Health Label and Behavioral Feature Prediction Using Bayesian Hierarchical Vector Autoregression Models.

Area of Science:

• Computational Health
• Digital Phenotyping
• Behavioral Science

Background:

• Wearable devices and sensors generate vast amounts of health data.
• Predictive models are crucial for understanding patient health outcomes.
• Multidimensional data, including self-reports and sensor data, enhance predictive accuracy.

Purpose of the Study:

• To predict behavioral and self-reported health outcomes in college students.
• To evaluate the performance of a Bayesian Hierarchical Vector Autoregression (BHVAR) model.
• To assess the value of augmenting self-reported data with passively collected sensor data.

Main Methods:

• Utilized a Bayesian Hierarchical Vector Autoregression (BHVAR) model.
• Collected data from smartphones, wearables, environmental sensors, and self-reports.
• Trained models using 3, 7, 11, and 13 features, including behavioral data.
• Performed clustering analysis on patient-level coefficients for insights.

Main Results:

• BHVAR models demonstrated robustness with increased variable counts.
• Reduced Mean Squared Error (RMSE) improved significantly compared to maximum likelihood estimate (MLE) models.
• Augmenting self-reported data with sensor data did not significantly impact accuracy.
• Clustering analysis provided valuable patient-level insights.

Conclusions:

• BHVAR is an effective model for predicting health outcomes from multidimensional data.
• Integrating passively collected sensor data enhances predictive capabilities.
• Digital phenotyping offers a powerful approach for personalized health monitoring and intervention.