ARTICLE AD BOX
Google Research just unveiled SensorFM, a powerhouse foundation model for wearable health that’s trained on more than a trillion minutes of sensor data. That data came from about five million users around the world. Instead of focusing on just one medical condition, SensorFM aims to capture the bigger picture—a unified snapshot of human physiology flexible enough for all sorts of health and wellness applications.
Most AI health models stick to one job, like spotting heart problems or managing diabetes. SensorFM is different. It takes on a bunch of prediction tasks using data straight from Fitbit and Pixel Watch devices, pulling inputs from over a hundred countries. The model looks at 34 different aggregated features, each rolling up a minute’s worth of data. It reads five sensor types, including heart rate, blood oxygen levels, skin temperature, movement, and a batch of other physiological signals.
Google put SensorFM through its paces with data from three independent clinical studies totaling nearly 14,000 participants. The results? It beat traditional supervised models—ones that rely on handpicked features—in 34 out of 35 health prediction scenarios. And when they paired SensorFM with a lightweight linear classifier, the gains got even clearer.
What stood out most was how well SensorFM picked up subtle physiological changes linked to depression and anxiety—patterns that standard wearable analytics usually miss.
The research, led by senior scientists Xin Liu and Daniel McDuff, looked at what happens when you scale things up. Turns out, a bigger model trained on a larger pool of data can make a real difference. Their largest version, SensorFM-B, drew from the full five million user dataset. It slashed reconstruction loss by 31% versus their smallest model, and downstream prediction accuracy jumped by an average of 9%.
To help developers build new health apps faster, the team also created an automated system that leans on large language model agents. This system can dream up, test, and tweak thousands of new prediction models built on SensorFM’s base. Their tests ran through over 30,000 candidate prediction heads resting on top of SensorFM’s embeddings.
SensorFM isn’t just a research showpiece—it’s working its way into AI-powered tools. Google demoed it inside a kind of Personal Health Agent. Clinicians checked AI-generated health summaries and said they stacked up well against traditional clinical assessments. In the research paper, doctors found no statistically significant differences between what the AI delivered and what came directly from clinical data.
All these findings appeared in the paper “Towards a General Intelligence and Interface for Wearable Health Data,” which popped up on arXiv in May before Google officially announced it in July.
Even with all this momentum, there are still big questions to answer—mostly around privacy, user consent, and whether the data properly represents everyone. Google mentioned their data came from users in more than 100 countries, but they didn’t break down demographics or regional splits.
Experts think models like SensorFM could be a boost for remote monitoring, preventive care, and personalized wellness. Still, they warn that systems trained mainly on consumer wearables need a lot of rigorous clinical validation before anyone should use them to diagnose or make treatment decisions.
There are regulatory hurdles, too—especially in places like India, where rules around personal health data get pretty strict under the Digital Personal Data Protection Act.
Right now, Google hasn’t said it’ll release the full SensorFM model or its code to the public. But they have hinted at more research tools on the way as the project evolves.
English (US) ·