Predicting Disengagement to Better Support Outcomes in a Web-Based Weight Loss Program Using Machine Learning Models:...

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Behavioral Science

Background:

• Engagement is crucial for behavior change and health improvement in interventions.
• Limited research exists on using machine learning (ML) to predict disengagement in commercial weight loss programs.
• Predicting disengagement can help participants achieve their health goals.

Purpose of the Study:

• To utilize explainable machine learning (ML) to forecast member disengagement risk weekly over a 12-week period.
• To identify key predictors of disengagement in a web-based weight loss program.

Main Methods:

• Developed and validated predictive models (Random Forest, Extreme Gradient Boosting, Logistic Regression) using data from 59,686 participants.
• Employed 10-fold cross-validation and temporal validation on a separate cohort.
• Utilized Shapley values for feature importance and prediction explanation.

Main Results:

• Extreme Gradient Boosting models demonstrated the best predictive performance, with Area Under the Receiver Operating Characteristic Curve (AUC-ROC) ranging from 0.85 to 0.93.
• Area Under the Precision-Recall Curve (AUC-PR) ranged from 0.57 to 0.95, showing significant improvement (20% at week 3).
• Key predictors of disengagement included overall platform activity and previous weight entry frequency.

Conclusions:

• Machine learning algorithms show potential for predicting and understanding participant disengagement in online weight loss programs.
• These insights can inform targeted support strategies to enhance user engagement.
• Improved engagement is linked to better health outcomes and greater weight loss success.