Google Research introduces SensorFM, a model trained on over a trillion minutes of data from five million Fitbit and Pixel Watch users
Google Research has released SensorFM, a foundation model designed to extract general patterns from wearable sensor data. The system processes inputs from more than one trillion minutes of unlabeled data collected across five million people. It currently supports 35 distinct health and behavioural tasks.
Current wearable features usually serve a single function. One tool tracks sleep stages, another estimates cardiovascular risk, and a third analyses stress levels. Google aims to replace these isolated approaches with a shared AI foundation. This system would make sense of continuous, often incomplete sensor data across many health questions, reduce the need for expensive labelled training data, and eventually feed personalised context into AI health assistants.
The researchers published details in a blog post and an accompanying paper. SensorFM learns a reusable representation of physiological and behavioural patterns from large volumes of unlabeled data. The training set included over 100 countries and used more than 20 different Fitbit and Pixel Watch models. The authors state this is the largest and most diverse wearable dataset ever used to train a model of this kind.
More data and larger models improve performance
SensorFM processes 34 features drawn from five sensor types: optical heart rate monitoring, acceleration, skin conductance, skin temperature, and barometric altitude. These include heart rate, heart rate variability, blood oxygen saturation, sleep stages, and motion data.
The model uses a self-supervised approach called Adaptive and Inherited Masking. This technique flags both genuinely missing values and values artificially hidden during training. The system learns to handle both types of data gaps.
Performance improves as model size and data volume grow. The four tested variants range from about 100,000 to 100 million parameters. Training datasets span from 5,000 to five million people. On the largest 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.
SensorFM outperforms comparison models on 34 out of 35 tasks
Researchers tested the model on data from three separate studies involving 13,985 participants. SensorFM had never seen this data during pretraining. The evaluation covered 35 prediction tasks spanning cardiovascular and metabolic health, mental health, sleep, demographics, and lifestyle.
Simple task-specific head models built on top of SensorFM’s learned representations beat supervised baselines using hand-crafted wearable features on 34 of 35 tasks. Scaled pretraining also made the model more label-efficient. It could adapt to new tasks with relatively few labelled examples. As the model grew larger, it relied less on extra demographic information. The authors believe scaled pretraining could be useful for hard-to-measure traits that vary widely between individuals, such as depression and anxiety symptoms.
To adapt the learned representations to new tasks, the researchers set up a classroom of competing and collaborating LLM agents. These agents repeatedly generated, tested, and refined code for downstream prediction models. They ran more than 30,000 experiments. The resulting models outperformed simple linear head models based on the same SensorFM representations on 28 of 35 prediction tasks.
Health agent answers improve with SensorFM data
The team integrated SensorFM into a personal health agent and compared three variants. All three received demographic information and daily summaries from wearable data, covering activity, sleep, blood oxygen, and skin temperature. One variant received SensorFM predictions for various health markers. A second received the actual known values for those same markers. The third received none of this extra information and served as the baseline.
Four clinicians evaluated 93 health summaries for 31 real participant profiles. They spent more than 40 hours and produced 1,860 individual ratings. Summaries augmented with SensorFM predictions scored significantly higher than the baseline across all five dimensions measured: context, personalization, justifiability, relevance, and safety. There was no statistically significant difference overall between summaries using SensorFM predictions and those using actual known health data. This does not mean SensorFM can replace clinical measurements or diagnoses.
SensorFM remains a research model
Several limitations apply. SensorFM was trained and tested only on data from Fitbit and Pixel Watch devices. Whether results transfer to other wearables is an open question. The model does not work with high-resolution raw signals but with data aggregated at the minute level. Very short or fine-grained patterns can get lost.
Many health markers studied are based on self-reports, medication records, or questionnaires rather than clinically confirmed findings. The study population does not fully represent the general population. The health agent was only evaluated in a static setup with single responses, not in longer conversations with follow-up questions.
SensorFM is purely a research model for now. Google already offers the Gemini-based Google Health Coach, which provides personalised tips on fitness, sleep, recovery, and other health topics. SensorFM could eventually serve as 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.
More details are available in the Google Research blog post and the open-access paper on arXiv.
What it means
For people using wearables, the immediate change is minimal. The technology is not yet available in consumer apps. However, the approach changes how future health summaries might be written. Instead of generic advice based on raw numbers, future summaries could include context derived from patterns the model has learned from millions of other users. This might make advice feel more tailored, though the model cannot yet replace a doctor’s diagnosis.
