The explosion of brain imaging studies over the last 2 decades has raised the prospect that the next stage of psychiatric practice should, or will, include biologically informed approaches to disease prognosis and diagnosis. Indeed, the ability to measure mental state from the body rather than infer through signs and symptoms would likely transform the practice to align it with known, underlying biology.1 The public availability of data and the growth of machine learning tools and computational infrastructures have also opened many avenues for humanitarian uses of neuroimaging.2 This unprecedented focus on using neuroimaging to understand human cognition in health and in disease has led naturally to the question of whether imaging will produce a tool for clinical use to diagnose or treat psychiatric conditions.
Brain Imaging in Psychiatry
It is no longer controversial to assert that psychiatric illnesses are manifestations of brain dysfunction. As neuroimaging technologies that can measure brain structure and function have matured, a significant focus has been on uncovering biomarkers of psychiatric illnesses with the goal of guiding treatment selection and predicting treatment outcome. After all, such tools, if they are accurate and based on sound clinical-neuropathological associations, would help clinicians resolve substantial uncertainty with regard to heterogeneity in psychiatric symptomatology and help further define the exact nature of illness etiology and help further define the exact nature of illness etiology, thereby enabling more systematic strategies to optimize treatment selection on an individual case-by-case basis.
Tools that capture the brain's structure or function noninvasively broadly include a wide range of technologies such as electroencephalography (EEG), magnetoencephalography (MEG), positron emission tomography (PET), magnetic resonance imaging (MRI), functional near infrared spectroscopy, and others. In general, these technologies allow many different facets of exploration, with different technologies measuring different aspects of cognition at different spatial and temporal resolutions, as well as with different proxies for neural function. EEG and MEG measure currents and magnetic fields generated by neural circuits in the brain, PET measures cerebral blood flow or responses to binding sites of specific radioactive markers, and MRI can measure blood oxygenation levels, blood flow, proton density, water diffusion, and chemical spectral signatures. In many ways, these measurements are proxies for neural function and dynamics. Yet, these measurements have proven helpful in unravelling many aspects of cognition and many aspects of neurological and psychiatric disorders. The next two sections of this article focus on a subset of neuroimaging discoveries related to MRI and targeting diagnoses, prognoses, and biotyping (Table 1), as well as the advent of neuroimaging-based therapeutics that have shown promise in the treatment of psychiatric disorders.
Diagnoses, Prognoses, and Biotyping Using Neuroimaging
Most neuroimaging studies in psychiatry have focused on distinguishing brain measurements between groups, with one group having a psychiatric illness whereas the other does not. Greater accessibility to mental health datasets, the pooling across studies, and the increase in study sizes have contributed to more studies incorporating machine learning into such analyses. The use of machine learning to create predictive models for diagnoses, treatment outcome, or biotyping has increased, stemming from the cross-disciplinary engagement of researchers across machine learning, brain imaging, and psychiatry, coupled with the availability of open resources (both data and computational platforms). One reason machine learning approaches are so important is they move beyond the status quo of focusing on group-level inferential statistics into predictive statistics in unseen data, testing how well an algorithm works when applied to new individual cases.
Neuroimaging studies have uncovered many relations between measures of brain structure (eg, regional volume, thickness, diffusion-based measures) and brain function (eg, task responses and resting state measures) and related psychiatric diagnoses. Using machine learning on brain imaging data, several studies have developed mathematical algorithms that can discriminate healthy people from people with mental illnesses. Such methods have been applied to major depression,3 schizophrenia,4 autism and attention-deficit/hyperactivity disorder (ADHD),5 and bipolar disorder.6 Such methods have also been applied across disorders to diagnose chronic psychiatric illnesses.7
Neuroimaging-based diagnosis alone, especially at a potentially higher cost relative to a clinical visit, can only be justified if such a diagnosis is more consistent or can lead to better prognosis, for instance by guiding treatment selection for optimal clinical outcome. A review by Janssen et al.8 noted a few efforts across different psychiatric disorders, such as major depressive disorder9 and social anxiety disorder.10
Some studies have used neuroimaging to define new phenotypic profiles or biotypes that may respond more to one treatment option over another. These biotypes may also provide a different perspective on how psychiatric illnesses are characterized based on biology in addition to symptoms. Some examples include work on schizophrenia11 and depression,9 with this latter work also demonstrating that these biotypes of depression also respond differentially to treatment.
Translational Challenges of Psychiatric Neuroimaging
The promise of these technologies has been evident from the earliest days, and yet the realization of this promise (ie, translation to clinical practice) has lagged behind, having been met with a host of valid concerns that are still being addressed: (1) What is the specificity and sensitivity of any brain-based measure of a psychiatric state or trait feature? (2) What are the relative costs of employing neuroimaging to inform diagnostic or prognostic evaluation, and what is the evidence for a favorable cost-benefit ratio, relative to treatment as usual? (3) How well does a brain-based measure generalize across disease diagnoses?
Most studies have focused on a particular disorder rather than choosing among many for diagnosis, or choosing among many treatment options for prognosis, or considering shared biotypes relative to classic categories of mental illnesses. These studies have largely been unimodal, focusing on a specific imaging modality rather than trying to combine information from different types of imaging, although multimodal approaches are emerging.12 Even promising biotype reports, such as one that recently reported subtypes of major depression based on neuroimaging findings,9 have had difficulties with replication.13 In many cases, the focus has been on just imaging rather than a complementary scheme that combines imaging with other sources of information, such as genetics, medical information, family history, or more objective ways of quantifying behavior. The number of people used in these studies has also been limited relative to the variation of symptomatology present in many of these disorders.
Advances in quantitative imaging have not yet been applied systematically to the study of mental illnesses. The spotlight has been on the clinical diagnosis or prognosis, and not on potentially confounding factors such as the diversity of software and hardware options that affect processing of brain imaging data across scanners and sequences, participants at different age, and clinical variations in assigning diagnosis labels or severity levels to conditions. These issues are addressable and, given that the stakes remain high to push the technology to its limits, the field will likely address many of them. How quickly it happens will depend on the level of attention paid to relevant clinical, technical, and operational details.
Neuroimaging in psychiatry is not just about diagnosis and prognosis. Another avenue is the ability to use neuroimaging-based biofeedback,14 such as real-time functional MRI (RT-fMRI) in which signals from a person's brain can be processed on the fly and delivered back to them, as a way of actually treating people with a mental illness. In general, this method of “neurofeedback” therapy involves training a person to learn to regulate brain activity, typically in a brain region or a network of regions, based on visual or auditory feedback of the current brain state of the person. Alongside a clinician or researcher coaching the person to adopt certain mental strategies (eg, theory of mind, mental imagery, meditation), clinical trials have used neurofeedback to address autism spectrum disorder15 and major depression disorder. Numerous current trials are addressing aggression, auditory hallucination in psychosis, Tourette syndrome, anxiety, addiction, stress, and major depressive disorder. In a completed, randomized trial of 36 participants, feedback from the amygdala was used successfully to reduce a quantitative measure of depression severity.16 The response from early trials is encouraging, but significant effort is needed to understand interactions in the complex parameter space involving MRI sequence options, brain targets, the various analytical methods used to compute the feedback signal, and the diversity of psychiatric disorders. Combining therapy with biotype detection may provide a way to optimize these parameters systematically.
Validating Machine Learning Approaches in Psychiatry
Many publications apply machine learning to brain imaging data in psychiatry. However, in the absence of published models and validation of these models on new data, evaluating their generalization potential in new practical situations (eg, different scanner, different pulse sequence, different participant pools) is difficult, if not impossible. Open machine learning challenges, such as the Netflix prize for content recommendation and the software competitions hosted at Kaggle.com, alter this landscape, allowing consolidation of both models and rigorous validation in a controlled setting. These competitions have enabled a new scalable and collaborative approach to solving practical problems using machine learning and generated new algorithms and software. In brain imaging and psychiatry, there are only a handful of examples of software competitions focused on a few isolated disorders: ADHD,17 schizophrenia,18 autism,19 and major depression.20 These competitions help reduce overfitting (ie, the tendency of any one model or algorithm to perform well initially but then fail to generalize, a common occurrence in machine-learning publications) by holding out test data to validate the models. In sum, validation in the modern era will depend increasingly on our collectively ability to share data and models in ways that minimize bias, such as open competitions, by enabling repeated testing in new samples and by new groups of researchers who lack any ties or vested interests to initial reports.
Aggregating Data to Enhance Clinical Use and Develop New Technologies
To understand human variation across psychiatric disorders and treatment response and to apply machine learning to psychiatry, there needs to be a continuous supply of cross-sectional and longitudinal data. There have been significant efforts in pooling data, both open access and restricted access, in autism21 and ADHD,17 and the releases of Connectomes for Disease will soon occur.22 Efforts such as the ENIGMA (Enhancing Neuro Imaging Genetics through Meta Analysis) consortium23 and GENUS (Genetics of Endophenotypes of Neurofunction to Understand Schizophrenia) consortium24 have focused on academic collaborations to scale sample sizes, but such consortia are still generally focused on specific mental illnesses and are not generally accessible for crowd-sourcing exploration. Even with these efforts and the number of people being imaged across the planet in various research and clinical studies, there are no large-scale databases of neuroimaging data across psychiatric disorders that match their prevalence, and open sources of longitudinal data are rarer still. This limits the possibilities enabled by a plethora of modern machine learning tools, and more importantly limits reproducibility and validation of current models. But sharing data is only a starting point to engage communities to work together to build better clinical tools that leverage current advances in statistical learning and the new methods that will need to be created to address the many challenges with these data. These data will come from different scanners and sequences, with each scanning site contributing some combination of neuroimaging, cognitive and neuropsychological assays, and genetic information. As such, there may not be consistency across all samples, but integrating information across sparse data is a current data science challenge. Given current directions in data protection, there need to be mechanisms and policies in place that ensure that such data can be used while respecting individual privacy. Simply building walls around data is an insufficient approach to solving a pressing health problem.
The Need for Computational Models When Imaging Complex Systems
With the combining of multiple imaging modalities and analytical pipelines, the field is now able to generate hundreds of thousands of features, far exceeding the number of samples in some of the largest datasets in existence. This requires development of computational models that can help guide the selection of features, predict observations in neuroimaging studies, and create a computational basis for symptoms associated with psychiatric disorders.25 Our ability to synthesize information from these neuroimaging studies into computational models of psychiatric disorders has lagged behind the general growth in brain imaging studies. Such models will serve as essential links between psychiatric disorders and recently identified interacting physiological processes such as inflammation, the microbiome, and others.26 Computational models, just like statistical learning models, require data to validate; thus, even these models will improve with greater sharing of data.
Integrating Neuroimaging with Other Phenotyping Approaches and Genomics at Scale
Mental illnesses are complex. No single modality, whether genomics, neuroimaging, neuropsychological assessments, or other forms of phenotyping (eg, wearables, voice, and language) is likely to provide robust markers of disease or pathways to treatment. Except in a few disorders, most studies have focused on one of these modalities. For example, voice analysis has been linked to suicide27 and depression,28 and gait abnormalities have been noted in depression29 and autism.30 Few efforts have tried to combine information across assays and disorders. Efforts such as the Adolescent Brain Cognitive Development study31 to collect longitudinal imaging, behavioral, and clinical data at scale have only just started but are focused primarily on substance abuse.
There is a need to integrate genomics, brain imaging, and other digital phenotyping to create as complete a picture as possible. This integration will require new technologies and models to be developed and validated, with cooperation at a global scale. Mental illness as a human problem needs to be studied across languages and cultures. If certain groups of people show lower incidence rates given a consistent diagnostic criterion, it would be useful to understand why there are differences in incidence rates, whether our symptomatic understanding needs modification, or whether there are genetic and/or cultural traits that make people in such groups more or less susceptible to mental illnesses.
Neuroimaging provides visibility to and quantification of mental illnesses that go beyond symptomatic understanding; however, such measures are in their infancy and still have limited clinical utility. The emergence of machine learning methods alongside the growth of neuroimaging datasets offers an unprecedented opportunity to move beyond classical diagnoses and improve prognosis. Neuroimaging can help create a set of biologically derived subtypes that respond better to specific treatment options and can itself serve a novel therapeutic role via neurofeedback. Using neuroimaging alongside clinical practice, and coupling with other emerging areas such as digital phenotyping and genotyping, may provide the contextual integration necessary to create tools for clinical use. As the future of psychiatry continues to evolve toward the ideals of precision medicine, we expect neuroimaging to emerge as an integral component of that future.
- Kapur S, Phillips AG, Insel TR. Why has it taken so long for biological psychiatry to develop clinical tests and what to do about it?Mol Psychiatry. 2012;17(12):1174–1179. doi:. doi:10.1038/mp.2012.105 [CrossRef]
- Gabrieli JDE, Ghosh SS, Whitfield-Gabrieli S. Prediction as a humanitarian and pragmatic contribution from human cognitive neuroscience. Neuron. 2015;85(1):11–26. doi:. doi:10.1016/j.neuron.2014.10.047 [CrossRef]
- Patel MJ, Khalaf A, Aizenstein HJ. Studying depression using imaging and machine learning methods. Neuroimage Clin. 2016;10:115–123. doi:. doi:10.1016/j.nicl.2015.11.003 [CrossRef]
- Zarogianni E, Moorhead TWJ, Lawrie SM. Towards the identification of imaging biomarkers in schizophrenia, using multivariate pattern classification at a single-subject level. Neuroimage Clin. 2013;3:279–289. doi:. doi:10.1016/j.nicl.2013.09.003 [CrossRef]
- Uddin LQ, Dajani DR, Voorhies W, Bednarz H, Kana RK. Progress and roadblocks in the search for brain-based biomarkers of autism and attention-deficit/hyperactivity disorder. Transl Psychiatry. 2017;7(8):e1218. doi:. doi:10.1038/tp.2017.164 [CrossRef]
- Librenza-Garcia D, Kotzian BJ, Yang J, et al. The impact of machine learning techniques in the study of bipolar disorder: a systematic review. Neurosci Biobehav Rev. 2017;80:538–554. doi:. doi:10.1016/j.neubiorev.2017.07.004 [CrossRef]
- Bansal R, Staib LH, Laine AF, et al. Anatomical brain images alone can accurately diagnose chronic neuropsychiatric illnesses. PLoS One. 2012;7(12):e50698. doi:. doi:10.1371/journal.pone.0050698 [CrossRef]
- Janssen RJ, Mourão-Miranda J, Schnack HG. Making individual prognoses in psychiatry using neuroimaging and machine learning. Biol Psychiatry Cogn Neurosci Neuroimaging. 2018;3(9):798–808. doi:. doi:10.1016/j.bpsc.2018.04.004 [CrossRef]
- Drysdale AT, Grosenick L, Downar J, et al. Resting-state connectivity biomarkers define neurophysiological subtypes of depression. Nat Med. 2017;23(1):28–38. doi:. doi:10.1038/nm.4246 [CrossRef]
- Doehrmann O, Ghosh SS, Polli FE, et al. Predicting treatment response in social anxiety disorder from functional magnetic resonance imaging. JAMA Psychiatry. 2013;70(1):87–97. doi:. doi:10.1001/2013.jamapsychiatry.5 [CrossRef]
- Clementz BA, Sweeney JA, Hamm JP, et al. Identification of distinct psychosis biotypes using brain-based biomarkers. Am J Psychiatry. 2015;173(4):373–384. doi:. doi:10.1176/appi.ajp.2015.14091200 [CrossRef]
- Whitfield-Gabrieli S, Ghosh SS, Nieto-Castanon A, et al. Brain connectomics predict response to treatment in social anxiety disorder. Mol Psychiatry. 2016;21(5):680–685. doi:. doi:10.1038/mp.2015.109 [CrossRef]
- Dinga R, Schmaal L, Penninx B, et al. Evaluating the evidence for biotypes of depression: attempted replication of Drysdale et.al. 2017. bioRxiv. doi:10.1101/416321 [CrossRef]. https://www.biorxiv.org/content/10.1101/416321v1.article-info. Accessed April 12, 2019.
- Stoeckel LE, Garrison KA, Ghosh S, et al. Optimizing real time fMRI neurofeedback for therapeutic discovery and development. Neuroimage Clin. 2014;5:245–255. doi:. doi:10.1016/j.nicl.2014.07.002 [CrossRef]
- US National Library of Medicine. ClinicalTrials.gov. Rt-fMRI NF intervention study in ASD (NF-ASD). https://clinicaltrials.gov/ct2/show/study/NCT02440451. Accessed April 25, 2019.
- Young KD, Siegle GJ, Zotev V, et al. Randomized clinical trial of real-time fMRI amygdala neurofeedback for major depressive disorder: effects on symptoms and autobiographical memory recall. Am J Psychiatry. 2017;174(8):748–755. doi: . doi:10.1176/appi.ajp.2017.16060637 [CrossRef]
- Milham MP, Fair D, Mennes M, Mostofsky SH, et al. The ADHD-200 consortium: a model to advance the translational potential of neuroimaging in clinical neuroscience. Front Syst Neurosci. 2012;6:62. doi:10.3389/fnsys.2012.00062 [CrossRef].
- MLSP 2014 schizophrenia classification challenge. https://www.kaggle.com/c/mlsp-2014-mri. Accessed April 12, 2019.
- IMPAC - IMaging-PsychiAtry Challenge: predicting autism. https://paris-saclay-cds.github.io/autism_challenge/. Accessed April 12, 2019.
- Predictive analytics competition 2018. https://www.photon-ai.com/pac. Accessed April 12, 2019.
- Di Martino A, Yan C-G, Li Q, et al. The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain architecture in autism. Mol Psychiatry. 2014;19(6):659–667. doi:. doi:10.1038/mp.2013.78 [CrossRef]
- Connectome Coordination Facility. Human connectome studies related to disease. https://www.humanconnectome.org/disease-studies. Accessed April 12, 2019.
- Thompson PM, Stein JL, Medland SE, et al. Alzheimer's Disease Neuroimaging InitiativeEPIGEN ConsortiumIMAGEN ConsortiumSaguenay Youth Study (SYS) Group. The ENIGMA Consortium: large-scale collaborative analyses of neuroimaging and genetic data. Brain Imaging Behav. 2014;8(2):153–182. doi:10.1007/s11682-013-9269-5 [CrossRef].
- Blokland GAM, Del Re EC, Mesholam-Gately RI, et al. The Genetics of Endophenotypes of Neurofunction to Understand Schizophrenia (GENUS) consortium: a collaborative cognitive and neuroimaging genetics project. Schizophr Res. 2018;195:306–317. doi:. doi:10.1016/j.schres.2017.09.024 [CrossRef]
- Cooper JA, Arulpragasam AR, Treadway MT. Anhedonia in depression: biological mechanisms and computational models. Curr Opin Behav Sci. 2018;22:128–135. doi:. doi:10.1016/j.cobeha.2018.01.024 [CrossRef]
- Kaplan BJ, Rucklidge JJ, Romijn A, McLeod K. The emerging field of nutritional mental health: inflammation, the microbiome, oxidative stress, and mitochondrial function. Clin Psychol Sci. 2015;3(6):964–980. doi:. doi:10.1177/2167702614555413 [CrossRef]
- France DJ, Shiavi RG, Silverman S, Silverman M, Wilkes DM. Acoustical properties of speech as indicators of depression and suicidal risk. IEEE Trans Biomed Eng. 2000;47(7):829–837. doi:. doi:10.1109/10.846676 [CrossRef]
- Mundt JC, Vogel AP, Feltner DE, Lenderking WR. Vocal acoustic biomarkers of depression severity and treatment response. Biol Psychiatry. 2012;72(7):580–587. doi:. doi:10.1016/j.biopsych.2012.03.015 [CrossRef]
- Lemke MR, Wendorff T, Mieth B, Buhl K, Linnemann M. Spatiotemporal gait patterns during over ground locomotion in major depression compared with healthy controls. J Psychiatr Res. 2000;34(4–5):277–283. doi:10.1016/S0022-3956(00)00017-0 [CrossRef]
- Vilensky JA, Damasio AR, Maurer RG. Gait disturbances in patients with autistic behavior: a preliminary study. Arch Neurol.1981;38(10):646–649. doi:10.1001/archneur.1981.00510100074013 [CrossRef]
- Bjork JM, Straub LK, Provost RG, Neale MC. The ABCD study of neurodevelopment: identifying neurocircuit targets for prevention and treatment of adolescent substance abuse. Curr Treat Options Psychiatry.2017;4(2):196–209. doi:. doi:10.1007/s40501-017-0108-y [CrossRef]
||A field of study that focuses on development of statistical models to uncover and learn patterns in data. In general, models are learned on one dataset and tested or validated on a different dataset. This allows such models to predict or associate information in a more generalizable manner. Practically speaking, applying machine learning to a particular problem may require new algorithms to extract novel features from data, in addition to selecting features and models in an unbiased manner, training these models on datasets capable of representing as much of the diversity of clinical population(s) under study, and then validating these models using previously unseen data
||The identification of a psychiatric illness with the intent to determine if any, and what type of, intervention is needed to reduce, and more preferably eliminate, the illness. In current psychiatric practice, the diagnosis is carried out via an evaluation of symptomsalongside patient history and familial context
||The ability to determine the likely course of a psychiatric illness, including, but not limited to, duration, effect on daily life, and response to or remission from treatment
||The ability to use genomic, endophenotypic, and other clinical measurements (eg, neuroimaging, wearables, voice and language, neuropsychological, cognitive tests, blood-based) as a means to better relate people to each other, to create groups that will respond favorably to treatments, or interventions tailored for that group. The extension of this idea is precision psychiatry, in which every person can be treated optimally based on their own genotype and phenotype, potentially taking into account relatedness to other people