Introduction
Google Research has published details of SensorFM, a foundation model built to turn raw wearable sensor streams — heart rate, movement, skin temperature, and more — into a single, general-purpose representation of a person's physiology. Rather than training a separate model for each health outcome, Google's team trained one model on more than a trillion minutes of data from over 5 million consenting device users, then tested how well it transfers to 35 different health-prediction tasks. The results, described in an accompanying paper, suggest that a single wearable foundation model can match or beat purpose-built systems across most of them.
The Problem: One Model per Health Outcome Doesn't Scale
Wearables like Fitbit and Pixel Watch already collect continuous physiological data, but turning that data into useful health predictions has historically meant building a separate, hand-engineered model for each outcome — one for sleep quality, another for cardiovascular risk, another for stress. Google's researchers point to three recurring obstacles:
- Individual variability: resting heart rate, sleep patterns, and stress responses differ enough between people that population-wide thresholds are unreliable.
- Label scarcity: confirmed diagnoses or lab-validated measurements to train against are expensive to collect and often impossible to obtain retrospectively.
- Fragmentation: bespoke models for each outcome mean shared physiological patterns — the way heart rate, movement, and skin temperature interact — never get reused across tasks.
A foundation model approach — training one large model on broad data, then adapting it cheaply to many downstream tasks — is Google's answer to all three.
How SensorFM Works
SensorFM ingests 34 one-minute aggregate features drawn from five sensor modalities: photoplethysmography (PPG, used for heart rate), accelerometry, electrodermal activity (EDA), skin temperature, and altimetry. Training uses a missing-aware masked reconstruction method built on Google's earlier LSM-2 framework — a form of self-supervised learning where the model learns by reconstructing deliberately hidden parts of its own input.
The key design choice is how it treats gaps in the data. Wearable data is rarely complete — a device gets removed to charge, a sensor loses skin contact — and prior approaches either discarded incomplete windows or imputed the missing values, both of which can distort what the model learns. SensorFM instead treats missingness itself as a training signal, learning to predict what's absent rather than working around it.
The largest published variant, SensorFM-B, has 100 million parameters and was trained on the model's full dataset: over a trillion minutes of data from more than 5 million people across over 100 countries and all 50 U.S. states, spanning more than 20 Fitbit and Pixel Watch models, collected between September 2024 and September 2025.
Results
Google evaluated SensorFM on 35 downstream health tasks spanning cardiovascular, metabolic, sleep, mental health, demographic, and lifestyle prediction, using data from 13,985 participants.
| Metric | Result |
|---|---|
| Linear probes vs. supervised, feature-engineered baselines | Won on 34 of 35 tasks |
| SensorFM-B vs. smaller SensorFM variants | Won on 33 of 35 tasks |
| Reconstruction loss, smallest → largest variant | −31% |
| Average classification AUC, smallest → largest variant | +9% |
| Regression performance (Pearson correlation), smallest → largest variant | +21% |
| Candidate prediction heads explored by Google's LLM-agent pipeline | 30,000+ |
| Agent-designed prediction heads vs. a simple linear probe | Won 16 of 20 classification tasks, 12 of 15 regression tasks |
The model also adapted to new tasks with only a handful of labeled examples — notably on conditions that are typically hard to measure from wearable signals alone, such as depression and anxiety, where labeled training data is especially scarce.
Google additionally tested SensorFM as the sensing backbone for its Personal Health Agent, an LLM-based coaching system, comparing three ways of generating health summaries from 31 real participant profiles: demographics and wearable metrics alone, that data plus SensorFM's predictions, or that data plus actual ground-truth lab measurements. A panel of clinicians, blinded to which condition they were rating, produced 1,860 ratings across five dimensions (context, relevance, justifiability, personalization, and potential for harm) over more than 40 hours of review. Adding SensorFM's predictions significantly improved responses over the baseline on every rubric dimension, and there was no statistically significant difference between grounding the agent in SensorFM's predictions versus actual ground-truth measurements — meaning the model's inferred physiological state was, in this evaluation, about as useful to the agent as directly measured lab data.
How This Fits the Wearable AI Landscape
Most health features shipped in consumer wearables today — Fitbit's Daily Readiness Score, Apple Watch's irregular-rhythm notifications, Whoop's strain and recovery metrics — are built as separate, task-specific models, each engineered and validated for one outcome. SensorFM's pitch is architectural: one pretrained backbone, adapted cheaply to many outcomes, rather than a new model per feature. That mirrors how large language models reorganized NLP around one pretrained backbone instead of a model per task, and Google frames SensorFM as the sensor-data equivalent, building on its own prior LSM-2 pretraining work.
What the paper does not include is a head-to-head benchmark against those shipped systems, or against other published wearable foundation models — its comparisons are against supervised baselines trained on the same evaluation data, not against competitors' production models.
What This Doesn't Show Yet
The research blog post and paper describe evaluation against internal baselines and a research-stage agent, not a released product. Google has not announced a public API, a consumer-facing feature, or a timeline for shipping SensorFM into the Fitbit or Pixel Watch apps people actually use. The clinician evaluation, while substantial, assessed AI-generated health summaries rather than clinical diagnoses, not regulatory-grade outcomes.
Conclusion
SensorFM is a demonstration that a single foundation model, trained on enough wearable data, can replace a fleet of task-specific models across most of the health outcomes Google tested — from sleep and cardiovascular metrics to conditions as hard to measure as depression and anxiety. Whether that translates into features people can use on their wrist is a separate question the research doesn't yet answer, but it adds to a growing body of work — including precision medicine applications more broadly — showing that self-supervised learning at scale generalizes well beyond text and images into continuous physiological signals.
Sources
- SensorFM: Towards a general intelligence and interface for wearable health data — Google Research blog, July 9, 2026
- SensorFM paper (arXiv:2605.22759)