SensorFM - Google launches a universal foundational model for wearable health data

What is SensorFM?

SensorFM is a large-scale sensor foundation model for wearable health data launched by Google Research. It is pre-trained on over 1 trillion minutes of unlabeled multimodal sensor data from 5 million participants to learn general physiological representations. It can be transferred to 35 health prediction tasks, including cardiovascular, metabolic, sleep, and mental health. It supports data filling and efficient label adaptation, serving as a grounding tool for personal health agents.

SensorFM's main functions

• General physiological representation learningWe extracted 34 one-minute aggregated features from five sensor modalities: PPG, accelerometer, EDA, skin temperature, and altimeter, and learned a general physiological representation across individuals and devices.
• Multi-domain health predictionIt covers 35 downstream discrimination tasks across six major categories: cardiovascular, metabolism, sleep, mental health, lifestyle, and demographics.
• Missing data fillingUsing generative pre-training capabilities, time interpolation and extrapolation are performed on missing segments in sensor data. For example, even with 60 minutes of missing data, the daily step count prediction accuracy can still be maintained at 99.7%.
• Tag-based high-efficiency adaptationIt can quickly adapt to new tasks with only a small number of labeled samples and outperform fully supervised baselines in data-scarce scenarios.
• Agentic Prediction Head DesignDeploy multiple LLM Agents to collaborate and compete, automatically search and generate the best downstream prediction header code, and explore over 30,000 candidate solutions.
• Personal health agent integrationBy embedding downstream predictors as tools into personal health agents, the context awareness, personalization, and security of LLM health responses can be improved.

The technical principle of SensorFM

• Self-supervised reconstruction pre-trainingSensorFM employs an LSM-2-based masked autoencoder framework to learn the intrinsic structure of physiological signals from trillions of minutes of unlabeled data by minimizing the reconstruction mean square error of the masked region, without relying on expensive health labeling.
• Adaptive and Inherited Masking (AIM)Treating real device omissions as equivalent to manual masks, the model learns directly from incomplete records, giving it an innate ability to perceive omissions and avoiding the waste caused by biases or data loss in traditional interpolation.
• Joint extension of data and modelsThe system verified the scaling patterns from 100K to 100M parameters and from 2 million to 2 billion sensor hours. It was found that when both scaled up simultaneously, the pre-training loss and downstream performance improved nearly linearly without saturation. The largest model, SensorFM-B, achieved the best performance on tasks 33/35.
• Linear detection evaluationWe froze the SensorFM encoder and trained a lightweight linear head solely on it to directly probe embedding quality. Results showed that it outperformed the fully supervised baseline with manually engineered features on tasks 34/35, and the dependence on demographic features gradually decreased as the model size increased, indicating that larger models implicitly learned physiologically relevant traits.
• LLM Agent Auto-OptimizationThe system constructs a "classroom"-style multi-agent system, allowing multiple LLM agents to iteratively generate, test, and optimize prediction header code. The final solution outperforms simple linear probing on 16/20 classification tasks and 12/15 regression tasks, and the stronger the agent, the better the solution it finds.

How to use SensorFM

SensorFM is currently in the technology preview stage and can be used as an academic research reference. It cannot be directly experienced or deployed yet.

SensorFM's core advantages

• Leading in scaleBased on pre-training with data from 5 million people and 1 trillion minutes, it is currently the largest and most diverse wearable dataset.
• High versatilityA single representation can simultaneously cover 35 tasks across six major health domains, eliminating the need to build a separate pipeline for each endpoint.
• Lack of perceptionThe AIM mechanism transforms missing data into learning signals, rather than noise or burden.
• Tagging efficiencyA small number of annotations can surpass the fully supervised baseline, solving the pain point of scarce annotations in the medical field.
• Automated adaptationThe LLM Agent automatically designs prediction heads, lowering the barrier to entry for downstream task development.
• Clinical validationVerified by ratings from 1,860 clinicians, the responses integrated into the personal health agent showed significant improvements in relevance and safety.

SensorFM's project address

• Project official websitehttps://research.google/blog/sensorfm-towards-a-general-intelligence-and-interface-for-wearable-health-data/
• arXiv technical paper: https://arxiv.org/pdf/2605.22759

Application scenarios of SensorFM

• Cardiovascular health monitoringBased on PPG and heart rate variability characteristics, it continuously assesses cardiovascular risks such as arrhythmias and blood pressure trends to achieve early warning.
• Metabolic risk screeningBy combining signals such as sleep, activity level, and skin temperature, metabolic risk indicators such as diabetes and obesity can be inferred.
• Sleep Disorder AnalysisBy utilizing multimodal sensor fusion, it can accurately identify sleep stages and breathing interruption patterns, assisting in the screening of disorders such as sleep apnea.
• Mental health trackingWe can extract weak signals of psychological states such as depression and anxiety from heart rate variability, skin conductance, and activity patterns to support long-term emotion monitoring.