SensorFM: A Foundation Model for Wearable Health Data

Wearable devices generate enormous amounts of multimodal sensor data (heart rate, motion, skin conductance, temperature, etc.), but most AI models for health are narrow, task-specific, and require large amounts of labeled data. This limits scalability and generalization in real-world settings where data is often fragmented or incomplete.

Google researchers introduce SensorFM, a large sensor foundation model pre-trained self-supervised on over one trillion minutes of wearable data from more than five million consented participants across diverse devices and geographies. It learns a general-purpose representation of human physiology using a missingness-aware framework (AIM) that treats real-world data gaps as meaningful signals rather than noise.

The model demonstrates strong scaling behavior: larger models trained on more data consistently improve representation quality. Frozen SensorFM embeddings with simple linear probes outperform feature-engineered baselines on 34 of 35 diverse health prediction tasks spanning cardiovascular, metabolic, mental health, sleep, and lifestyle domains. An agentic system of LLM agents further improves performance by automatically generating and refining task-specific prediction heads.

When integrated into a Personal Health Agent, SensorFM-based predictions produce clinician-rated summaries that approach the quality of those using ground-truth measurements.

This Ledger Entry expands how readers think about foundation models in health by showing that self-supervised pre-training on massive, real-world wearable sensor data can produce a general-purpose physiological representation that generalizes across dozens of health tasks, handles fragmented data, and effectively grounds AI health agents — moving beyond narrow task-specific models toward scalable, label-efficient wearable intelligence.