Google Research Unveils SensorFM Health Foundation Model

Google Research introduced SensorFM, a foundation model for wearable health, pre-trained on more than 1 trillion minutes of sensor data collected from 5 million individuals. This model is designed to address the limitations of building individual models for each health endpoint, a process that becomes inefficient with a large number of outcomes. SensorFM ingests 34 one-minute aggregate features derived from five sensors: photoplethysmography (PPG), accelerometer, electrodermal activity (EDA), skin temperature, and altimeter. These features are organized into seven categories and processed within a 24-hour context window. The model's architecture utilizes a ViT-1D encoder trained with a masked-autoencoder objective and a patch size of [20, 1].

The extensive pretraining corpus, spanning from September 2024 to September 2025, encompasses data from 5,000,000 consented participants across over 100 countries, all 50 U.S. states, and more than 20 Fitbit and Pixel Watch models. This dataset totals over two billion hours, equivalent to more than one trillion minutes. Google Research offers four variants of SensorFM, each scaled with proportional data volumes: XXS, XS, S, and B. The largest variant, SensorFM-B, features 110,763,412 parameters and was trained on the full 5 million subject corpus, accumulating 2x10⁹ sensor-hours.

Evaluation of SensorFM was conducted using separate datasets from three prospective Institutional Review Board (IRB)-approved studies, covering 13,985 subjects. These studies focused on metabolic, cardiac, and respiratory health (N=1,655), sleep (N=6,377), and mental health (N=5,953). The model was tested across 35 distinct tasks, categorized into cardiovascular (6), metabolic (8), mental health (8), sleep (3), demographics (4), and lifestyle (6) domains. The research team investigated the impact of scale by comparing four model sizes against four data volumes. Results indicated that SensorFM-B, trained on the 5 million participant corpus, achieved a 31% reduction in reconstruction validation loss compared to the SensorFM-XXS variant. Generative loss also decreased by an average of 28%, demonstrating the benefits of increased scale.