High-Resolution Digital Phenotypes From Consumer Wearables and Their Applications in Machine Learning of...

Area of Science:

• Digital Health
• Cardiovascular Research
• Wearable Technology

Background:

• Consumer wearables capture detailed heart rate and step data during daily activities.
• Summary statistics from wearables show potential for health monitoring.
• High-resolution physiological dynamics from wearables and their link to disease markers are understudied.

Purpose of the Study:

• To extract high-resolution digital phenotypes from wearable data.
• To investigate associations between these phenotypes and cardiometabolic disease risk markers.

Main Methods:

• Developed a framework to extract 66 interpretable features from wearable data (heart rate, step count) across activity, sedentary, and sleep states.
• Applied the framework to the SingHEART study dataset (NCT02791152) including clinical and genomic data from 692 volunteers.
• Utilized machine learning to model relationships between phenotypes and risk markers, comparing predictive value using Brier scores.

Main Results:

• High-resolution features improved prediction of cardiometabolic risk markers by 7.36%–17.9% over baseline models (age, gender, resting heart rate).
• Sedentary state heart rate dynamics predict lipid abnormalities and obesity; active state dynamics predict blood pressure abnormalities.
• Wearable phenotypes better represent genomic risk for cardiometabolic disease (11.9%–22.0% Brier score improvement) and correlate with clinical events.

Conclusions:

• High-resolution digital phenotypes from consumer wearables can significantly enhance cardiometabolic disease risk prediction.
• These detailed physiological insights enable more proactive and personalized health management strategies.