SensorFM: A Foundation Model for Wearable Health Data

Researchers introduced SensorFM, a Large Sensor Foundation Model trained at population scale on unlabeled wearable sensor data. Pre-trained on over a trillion minutes (two billion+ hours) of minute-resolution signals from five million consented participants across 100+ countries and 20+ Fitbit and Pixel Watch models, the model learns from 34 aggregate features spanning PPG, accelerometry, electrodermal activity, skin temperature, and altimetry — capturing heart rate, blood oxygen, sleep stages, and movement.

A key innovation is the Adaptive and Inherited Masking (AIM) framework, which treats the missing and fragmented data common in real-world wearable recordings as a natural signal rather than a problem, allowing the model to learn productively from incomplete inputs instead of imputing or discarding gaps. Systematic scaling experiments across four orders of magnitude in both data and model size showed near-linear, unsaturated gains when both dimensions were scaled together, with the largest model winning on 33 of 35 downstream tasks.

Evaluated across 35 health tasks spanning cardiovascular, metabolic, mental health, sleep, demographic, and lifestyle categories, SensorFM's frozen embeddings — paired with only a simple linear probe — outperformed supervised baselines built on engineered features in 34 of 35 tasks. An agentic 'classroom' of collaborating LLMs also automatically designed prediction heads by exploring over 30,000 candidate solutions, often beating simple linear probes. Finally, the model was integrated into a Personal Health Agent, demonstrating its value for generating personalized health summaries grounded in real physiological data.