Google's SensorFM Beats Rivals on 34 of 35 Health Tasks

Google Research trained a foundation model on more than one trillion minutes of wearable sensor data pulled from five million Fitbit and Pixel Watch users, then tested it against 35 separate health and behavioral tasks. It won 34 of them, beating specialized models built with hand-tuned features for single jobs like sleep staging or cardiovascular risk scoring.

The model is called SensorFM, and it's aimed at a problem that's been baked into consumer health tech since the first fitness tracker shipped: every feature gets its own narrow model. Sleep detection runs separately from stress estimation, which runs separately from metabolic tracking. Each one needs its own labeled training set, which is expensive and slow to produce, and none of them talk to each other.

Google's pitch is a shared base layer that learns directly from raw, often incomplete sensor streams, the kind of gappy, noisy data wearables actually produce, without needing labels for every task upfront. That reusable representation could then get applied wherever it's needed, from clinical-style summaries to whatever an AI health assistant might ask of it next.