Google Research has introduced SensorFM, a foundation model for wearable health data pre-trained on more than one trillion minutes of sensor data from five million people. The model learns a general-purpose representation of human physiology that transfers across 35 health prediction tasks spanning cardiovascular, metabolic, sleep, and mental health domains.
SensorFM ingests 34 one-minute aggregate features from photoplethysmography, accelerometry, electrodermal activity, skin temperature, and altimetry. It uses a self-supervised, missingness-aware training approach called Adaptive and Inherited Masking that learns directly from incomplete and fragmented wearable recordings rather than discarding them.
In evaluations across three independent studies with nearly 14,000 participants, linear probes on frozen SensorFM embeddings outperformed feature-engineered supervised baselines on 34 of 35 tasks. The model is designed to support label-efficient adaptation, automated prediction head generation via an agentic classroom of LLMs, and grounding for a Personal Health Agent.
The research represents a shift from many single-outcome wearable models toward a single generalist representation of human physiology. The team published details in a paper available on arXiv.
