Google SensorFM: What the Wearable Health AI Model Could Mean

Google Research introduced SensorFM as a foundation model for physiological and behavioral patterns from wearable sensor data. Instead of building separate models for sleep, stress, cardiovascular risk, or other single-purpose features, the model is designed to create a shared representation that can support many health questions. It was pretrained on more than a trillion minutes of unlabeled wearable data from five million Fitbit and Pixel Watch users.

The core promise is practical: wearable data is continuous, messy, and often incomplete, and SensorFM is built to make that data more useful across health and behavioral tasks.

SensorFM processes 34 features from five sensor-data types, including optical heart rate monitoring, acceleration, skin conductance, skin temperature, and barometric altitude. Its inputs include heart rate, heart rate variability, blood oxygen saturation, sleep stages, and motion data.

The researchers tested model variants ranging from about 100,000 to 100 million parameters, with training datasets spanning 5,000 to five million people. The report says performance improved as model size and pretraining data grew together, with the largest configuration performing best on most downstream prediction tasks.

SensorFM was evaluated on data from three separate studies with a total of 13,985 participants that the model had not seen during pretraining. The evaluation covered 35 prediction tasks across cardiovascular and metabolic health, mental health, sleep, demographics, and lifestyle.

According to the report, simple task-specific models built on SensorFM representations outperformed supervised baselines with hand-crafted wearable features on 34 of 35 tasks. The researchers also used a “classroom” of LLM agents to generate, test, and refine downstream prediction models through more than 30,000 experiments.

The researchers also tested SensorFM as added context for a personal health agent. In an experimental setup, clinician ratings found summaries augmented with SensorFM predictions scored significantly higher than the baseline across context, personalization, justifiability, relevance, and safety.

Important limits remain: the model was trained and tested only on Fitbit and Pixel Watch data, uses minute-level aggregated data rather than high-resolution raw signals, and many studied health markers came from self-reports, medication records, or questionnaires. Google has not announced concrete plans to integrate SensorFM into Fitbit, Pixel Watch, or its AI health coach.