Beyond Self-Report: Adding Wearable Physiological Data to Predict Dietary Restriction in Anorexia Nervosa

Restricting food intake during meals is the core behavioral mechanism maintaining anorexia nervosa (AN) yet predicting when restriction occurs in daily life remains largely unexplored. Ecological momentary assessment (EMA) captures negative affect preceding restriction but relies on intermittent self-report, whereas wearables can passively measure physiological signals (EPA). We investigated whether combining EMA and EPA improves prediction of restriction beyond EMA alone, hypothesizing that EMA + EPA models would outperform EMA-only models, which would outperform EPA-only models.

Individuals with AN (n = 78) completed a 2-week EMA protocol while wearing an Empatica watch, completing 7151 surveys and reporting 4005 meals. Features derived from heart rate, electrodermal activity, and skin temperature were extracted across the hour before each meal. As preregistered, machine learning models were developed to predict both the occurrence and severity of self-reported restriction (logistic/linear regression, random forest, support vector machines) using EMA predictors, EPA predictors, or their combination.

Contrary to our hypothesis, adding EPA predictors did not consistently improve performance. Combined models outperformed EMA-only models for some classifiers but not for logistic regression and provided no benefit for predicting severity. Models containing EMA predictors consistently outperformed EPA-only models. The best classification model was EMA-only logistic regression (AUC = 0.734), while the best regression model was EMA-only linear regression (RMSE = 1.894).

Self-reported negative affect provided more information for predicting restriction than the physiological signals examined in this study. These findings suggest that self-report currently offers greater utility than wearable-derived stress physiology for identifying affect-related risk for restriction.