Google Research has introduced SensorFM, a foundation model trained on sensor data from five million wearable device users, which can perform 35 different health prediction tasks.
When tested on new data, the model outperformed traditional supervised approaches on 34 out of 35 tasks, even with limited labeled examples, and when integrated into a health assistant, produced summaries that clinicians rated as significantly more personalized and contextually relevant than baselines.
What happened
Google Research unveiled SensorFM, a foundation model trained on more than a trillion minutes of sensor data from five million Fitbit and Pixel Watch users across over 100 countries. The model can handle 35 different health and behavioral prediction tasks, outperforming supervised baselines on 34 of them.
Why it matters
Instead of building separate AI models for each health feature (sleep detection, stress analysis, cardiovascular risk), SensorFM learns one shared representation from messy, incomplete sensor data that can adapt to many tasks with fewer labeled examples. This approach may make personalized health insights cheaper and faster to deploy across wearables.
What to watch
SensorFM is currently a research model only; Google has not announced plans to integrate it into Fitbit, Pixel Watch, or its existing Google Health Coach. The model was trained only on Fitbit and Pixel Watch data, and it works with minute-level aggregates rather than raw signals, leaving questions about generalization to other devices.
Ask the AI about this article →
Google's SensorFM addresses a longstanding inefficiency in wearable health technology: today, each health feature requires its own specialized model, trained on expensive labeled data. By training a single foundation model on unlabeled sensor data from millions of devices, the researchers show that a shared representation can outperform these siloed approaches across dozens of downstream tasks. The scale of the training dataset—more than a trillion minutes from five million users—appears to be the largest and most diverse wearable dataset used for this purpose, and the paper demonstrates that performance systematically improves as both model size and data volume increase.
The integration into a health AI agent illustrates the practical motivation: when SensorFM predictions were added as context to summaries generated by Gemini, clinicians rated them significantly higher across five dimensions (context, personalization, justifiability, relevance, and safety) than a baseline without this information. Notably, the SensorFM-augmented summaries performed nearly as well as summaries using actual known health data, suggesting the model's predictions are therapeutically useful even without ground truth.
However, SensorFM faces meaningful limitations. It was trained only on Fitbit and Pixel Watch devices, so its generalization to other wearables remains uncertain. The model operates on minute-level aggregates rather than raw signals, which means fine-grained or very short-duration patterns are lost. Many of the health outcomes studied relied on self-reports or questionnaires rather than clinical confirmation, and the study population does not fully represent the general population. For now, Google has no announced timeline or concrete product roadmap for SensorFM, keeping it firmly in the research phase.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
