Google Research has unveiled SensorFM, a foundation model designed to extract meaningful health insights from wearable devices. The model was pre-trained on over one trillion minutes of sensor data collected from five million consented participants across more than 100 countries and over 20 device models from Fitbit and Pixel Watch. The dataset represents the largest and most diverse wearable corpus ever used to train such a model, capturing continuous physiological signals including heart rate, blood-oxygen levels, sleep patterns, movement, and skin conductance.
A key innovation in SensorFM's design is its use of self-supervised learning through missing-aware masked reconstruction, an approach that solves a fundamental challenge in wearable health data: real-world gaps caused by device power cycling, battery constraints, or simply the watch being removed from the wrist. Rather than discarding incomplete data or introducing bias through imputation, the model treats missing data as a natural artifact and learns from it directly. This enables SensorFM to build a generalizable representation of human physiology that transfers across 35 different health prediction tasks spanning cardiovascular health, metabolic function, sleep quality, and mental health.
Systematic scaling experiments demonstrate that SensorFM's capabilities improve predictably as both data volume and model size increase together, with no sign of saturation. The largest model variant achieved 9% average performance gains on classification tasks and 21% performance gains on regression tasks compared to smaller variants. Researchers suggest SensorFM can serve as a foundation for a Personal Health Agent capable of label-efficient adaptation and data infilling across diverse health domains.
Key Points
SensorFM foundation model trained on 1+ trillion minutes of wearable health data from 5 million people across 100+ countries
Uses missing-aware masked reconstruction to handle real-world data gaps common in wearable device operation
Learned representation transfers across 35 health prediction tasks including cardiovascular, metabolic, sleep, and mental health domains
Scaling experiments show continuous performance gains (9% classification, 21% regression improvements) as model size and data volume increase together