Using neural network-based cognitive models to quantify individual differences and predict psychiatric symptoms

/ABSTRACT Despite advances in collecting large-scale behavioral datasets, our ability to gain insights into an individual’s learning and decision-making processes remains limited. This is particularly true for characterizing individual dif- ferences in task performance, or how behavior in psychological tasks relates to psychiatric symptoms. Progress towards this ambitious goal depends on computational models that formalize the relationships between behavioral observations, the underlying latent cognitive processes, and individual differences in behavior. Unfortunately, ex- isting modeling approaches are either too simple to handle the highly variable nature of behavior, or too complex to yield interpretable insights into the cognitive processes of interest. An approach combining flexibility and inter- pretability could transform our understanding of healthy decision-making and psychiatric conditions. This proposal addresses this critical need by developing a novel computational framework to model an individual’s learning and decision-making processes in a flexible and interpretable manner. The proposal focuses on reward learning due to its critical role in healthy and dysfunctional decision-making, as well as its prevalence in psychology. Critically, our approach captures behavioral idiosyncrasies in individual subjects, instead of focusing on group averages. To achieve this specificity without undue sacrifices in interpretability, our framework relies on two techniques: very small recurrent neural networks (RNNs) trained to imitate an individual’s behavior, and dynamical systems theory to interpret how the RNN converts observations into decisions. Our prior research shows these tiny RNNs predict individual choices more accurately than classical models while revealing complex, previously unobserved learning strategies. Preliminary analyses suggest this approach discovers relationships in strategy use across tasks and identifies distinct patterns of decision-making based on clinical diagnosis. The proposed work has two primary aims. First, we will validate the stability of individual differences across multiple decision-making tasks by relating subject-specific strategies across tasks. Second, we will relate cognitive processes to psychiatric symp- toms by examining how strategies vary with symptom severity. We will also predict psychiatric symptoms based on individual differences in strategies derived from the fitted RNN models. Both analyses will use a large dataset (N = 815) currently under acquisition in the research lab of co-investigator Dr. Catherine A. Hartley, which in- cludes data from three decision-making tasks and an array of psychiatric symptom assessments. Our approach is a novel integration of data-driven and theory-driven approaches for computational psychiatry, offering a frame- work that can benefit from large datasets while still providing theoretical insights. This ability to generate cognitive theories from data alone could accelerate the study of individual cognitive differences, and particularly benefit the study of mental health. Ultimately, this could lead to more precise diagnostic tools and targeted interventions for psychiatric conditions by providing deeper insights into the cognitive mechanisms underlying decision-making. Project Number: 1R21MH142970-01 | Fiscal Year: 2026 | NIH Institute/Center: National Institute of Mental Health (NIMH) | Principal Investigator: Marcelo Mattar | Institution: NEW YORK UNIVERSITY, NEW YORK, NY | Award Amount: $435,974 | Activity Code: R21 | Study Section: Special Emphasis Panel[ZRG1 BP-V (91)] View on NIH RePORTER: https://reporter.nih.gov/project-details/11288358