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

Abstract

Background

Psychiatric disorders, including eating disorders (EDs), have clinical outcomes that range widely in severity and chronicity. The ability to predict such outcomes is extremely limited. Machine-learning (ML) approaches that model complexity may optimize the prediction of multifaceted psychiatric behaviors. However, the investigations of many psychiatric concerns have not capitalized on ML to improve prognosis. This study conducted the first comparison of an ML approach (elastic net regularized logistic regression) to traditional regression to longitudinally predict ED outcomes.

Methods

Females with heterogeneous ED diagnoses completed demographic and psychiatric assessments at baseline ( n= 415) and Year 1 ( n= 320) and 2 ( n= 277) follow-ups. Elastic net and traditional logistic regression models comprising the same baseline variables were compared in ability to longitudinally predict ED diagnosis, binge eating, compensatory behavior, and underweight BMI at Years 1 and 2.

Results

Elastic net models had higher accuracy for all outcomes at Years 1 and 2 [average Area Under the Receiving Operating Characteristics Curve (AUC) = 0.78] compared to logistic regression (average AUC = 0.67). Model performance did not deteriorate when the most important predictor was removed or an alternative ML algorithm (random forests) was applied. Baseline ED (e.g. diagnosis), psychiatric (e.g. hospitalization), and demographic (e.g. ethnicity) characteristics emerged as important predictors in exploratory predictor importance analyses.

Conclusions

ML algorithms can enhance the prediction of ED symptoms for 2 years and may identify important risk markers. The superior accuracy of ML for predicting complex outcomes suggests that these approaches may ultimately aid in advancing precision medicine for serious psychiatric disorders.