Google Research is testing a broader way to turn wearable sensor streams into health intelligence. Its new SensorFM model is designed to learn reusable patterns from Fitbit and Pixel Watch data, then apply those patterns across many different health and behavioral questions.
A foundation model for wearable signals
Most wearable health features are built for one narrow purpose. A watch may use one model for sleep stages, another for stress, another for cardiovascular risk, and another for metabolic markers. SensorFM takes a different approach: it tries to learn a shared representation of human physiology and behavior from continuous sensor data.
The model was pretrained on more than a trillion minutes of multimodal wearable data from five million people. The data came from Fitbit and Pixel Watch devices, spanning over 100 countries and more than 20 different Fitbit and Pixel Watch models.
According to the authors, this is the largest and most diverse wearable dataset ever used to train a model of this kind. The goal is not simply to improve one feature, but to create a general-purpose layer that can support many downstream tasks.
What SensorFM learns from
SensorFM processes 34 features drawn from five types of sensor data: optical heart rate monitoring, acceleration, skin conductance, skin temperature, and barometric altitude. Those features include heart rate, heart rate variability, blood oxygen saturation, sleep stages, motion data, and other signals.
Wearable data is often incomplete. Devices are removed, sensors miss readings, and real-world use rarely produces clean laboratory-style records. SensorFM is trained to deal with that problem directly.
The model uses a self-supervised training method in which it reconstructs deliberately hidden segments of data. The technique, called "Adaptive and Inherited Masking" (AIM), marks both truly missing values and values that were artificially masked during training. That forces the model to learn from gaps instead of treating them as an afterthought.
Google Research tested four model variants ranging from about 100,000 to 100 million parameters. Training datasets ranged from 5,000 to five million people. The researchers report that performance improved when model size and data volume grew together.
On the largest training dataset, the biggest model's reconstruction error was 31 percent lower than the smallest model's. The largest configuration also performed best on most downstream prediction tasks.
Strong results across health and behavior tasks
The researchers evaluated SensorFM on data from three separate studies with a total of 13,985 participants. This data had not been seen by the model during pretraining.
The evaluation covered 35 prediction tasks across several areas:
- Cardiovascular and metabolic health
- Mental health
- Sleep
- Demographics
- Lifestyle
Even simple task-specific models built on top of SensorFM's learned representations outperformed supervised baselines using hand-crafted wearable features on 34 of 35 tasks, according to the paper. The model also became more label-efficient with scaled pretraining, meaning it could adapt to new tasks with relatively few labeled examples.
As SensorFM grew larger, it also relied less on extra demographic information. The authors argue that this kind of scaled pretraining could be especially useful for traits that are difficult to measure and vary widely between individuals, such as depression and anxiety symptoms.
The team also used a "classroom" of competing and collaborating LLM agents to adapt SensorFM representations to new tasks. These agents generated, tested, and refined code for downstream prediction models across more than 30,000 experiments. The resulting models outperformed simple linear head models using the same SensorFM representations on 28 of 35 prediction tasks.
Better health summaries, with limits
Google Research also tested SensorFM inside a personal health agent. The researchers compared three versions of the agent. All three received demographic information and daily summaries from wearable data, including activity, sleep, blood oxygen, and skin temperature.
One version also received SensorFM predictions for various health markers. Another received the actual known values for those markers. The baseline version received no extra marker information.
Four clinicians evaluated 93 health summaries for 31 real participant profiles. They spent more than 40 hours on the review and produced 1,860 individual ratings.
Summaries augmented with SensorFM predictions scored significantly higher than the baseline across all five measured dimensions: context, personalization, justifiability, relevance, and safety. Overall, there was no statistically significant difference between summaries using SensorFM predictions and summaries using actual known health data.
That finding does not mean SensorFM can replace clinical measurements or diagnoses. It suggests that richer wearable-derived context can improve how a health agent frames a response, but the source is clear that this remains a research setup.
Why SensorFM is not a product yet
SensorFM is still a research model. Google Research notes several important limitations that shape how the results should be read.
The model was trained and tested only on Fitbit and Pixel Watch data. Whether the same approach transfers well to other wearables remains an open question.
SensorFM also does not work with high-resolution raw signals. It uses data aggregated at the minute level, which means very short or fine-grained patterns can be lost.
Many of the health markers studied by the team are based on self-reports, medication records, or questionnaires rather than clinically confirmed findings. The study population also does not fully represent the general population.
The health agent evaluation was limited as well. It tested static, single-response summaries, not longer conversations with follow-up questions.
Google already offers the Gemini-based Google Health Coach, which provides personalized tips on fitness, sleep, recovery, and other health topics. SensorFM could eventually become a technical foundation for features like these, but Google has not announced concrete plans to integrate it into Fitbit, Pixel Watch, or the AI coach.
For now, the significance of SensorFM is in the direction it points: wearable AI systems that learn from broad, messy, real-world sensor histories instead of relying only on isolated, purpose-built models.