Research Explainer · XiaoHu Explains

Google Research unveils SensorFM: trained on a trillion minutes of wearable data, wins 33 of 35 health tasks

Pretrained on 5 million people, 2 billion hours of wearable data — a frozen encoder with a single linear head beats supervised baselines on 34 of 35 health tasks.
60-Second Rundown
  • Google Research has released SensorFM, a health foundation model pretrained on over a trillion minutes of wearable sensor data, detailed in a paper at arXiv:2605.22759.
  • The training data comes from 5 million consenting users, spanning over 100 countries and all 50 U.S. states, covering 20+ Fitbit and Pixel Watch device models, from September 2024 to September 2025.
  • In scaling experiments spanning four orders of magnitude, the largest model, SensorFM-B, won 33 of 35 health prediction tasks; freezing the encoder and attaching just a linear head still beat traditional feature-engineering-plus-supervised-learning baselines on 34 of 35 tasks.
  • The core technique is AIM (Adaptive Inheritance Masking): it treats data genuinely dropped by the device and data artificially masked during training as the same problem, so the model is inherently built to handle incomplete information.
  • In a blinded clinician evaluation using a personal health agent, health summaries generated from SensorFM predictions scored no statistically significant difference from summaries generated using real medical measurements.
⚑ This is Google Research's own official blog post introducing its own model, SensorFM. The data scale, task win rates, and blind-test conclusions below all come from Google's official account and its paper, and have not yet been independently reproduced by a third party.
1Background

Wearables have piled up a decade of data, most of it never really put to use

Over the past decade, watches and wristbands have logged massive amounts of heart rate, blood oxygen, sleep, and activity data. Yet most of this data just sits dormant on the device, or gets used for one narrow purpose: counting today's steps, checking how well you slept. Getting it to actually predict health risk runs into three walls with the traditional approach.

Dilemma 01

One task, one model

Want to predict diabetes? Train a dedicated diabetes model. Want to predict depression? Train a separate depression model. Each model needs its own set of labeled data, and knowledge doesn't transfer between them.

Dilemma 02

Labels are expensive and hard to get

A health label (say, "this person has been diagnosed with diabetes") sits behind medical exams, lab work, and diagnosis — slow and costly. Many labels simply can't be retrieved retroactively; you can't send someone back three years to get a checkup redone.

Dilemma 03

Individual variation is huge

The exact same heart-rate trace can be perfectly normal in one person and a risk signal in another. A one-size-fits-all model can't handle that kind of person-to-person difference.

2Enter

Google Research's answer: SensorFM

On July 9, 2026, Google Research published a blog post introducing SensorFM, a foundation model for wearable health data, with the corresponding paper on arXiv (number 2605.22759).

It aims to use a single, unified model to make sense of the body's various physiological signals, then transfer that capability across dozens of health prediction tasks — from cardiovascular to metabolic, from sleep to mental health. It's built to tear down exactly the three walls from the previous section.
💡
This is the largest dataset ever used to train a wearable health model: over a trillion minutes, 5 million people. The largest model won 33 of 35 health prediction tasks; freezing the encoder and adding just a linear head still beats traditional feature-engineering-plus-supervised-learning on 34/35 tasks.