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

Details

Google Research has introduced SensorFM, a large foundation model for wearable health built on an unprecedented training corpus: over one trillion minutes of sensor data from five million consented participants, collected between September 2024 and September 2025. The dataset spans more than 100 countries, all 50 US states, and over 20 Fitbit and Pixel Watch device models — to Google's knowledge, the largest and most diverse wearable dataset used to train a model to date. SensorFM ingests 34 one-minute aggregate features from five sensor modalities including photoplethysmography (PPG) and accelerometer, learning a single reusable representation of human physiology. That representation transfers to 35 downstream health prediction tasks spanning cardiovascular, metabolic, sleep, and mental health, as well as lifestyle and demographic factors — without requiring task-specific training pipelines. Key capabilities include label-efficient adaptation (working well with minimal labeled data), data infilling, and serving as a grounding layer for a Personal Health Agent. The practical impact is significant: most previous wearable health models were narrow, one-outcome-at-a-time systems requiring expensive bespoke training. SensorFM's generalist architecture could enable health monitoring applications that were previously impractical. For health systems exploring remote patient monitoring, insurers building wellness programs on wearable data, or developers building on Fitbit and Pixel Watch ecosystems, this represents a meaningful capability advance. Privacy implications are notable given the scale — five million people's continuous physiology data — though Google states all participants provided informed consent.