Trillion-minute pretraining across five million people

The model is pretrained on more than one trillion minutes of unlabeled wearable sensor signals drawn from a cohort of approximately five million participants . These data come from over 100 countries (including all 50 U.S. states) and more than 20 device types, including Fitbit and Pixel Watch .

Inputs are per-minute sensor features including: accelerometer data, heart rate, skin temperature, SpO₂, and other modalities . The pretraining objective is a self-supervised reconstruction task: the model sees 24-hour windows of data, intentionally hides portions, and learns to reconstruct the missing signals. This teaches SensorFM general patterns of human physiology and daily behavior without requiring labeled health outcomes .

35-task evaluation across six health domains

SensorFM was evaluated on 35 health prediction tasks spanning six domains: cardiovascular health, metabolic risk, sleep disorders, mental health, lifestyle choices, and demographics . In 34 of those classification tasks, SensorFM outperformed traditional methods that rely on engineered features . On generative tasks, it achieved a 28% reduction in mean squared error compared to supervised baselines .

Scaling characteristics: bigger data, better performance

The research systematically demonstrates that joint scaling of model capacity and pretraining data volume leads to consistent improvements across downstream tasks . This mirrors the scaling laws observed in language and vision models — now validated for wearable sensor data. In earlier precursor models like Google's Large Sensor Model (LSM), scaling experiments showed performance gains of up to 38% over traditional imputation methods with increased data and compute .

How SensorFM advances general-purpose AI for health monitoring

General sensor representations

Traditional health AI builds separate algorithms for each condition — one for detecting atrial fibrillation, another for monitoring sleep apnea, and so on. SensorFM instead learns a shared representation of human physiological and behavioral patterns from unlabeled data. This representation can then be adapted to specific tasks with limited labeled data — a hallmark of foundation-model generality .

Integration with LLM agents

Rather than using a single fixed prediction head for each task, SensorFM deploys autonomous LLM agents — including those grounded in Gemini — to navigate embedding spaces and perform predictive tasks . This architecture, described as an "AI classroom" of collaborating virtual agents, directly connects sensor understanding with language model reasoning .

Path to a Personal Health Agent

SensorFM does not operate in isolation. Google Research has published separately on the Personal Health Insights Agent (PHIA) — a system that uses multistep reasoning, code generation, and information retrieval to analyze wearable data and answer health questions . In evaluations against over 4,000 health-insight questions, PHIA achieved 84% accuracy on objective, numerical questions and earned 83% favorable ratings from human experts on open-ended responses .

Google has also described a broader Personal Health Agent (PHA) — a multi-agent framework that coordinates specialized sub-agents for data analysis, clinical reasoning, and behavioral coaching . SensorFM provides the foundational sensor-understanding layer that such agents would draw upon to deliver continuous, personalized health insights.

Key limitations and open questions

While SensorFM represents a significant research advance, it is important to note:

• The work is research, not a product. Google has not announced a commercial launch of SensorFM.
• The pretraining data are de-identified and drawn from IRB-approved observational studies , which means real-world deployment would require additional privacy and regulatory work.
• The model's performance on 35 tasks is impressive, but real-world clinical validation would require testing across more diverse populations and longer time horizons.
• The integration with LLM agents introduces potential for hallucination or incorrect reasoning, as with any language model, and the research does not fully characterize failure modes in clinical settings.

Summary

SensorFM scales sensor-based AI to a trillion-minute, five-million-person regime, demonstrates systematic performance gains with size, generalizes across 35 health tasks, and connects to LLM-based personal health agents. It is the latest — and largest — in a series of Google wearable foundation models that includes LSM (2025), LSM-2 (2026), and SensorLM (2026) . Together, these represent a concrete research trajectory toward a single general-purpose AI system that continuously monitors and interprets human health from wearable devices.

Trillion-minute pretraining across five million people

The model is pretrained on more than one trillion minutes of unlabeled wearable sensor signals drawn from a cohort of approximately five million participants . These data come from over 100 countries (including all 50 U.S. states) and more than 20 device types, including Fitbit and Pixel Watch .

Inputs are per-minute sensor features including: accelerometer data, heart rate, skin temperature, SpO₂, and other modalities . The pretraining objective is a self-supervised reconstruction task: the model sees 24-hour windows of data, intentionally hides portions, and learns to reconstruct the missing signals. This teaches SensorFM general patterns of human physiology and daily behavior without requiring labeled health outcomes .

35-task evaluation across six health domains

SensorFM was evaluated on 35 health prediction tasks spanning six domains: cardiovascular health, metabolic risk, sleep disorders, mental health, lifestyle choices, and demographics . In 34 of those classification tasks, SensorFM outperformed traditional methods that rely on engineered features . On generative tasks, it achieved a 28% reduction in mean squared error compared to supervised baselines .

Scaling characteristics: bigger data, better performance

The research systematically demonstrates that joint scaling of model capacity and pretraining data volume leads to consistent improvements across downstream tasks . This mirrors the scaling laws observed in language and vision models — now validated for wearable sensor data. In earlier precursor models like Google's Large Sensor Model (LSM), scaling experiments showed performance gains of up to 38% over traditional imputation methods with increased data and compute .

How SensorFM advances general-purpose AI for health monitoring

General sensor representations

Traditional health AI builds separate algorithms for each condition — one for detecting atrial fibrillation, another for monitoring sleep apnea, and so on. SensorFM instead learns a shared representation of human physiological and behavioral patterns from unlabeled data. This representation can then be adapted to specific tasks with limited labeled data — a hallmark of foundation-model generality .

Integration with LLM agents

Rather than using a single fixed prediction head for each task, SensorFM deploys autonomous LLM agents — including those grounded in Gemini — to navigate embedding spaces and perform predictive tasks . This architecture, described as an "AI classroom" of collaborating virtual agents, directly connects sensor understanding with language model reasoning .

Path to a Personal Health Agent

SensorFM does not operate in isolation. Google Research has published separately on the Personal Health Insights Agent (PHIA) — a system that uses multistep reasoning, code generation, and information retrieval to analyze wearable data and answer health questions . In evaluations against over 4,000 health-insight questions, PHIA achieved 84% accuracy on objective, numerical questions and earned 83% favorable ratings from human experts on open-ended responses .

Google has also described a broader Personal Health Agent (PHA) — a multi-agent framework that coordinates specialized sub-agents for data analysis, clinical reasoning, and behavioral coaching . SensorFM provides the foundational sensor-understanding layer that such agents would draw upon to deliver continuous, personalized health insights.

Key limitations and open questions

While SensorFM represents a significant research advance, it is important to note:

• The work is research, not a product. Google has not announced a commercial launch of SensorFM.
• The pretraining data are de-identified and drawn from IRB-approved observational studies , which means real-world deployment would require additional privacy and regulatory work.
• The model's performance on 35 tasks is impressive, but real-world clinical validation would require testing across more diverse populations and longer time horizons.
• The integration with LLM agents introduces potential for hallucination or incorrect reasoning, as with any language model, and the research does not fully characterize failure modes in clinical settings.

Summary

SensorFM scales sensor-based AI to a trillion-minute, five-million-person regime, demonstrates systematic performance gains with size, generalizes across 35 health tasks, and connects to LLM-based personal health agents. It is the latest — and largest — in a series of Google wearable foundation models that includes LSM (2025), LSM-2 (2026), and SensorLM (2026) . Together, these represent a concrete research trajectory toward a single general-purpose AI system that continuously monitors and interprets human health from wearable devices.