Google: SensorFM — foundation model trained on one trillion minutes of wearable data wins on 34 of 35 health tasks

On July 9, 2026, Google Research introduced SensorFM, a foundation model for health data from wearable devices. A foundation model is a network trained on a massive unlabeled corpus that is then adapted to dozens of tasks — an approach that transformed language processing, and SensorFM applies it to sensor signals for the first time at this scale.

An unprecedented data scale

SensorFM was trained on more than one trillion (10¹²) minutes of sensor data from Fitbit and Pixel Watch devices: heart rate, heart rate variability, steps, sleep, and skin temperature. The data comes from approximately 5 million users across more than 100 countries, collected between September 2024 and September 2025. For comparison, previous academic models for wearable data were typically trained on thousands of users — SensorFM uses three orders of magnitude more.

How does the model learn from incomplete signals?

The key innovation is adaptive masking: wearable devices regularly have gaps in data (watch taken off, dead battery), so the model learns during training to reconstruct intentionally masked segments. This turns incomplete real-world data into a training signal rather than a problem. Google trained variants ranging from 100 thousand to 100 million parameters — tiny compared to language models, but sufficient for sensor patterns.

Results against specialized approaches

Across 35 health tasks — from apnea detection to fitness estimation — SensorFM outperforms specialized baseline approaches on 34 of them. Specifically: +9% AUC on classification tasks and +21% Pearson correlation on regression tasks, and clinicians in a blind test could not distinguish the model’s predictions from actual measurements.

What this means for health AI

SensorFM opens the path to AI assistants that continuously interpret biometrics without laboratory tests. The announcement comes the same day Microsoft presented Aurora 1.5 for weather — the foundation paradigm is visibly expanding beyond text to physical and biological signals.