Google Research released SensorFM — a model for 35 health forecasting tasks

On July 10, 2026, Google Research introduced SensorFM — a foundation model for predicting health outcomes from wearable device data. It was trained on 1 trillion minutes (over 2 billion hours) of sensor data from 5 million volunteers and transfers to 35 medical forecasting tasks.

What is SensorFM

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

At its core is a ViT-1D encoder (one-dimensional Vision Transformer) trained with masked autoencoder methods. The model comes in four scale variants, from XXS (138K parameters) to B (110M parameters).

• Architecture: ViT-1D with 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: forecasting 35 health outcomes

What data was used for training

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

Evaluation used separate data from 13,985 subjects from three prospective IRB-approved studies. They cover three domains: metabolism, cardiovascular, and respiratory.

Why this matters

Traditional health models are built one outcome at a time. With 35 endpoints, this approach becomes inefficient: labeling is expensive, retrospective annotation is infeasible. SensorFM uses a foundation model paradigm: training on large raw unlabeled data, then transfer to 35 specific tasks. This saves labeling costs and improves generalization.

Wearables (Fitbit, Apple Watch, Pixel Watch) are a powerful source of real-world health data. Lab models rarely work on commercial devices. SensorFM was trained on them directly — critical for clinical deployment.

What it means

Foundation models demonstrate that data scale is a factor of quality and universality. SensorFM is the first serious foundation transformer for wearable health sensors: a building block that telehealth startups, pharma companies, and clinics can build on. One trillion minutes is an almost unreachable benchmark that competitors are unlikely to match without access to comparable data ecosystems.