A Reinforcement Learning Approach to Examining the Relationship Between Risk-Sensitive Learning and Anxiety

Abstract

In light of the widely acknowledged shortcomings of the categorical classification of psychiatric disorders in the Diagnostic and Statistical Manual of Mental Disorders (DSM), there has been an effort to evaluate functioning in healthy and psychiatric behaviors on a dimensional scale. Abnormalities in one such dimension, risk sensitivity, have been observed in various psychiatric disorders, including anxiety and depression. Thus, risk sensitivity offers a promising glimpse into the underlying mechanisms of psychiatric dysfunction. One way we can experimentally assess risk sensitivity is by comparing asymmetric learning rates for positive and negative prediction errors in reinforcement learning processes. Garrett et al. (2018) revealed a change in asymmetric learning rates under state-induced anxiety, suggesting a link between anxiety symptoms and risk sensitivity. Here, we performed a conceptual replication of the Garrett et al. (2018) study using a risk-sensitive learning task to probe the correlation between risk attitudes (via asymmetric learning rates) and state anxiety in individual participants. Using a risk-sensitive temporal difference (RSTD) model for learning and decision making in the experimental task, we first show that asymmetric learning rates are significantly correlated with participants’ risk attitudes in the experimental task, reaffirming the value of the RSTD learning model in learning and decision-making research. Though we expected that asymmetric learning rates and risk attitudes in the experimental task would be correlated with self-reported anxiety measures, we did not find evidence of this correlation in our sample. This null result highlights the challenge of ecological validity in laboratory-based computational psychiatry research and motivates important future directions for research in both the theoretical reinforcement learning framework and in computational psychiatry research.