Predicting anorexia nervosa treatment efficacy: an explainable machine learning approach

Introduction: Anorexia nervosa (AN) is a psychopathology with an alarmingly high mortality rate. The growing number of individuals seeking help, coupled with the limited resources of clinics, highlights the critical need to identify factors that can predict treatment efficacy. Machine learning (ML) techniques hold great promise in this regard. This data-driven approach offers an unbiased means to uncover predictors of specific outcomes, advancing the understanding and management of this challenging condition. Objective: Six supervised ML algorithms (e.g., Decision Tree and Random Forest) were applied to develop a binary classification model predicting short-term weight recovery/stabilization in AN inpatients and identify the most critical factors influencing this outcome. Methods: Change in Body Mass Index (BMI) from admission to discharge (ΔBMI) was used as the outcome, allowing to classify patients into "improved" (BMI stability or increase) and "aggravation" (BMI decrease). Predictors included clinically relevant psychological tests and physical parameters. Scikit-learn features importance, and SHAP (SHapley Additive exPlanations) analyses were used to investigate predictor importance. Results: The Random Forest model achieved an accuracy of 0.77, an AUC-ROC of 0.72, and a PR curve score of 0.88. Body Uneasiness, Personal Alienation, and Interpersonal Problems subscales emerged as best predictors. SHAP analysis confirmed these results at the individual prediction level. Discussion: Results encouraged interventions focused on body-self experience in addition to interpersonal relationships, including body-swapping experiences and metaverse activities, respectively. This could maximize treatment efficacy, effectively allocating limited resources to achieve clinically relevant outcomes.