Google Research and Google DeepMind, the Alphabet-owned AI research organization, have developed "SensorFM," a foundation model designed to process health data from wearable devices. The research was published on arXiv on May 28, 2026. SensorFM was pre-trained using sensor signals from 5 million participants and over 1 trillion minutes of data collected from devices such as Fitbit and Pixel Watch. The model processes multimodal signals, including heart rate, blood oxygen levels, skin temperature, and accelerometer data, to predict 35 different health indicators across 6 domains: cardiovascular, metabolic, mental health, and sleep.

The model was trained using self-supervised learning, accounting for common wearable data gaps caused by battery depletion or sensor noise. In evaluations, it outperformed traditional supervised learning models using manual feature engineering in 34 out of 35 tasks. The model demonstrates high robustness against signal loss; even with one hour of missing data, it maintained 99.7% accuracy for step count, 99.9% for deep sleep estimation, and 99.2% for light exercise tracking.

The study introduced a "classroom" exploration method utilizing LLM agents to design predictive heads for downstream tasks. Across over 30,000 experiments, agent-designed heads outperformed simple linear heads in 16 of 20 classification tasks and 12 of 15 regression tasks. When integrated with Gemini 3 Flash to generate health summaries for personal agents, clinical ratings significantly improved compared to using daily indicators alone.

SensorFM is intended for risk assessment and long-term monitoring rather than replacing clinical diagnosis. The team noted limitations, including that data is restricted to Fitbit and Pixel Watch users, which does not fully represent the general population, and that input data relies on aggregated 1-minute feature windows. Future work requires clinical validation through comparison with electronic health records and standard physiological measurements.

Google Research and Google DeepMind, the Alphabet-owned AI research organization, have developed "SensorFM," a foundation model designed to process health data from wearable devices. The research was published on arXiv on May 28, 2026. SensorFM was pre-trained using sensor signals from 5 million participants and over 1 trillion minutes of data collected from devices such as Fitbit and Pixel Watch. The model processes multimodal signals, including heart rate, blood oxygen levels, skin temperature, and accelerometer data, to predict 35 different health indicators across 6 domains: cardiovascular, metabolic, mental health, and sleep.

The model was trained using self-supervised learning, accounting for common wearable data gaps caused by battery depletion or sensor noise. In evaluations, it outperformed traditional supervised learning models using manual feature engineering in 34 out of 35 tasks. The model demonstrates high robustness against signal loss; even with one hour of missing data, it maintained 99.7% accuracy for step count, 99.9% for deep sleep estimation, and 99.2% for light exercise tracking.

The study introduced a "classroom" exploration method utilizing LLM agents to design predictive heads for downstream tasks. Across over 30,000 experiments, agent-designed heads outperformed simple linear heads in 16 of 20 classification tasks and 12 of 15 regression tasks. When integrated with Gemini 3 Flash to generate health summaries for personal agents, clinical ratings significantly improved compared to using daily indicators alone.

SensorFM is intended for risk assessment and long-term monitoring rather than replacing clinical diagnosis. The team noted limitations, including that data is restricted to Fitbit and Pixel Watch users, which does not fully represent the general population, and that input data relies on aggregated 1-minute feature windows. Future work requires clinical validation through comparison with electronic health records and standard physiological measurements.