Leveraging Machine Learning to Develop Digital Engagement Phenotypes of Users in a Digital Diabetes Prevention...

Area of Science:

• Digital Health
• Machine Learning
• Diabetes Prevention

Background:

• Digital diabetes prevention programs (dDPPs) show promise but suffer from high attrition rates.
• Personalized automatic messaging systems (PAMS) aim to improve dDPP engagement through enhanced patient-provider communication.
• Optimizing PAMS requires understanding user preferences and predicting engagement through data analysis.

Purpose of the Study:

• To evaluate machine learning (ML) for developing digital engagement phenotypes in dDPP users.
• To assess ML accuracy in predicting engagement with dDPP activities.
• To inform PAMS personalization by incorporating engagement prediction and digital phenotyping.

Main Methods:

• Utilized a gradient-boosted forest model to predict engagement in dDPP activities (physical activity, lessons, social, weigh-ins) and general activity.
• Employed metrics like AUC-ROC, AUC-PR, and Brier score to evaluate model performance.
• Used Shapley values for feature importance and latent profile analysis for user profiling.

Main Results:

• Developed two ML models (weekly and daily data) with over 90% predictive accuracy.
• The daily model was crucial for creating "digital phenotypes" by predicting daily activity changes.
• Identified 6 user profiles (high engagement, minimal engagement, attrition) based on 2-week engagement data.

Conclusions:

• ML methods can effectively tailor and optimize messaging interventions for digital prescriptions.
• Preliminary results support the use of ML for enhancing patient engagement and adherence in dDPPs.
• This methodology can be expanded to optimize existing platforms and other clinical domains.