Leveraging Mobile Phone Sensors, Machine Learning, and Explainable Artificial Intelligence to Predict Imminent...

Area of Science:

• Digital health interventions
• Machine learning in public health
• Mobile sensing for behavioral health

Background:

• Digital just-in-time adaptive interventions can reduce binge-drinking events (BDEs) in young adults.
• Optimizing intervention timing and content is crucial for effectiveness.
• Proactive support messages delivered before BDEs may enhance intervention impact.

Purpose of the Study:

• To assess the feasibility of a machine learning (ML) model for predicting same-day BDEs 1-6 hours in advance using smartphone sensor data.
• To identify key phone sensor features associated with BDEs on weekdays and weekends.
• To understand features driving prediction model performance.

Main Methods:

• Collected smartphone sensor data from 75 young adults (aged 21-25) over 14 weeks.
• Developed ML models (XGBoost, decision tree) to predict same-day BDEs using sensor data.
• Tested prediction distances (1-6 hours) and analysis time windows (1-12 hours); employed explainable AI.

Main Results:

• The XGBoost model achieved 95% accuracy (weekends) and 94.3% (weekdays) in predicting same-day BDEs.
• Informative features included time of day and GPS-derived data (e.g., radius of gyration).
• Interactions between time and GPS features significantly contributed to BDE prediction.

Conclusions:

• Smartphone sensor data and ML can feasibly and accurately predict imminent BDEs in young adults.
• The prediction model identifies "windows of opportunity" for intervention.
• Explainable AI highlights key features for triggering timely adaptive interventions to mitigate BDEs.