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

Summary

StepFM is a foundation model built exclusively on step-counter data for broad-spectrum health risk prediction. The motivation is a practical objection to the prevailing wearable FM recipe: most existing models consume high-frequency raw sensor streams (PPG, ECG, tri-axial accelerometry), which raises privacy concerns, imposes real computational overhead, and limits portability across the heterogeneous device population that actually exists in the wild. Step count, by contrast, is ubiquitous — essentially every wearable and every smartphone reports it — extremely low-dimensional, and far less privacy-sensitive than a raw waveform.