The Future of Data-Driven Approaches to Mental Health
How computational psychiatry is building the next generation of mental health treatments
Despite humanity’s best efforts over many years, healing the mind remains a herculean task. Today, many researchers and clinicians seek clues in the brain, an organ whose complexity seemingly exceeds all else in the known universe, to understand mental illness and how it can be treated.
Computational psychiatry is a relatively young field that’s progressing this ancient endeavor. In this article I’ll lay out the issues facing psychiatry, explain computational psychiatry approaches to finding solutions, and give some few examples of work that’s enacting those approaches. Hopefully, by the end, you’ll have an idea of the state of things and what might lie ahead for the world of mental health treatments.
What impact does mental illness have on the world?
It’s not uncommon to know someone who is affected by mental illness. Mental disorders are estimated to affect 1 in 8 people worldwide. In the US particularly, 1 in 5 adults and 1 in 7 children are affected. However, despite millennia of theories and remedies, mental illness still wreaks enormous pain and damage to society. Over 700,000 people worldwide die from suicide every year, amounting to about one death every 45 seconds. Globally, depression is the leading cause of disability. Untreated disorders exacerbate systemic failures to provide support for marginalized communities, like people experiencing homelessness and incarcerated people. It’s estimated that the combined cost of depression and anxiety to the global economy is trillions (in USD) each year.
Why is mental illness so difficult to remedy?
There’s clearly a lot to gain from better diagnostics, treatment, and prevention of mental illness. Psychiatry, the field of medicine tasked with this goal, faces several challenges doing so. Let’s unpack them.
- Many symptoms are shared across disorders, and different subsets of symptoms can lead to the same diagnosis. A lack of unique diagnostic symptoms and biomarkers make it hard to diagnose and take into account differences between individuals who share a diagnosis.
- It is difficult to predict what treatments a patient will respond well to. It often takes trial and error to find a treatment that alleviates a patient’s symptoms. Patients can face long and arduous journeys to find relief while their clinicians remain uncertain about what treatment will be most effective.
- How disorders evolve is poorly understood. For many disorders, it’s not clear what critical risk factors cause symptoms to worsen. Lives could be saved with a better understanding of how disorders progress and when an intervention might prevent relapse.
- Treating symptoms is less effective than treating their causes. Without an understanding of the mechanisms that cause a disorder, it is challenging to develop more effective treatments.
How does computational psychiatry help?
Computational psychiatry, sometimes referred to as precision psychiatry or biological psychiatry, is a field that aims to improve how we understand mental illness by examining how information processing in the brain produces perceptions, behavior, and emotional states that underlie mental illness. It uses methods from disciplines like biology, genetics, neuroscience, theoretical physics, mathematics, and machine learning. The field accomplishes its aim through the construction of computational models that explain how illnesses manifest, evolve, and can be treated.
So how exactly does computational psychiatry help solve psychiatry’s problems?
Approach 1 — Biologically informed classification of psychiatric disorders
The Diagnostic Statistical Manual (DSM) and the International Classification of Diseases (ICD) are used by clinicians to diagnose psychiatric disorders. Both use diagnostic criteria — a list of symptoms, some of which must be present — to determine whether someone has a mental disorder. However, these manuals’ approach to nosology (the classification of disease) uses patients’ self-reported behavior, mood, social interactions, and, to a lesser extent, clinician assessment. By nature of their construction, psychiatric disorders are diagnosed solely by subjectively reported symptoms and not by biomarkers or causal mechanisms. Two problems emerge here:
- The problem of heterogeneity: Wide variation in a disorder’s qualifying diagnostic criteria means that two people might have very different symptoms but still be diagnosed with the same disorder. For example, consider the 636,120 different symptom presentations that can all lead to a diagnosis of Post-Traumatic Stress Disorder in the fifth edition of the DSM.
- The problem of comorbidity: If you’ve been diagnosed with some specific disorders, you’re more likely to be diagnosed with certain others. This motivates scrutiny toward the elements shared by disorders as well as the utility of considering disorders as independent.
Some research in the field attempts to find new ways to conceptualize the nature of disorders. The Research Domain Criteria (RDoC) framework proposed by the National Institute of Mental Health (NIMH) is a coordinated attempt to promote basic research that conceptualizes mental illness in the context of domains, fundamental neurological and behavioral functions. Symptoms can be measured with various units of analysis and can be grouped into domain-specific constructs. For example, in attention-deficit/hyperactivity disorder (ADHD), dysfunctional attention, a construct in the cognitive domain, might manifest at the physiological unit of analysis as abnormal connectivity between brain networks responsible for filtering salient events and mind wandering. Understanding dimensions of mental illness beyond those considered by the ICD and DSM might lead to the development of improved diagnostic systems and treatment decisions.
Examples of research methods include determining constructs that contribute to a disease dimension strongly associated with functional impairments and quality of life, mapping data-driven symptom phenotypes onto brain networks, and using game performance to differentiate atypical processing of rewards and social interactions.
Approach 2 — Development of personalized treatments
We still lack the necessary understanding to effectively treat many patients who have psychiatric disorders. Many patients don’t respond to first-line treatment and some are further resistant to other currently available treatments. Early, computationally derived signs that a treatment is working or will work could ensure better safety for patients and save patients, clinicians, and insurers alike time and money.
We’ve seen several advances in neuroimaging to predict treatment response in various disorders. Functional magnetic resonance imaging (fMRI), which can measure whole-brain activity associated with changes in blood flow, has been used to predict how patients with internalizing disorders respond to cognitive behavioral therapy (CBT) as well as how patients with depressive disorders respond to electroconvulsive therapy and diferent types of antidepressants. Electroencephalography (EEG), which measures electrical neural activity in the cortex, has been used to predict changes in specific symptoms in response to various antidepressants. Functional near-infrared spectroscopy (fNIRS), which measures oxygenation-related changes in hemoglobin absorption of light as a proxy for neural activity, may predict ADHD medication efficacy in children as well.
Approach 3 — Better understanding of how disorders develop
The evolution of many psychiatric disorders over time is not well understood. A patient might get better soon after a treatment but develop even worse symptoms some time later. Their care team might not understand why this happens, and crucially, not know how or when to intervene to prevent another relapse in the future.
For instance, it’s estimated that 50 to 85% of people with Major Depressive Disorder (MDD) who recover from a first episode will have another, as will about 24% of adults with anxiety disorders. Some key questions we can ask here are: what are risk factors for developing symptoms? How long does post-treatment symptom reduction last? And what are the consequences of inaction, if certain symptoms are left untreated? The prediction of risk provides an opportunity to intervene before someone is no longer able to take care of themselves.
In this vein, both neuroimaging and smartphone-based assessments have been used to identify distinct long-term developmental patterns of posttraumatic symptoms. Changes in affect and behavior over the school year, as measured from smartphone-based surveys and actigraphy wristbands, have been used to identify high-risk students and risk factors for psychopathological symptom development. Developmental trajectories of suicidal ideation have also been identified and researchers are developing a web tool to detect suicide attempt risk among people in grief.
Approach 4 — Better understanding of mechanisms that underlie psychiatric disorders
Symptoms do not exist in isolation, without origin. They arise via processes, or mechanisms, which are often unobserved by both clinicians and patients. Under this assumption, computational psychiatry attempts to find the neurobiological basis of pathological mechanisms, with the hope that discoveries might inform the development of future therapies.
This approach isn’t new to psychiatry. Selective Serotonin Reuptake Inhibitors (SSRIs), a class of antidepressant drugs, were developed with the intention of leveraging mechanistic knowledge of serotonin’s role in depression. While SSRIs have been effective for some, many people only find relief from their depression after trying several different SSRIs, and some do not find relief at all. Although the monoamine hypothesis of depression propelled SSRIs forward in their early development, it can’t explain why it often takes weeks after the initial change in serotonin levels for some patients to experience symptom relief. A better mechanistic understanding of the underlying causes of a disorder shows promise for developing better therapies.
In practice, dynamic causal modeling (DCM), a framework for testing competing mechanistic hypotheses, has been used to identify a cascade of putative mechanisms in cortical microcircuits that contribute to psychotic symptoms as well as synaptic mechanisms that seem to underpin ketamine’s antidepressant effects. Researchers have leveraged knowledge of neural mechanisms like hemodynamic response, gene expression, and excitation-inhibition balance to simulate cortical activity and predict individual’s neural and subjective responses to LSD, which has relevance for psychedelic-assisted psychotherapies.
What still needs to be done?
In the advent of prescription digital therapeutics, clinical trial simulation, and precision medical care, the role of computational psychiatry can’t be understated. Though approaches to mental health are experiencing improvements, there’s a lot to be done for the continued advancement of psychiatry.
To validate initial research findings, large amounts of data from many people are needed. Greater collaboration between researchers will be crucial to achieve this. A shared practice in what and how data are collected will greatly benefit future projects that aim to create robust models at scale. Data consortia, organizations where many researchers with a unified aim pool data they’ve collected, are one way that researchers are approaching this. Data collected in naturalistic environments, like real world data and ecological momentary assessments, also improve the ecological validity of research, making findings more generalizable to real-life situations.
Clearer regulations are needed to determine how computational psychiatry tools should be validated in order to ensure safety and efficacy, serve patients at scale, and be covered by patients’ insurance. Germany’s Digital Healthcare Act of 2019 led development on this front in the EU, and the FDA’s implementation of the Software as a Medical Device regulatory framework in 2017 was a step in this direction in the US. In the past few years we have seen an explosion of mental health apps and several guiding legislative attempts, but definitive regulations around the use of sensitive health data must be better established in order to safeguard people struggling with mental illness from further potential harm and exploitation.
Taking into account each individual patient’s biology and behaviors holds a lot of promise for improving mental healthcare. However it’s important we acknowledge that mental disorders often arise from from systemic social factors like financial insecurity, housing instability, war, and gender and race discrimination. I believe greater consideration of these factors will lead the field to develop more powerful models, better identify where intervention needs to occur and further improve accessibility of health services to those of us in need
The good news is computational psychiatry is a burgeoning field, with published research growing rapidly in the past twenty years and the global market expected to grow even more this decade. The NIMH has been increasingly funding research into computational psychiatry as well, based on estimates from NIH Reporter. Funding for the Computational Psychiatry and Statistical Methods in Psychiatry programs has increased from 0.03% (~440k) to 1.03% (~1.95mil) of the NIMH’s total yearly budget in the past 5 years alone. It’s clear there is work to be done.
In sum…
As the devastating burden of mental illness continues to weigh on the world, computational psychiatry will be crucial to improve the tools we have to identify, predict the development of, and treat psychopathology. The field’s advances are a promising for the improvement of psychiatric care but further research and regulations are necessary in order for us to move toward a world where mental health can be better addressed for everyone.
Hi, I’m Adú, a data scientist who’s passionate about improving mental health awareness, accessibility, and treatments. Have questions or want to connect? Feel free to contact me at adu@adumatory.com.
