One trillion sensor minutes for a generalist health model

Google Research introduces SensorFM, a health foundation model trained on one trillion sensor minutes from five million Fitbit and Pixel Watch users.

Google Research introduces SensorFM, a foundation model for health data from wearable devices. Its defining feature is its scale: it is pre-trained on over one trillion minutes of sensor signals, collected from five million consenting participants via Fitbit and Pixel Watch devices, which Google describes as the largest dataset of its kind to date.

This approach breaks from common practice. Until now, most wrist-worn health models targeted a single outcome, using a supervised pipeline tailored to a narrow objective. In contrast, SensorFM learns a general representation of human physiology, without labels, through self-supervised reconstruction from 34 features derived from five sensor types (photoplethysmography, accelerometer, electrodermal activity, skin temperature, altimeter). This single representation then transfers to 35 prediction tasks covering cardiovascular, metabolic, sleep, and mental health domains, as well as lifestyle factors.

One design choice stands out: sensor data is inherently sparse (battery depletion, watch removed, standby mode), and where traditional methods fill in these gaps or discard incomplete windows, SensorFM learns directly from fragmented recordings by treating the absence of data as a signal in its own right. Google also reports a clear scaling effect: jointly increasing model size and data volume improves performance in a near-linear fashion, with no signs of saturation.

Two extensions are highlighted. A "classroom" of LLM agents, powered by Gemini models, automatically writes and refines the code for prediction heads, exploring over thirty thousand solutions. And integrated into a "Personal Health Agent," SensorFM served as a physiological anchor: according to a blinded panel of clinicians, summaries based on its inferences were not judged to be significantly inferior to those based on real measurements. This is research work, not an accessible product.