Google Research releases SensorFM, a model for 35 health prediction tasks

On July 10, 2026, Google Research introduced SensorFM, a foundation model that predicts health outcomes from wearable-device data. It was trained on 1 trillion minutes (more than 2 billion hours) of sensor data from 5 million volunteers, and it transfers to 35 medical prediction tasks.

What is SensorFM

SensorFM (Large Sensor Foundation Model) is a specialized foundation transformer for analyzing time series from wearables. It processes 34 features, aggregated per minute, from five sensors: PPG (heart rate), accelerometer, electrodermal activity, skin temperature and altimeter. The features are organized into seven categories with a 24-hour context window.

It is built on a ViT-1D encoder (a one-dimensional Vision Transformer) trained with a masked autoencoder method. The model ships in four sizes, from XXS (138K parameters) to B (110M parameters).

• Architecture: ViT-1D with a masked autoencoder
• Input data: 34 features from 5 sensors (PPG, accelerometer, EDA, temperature, altimeter)
• Context window: 24 hours
• Four variants: XXS, XS, S, B (from 138K to 110M parameters)
• Target tasks: prediction of 35 health outcomes

What data it was trained on

Training used data from 5 million volunteers, collected between September 2024 and September 2025. The dataset spans 100+ countries, all 50 US states and more than 20 Fitbit and Pixel Watch models. The total volume exceeds 2 billion hours, or more than 1 trillion minutes.

For evaluation, the team used separate data from 13,985 subjects in three prospective studies approved by an ethics committee. They cover three areas: metabolism, the cardiovascular system and breathing.

Why it matters

Traditional health models are built for one outcome at a time. With 35 endpoints, that approach becomes inefficient: labeling is expensive and retrospective annotation is unrealistic. SensorFM uses the foundation-model paradigm: it trains on large volumes of raw, unlabeled data, then transfers to 35 specific tasks. This cuts labeling costs and improves generalization.

Wearables (Fitbit, Apple Watch, Pixel Watch) are a powerful source of real-world medical data. Lab-built models rarely work on commercial devices. SensorFM was trained on exactly those devices, which is critical for clinical deployment.

What this means

Foundation models show that data scale drives both quality and versatility. SensorFM is the first serious foundation transformer for wearable health sensors: a building block that telemedicine startups, pharma companies and clinics can build on. A trillion minutes is a nearly unreachable benchmark, and competitors are unlikely to beat it without access to comparable data ecosystems.