State, Trait and Recovery-Related Differences in Reinforcement Learning and Value-Based Decision Making in Major Depressive Disorder

ImportanceAltered reinforcement learning and decision-making under uncertainty have been implicated in major depressive disorder. However, it remains unclear whether these alterations reflect vulnerability or state effects of active depression, and which behaviours support sustained recovery after remission.ObjectiveTo dissociate state-like, trait-like, and remission-related alterations in reinforcement learning and value-based decision making in depression.Design, Setting, and ParticipantsUnmedicated adults (N=195) were recruited between 2013 and 2019, including individuals with current depression (n=49), individuals remitted from depression (n=50), first-degree non-depressed relatives of individuals with depression (n=36), and healthy controls (n=60).Main Outcomes and MeasuresParticipants completed a 4-armed bandit (4AB) reinforcement learning task under volatile uncertainty and a gambling task assessing value-based decision making under risk. Computational parameters indexing reward and punishment learning rates, outcome sensitivity, and decision noise (4AB), and risk aversion, loss aversion, and inverse temperature (gambling) were estimated using hierarchical Bayesian modelling with Bayesian model averaging. Group differences and associations with symptom dimensions derived from confirmatory factor analysis were examined.ResultsParticipants were 68.7% female, mean age=27.3 (SD=8.9). Punishment learning rate differed significantly across groups (F₍3,191₎=5.43, p=.001), driven by lower punishment learning rate in the remitted participants compared with all other groups. The remitted group also had lower lapse rates, indicating less random choice (F₍3,191₎=4.37, p=.005). By contrast, in the gambling task depressed and at-risk participants had higher inverse temperature relative to controls (F₍3,191₎=2.93, p=.03), suggesting more consistent use of values to guide choices. Reward and punishment sensitivity (4AB) and risk and loss aversion (gambling) did not differ across groups. Among non-depressed participants, lower punishment learning rates were associated with greater apathy (r=-0.18, p=.033) and anhedonia (r=-0.18, p=.035).Conclusions and RelevanceThese findings suggest a dissociation between state-, trait-, and recovery-associated computational phenotypes in depression. Reduced learning from negative outcomes appears specific to remitted depression, possibly reflecting adaptive changes that support sustained recovery without medication. Computational measures of learning dynamics may help clarify mechanisms of recovery and inform targets for treatment and relapse prevention.