Google AI’s Sensor FM opens a new chapter in wearable analytics by training a Large Sensor Foundation Model on over one trillion minutes of data from five million users across more than 100 countries.

The model ingests 34 one‑minute aggregate features from five sensor modalities—PPG, accelerometry, EDA, skin temperature, and altimetry—capturing heart‑rate variability, blood‑oxygen saturation, sleep stages, motion, skin conductance, and temperature. Unlike conventional self‑supervised methods, Sensor FM’s Adaptive and Inherited Masking (AIM) framework treats real‑world data gaps as natural, training directly on incomplete sequences. This missingness‑aware design eliminates bias from imputation and preserves valuable data.

Scaling experiments proved that doubling data and model size yields near‑linear gains. The largest variant, Sensor FM‑B, reduces reconstruction loss by 31% and lifts downstream AUC by 9% and Pearson R by 21% across 35 health tasks, including cardiovascular risk, metabolic syndrome, depression, and sleep quality. Linear probes on frozen embeddings beat engineered‑feature baselines on 34 of 35 tasks, while agent‑generated prediction heads further improve performance.

Practical impact follows: Sensor FM enables rapid, label‑efficient deployment of health predictors without bespoke pipelines, and its agentic head‑generation system automates model adaptation. For developers, the model offers a reusable, privacy‑respectful backbone that scales with data volume and model capacity, promising more accurate, personalized health insights from existing wearable streams.