A wearable accelerometry landscape of human health and disease.

Abstract

Digital wearable devices provide objective characterization of human behaviors in real-world settings, yet their implications for health and diseases remain incompletely understood. Here, we analyzed a large-scale accelerometry-based cohort, including 70,473 participants with up to 10 years of follow-up, yielding a total of 11.8 million person-hours of data. From this cohort, we constructed a comprehensive atlas of 210 accelerometer-derived phenotypes (ADPs), encompassing physical activity, sedentary behavior, sleep, circadian rhythm, step counts, and step intensity. The study identified 8149 ADP-disease and 36,614 ADP-trait associations across 406 diseases and 956 health-related traits. Step intensity and circadian rhythm emerged as the most prominent ADP domains associated with diseases. In addition, we developed an artificial intelligence (AI) framework, named Wearable-Accelerometry-based Temporal-Structural Network (WATSNet), achieving high discrimination (AUC >0.7) for 112 (28.1%) incident and 65 (34.0%) prevalent diseases, e.g., incidence of Parkinson's disease (AUC = 0.873) and dementia (AUC = 0.841). We provided evidence-informed behavioral benchmarks associated with lower disease risks: at least 0.8 h per day of moderate-to-vigorous physical activity, 10,000 daily steps, and 7.8 h of sleep per day. Furthermore, we evaluated the generalizability of both ADP-disease associations and the performance of WATSNet in an external dataset (n = 4406). Together, these findings provide a comprehensive landscape linking digital behaviors to diverse health outcomes (https://digital-phenome-atlas.com) and a foundation for precision health monitoring.