Predicting anorexia nervosa treatment efficacy: an explainable machine learning approach.

Area of Science:

• Psychiatry and Mental Health
• Computational Medicine
• Machine Learning Applications

Background:

• Anorexia nervosa (AN) has a high mortality rate, necessitating improved treatment efficacy prediction.
• Limited clinical resources require identification of factors that predict treatment outcomes.
• Machine learning (ML) offers a data-driven approach to uncover predictors for AN management.

Purpose of the Study:

• Develop a binary classification model using supervised ML algorithms to predict short-term weight recovery in AN inpatients.
• Identify critical psychological and physical factors influencing weight recovery in AN.
• Utilize ML to enhance understanding and management of AN.

Main Methods:

• Applied six supervised ML algorithms, including Random Forest, for binary classification.
• Used change in Body Mass Index (BMI) from admission to discharge (ΔBMI) as the outcome measure.
• Employed Scikit-learn feature importance and SHAP analyses to identify key predictors.

Main Results:

• The Random Forest model achieved 0.77 accuracy, 0.72 AUC-ROC, and 0.88 PR curve score.
• Body Uneasiness, Personal Alienation, and Interpersonal Problems subscales were identified as significant predictors.
• SHAP analysis validated predictor importance at the individual prediction level.

Conclusions:

• Interventions targeting body-self experience and interpersonal relationships are recommended.
• Novel therapeutic approaches like body-swapping and metaverse activities may enhance treatment.
• Optimizing resource allocation through predictive modeling can improve clinically relevant outcomes in AN treatment.