Reinforcement learning in women remitted from anorexia nervosa: Preliminary examination with a hybrid reinforcement learning/drift diffusion model.

Objective: Altered reinforcement learning (RL) and decision-making have been implicated in the pathophysiology of anorexia nervosa. To determine whether deficits observed in symptomatic anorexia nervosa are also present in remission, we investigated RL in women remitted from anorexia nervosa (rAN).

Methods: Participants performed a probabilistic associative learning task that involved learning from rewarding or punishing outcomes across consecutive sets of stimuli to examine generalization of learning to new stimuli over extended task exposure. We fit a hybrid RL and drift diffusion model of associative learning to model learning and decision-making processes in 24 rAN and 20 female community controls (cCN).

Results: rAN showed better learning from negative outcomes than cCN and this was greater over extended task exposure (p < .001, η_p² = .30). rAN demonstrated a reduction in accuracy of optimal choices (p = .007, η_p² = .16) and rate of information extraction on reward trials from set 1 to set 2 (p = .012, η_p² = .14), and a larger reduction of response threshold separation from set 1 to set 2 than cCN (p = .036, η_p² = .10).

Conclusions: rAN extracted less information from rewarding stimuli and their learning became increasingly sensitive to negative outcomes over learning trials. This suggests rAN shifted attention to learning from negative feedback while slowing down extraction of information from rewarding stimuli. Better learning from negative over positive feedback in rAN might reflect a marker of recovery.