Leveraging Machine Learning Algorithms to Predict Self-Rated Physical Health: Insights from the All of Us Dataset

Abstract

Aging is associated with cognitive and physical declines; however, the process varies significantly across individuals due to lifestyle and physiological factors. Non-chronological aging, i.e., aging not strictly tied to years lived, can be assessed through variables such as physical activity, sleep, and heart rate. This study explores the relationship between these metrics and self-rated physical health (SRH-P), a subjective but powerful predictor of health status throughout the aging process. Using data from the All of Us Research Program, including Fitbit-derived activity, heart rate, and sleep data from 15,620 participants, we evaluated whether machine learning (ML) models could predict SRH-P levels. SRH-P was binarized into two categories: positive (very good/excellent) and negative (poor/fair/good). We applied five ML algorithms, i.e., logistic regression (LR), support vector machine (SVM), decision tree (DT), random forest (RF), and gradient boosting decision tree (GBDT), and used AUC (area under the curve) as the performance metric. Physical activity metrics showed the strongest predictive power (AUC up to 0.7484), followed by heart rate and sleep features. Combining physical activity and heart rate features yielded the highest AUC of 0.7809 with the GBDT classifier. These results suggest that SRH-P can be meaningfully predicted from wearable data and may serve as a proxy for non-chronological aging. The two-category SRH-P framework provides a foundation for developing a standardized unit of non-chronological aging, potentially enabling personalized health tracking. Future research should incorporate additional predictors and longitudinal SRH-P data to further refine aging models and enhance their clinical applicability.