Google’s SensorFM turns 1T wearable minutes into one health model

The interesting shift in wearable health AI is not another single-purpose sleep or heart-risk classifier. It is the attempt to turn messy, interrupted sensor streams into one reusable representation of human physiology. Google Research’s SensorFM is a large step in that direction, trained on more than one trillion minutes of wearable signals from five million consented participants.

The corpus is unusually broad. Google sampled de-identified Fitbit and Pixel Watch data collected between September 2024 and September 2025, spanning more than 100 countries, all 50 U.S. states, and over 20 device models. In total, the training set covers more than two billion hours of minute-resolution signals.

SensorFM ingests 34 one-minute aggregate features across five sensor modalities: photoplethysmography, accelerometry, electrodermal activity, skin temperature, and altimetry. Those signals cover heart rate, heart-rate variability, blood oxygen, sleep stages, motion, steps, skin conductance, and temperature across a 24-hour window.

The technical bet is that missing data should be treated as part of the signal, not swept away before training. Wearable streams routinely break when devices are charging, removed, power-saving, or switching sensors. SensorFM uses Adaptive and Inherited Masking so naturally missing tokens and deliberately masked tokens are handled together during self-supervised reconstruction.

The reported scaling results are notable. SensorFM-B, the largest model, reduced reconstruction loss by 31% versus the smallest variant. On downstream tasks it delivered an average 9% AUC gain for classification and a 21% Pearson coefficient gain for regression. Across 35 health prediction tasks from three independent prospective studies with 13,985 participants, frozen SensorFM embeddings with a lightweight linear head outperformed feature-engineered supervised baselines on 34 tasks.

Google also tested an agentic “classroom” that used LLM agents to design prediction heads on top of SensorFM embeddings. After exploring more than 30,000 candidate solutions, the agent-designed heads beat a simple linear probe on 16 of 20 classification tasks and 12 of 15 regression tasks. The result is not a clinical deployment claim, but it gives wearable health research a sharper target: one foundation representation that can adapt across cardiovascular, metabolic, sleep, mental health, lifestyle, and demographic tasks.