Physical activities enable scalable foundation modelling for broad-spectrum health prediction

As wearable health monitoring scales, researchers face a trade-off between sensing fidelity and practical deployment. While high-frequency raw sensor data (like IMU signals) provides rich information, it introduces significant privacy risks, high computational costs, and device-specific constraints. This paper investigates whether a foundation model trained exclusively on simple, ubiquitous step-count data can achieve broad-spectrum health prediction without the overhead of raw sensor processing.

The authors developed StepFM, a foundation model built on a Mamba-based architecture. Unlike traditional models that require raw motion signals, StepFM processes hourly step counts through a multi-stage pipeline: a log-scaled tokenizer for activity resolution, Fourier-based temporal rhythm encoding, and a micro-stream module that uses 1D convolutions to capture minute-level activity dynamics. The model was pre-trained on 141.12 million minute-level observations and evaluated across 21 health risk prediction tasks, ranging from cardiovascular diseases to mental health conditions.

StepFM consistently outperformed both generic time-series models and specialized wearable foundation models, achieving a mean AUROC of 0.7318. The study demonstrates that the model effectively captures both known epidemiological indicators (such as circadian regularity and activity volume) and latent, complex behavioral patterns that are difficult to engineer manually. The authors found that the predictability of a health risk is primarily driven by its inherent biological link to physical activity, and StepFM successfully amplifies these signals across heterogeneous datasets and devices.