Applying Machine Learning to Predict Complex Clinical Course in Youth With Eating Disorders

Area of Science:

• Child and Adolescent Psychiatry
• Data Science in Healthcare
• Eating Disorder Research

Background:

• Identifying youth with eating disorders (EDs) at risk of a complex clinical course is crucial for timely intervention.
• Traditional statistical methods may have limitations in predicting complex disease trajectories from multifaceted clinical data.

Purpose of the Study:

• To compare the predictive performance of supervised machine learning (ML) models against logistic regression.
• To identify youth with EDs at risk of a complex clinical course using clinical characteristics from their first treatment episode.

Main Methods:

• Utilized clinical data from 327 youth treated for EDs.
• Defined complex clinical course by readmission or non-step-down treatment trajectory.
• Trained seven ML models and logistic regression on 34 intake and discharge variables using nested cross-validation.

Main Results:

• The Random Forest model, using both intake and discharge data, achieved the highest performance (AUC=0.723, Brier=0.176), outperforming logistic regression.
• Models using only intake data showed poor predictive discrimination (AUCs < 0.6).
• Inclusion of discharge data improved performance across all ML algorithms; weight change was the most significant predictor.

Conclusions:

• Supervised ML models offer superior predictive performance for ED disease course outcomes compared to traditional methods.
• These findings support the use of ML in analyzing complex biopsychosocial data for precision medicine in ED treatment.
• Further application of ML can enhance understanding of ED etiology and disease trajectory.