Historically, psychiatry has been a challenging discipline, and these challenges have only worsened with modern capitalism—computational psychiatry may help to mitigate these. Photo credit: Natasha Connell on Unsplash
From the era of asylums in the 18th century to our age of Prozac and Adderall, the history of psychiatry is stained with prejudices and misunderstandings entrenched in our socioeconomic and cultural conceptions.
Under the social circumstances and cultural trends surrounding each period, certain mental disorders have emerged, emphasised, and on rare occasions, dissipated. Throughout, patients with mental illnesses were victimised by biases, societal norms, and the manipulations of capitalistic marketing strategies. (For more details on the history of psychiatry, please refer to a book by Edward Shorter.)
It is striking to observe how disorders such as social phobia and attention deficit hyperactivity disorder (ADHD), which now constitute a significant portion of psychiatric diagnoses, owe much of their dramatic increase to the rise of a hypercompetitive, capitalistic society in the late 20th century.
Beneath the veneer of cosmetic psychopharmacology, conditions like anxiety and restlessness that might be considered normal or subclinical have been medicalised and absorbed into psychiatric practice. By 2021, reports indicated that about 14% of teenagers and 48% of young adults are suffering from one or more mental health conditions.
This grim history is partly attributable to the complex nature of psychiatric disorders arising from the intricacies of the human mind; as much as it is challenging to understand the universe within one, in studying our mind we are intrinsically confined by the very constraints our consciousness imposes on us.
Despite many efforts to rectify our historical faults in mending marred human minds, psychiatry nowadays remains as arguably one of the most unstructured medical disciplines. Its practice relies heavily on self-reported symptoms and their descriptions outlined in the Diagnostic and Statistical Manual of Mental Disorders (DSM-5), rather than on a systematic classification of diseases, or nosology, grounded in biological mechanisms.
Recognising this problem, the field of psychiatry gradually departed from a symptom-based approach to a more biology-oriented framework over the last few decades. By delineating the genetic and pathophysiological underpinnings of mental disorders, researchers endeavoured to objectively categorise patients into diagnostic cohorts and guide them toward customised therapeutic strategies. Such efforts gave rise to the modern era of biological psychiatry, dedicated to unravelling dysfunctions of the nervous system that result in cognitive abnormalities.
Challenges in modern psychiatry
In its most reductive form, the process of medical intervention can be put as a linear progression from diagnosis to treatment. Within this frame, the biggest challenge confronting modern psychiatry is the lack of a definitive nosology and a comprehensive understanding of aetiology, which concerns the causes and origins of diseases.
In addressing the former, numerous studies have identified biomarkers (i.e., biological phenotypes) at various levels—including molecular, cellular, cognitive, and behavioural—to stratify mental disorders under a robust scientific lexicon. For the latter, researchers have constructed various models of our cognitive faculties, such as memory and decision-making, to explain their inner workings in both healthy and afflicted individuals.
…the biggest challenge confronting modern psychiatry is the lack of a definitive nosology and a comprehensive understanding of aetiology…
However, these ventures call for considerable refinement. Firstly, the phenotypes identified to date are not translating into clinical efficacy. This is partly due to a multitude of mental disorders exhibiting transdiagnostic traits and a significant degree of comorbidity, further compounded by heterogeneity in disease manifestation and development.
Consequently, we are yet to discover biomarkers capable of demarcating disorders and their subtypes with high precision and clinical utility. Secondly, the absence of such biological phenotypes hampers the formation of hypotheses aimed at pinpointing the underlying mechanisms of mental disorders. This impediment in turn deters the development of targeted therapeutics. Despite an abundance of data and a proliferation of novel techniques, there also remains a dearth of theoretical explanations that can coherently integrate and make sense of this wealth of high-quality data.
Another way to view this quandary is that there is no clear starting point. As the definition of mental disorders is inherently symptomatic, it casts doubt on whether the patients under study genuinely represent distinct pathologies.
Although certain major psychiatric disorders like schizophrenia and bipolar disorder may be more readily identifiable through behavioural conditions and psychological evaluations, they, along with many others, exhibit overlapping characteristics, not to mention the potential for unknown subtypes. For example, a study reports that the diagnostic criteria for post-traumatic stress disorder (PTSD) can be combined in 636,120 different ways.
Hence, the efficacy of biomarkers in mental health research is subject to a degree of scepticism, often inflicted by inadequate statistical power and uncertainties surrounding effect sizes. To draw a rather extreme but stark analogy, the dilemma we encounter is akin to attempting to understand social ethology without taxonomic classifications or studying metallurgy devoid of knowledge of the periodic table.
So, what is computational psychiatry?
Amidst these daunting challenges, computational psychiatry emerged as a beacon of promise in the 2010s. As a burgeoning field, computational psychiatry seeks to understand mental health conditions by developing mathematical models that encapsulate cognitive processes and employing computational tools to scrutinise latent neural information.
According to a landmark paper in the field, computational psychiatry can be described as a two-body system consisting of the data-driven and theory-driven approaches. These have also been characterised as “the two cultures of computational psychiatry”—namely, the culture of machine learning and the culture of explanatory modelling.
The data-driven approach utilises statistical techniques to analyse available data, typically for diagnostic classification, prediction of treatment response, and treatment selection. Conversely, the theory-driven approach focuses on elucidating disorders through mechanistic models that clarify the relationships between biologically relevant variables. This latter approach has historically included models of biophysically plausible neural networks, reinforcement learning (RL), and Bayesian inference, providing a comprehensive picture of the neural computations at play. It is important to note that these two approaches are not mutually exclusive and can be combined to address multifaceted challenges posed by psychiatric conditions.
…computational psychiatry seeks to understand mental health conditions by developing mathematical models that encapsulate cognitive processes and employing computational tools to scrutinise latent neural information.
Ultimately, computational psychiatry stands at the confluence of neural measurement and clinical observation. It strives to connect neural substrates at the cellular and circuit level with the clinical symptoms and behaviours observed in patients. By harnessing computational resources, the field affords an opportunity to evaluate competing theories, parameterise computational phenotypes of human cognition, and enhance spatiotemporal precision in neuroimaging techniques. These benefits contribute unique values to our ongoing quest for biomarkers and the origins of psychiatric disorders without relying on subjective symptomatology, offering a new avenue for testing the clinical utility of the discoveries made.
A glimpse into the past and present of computational psychiatry
While the earlier development of computational psychiatry can be traced back to the early connectionist models of mental dysfunctions from the 1980s, for introductory purposes, a few examples for each of the two approaches would suffice.
A representative example of the data-driven approach involves studies that apply statistical models to decode given data to identify characteristic features of disorders. These features can then be used to classify patients from controls.
For instance, a study by Sitnikova and her group used a generative model called the Hidden Markov Model (HMM) to segment electrophysiological data of the human brain into a sequence of transiently activating brain networks. The authors found that a specific brain network called the default mode network (DMN), which is known to be associated with wakeful rest, mind-wandering, and self-referencing, activates less frequently and for shorter periods of time in patients with Alzheimer’s disease than in healthy controls.
Similarly, research by Higgins and his colleagues employing the HMM revealed that “replay bursts”—the patterns of neural activity associated with specific items that are reinstated during sleep or rest—coincide with DMN activation. In either case, the likelihood and stability of DMN activation were compromised in these disorders, highlighting its potential as a computational phenotype for multiple neuropsychiatric conditions.
For the theory-driven approach, the past decade has seen a focus on decision-making processes, which questions how people assess possible outcomes and make their choices accordingly. According to Montague and his colleagues:
‘Key to the initial form of computational psychiatry is the premise that, if the psychology and neurobiology of normative decision-making can be characterized and parameterized via a multi-level computational framework, it will be possible to understand the many ways in which decision-making can go wrong.’
Hence, by having decision-making as a prototypical testbed, scientists have been trying to uncover how our cognitive abilities alter under psychiatric conditions. One example of such effort is a study by Huys and his team that modelled major depressive disorder (MDD) and anhedonia using RL; these conditions were shown to be correlated with a model parameter corresponding to reward sensitivity but not with one representing the rate of learning from this reward.
Today, the rapid advancement in statistical techniques, especially in deep learning and RL, is broadening the scope of computational psychiatry more than ever.
The advent of large language models is paving the way for more accurate transdiagnostic classification and customised treatment selection, using various data sources including electronic health records. Its unprecedented efficacy enables us to model specific symptoms, such as disrupted language observed in psychotic disorders, and to develop novel generative models for interpreting a wide range of neuroimaging data. Concurrently, the field is evolving to acknowledge its previous focus on decision-making and began to examine additional cognitive traits and behaviours.
However, it is important for readers to recognise that the advancements in computational psychiatry mentioned here are but one direction the field is heading toward. While we have highlighted instances of computational models utilising neuroimaging that measures neural activity at the circuit or network level, there is another active body of research concentrating on the genetic and molecular dimensions of psychiatric disorders. The potential for ground-breaking work in understanding the pharmacological and immunological effects in psychiatry is vast. Furthermore, the recent developments in mRNA technologies post-COVID-19 and CRISPR (Clustered Regularly Interspaced Short Palindromic Repeats) genome editing herald a promising future for treatments that are on the horizon, especially with the aid of computational psychiatry in pinpointing precise neural substrates to target.
It is hoped, therefore, that this brief overview of computational psychiatry sparks interest in the field’s future direction and its synergistic possibilities with the latest computational and data science innovations.
