Abstract
Computational neuroscience has contributed to understanding compulsive behavior by distinguishing habitual from goal-directed choice through model-free and model-based learning. Yet, questions remain about applying this approach to psychiatric conditions that are characterized by complex behaviors that occur outside the laboratory. Here, we compared individuals with anorexia nervosa (AN), whose self-starvation appears both excessively goal-directed and habitual, with healthy controls (HC) to assess: 1) whether their behavior is characterized by enhanced or diminished model-based behavior, 2) the domain specificity of any abnormalities by comparing learning in a food-specific context as well as in a monetary context, and 3) whether impairments are secondary to starvation by comparing learning before and after initial treatment. Across all conditions, individuals with AN showed an impairment in model-based, but not model-free, learning, suggesting a general and persistent contribution of habitual over goal-directed control, across domains and timepoints.
Introduction: The Control Paradox
Anorexia Nervosa (AN) presents a unique puzzle for computational psychiatry. On one hand, the "unrelenting pursuit of thinness" appears to be the ultimate expression of goal-directed behavior. On the other hand, the rigidity of self-starvation resembles a deep-seated, maladaptive habit. Researchers often distinguish between two types of decision-making: model-free (habitual, repeating what worked before) and model-based (goal-directed, using a mental map to plan). While most compulsive disorders like OCD or addiction are associated with a shift away from model-based planning, the profile of AN remained unclear. This study sought to determine if AN is characterized by excessive or deficient goal-directed control and whether these traits are specific to food or a result of starvation.
Methods: Dissociating Learning Systems
To measure these cognitive processes, the researchers utilized a Two-Step Markov Decision Task. This task is designed to dissociate model-based and model-free learning by introducing probabilistic transitions between stages.
In this task, participants make a first-stage choice that leads to one of two second-stage states with a specific probability (70% "common" or 30% "rare"). A model-free learner simply repeats an action if it was rewarded. In contrast, a model-based learner understands the transition structure; if a "rare" transition leads to a reward, they will switch their first-stage choice to the action that usually leads to that state. The study innovated by creating two versions: one for monetary rewards and one for food rewards (snack tokens), allowing for the assessment of domain-general vs. domain-specific deficits.
Participant Characteristics and Study Design
The study compared 41 women with AN (including restricting and binge-eating/purging subtypes) against 53 healthy controls (HC). A critical component of the design was testing patients twice: once at hospital admission (Time 1, acute starvation) and again after significant weight restoration (Time 2).
As shown in the clinical data, the AN group successfully achieved weight restoration between sessions, with their mean BMI rising from 16.0 kg/m² to 20.4 kg/m², reaching parity with the healthy control group ($p = 0.172$ at Time 2).
Key Insights and Results
1. Model-Based Attenuation
Contrary to the stereotype of extreme self-discipline, individuals with AN showed significantly lower model-based learning scores than healthy controls (Estimate: 0.15, $p = 0.023$). Their model-free (habitual) learning, however, remained intact ($p = 0.174$), suggesting that the deficit is specific to the planning system rather than a general learning failure.
2. Domain Generality and Persistence
The deficit was not limited to food-related decisions; patients struggled with model-based planning even when playing for monetary rewards. Furthermore, the impairment did not improve after weight restoration. This suggests that the planning deficit is a stable "trait" of the disorder rather than a "state" caused by a malnourished brain.
3. Subtype Differences
The impairment was most pronounced in the Binge-Eating/Purging (AN-BP) subtype ($p = 0.0008$ compared to HCs). While the Restricting (AN-R) group also showed lower scores, they fell between the HCs and the AN-BP group, suggesting potential computational differences between clinical presentations.
Discussion and Future Directions
These findings suggest that the "self-control" observed in Anorexia may be less about sophisticated, goal-directed planning and more about the entrenchment of rigid, habitual behaviors. The brain may become locked into certain patterns that it can no longer "plan" its way out of, even when those patterns are life-threatening.
One remaining paradox is that AN patients often perform well on delay discounting tasks (waiting for long-term rewards), which typically correlates with high model-based control. Reconciling these two findings—high patience but low planning—remains a primary goal for future research. Ultimately, this work highlights that what looks like extreme goal pursuit may actually be a failure of the very systems meant to guide flexible, goal-directed behavior.
Summary updated