Machine learning v. traditional regression models predicting treatment outcomes for binge-eating disorder from a randomized controlled trial.

BACKGROUND: While effective treatments exist for binge-eating disorder (BED), prediction of treatment outcomes has proven difficult, and few reliable predictors have been identified. Machine learning is a promising method for improving the accuracy of difficult-to-predict outcomes. We compared the accuracy of traditional and machine-learning approaches for predicting BED treatment outcomes.
METHODS: Participants were 191 adults with BED in a randomized controlled trial testing 6-month behavioral and stepped-care treatments. Outcomes, determined by independent assessors, were binge-eating (% reduction, abstinence), eating-disorder psychopathology, and weight loss (% loss, ⩾5% loss). Predictors included treatment condition, demographic information, and baseline clinical characteristics. Traditional models were logistic/linear regressions. Machine-learning models were elastic net regressions and random forests. Predictive accuracy was indicated by the area under receiver operator characteristic curve (AUC), root mean square error (RMSE), and R2. Confidence intervals were used to compare accuracy across models.
RESULTS: Across outcomes, AUC ranged from very poor to fair (0.49-0.73) for logistic regressions, elastic nets, and random forests, with few significant differences across model types. RMSE was significantly lower for elastic nets and random forests v. linear regressions but R2 values were low (0.01-0.23).
CONCLUSIONS: Different analytic approaches revealed some predictors of key treatment outcomes, but accuracy was limited. Machine-learning models with unbiased resampling methods provided a minimal advantage over traditional models in predictive accuracy for treatment outcomes.