(PS8-B34) Habit and Goal-directed Learning in Eating Disorders: A Computational Approach

Aberrant decision-making has been recognized as an important feature of eating disorder (ED) severity. Theoretical models of decision making outline the influence of two learning systems: goal-directed learning, which reflects controlled actions based on anticipated outcomes, and habit learning, which reflects automatic and inflexible choices established by previously reinforced actions (Dolan, 2013; Morris, 2016; Doll, 2016). It is unclear which of the two learning systems is more impacted in EDs, with maladaptive goal-directed learning being reflected in altered inhibitory control and reward sensitivity, and maladaptive habit learning being reflected in compulsive or ritualistic behaviors. Moreover, deficits in learning may be influenced by behavioral phenotypes; specifically, overcontrol and greater inhibition (e.g., Restrictive Anorexia Nervosa) or under-control and greater impulsivity (e.g., Bulimia Nervosa). A recently published study reported differences in learning systems between restrictive anorexia nervosa and binge-purge anorexia nervosa (Foerde et al., 2019); thus, divergent ED presentations may reflect differences in learning systems that maintain pathology. This study examined whether reliance on goal-directed or habit learning systems differed between eating disorder phenotypes across ED diagnoses. Participants were 78 adult and adolescent ED patients, aged 17 to 66 (M = 24.51), in a partial hospitalization program. All participants completed a battery of self-report questionnaires and a two-step sequential decision-making task on the computer. We then separated the sample based on phenotype, such that one group included participants with a restrictive presentation (n = 37) and the other group included participants with a binge-purge presentation (n = 40). In order to examine the influence of goal-directed versus habit learning, we fit the task data to a reinforcement learning model that incorporates a weighted combination of model-free (habit) and model-based (goal-directed) learning (Daw et al., 2011; Otto et al., 2013). The model produced five free parameters, including weights ß^MF and ß^MB, which represented the relative contribution of model-free (habit) and model-based (goal-directed) learning, respectively. When comparing the weights between groups, we found no difference for ß^MF (t_59.47 = 0.50, p = 0.62), but we found a significant difference for ß^MB (t_42.31 = -2.16, p = 0.04). Our results indicate that ED patients with a restrictive presentation recruit habit learning at the same rate as patients with a binge-purge presentation; however, patients with a restrictive presentation demonstrate lower recruitment of goal-directed learning when compared to the binge-purge group, likely reflecting higher levels of compulsive behaviors. Overall, these findings tentatively suggest unique decision-making profiles for different phenotypes in ED populations.