SensorFM: Towards a general intelligence and interface for wearable health data

A foundation model called SensorFM has been trained on over one trillion minutes of wearable sensor data from five million people to learn general patterns of human physiology. The model addresses a key challenge in wearable health: that individuals' baseline physiology varies greatly, labeled health data is expensive to obtain, and most current models target only single health outcomes rather than generalizing broadly. SensorFM learns from unlabeled data using a method that treats missing or fragmented sensor readings as natural artifacts rather than problems to solve, and its learned representation transfers across 35 different health prediction tasks spanning cardiovascular, metabolic, sleep, and mental health domains. The research demonstrates that scaling both the model size and training data together produces consistent improvements in performance, with larger models increasingly capturing physiologically relevant traits without requiring explicit demographic information.