Research Deep Dive · XiaoHu Explains

Google Research Releases SensorFM: Trained on a Trillion Minutes of Wearable Data, Wins 33 of 35 Health Tasks

5 million users, 2 billion hours of wearable data pretraining — freeze the encoder, attach a linear head, and it beats supervised-learning baselines on 34 out of 35 health tasks.
60-Second Summary
  • Google Research has released SensorFM, a health foundation model pretrained on over a trillion minutes of wearable sensor data. The corresponding paper is at arXiv:2605.22759.
  • Training data comes from 5 million consenting users, spanning 100+ countries and all 50 US states, across 20+ Fitbit and Pixel Watch device models, from September 2024 to September 2025.
  • Across scaling experiments spanning four orders of magnitude, the largest model, SensorFM-B, wins 33 of 35 health prediction tasks. Freeze the encoder and add just a linear head, and it still beats traditional feature-engineering-plus-supervised-learning on 34/35 tasks.
  • The core technique is AIM (Adaptive Inheritance Masking): it treats data the device genuinely lost the same way it treats data deliberately masked during training, so the model natively learns to handle incomplete information.
  • In a blinded clinician evaluation paired with a personal health agent, health summaries generated from SensorFM predictions scored with no statistically significant gap compared to summaries generated from real medical measurements.
⚑ This is Google Research's own official blog 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 undergone independent third-party replication.
1The Problem

Wearables Have Piled Up Data Most of It Never Really Gets Used

Over the past decade, watches and bands have logged massive amounts of heart rate, blood oxygen, sleep, and activity data. Yet most of it just sits dormant on the device, or gets used for one narrow thing: tallying today's step count, or how well you slept. Turning that data into real health risk prediction runs into three walls under the traditional approach.

Wall 01

One Task, One Model

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

Wall 02

Labels Are Expensive and Hard to Get

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

Wall 03

Individual Variation Is Huge

The same heart rate curve can be perfectly normal for one person and a red flag for another. A one-size-fits-all model can't handle that kind of person-to-person variance.

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 one unified model to read 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 wins 33 of 35 health prediction tasks; freeze the encoder and add just a linear head, and it still beats traditional feature-engineering-plus-supervised-learning on 34/35 tasks.