Developing a Clinical Prediction Model for Nonimprovement of Depressive Symptoms at Discharge After Treatment of Eating Disorders

OBJECTIVE: Depressive symptoms are highly prevalent among people with eating disorders (ED). Although at the group level, depressive symptoms tend to improve alongside ED symptoms during treatment, many patients do not experience clinically meaningful reductions. Identifying at admission which patients are at risk for persistent depressive symptoms during ED treatment could support more personalized care and targeted treatment planning. METHOD: We analyzed routinely collected electronic health record data from 1412 persons receiving inpatient or day hospital ED treatment. Two outcomes were examined: (1) nonimprovement of depressive symptoms and (2) residual depression at discharge. Two machine learning (ML) models, namely elastic net regularized regression and extreme gradient boosting, were applied. Model performance was evaluated using standard classification metrics, feature importance, and decision curve analysis. RESULTS: Nonimprovement was predicted poorly (AUC = 0.64-0.65), whereas residual depression was predicted adequately (AUC = 0.73-0.77). Important predictors included phobic anxiety, resilience, life satisfaction, and baseline depression. Decision curve analysis indicated that all models provided greater net benefit than treating all or no patients across clinically relevant thresholds. CONCLUSIONS: A substantial portion of patients continued to experience notable depressive symptoms despite specialized ED treatment. Although predictive performance was moderate, our findings demonstrate the potential of preregistered and transparent ML approaches in ED settings.