184 also produce replicable, transparent and easy-to-use processing workflows, which will in turn 185 ensure the validity of inferences and the interpretability of results. Further, to analyze legacy 186 datasets using newer processing workflows (e.g., surface-base instead of volumetric registration), 187 novel tools are being developed that can integrate legacy datasets without acquisition of 188 accessory sequences (e.g., T2-weighted maps or fieldmaps) with data acquired using state-of-the- 189 art acquisition protocols (Dickie et al. 2019). Lastly, as data sharing grows, large-scale 190 computational frameworks are needed for processing such data. Psychiatric neuroimaging can 191 piggyback on the success of cloud computing for processing such large datasets (or individual 192 data slices) as a web-service; thereby collectively accelerating the advancement and rate of 193 breakthroughs (Madhyastha et al. 2017) (Doel et al. 2017). 194 195

6

196 197 Figure 1: Pushing the boundaries of psychiatric neuroimaging to anchor diagnosis in biological 198 features. Here, we broadly classified the push into four domains: (a) reliability of findings; (b) 199 effective clinical translation; (c) capturing mechanistic insights; and (d) enhancing ecological 200 validity of lab findings. 201 202 4.2 Improving translational outcomes 203 For improving translational outcomes, it has been suggested that researchers should not only 204 focus on increasing the sample size but also on collecting more than one (perhaps several) 205 samples from each individual (a.k.a. deep phenotyping or dense scanning). More data from each 206 individual has the obvious advantage to better capture stable traits, supporting the development 207 of personalized medicine (Gratton et al. 2018; Poldrack et al. 2015). However, to robustly study 208 individual differences, and to discern between state- and trait-related differences, one needs to 209 acquire large quantities (requiring longer scan durations or multiple scan sessions) of artifact-free 210 data (E. M. Gordon et al. 2017). Such data collection marathons are even more challenging to 211 acquire when dealing with patients with psychiatric illnesses. In the future, recent advances in 212 the development of software suites that provide scanner operators with head motion analytics in

7

213 real-time (Dosenbach et al. 2017) could be used to acquire artifact-free data. Similarly, to 214 increase participant compliance, especially in pediatric populations, novel protocols that use low 215 cognitive demand stimuli (e.g., abstract movies) that are still somewhat comparable to “resting” 216 state conditions can be used (Vanderwal et al. 2015). Importantly, the use of movie-based 217 paradigms has facilitated the measurement of functional connectivity in awake younger 218 participants (< 7 years; (Vanderwal, Eilbott, and Castellanos 2019)). Although the use of movies 219 as a stimulus (compared with traditional resting state paradigms) increases compliance in terms 220 of arousal and reduced head movement, certain caveats are important to note. First, it is unclear 221 to what degree the transitions in brain activity are associated with extrinsic stimuli (movie 222 watching) vs intrinsic signal fluctuations putatively associated with the wanderings of the mind 223 (during resting state). Second, as most movie paradigms are inherently social in nature, 224 researchers should keep in mind when interpreting results that social processing is invoked 225 (Vanderwal, Eilbott, and Castellanos 2019). Third, from the point of view of data aggregation 226 across sites/studies, it is unclear how such aggregation can be possible for different movie 227 paradigms or across movie and resting state paradigms. 228 229 For better bench-to-bedside outcomes, it is also important to invest in computational methods 230 that do not require averaging of data (across participants, space or time) at the outset (Figure 2). 231 Traditionally, the high spatiotemporal dimensionality and complexity of neuroimaging data has 232 required researchers to reduce the dimensionality of the dataset to increase the signal-to-noise 233 ratio at the cost of potentially useful information. A common example of such reduction in 234 dimensionality across participants is examining group averages. Averaging across the spatial 235 dimension has helped researchers define parcellation schemes and atlases (Schaefer et al. 2018) 236 that are beneficial for developing interventions (e.g., neurofeedback (Sitaram et al. 2017)) and 237 increasing interpretability and reproducibility of results across labs. Similarly, averaging across 238 the temporal dimension has been shown to benefit in examination of test-retest reliability of 239 functional connectivity measures. This can also aid in understanding to what degree the 240 functional organization of the brain is stable over time or is state-dependent (Gratton et al 2018). 241 Here, we argue that advances in machine learning and applied mathematics could provide novel 242 avenues to avoid such averaging of data at the outset while distilling complex neuroimaging data 243 into simple---yet vibrant and behaviorally relevant---representations that can be interactively 244 explored to discover new aspects of the data. One such approach was recently developed using 245 Topological Data Analysis (TDA) to generate interactive graphical representations of 246 neuroimaging data at the single participant level (n=1), with a spatiotemporal resolution of 247 individual voxels and time-frames to examine whole-brain activity transitions due to intrinsic or 248 extrinsic load (Saggar et al. 2018). One can imagine that such methods could be applied to 249 biologically characterize disorders of attention deficit as those with excessively “rapid” 250 transitions between brain activation patterns, while “inflexibility” or lack of transitions in brain 251 activation could be characteristic of ruminative tendencies such as those observed in depression.

8

252 Lastly, for better clinical translation of psychiatric neuroimaging findings, we suggest combining 253 data and insights across multiple units of analysis. Thus, instead of solely relying on measuring 254 the index of functional activity or connectivity associated with a domain or construct, we should 255 focus on also simultaneously assessing well-established measures of physiology (e.g., heart rate, 256 cortisol, etc.), behavior during the task and/or genetic predispositions (Uddin and Karlsgodt 257 2018). Such a multivariable approach could not only help reduce the impact of artifacts in 258 neuroimaging data (Power 2017), but could also provide a more holistic interpretation of 259 findings. 260

261 262 Figure 2. Improving clinical translation of psychiatric neuroimaging findings. Most 263 neuroimaging studies average the data in time, space and across individuals (red cube). In the 264 future, novel analytical methods (e.g., Topological Data Analysis) are needed to examine 265 neuroimaging data at the highest spatiotemporal resolution (blue cube), with the hope that such 266 methods can allow appropriate data-driven resolutions and insights. 267 268 4.3 Providing mechanistic insights 269 Psychiatric neuroimaging on its own is limited to providing descriptive information about brain- 270 behavior relationships. Thus, even with the best quality neuroimaging data we can only reveal 271 associations that require further testing to confirm cause and effect relationships (e.g., between 272 circuit perturbation and changes in behavior (Etkin 2018)). Although descriptive information is

9

273 vital, causal mechanistic information could provide the much-needed acceleration in developing 274 treatments or evaluating clinical risk factors. To this end, several techniques can be employed. 275 Here we briefly discuss four such methods – (a) performing mechanistic comparative trials; (b) 276 designing causal neurostimulation experiments; (c) developing computational models to generate 277 concrete and testable hypotheses for the putative causes of psychopathology; and (d) using 278 neurofeedback to confirm causal links. 279

280 Randomized clinical trials are often used to test the efficacy of one treatment over others, while 281 reducing selection and treatment biases (Guyatt et al. 2002). A similar approach, known as 282 comparative mechanistic trials, could also be used to isolate the neural mechanisms perturbed by 283 a specific treatment (Etkin 2018). In the case of mechanistic trials, randomization could be done 284 across groups to potentially isolate the neural mechanisms (or circuits) through which different 285 interventions influence changes in brain and behavior. Thus, such mechanistic trials could be 286 used to differentiate between two or more possible neural mechanisms underlying a particular 287 psychopathology. Further, as per the principle of clinical equipoise (London 2017), such 288 comparative mechanistic trials could be considered ethical as the neural mechanisms underlying 289 psychopathology are unknown. Clinical trials, with careful design of comparative treatments, 290 could not only help answer critical questions about the how a particular treatment perturbs neural 291 circuits, but could also reveal factors underlying treatment response (Etkin 2018).Combining 292 neuroimaging with neuromodulation (e.g., transcranial magnetic stimulation, TMS) or 293 pharmacology provides another avenue to examine how targeted perturbations affect brain 294 functioning. Such multimodal experimentation could allow for better understanding of why the 295 relation between a certain neuroimaging signal and phenotype of interest was observed. 296 Specifically, it could help answer whether the observed relation is a manifestation of the illness, 297 a compensatory process, or purely an epiphenomenon (Treadway and Leonard 2016). A recent 298 example of this line of research comes from the work of Brady et al. (2019), where the authors 299 demonstrated that TMS-modulated changes in connectivity between the cerebellum and the right 300 dorsolateral prefrontal cortex causally altered the presence of negative symptom severity in 301 patients with schizophrenia (Brady et al. 2019).

302 A third potential avenue to provide mechanistic insights associated with psychopathology comes 303 from computational modeling approaches (Friston, Redish, and Gordon 2017; Murray, Demirtaş, 304 and Anticevic 2018). Although a vast area of research, we specifically point out modeling 305 approaches grounded in biology. One such approach is that of large-scale bio-physical network 306 models (BNMs). A BNM is a system of differential equations describing how the state of each 307 local neuronal population (e.g. firing rate) changes over time in a globally connected network of 308 neuronal populations or brain regions. Here we focus on biologically realistic BNMs whose 309 parameters are constrained locally for each brain region to be consistent with microscopic 310 properties of neurons (e.g. membrane conductance, time constants of different receptors) and 311 globally for the whole brain by using human connectomes (derived from diffusion-weighted

10

312 imaging (DWI) data). The BNMs can be further personalized by fine-tuning model parameters to 313 fit each subject’s own resting state fMRI data (e.g. functional connectivity). Such BNMs are 314 already successfully generating concrete and testable hypothesis regarding neural mechanisms 315 underlying neurological disorders (Jirsa et al. 2016).

316 Lastly, another potential avenue for generating mechanistic insights about cognition in general 317 and psychopathology in particular is neurofeedback. Neurofeedback experiments entail 318 providing participants with feedback depicting their brain activity in real-time. With practice, 319 participants can learn to modulate this activity on command (Ordikhani-Seyedlar et al. 2016). 320 Although traditionally limited to brain-machine-interfaces in order to assist individuals 321 overcoming physical disabilities, recent work has shown that neurofeedback could help alleviate 322 symptoms of psychopathology as well (Nicholson et al. 2017; Kim and Birbaumer 2014). 323 Neurofeedback, being a perturbative approach, could also help in generating mechanistic 324 theories about the brain’s dynamical landscape in health and disease. Further, coupled with the 325 fields of connectomics and network control theory, neurofeedback could become the frontier in 326 understanding human cognition in health and disease (Bassett and Khambhati 2017).

327 4.4 Improving ecological validity 328 In the psychological sciences there has always been a schism between studies which are more 329 naturalistic and hence more ecologically valid versus those conducted in lab environments with 330 greater experimental control. For psychiatric neuroimaging to succeed, biological markers 331 obtained in laboratory environments under ideal conditions need to be equally reliable and 332 applicable to less ideal situations of real-world clinical care (Etkin 2019). For best bench-to- 333 bedside translation, several factors need to be reconsidered. One of the first factors to reconsider 334 is the cost and availability of high-end neuroimaging devices. MRI is perhaps the most widely 335 used neuroimaging device in the field of psychiatric neuroimaging. Although MRI provides the 336 best spatial resolution for non-invasive human brain imaging, its high cost to acquire and 337 maintain, non-portability, and low ecological validity makes it less than ideal solution for real- 338 world clinical care. Other neuroimaging modalities, like electroencephalogram (EEG) and near- 339 infrared spectroscopy (NIRS) are cost-effective, portable and more realistic solutions. Thus, 340 future work is needed to translate biological markers derived from MRI to other imaging 341 modalities. One potential venue is to use novel machine learning algorithms (e.g., Generative 342 Adversarial Networks; GANs) for performing cross-modality image synthesis (Hiasa et al. 343 2018). 344 345 To conclude, there is a growing recognition that the boundaries of human neuroimaging must be 346 pushed to ground psychiatric diagnosis in biology. 347 348 349 350

11

351 References 352 Bassett, Danielle S, and Ankit N Khambhati. 2017. “A Network Engineering Perspective on 353 Probing and Perturbing Cognition with Neurofeedback..” Edited by Michael B Miller and 354 Alan Kingstone. Annals of the New York Academy of Sciences 1396 (1). John Wiley & Sons, 355 Ltd (10.1111): 126–43. doi:10.1111/nyas.13338. 356 Bijl, R V, A Ravelli, and G van Zessen. 1998. “Prevalence of Psychiatric Disorder in the General 357 Population: Results of the Netherlands Mental Health Survey and Incidence Study 358 (NEMESIS)..” Social Psychiatry and Psychiatric Epidemiology 33 (12). Springer-Verlag: 359 587–95. doi:10.1007/s001270050098. 360 Bjork, James M, Lisa K Straub, Rosellen G Provost, and Michael C Neale. 2017. “The ABCD 361 Study of Neurodevelopment: Identifying Neurocircuit Targets for Prevention and Treatment 362 of Adolescent Substance Abuse..” Current Treatment Options in Psychiatry 4 (2). Springer 363 International Publishing: 196–209. doi:10.1007/s40501-017-0108-y. 364 Brady, Roscoe O, Jr., Irene Gonsalvez, Ivy Lee, Dost Ongür, Larry J Seidman, Jeremy D 365 Schmahmann, Shaun M Eack, Matcheri S Keshavan, Alvaro Pascual-Leone, and Mark A 366 Halko. 2019. “Cerebellar-Prefrontal Network Connectivity and Negative Symptoms in 367 Schizophrenia.” American Journal of Psychiatry 176 (7): 512–20. 368 doi:10.1176/appi.ajp.2018.18040429. 369 Button, Katherine S, John P A Ioannidis, Claire Mokrysz, Brian A Nosek, Jonathan Flint, Emma 370 S J Robinson, and Marcus R Munafo. 2013. “Power Failure: Why Small Sample Size 371 Undermines the Reliability of Neuroscience..” Nature Reviews. Neuroscience 14 (5). Nature 372 Publishing Group: 365–76. doi:10.1038/nrn3475. 373 Calhoun, Vince D, Robyn Miller, Godfrey Pearlson, and Tülay Adalı. 2014. “The 374 Chronnectome: Time-Varying Connectivity Networks as the Next Frontier in fMRI Data 375 Discovery..” Neuron 84 (2): 262–74. doi:10.1016/j.neuron.2014.10.015. 376 Cuthbert, Bruce N, and Thomas R Insel. 2013. “Toward the Future of Psychiatric Diagnosis: the 377 Seven Pillars of RDoC..” BMC Medicine 11 (1). BioMed Central: 126. doi:10.1186/1741- 378 7015-11-126. 379 Davis, Tyler, and Russell A Poldrack. 2013. “Measuring Neural Representations with fMRI: 380 Practices and Pitfalls..” Annals of the New York Academy of Sciences, June. 381 doi:10.1111/nyas.12156. 382 De Boeck, Paul, and Minjeong Jeon. 2018. “Perceived Crisis and Reforms: Issues, Explanations, 383 and Remedies..” Psychological Bulletin 144 (7): 757–77. doi:10.1037/bul0000154. 384 Di Martino, A, C G Yan, Q Li, E Denio, F X Castellanos, K Alaerts, J S Anderson, et al. 2014. 385 “The Autism Brain Imaging Data Exchange: Towards a Large-Scale Evaluation of the 386 Intrinsic Brain Architecture in Autism.” Molecular Psychiatry 19 (6): 659–67. 387 doi:10.1038/mp.2013.78. 388 Dickie, Erin W, Alan Anticevic, Dawn E Smith, Timothy S Coalson, Mathuvanthi Manogaran, 389 Navona Calarco, Joseph D Viviano, Matthew F Glasser, David C Van Essen, and Aristotle N 390 Voineskos. 2019. “Ciftify: a Framework for Surface-Based Analysis of Legacy MR 391 Acquisitions..” NeuroImage 197 (August): 818–26. doi:10.1016/j.neuroimage.2019.04.078. 392 Doel, Tom, Dzhoshkun I Shakir, Rosalind Pratt, Michael Aertsen, James Moggridge, Erwin 393 Bellon, Anna L David, Jan Deprest, Tom Vercauteren, and Sebastien Ourselin. 2017. “GIFT- 394 Cloud: a Data Sharing and Collaboration Platform for Medical Imaging Research..” 395 Computer Methods and Programs in Biomedicine 139 (February): 181–90. 396 doi:10.1016/j.cmpb.2016.11.004.

12

397 Dosenbach, Nico U F, Jonathan M Koller, Eric A Earl, Oscar Miranda-Dominguez, Rachel L 398 Klein, Andrew N Van, Abraham Z Snyder, et al. 2017. “Real-Time Motion Analytics During 399 Brain MRI Improve Data Quality and Reduce Costs.” NeuroImage 161 (November): 80–93. 400 doi:10.1016/j.neuroimage.2017.08.025. 401 Eklund, Anders, Thomas E Nichols, and Hans Knutsson. 2016. “Cluster Failure: Why fMRI 402 Inferences for Spatial Extent Have Inflated False-Positive Rates..” Proceedings of the 403 National Academy of Sciences, June. National Acad Sciences, 201602413. 404 doi:10.1073/pnas.1602413113. 405 Esteban, Oscar, Christopher J Markiewicz, Ross W Blair, Craig A Moodie, A Ilkay Isik, Asier 406 Erramuzpe, James D Kent, et al. 2019. “fMRIPrep: a Robust Preprocessing Pipeline for 407 Functional MRI.” Nature Methods 16 (1). Nature Publishing Group: 111–16. 408 doi:10.1038/s41592-018-0235-4. 409 Etkin, Amit. 2018. “Addressing the Causality Gap in Human Psychiatric Neuroscience..” JAMA 410 Psychiatry (Chicago, Ill.) 75 (1). American Medical Association: 3–4. 411 doi:10.1001/jamapsychiatry.2017.3610. 412 Etkin, Amit. 2019. “A Reckoning and Research Agenda for Neuroimaging in Psychiatry..” The 413 American Journal of Psychiatry 176 (7). American Psychiatric AssociationWashington, DC: 414 507–11. doi:10.1176/appi.ajp.2019.19050521. 415 Fisher, Aaron J, John D Medaglia, and Bertus F Jeronimus. 2018. “Lack of Group-to-Individual 416 Generalizability Is a Threat to Human Subjects Research..” Proceedings of the National 417 Academy of Sciences 115 (27). National Academy of Sciences: E6106–15. 418 doi:10.1073/pnas.1711978115. 419 Friston, Karl J, A David Redish, and Joshua A Gordon. 2017. “Computational Nosology and 420 Precision Psychiatry.” Computational Psychiatry, June. MIT Press One Rogers Street, 421 Cambridge, MA 02142-1209 USA [email protected], 1–22. 422 doi:10.1162/CPSY_a_00001. 423 Gandal, Michael J, Jillian R Haney, Neelroop N Parikshak, Virpi Leppa, Gokul Ramaswami, 424 Chris Hartl, Andrew J Schork, et al. 2018. “Shared Molecular Neuropathology Across Major 425 Psychiatric Disorders Parallels Polygenic Overlap..” Science 359 (6376). American 426 Association for the Advancement of Science: 693–97. doi:10.1126/science.aad6469. 427 Geschwind, Daniel H, and Jonathan Flint. 2015. “Genetics and Genomics of Psychiatric 428 Disease..” Science 349 (6255). American Association for the Advancement of Science: 429 1489–94. doi:10.1126/science.aaa8954. 430 Glasser, Matthew F, Stamatios N Sotiropoulos, J Anthony Wilson, Timothy S Coalson, Bruce 431 Fischl, Jesper L Andersson, Junqian Xu, et al. 2013. “The Minimal Preprocessing Pipelines 432 for the Human Connectome Project..” NeuroImage 80 (October): 105–24. 433 doi:10.1016/j.neuroimage.2013.04.127. 434 Goodkind, Madeleine, Simon B Eickhoff, Desmond J Oathes, Ying Jiang, Andrew Chang, Laura 435 B Jones-Hagata, Brissa N Ortega, et al. 2015. “Identification of a Common Neurobiological 436 Substrate for Mental Illness..” JAMA Psychiatry (Chicago, Ill.) 72 (4). American Medical 437 Association: 305–15. doi:10.1001/jamapsychiatry.2014.2206. 438 Gordon, Evan M, Timothy O Laumann, Adrian W Gilmore, Dillan J Newbold, Deanna J Greene, 439 Jeffrey J Berg, Mario Ortega, et al. 2017. “Precision Functional Mapping of Individual 440 Human Brains..” Neuron 95 (4): 791–97. doi:10.1016/j.neuron.2017.07.011. 441 Gorgolewski, Krzysztof J, Fidel Alfaro-Almagro, Tibor Auer, Pierre Bellec, Mihai Capotă, M 442 Mallar Chakravarty, Nathan W Churchill, et al. 2017. “BIDS Apps: Improving Ease of Use,

13

443 Accessibility, and Reproducibility of Neuroimaging Data Analysis Methods.” Edited by 444 Dina Schneidman. PLoS Computational Biology 13 (3). Public Library of Science: 445 e1005209. doi:10.1371/journal.pcbi.1005209. 446 Gratton, Caterina, Timothy O Laumann, Ashley N Nielsen, Deanna J Greene, Evan M Gordon, 447 Adrian W Gilmore, Steven M Nelson, et al. 2018. “Functional Brain Networks Are 448 Dominated by Stable Group and Individual Factors, Not Cognitive or Daily Variation..” 449 Neuron 98 (2): 439–452.e5. doi:10.1016/j.neuron.2018.03.035. 450 Guyatt, G, D Rennie, M Meade, and D Cook. 2002. Users' Guides to the Medical Literature: a 451 Manual for Evidence-Based Clinical Practice. 452 Hiasa, Yuta, Yoshito Otake, Masaki Takao, Takumi Matsuoka, Kazuma Takashima, Aaron 453 Carass, Jerry L Prince, Nobuhiko Sugano, and Yoshinobu Sato. 2018. “Cross-Modality 454 Image Synthesis From Unpaired Data Using CycleGAN.” In Simulation and Synthesis in 455 Medical Imaging, 11037:31–41. Effects of Gradient Consistency Loss and Training 456 Data Size. Cham: Springer, Cham. doi:10.1007/978-3-030-00536-8_4. 457 Jirsa, V K, T Proix, D Perdikis, M M Woodman, H Wang, J Gonzalez-Martinez, C Bernard, et 458 al. 2016. “The Virtual Epileptic Patient: Individualized Whole-Brain Models of Epilepsy 459 Spread..” NeuroImage, July. doi:10.1016/j.neuroimage.2016.04.049. 460 Kharabian Masouleh, Shahrzad, Simon B Eickhoff, Felix Hoffstaedter, Sarah Genon, 461 Alzheimer's Disease Neuroimaging Initiative. 2019. “Empirical Examination of the 462 Replicability of Associations Between Brain Structure and Psychological Variables..” eLife 8 463 (March): 187. doi:10.7554/eLife.43464. 464 Kim, Sunjung, and Niels Birbaumer. 2014. “Real-Time Functional MRI Neurofeedback: a Tool 465 for Psychiatry..” Current Opinion in Psychiatry 27 (5): 332–36. 466 doi:10.1097/YCO.0000000000000087. 467 London, Alex John. 2017. “Equipoise in Research: Integrating Ethics and Science in Human 468 Research..” JAMA : the Journal of the American Medical Association 317 (5). American 469 Medical Association: 525–26. doi:10.1001/jama.2017.0016. 470 Madhyastha, Tara M, Natalie Koh, Trevor K M Day, Moises Hernández-Fernández, Austin 471 Kelley, Daniel J Peterson, Sabreena Rajan, Karl A Woelfer, Jonathan Wolf, and Thomas J 472 Grabowski. 2017. “Running Neuroimaging Applications on Amazon Web Services: How, 473 When, and at What Cost?.” Frontiers in Neuroinformatics 11 (November). Frontiers: 144. 474 doi:10.3389/fninf.2017.00063. 475 McTeague, Lisa M, Julia Huemer, David M Carreon, Ying Jiang, Simon B Eickhoff, and Amit 476 Etkin. 2017. “Identification of Common Neural Circuit Disruptions in Cognitive Control 477 Across Psychiatric Disorders..” The American Journal of Psychiatry, March. American 478 Psychiatric AssociationArlington, VA, appiajp201716040400. 479 doi:10.1176/appi.ajp.2017.16040400. 480 Mumford, Jeanette A. 2012. “A Power Calculation Guide for fMRI Studies..” Social Cognitive 481 and Affective Neuroscience 7 (6): 738–42. doi:10.1093/scan/nss059. 482 Murray, John D, Murat Demirtaş, and Alan Anticevic. 2018. “Biophysical Modeling of Large- 483 Scale Brain Dynamics and Applications for Computational Psychiatry..” Biological 484 Psychiatry : Cognitive Neuroscience and Neuroimaging 3 (9): 777–87. 485 doi:10.1016/j.bpsc.2018.07.004. 486 Nicholson, Andrew A, Daniela Rabellino, Maria Densmore, Paul A Frewen, Christian Paret, 487 Rosemarie Kluetsch, Christian Schmahl, et al. 2017. “The Neurobiology of Emotion 488 Regulation in Posttraumatic Stress Disorder: Amygdala Downregulation via Real‐Time

14

489 fMRI Neurofeedback.” Human Brain Mapping 38 (1). John Wiley & Sons, Ltd: 541–60. 490 doi:10.1002/hbm.23402. 491 Open Science Collaboration. 2015. “PSYCHOLOGY. Estimating the Reproducibility of 492 Psychological Science..” Science 349 (6251). American Association for the Advancement of 493 Science: aac4716–16. doi:10.1126/science.aac4716. 494 Ordikhani-Seyedlar, Mehdi, Mikhail A Lebedev, Helge B D Sorensen, and Sadasivan 495 Puthusserypady. 2016. “Neurofeedback Therapy for Enhancing Visual Attention: State-of- 496 the-Art and Challenges.” Frontiers in Neuroscience 10 (August). Frontiers: 848. 497 doi:10.3389/fnins.2016.00352. 498 Picó-Pérez, M, M Alemany-Navarro, J E Dunsmoor, J Radua, A Albajes-Eizagirre, B Vervliet, N 499 Cardoner, et al. 2019. “Common and Distinct Neural Correlates of Fear Extinction and 500 Cognitive Reappraisal: a Meta-Analysis of fMRI Studies..” Neuroscience and Biobehavioral 501 Reviews 104 (July): 102–15. doi:10.1016/j.neubiorev.2019.06.029. 502 Poldrack, Russell A, and Krzysztof J Gorgolewski. 2014. “Making Big Data Open: Data Sharing 503 in Neuroimaging..” Nature Neuroscience 17 (11): 1510–17. doi:10.1038/nn.3818. 504 Poldrack, Russell A, and Krzysztof J Gorgolewski. 2016. “OpenfMRI: Open Sharing of Task 505 fMRI Data.” NeuroImage 144 (December). Elsevier Inc.: 1–3. 506 doi:10.1016/j.neuroimage.2015.05.073. 507 Poldrack, Russell A, Paul C Fletcher, Richard N Henson, Keith J Worsley, Matthew Brett, and 508 Thomas E Nichols. 2008. “Guidelines for Reporting an fMRI Study..” NeuroImage 40 (2): 509 409–14. doi:10.1016/j.neuroimage.2007.11.048. 510 Poldrack, Russell A, Timothy O Laumann, Oluwasanmi Koyejo, Brenda Gregory, Ashleigh 511 Hover, Mei-Yen Chen, Krzysztof J Gorgolewski, et al. 2015. “Long-Term Neural and 512 Physiological Phenotyping of a Single Human..” Nature Communications 6: 8885. 513 doi:10.1038/ncomms9885. 514 Power, Jonathan D. 2017. “A Simple but Useful Way to Assess fMRI Scan Qualities..” 515 NeuroImage 154 (July): 150–58. doi:10.1016/j.neuroimage.2016.08.009. 516 Power, Jonathan D, Kelly A Barnes, Abraham Z Snyder, Bradley L Schlaggar, and Steven E 517 Petersen. 2012. “Spurious but Systematic Correlations in Functional Connectivity MRI 518 Networks Arise From Subject Motion..” NeuroImage 59 (3): 2142–54. 519 doi:10.1016/j.neuroimage.2011.10.018. 520 Redish, A David, and Joshua A Gordon. 2016. Computational Psychiatry. MIT Press. 521 Saggar, Manish, Olaf Sporns, Javier Gonzalez-Castillo, Peter A Bandettini, Gunnar Carlsson, 522 Gary Glover, and Allan L Reiss. 2018. “Towards a New Approach to Reveal Dynamical 523 Organization of the Brain Using Topological Data Analysis..” Nature Communications 9 (1): 524 1399. doi:10.1038/s41467-018-03664-4. 525 Smith, Stephen M, Diego Vidaurre, Fidel Alfaro-Almagro, Thomas E Nichols, and Karla L 526 Miller. 2019. “Estimation of Brain Age Delta From Brain Imaging..” NeuroImage 200 527 (June): 528–39. doi:10.1016/j.neuroimage.2019.06.017. 528 Treadway, M T, and C V Leonard. 2016. “Isolating Biomarkers for Symptomatic States: 529 Considering Symptom–Substrate Chronometry.” Molecular Psychiatry 21 (9). Nature 530 Publishing Group: 1180–87. doi:10.1038/mp.2016.83. 531 Uddin, L Q, D R Dajani, W Voorhies, H Bednarz, and R K Kana. 2017. “Progress and 532 Roadblocks in the Search for Brain-Based Biomarkers of Autism and Attention- 533 Deficit/Hyperactivity Disorder..” Translational Psychiatry 7 (8). Nature Publishing Group: 534 e1218. doi:10.1038/tp.2017.164.

15

535 Uddin, Lucina Q. 2015. “Salience Processing and Insular Cortical Function and Dysfunction..” 536 Nature Reviews. Neuroscience 16 (1): 55–61. doi:10.1038/nrn3857. 537 Uddin, Lucina Q, and Katherine H Karlsgodt. 2018. “Future Directions for Examination of Brain 538 Networks in Neurodevelopmental Disorders..” Journal of Clinical Child and Adolescent 539 Psychology : the Official Journal for the Society of Clinical Child and Adolescent 540 Psychology, American Psychological Association, Division 53 47 (3). Routledge: 483–97. 541 doi:10.1080/15374416.2018.1443461. 542 van den Heuvel, Martijn P, Edward T Bullmore, and Olaf Sporns. 2016. “Comparative 543 Connectomics..” Trends in Cognitive Sciences 20 (5): 345–61. 544 doi:10.1016/j.tics.2016.03.001. 545 Van Essen, David C, Stephen M Smith, Deanna M Barch, Timothy E J Behrens, Essa Yacoub, 546 Kamil Ugurbil, WU-Minn HCP Consortium. 2013. “The WU-Minn Human Connectome 547 Project: an Overview..” NeuroImage 80 (October): 62–79. 548 doi:10.1016/j.neuroimage.2013.05.041. 549 Vanderwal, Tamara, Clare Kelly, Jeffrey Eilbott, Linda C Mayes, and F Xavier Castellanos. 550 2015. “Inscapes: a Movie Paradigm to Improve Compliance in Functional Magnetic 551 Resonance Imaging..” NeuroImage 122 (November): 222–32. 552 doi:10.1016/j.neuroimage.2015.07.069. 553 Milham, Michael P, R Cameron Craddock, Jake J Son, Michael Fleischmann, Jon Clucas, Helen Xu, 554 Bonhwang Koo, et al. 2018. “Assessment of the Impact of Shared Brain Imaging Data on the 555 Scientific Literature..” Nature Communications 9 (1): 2818. doi:10.1038/s41467-018-04976-1. 556 Schaefer, Alexander, Ru Kong, Evan M Gordon, Timothy O Laumann, Xi-Nian Zuo, Avram J Holmes, 557 Simon B Eickhoff, and B T Thomas Yeo. 2018. “Local-Global Parcellation of the Human Cerebral 558 Cortex From Intrinsic Functional Connectivity MRI..” Cerebral Cortex (New York, N.Y. : 1991) 28 559 (9): 3095–3114. doi:10.1093/cercor/bhx179. 560 Sitaram, Ranganatha, Tomas Ros, Luke Stoeckel, Sven Haller, Frank Scharnowski, Jarrod Lewis- 561 Peacock, Nikolaus Weiskopf, et al. 2017. “Closed-Loop Brain Training: the Science of 562 Neurofeedback..” Nature Reviews. Neuroscience 18 (2): 86–100. doi:10.1038/nrn.2016.164. 563 Vanderwal, Tamara, Jeffrey Eilbott, and F Xavier Castellanos. 2019. “Movies in the Magnet: Naturalistic 564 Paradigms in Developmental Functional Neuroimaging..” Developmental Cognitive Neuroscience 36 565 (April): 100600. doi:10.1016/j.dcn.2018.10.004. 566 Gratton, C. et al. Functional Brain Networks Are Dominated by Stable Group and Individual Factors, 567 Not Cognitive or Daily Variation. Neuron 98, 439–452.e5 (2018). 568

16