Predicting responsiveness to a dialectical behaviour therapy skills training app for recurrent binge eating: A...

Area of Science:

• Digital health interventions
• Clinical psychology
• Machine learning applications

Background:

• Smartphone applications (apps) offer scalable interventions for disordered eating, but user responsiveness varies.
• Predicting individual responses to app-based interventions remains a challenge.
• Machine learning (ML) shows potential for improving predictions of complex clinical outcomes.

Purpose of the Study:

• To apply ML techniques to predict user responsiveness to a dialectical behavior therapy-based smartphone app for recurrent binge eating.
• To evaluate the predictive power of baseline demographic, clinical, and psychological variables.
• To assess the impact of app usage data on prediction accuracy.

Main Methods:

• Utilized data from a randomized controlled trial (RCT) of 576 participants with recurrent binge eating.
• Employed 10 common classification and regression ML approaches to predict intervention uptake, adherence, dropout, and symptom change.
• Developed models using 69 baseline self-reported variables and app usage metrics (e.g., modules completed).

Main Results:

• ML models using only baseline predictors showed suboptimal performance in predicting engagement (AUCs=0.48-0.61) and symptom change (R²=0.00-0.07).
• Integrating app usage data significantly improved the prediction of study dropout (AUC=0.69-0.76).

Conclusions:

• ML models alone, based on baseline predictors, could not accurately forecast user responsiveness to the app.
• Enhanced prediction accuracy may necessitate considering time-varying predictors and user context.
• Tracking intervention usage patterns, as shown by improved dropout prediction, is valuable for identifying at-risk individuals.