Google SensorFM Trains on 1 Trillion Minutes of Wearable Data

Google Research launched SensorFM on July 9, 2026, shifting wearable analytics from single-outcome algorithms to a general-purpose foundation model. The system ingests raw physiological signals to build multi-task representations of human health. For developers building health AI agents, this approach bypasses the annotation bottleneck typically required for supervised medical models.

Data Architecture and Scale

SensorFM was pre-trained on more than one trillion minutes of de-identified sensor data, totaling over two billion hours. The dataset spans five million people across all 50 U.S. states and more than 100 countries. Google sourced the data from over 20 device models, including the Pixel Watch series and Fitbit devices like the Inspire 3, Charge 6, and Versa lines.

The model processes 34 one-minute aggregate features derived from five key sensor modalities. These include photoplethysmography (PPG) for heart rate, heart-rate variability, and blood-oxygen saturation. It also ingests accelerometry for motion, electrodermal activity (EDA) for skin conductance, skin temperature, and altimetry.

Handling Fragmented Sensor Input

Wearable data is inherently noisy due to sensor removal or device charging routines. SensorFM mitigates this using an Adaptive and Inherited Masking (AIM) framework, building on earlier LSM-2 approaches.

The architecture processes 24-hour data windows and uses self-supervised reconstruction to fill in fragmented inputs. This self-supervised approach allows the model to learn deep physiological representations without requiring perfectly continuous data streams.

Task Adaptation and Clinical Validation

SensorFM generalizes to 35 health prediction tasks across cardiovascular, metabolic, sleep, mental health, lifestyle, and demographic categories. To optimize performance across these disparate domains, researchers deployed a classroom of LLM agents to autonomously search for the best predictive adapters for specific downstream tasks. The performance gains from this method scaled directly with the size of the LLM used for the optimization.

Google integrated SensorFM into a Personal Health Agent to test its real-world viability. The clinical evaluation was validated through 1,860 ratings from a cohort of clinicians. In blinded tests, clinicians could not reliably distinguish between advice generated from real lab data and advice grounded strictly in SensorFM’s wearable predictions (p = 0.396).

If you build health-oriented applications, a foundation model approach fundamentally changes your system design. You no longer need separate pipelines for sleep tracking, stress alerts, and activity recognition. A single unified inference layer can ground conversational interfaces and contextual UIs in interconnected physiological states, replacing generic health recommendations with highly specific, data-driven outputs.