Towards a General Intelligence for Wearable Health Data — Inside Google's SensorFM
— by Vivax
Google Research has introduced SensorFM, a foundation model for wearable health data pre-trained on more than one trillion minutes of raw sensor data from…
Hundreds of millions of people wear a smartwatch or fitness band, and every one of them generates a continuous stream of physiological signals — heart rate, movement, skin temperature, sleep. Yet for years each new health question asked of that data required its own bespoke model, hand-built features and a freshly labeled dataset. In 'Towards a General Intelligence and Interface for Wearable Health Data' (arXiv:2605.22759, published 9 July 2026, with an accompanying Google Research blog post), researchers at Google argue the field is ready for the same shift language and vision went through: one foundation model, pre-trained once at enormous scale, then adapted cheaply to many tasks. Their model is SensorFM — a foundation model that treats raw wearable sensor streams as a first-class data modality.
The scale is the headline. SensorFM is pre-trained, self-supervised, on more than one trillion minutes of wearable sensor data drawn from roughly five million people — multimodal streams learned directly from the raw signals rather than from hand-engineered summaries. As with other foundation models, the pretraining objective needs no human labels: the model learns the structure of human physiology by predicting held-out portions of the signal itself, which is exactly what makes a wearable corpus of this size usable in the first place.
What that buys is generality. The paper evaluates SensorFM across 35 health tasks — spanning cardiovascular, metabolic, sleep, activity and related endpoints — and finds that a single pre-trained backbone transfers across all of them, matching or beating task-specific baselines while adapting with far fewer labeled examples — the label-efficiency that matters most in health, where ground truth is scarce and expensive. Because the model learns the underlying dynamics of the signals, it can also infill missing stretches of data — the gaps left when a watch comes off the wrist or a battery dies — turning fragmented real-world wear into usable, continuous records.
The second half of the title — 'and Interface' — is just as deliberate. SensorFM grounds a Personal Health Agent: a conversational system that can reason over an individual's own longitudinal sensor history, answer questions about their sleep, fitness and daily patterns, and translate raw physiology into plain language. The authors are careful about scope: this is a research direction, not a medical device — clinical claims still need prospective validation, and the agent is framed as an interface for personal health understanding rather than diagnosis. But the architecture is telling: a grounded perception model underneath, a language interface on top.
The resonance with the Vivax thesis is hard to miss. A general model of health was not scraped from the open web — it was learned from the data health itself generates, at the source, the same conviction behind the Vivax Data Layer and our clinical world models. NeuroVFM showed it for hospital neuroimaging; SensorFM now shows it for the continuous physiological data people generate outside the clinic. Between the two sits the full arc of grounded medical AI: models that learn the territory of human health — from the operating room to the wrist — rather than second-hand descriptions of it. Read the paper on arXiv and the announcement on the Google Research blog.