Computational psychiatry: How can mathematical psychology contribute to this emerging field?

Obsessive compulsive disorder (OCD) is a neuropsychiatric disorder characterized by recurrent unwanted thoughts (obsessions) and repetitive stereotyped behaviors (compulsions) aimed to relieve anxiety. The strength and persistence of compulsions can severely impact patients' ability to function independently, and up to 30% of treatment interventions (including combinations of medication and therapy) may fail to relieve symptoms to a significant degree. Part of the difficulty of treating compulsions lies in the fact that the specific mechanism by which they develop remains yet unknown. Previous research has offered mixed findings on whether they link to failures in learning or to failures in goal-directed behavior, and it is unclear how they work to relieve anxiety, or why treating one compulsion can still lead to a different one arising to replace it. We used computational modeling in a predictive inference task that requires integrating information at a "local" level into the wider knowledge about the structure of the "global" world. In a sample of 20 OCD patients and 23 healthy, age-matched controls, we showed similar local learning (e.g. the ability to successfully reduce uncertainty about an underlying generative process by observing sequential samples from that process) in patients and controls, but impaired ability to integrate the local knowledge into representing the wider world structure in patients. Our model proposes a hierarchical goal structure that allows for local, short-term goals (e.g. "I will wash my hands to avoid a dangerous virus") into global, longer-term goals (e.g. "I want to stay healthy and avoid disease, accidents, crimes etc."), and shows how the intact ability to acquire information to resolve local goals ("I have washed my hands and now they're clean") but the impaired ability to integrate those into the global goal leads can produce compulsive-like behaviors.

Pathological gambling is the only behavioral addiction formally recognized by the American Psychiatric Association. It has an immense potential to serve as a blueprint of addiction in the absence of substance abuse. Yet, little is known about the factors that contribute to persistent gambling. We are specifically interested in two questions: First, what gets us hooked? Why do we engage for hours in gaming, gambling or with our smartphones? And second, why do some of us stay hooked? Or what are the factors that contribute to a behavioral addiction? We test these questions by assessing the differential roles of reward and control in gambling using a novel slot machine paradigm in combination with computational modeling in pathological and matched recreational gamblers.

Auditory verbal hallucinations (AVH) are among the most distressing symptoms in psychosis, and up to 30% of patients exhibit little to no response to current treatments. This is especially concerning given that the presence of hallucinations alone increases risk of suicide in patients with psychosis. Recent advances in computational psychiatry have identified latent cognitive and perceptual states that predispose to hallucinations. Behavioral data fit to Bayesian models have demonstrated an over-reliance on priors (i.e., prior over-weighting) during perception in select samples of individuals with hallucinations. Ongoing work demonstrates that this over-reliance reflects recent symptom severity, is sensitive to the sensory modalities affected, and may be impacted by environmental exposures known to increase risk for psychosis. Taken together, this work demonstrates the potential utility of formal mathematical frameworks for understanding the generation of symptoms in psychiatric illness.

Reinforcement learning models typically assume biological agents make decisions to maximize reward. However, humans and other animals frequently behave suboptimally, suggesting they may be optimizing a different objective function. The framework of policy compression may provide some insight, by assuming agents not only seek to maximize reward but also seek to minimize cognitive cost, which is formalized as policy complexity (the mutual information between actions and states of the environment). I will first show how policy compression explains undermatching, a ubiquitous behavioral suboptimality observed in generalized matching tasks. I will then attempt to extend policy compression towards explaining psychopathology. Taken together, policy compression and other frameworks of resource-rational decision making may provide insight into psychiatric disease.

Understanding how semantic memory changes because of cognitive impairment is a basic challenge for cognitive science, and an important question for society. A rich source of real-world behavioral evidence to address this challenge is provided by memory tests routinely administered in clinical care settings. We use tens of thousands of test results from the triadic comparison task in the Mild Cognitive Impairment Screen (MCIS). This task requires people to identify the "odd one out" of a set of three animal names, which provides information about how they represent the semantic relationships between the animals, and allows inferences about the underlying mental representations. We develop a novel cognitive model of the task, using classic theories from mathematical psychology including Tversky's contrast model and Luce's choice rule. This model-based approach allows us to test different hypotheses about whether and how semantic memory changes as impairment increases. Contrary to previous claims, we find no evidence that the semantic representation of the animals changes. Instead, changes in performance can be explained in terms of worsening access to memory and the use of compensating response strategies. We emphasize how the use of cognitive models increases the theoretical insight into the changes in semantic memory, and provides a fine-grained clinical measurement capability that can be used in detection, diagnosis, and treatment.

Reinforcement learning (RL) - the process by which we learn about the environment is dysregulated in many psychiatric disorders but is especially impaired in major depressive disorder (MDD). Understanding the neurobiological correlates of RL is, therefore, a promising avenue to parse depression pathophysiology. However, RL is a multifaceted construct involving several sub-processes ranging from valuation, accumulating evidence for these options (sequential sampling), choosing the best option (explore-exploit behavior), salience attribution and lastly feedback integration (learning rate). Using computational modeling we can quantify these sub-processes and elucidate the underlying latent behavioral constructs. Interestingly, animal work has shown that these different sub-processes have different biological underpinnings, suggesting that RL sub-processes can be utilized to parse MDD heterogeneity and develop more targeted interventions. The goal of this study is to identify the functional and structural connectome of these RL subconstructs using multimodal data fusion. 46 (15 healthy, 31 clinical) subjects completed a structural T1-weighted MPRAGE scan and an RL task where they have to learn to choose the stimulus associated with rewards. A combined Q-learning/Drift Diffusion model was used to estimate RL parameters including drift rate (DR), boundary threshold (BT) and learning rate (LR) for each subject. The boundary threshold is the amount of evidence needed until a decision threshold is reached. Wider decision boundaries lead to slower and more accurate decisions, whereas narrower boundaries lead to faster but more error-prone decisions. The drift rate reflects the average speed with which the decision process approaches the response boundaries. High drift rates lead to faster and more accurate decisions. Learning rate represents the degree to which the expected values are updated and how we adjust the decisions in changing circumstances. We performed a Linked Independent Component Analysis (LICA) of 1) modulated grey matter (GM) images generated by FSLVBM, 2) vertex-wise cortical thickness (CT) and pial surface area (PSA) maps estimated using FreeSurfer across all subjects. LICA is a data-driven multivariate approach that identifies a set of multimodal spatial patterns, each comprised of morphometric properties linked across modalities, and subject loadings for each that capture inter-subject variability. LICA identified three components uniquely associated with the three RL parameters. The LR component comprised of GM density in the ventromedial and dorsolateral prefrontal, visual cortices; PSA and CT in the amygdala/hippocampus, whereas BT and DR components showed different spatial patterns. Critically, component loadings correlated with clinical symptoms. Lower structural covariance (SC) in LR component was associated with higher anxiety and lower anhedonia, suggesting different mechanisms of action. Similarly, lower SC in BT component was associated with negative affect. Multimodal data-fusion disentangles the structural connectomes of RL sub-constructs providing insight into MDD heterogeneity. Other studies utilizing these methods and prediction models will also be discussed.