Machine learning enhances prediction of illness course: a longitudinal study in eating disorders

Area of Science:

• Psychiatry
• Computational Biology
• Data Science

Background:

• Predicting clinical outcomes for psychiatric disorders, including eating disorders (EDs), remains challenging.
• Machine learning (ML) offers potential for modeling complex behaviors and improving prognostic accuracy.
• This study is the first to compare ML with traditional regression for predicting longitudinal ED outcomes.

Purpose of the Study:

• To compare the predictive accuracy of an ML approach (elastic net) versus traditional logistic regression for longitudinal ED outcomes.
• To identify key predictors for ED prognosis using ML techniques.
• To assess the potential of ML in advancing precision medicine for psychiatric conditions.

Main Methods:

• Longitudinal study of females with ED diagnoses (n=415 at baseline).
• Collected demographic and psychiatric data at baseline and Years 1 and 2 follow-ups.
• Compared elastic net regularized logistic regression with traditional logistic regression for predicting ED diagnosis, binge eating, compensatory behavior, and underweight BMI.

Main Results:

• Elastic net models demonstrated higher accuracy (AUC=0.78) than logistic regression (AUC=0.67) for all outcomes at Years 1 and 2.
• Model performance remained robust even when key predictors were removed or alternative ML algorithms were used.
• Baseline ED diagnosis, psychiatric history (e.g., hospitalization), and demographic factors (e.g., ethnicity) were identified as significant predictors.

Conclusions:

• ML algorithms can significantly improve the 2-year prediction of ED symptoms and identify crucial risk markers.
• The enhanced accuracy of ML for complex outcomes supports its utility in developing personalized treatment strategies for severe psychiatric disorders.