Step-Counter Data Powers Scalable Health Prediction Foundation Model

Wearable and mobile sensors hold great promise for health inference, but most existing models are tailored to specific diseases, limiting their broader applicability. While wearable foundation models offer a more generalizable approach, they often rely on high-frequency raw sensor data, which raises concerns about privacy, computational demands, and scalability across diverse devices and populations. This research proposes StepFM, a novel foundation model that addresses these limitations by exclusively using step counter data. Step data is widely available and low-dimensional, making StepFM a practical, privacy-conscious, and efficient alternative. StepFM employs a scalable pre-training framework to capture temporal dynamics and behavioral patterns from vast step sequences. This enables it to transfer effectively across more than 20 different health risk prediction tasks, spanning various devices, geographical regions, and novel disease types. Experiments confirm StepFM's strong performance and robustness, while also revealing interpretable links between physical activity and health risks.