source: google research: sensorfm: towards a general intelligence and interface for wearable health data
level: research
sensorfm is a foundation model for wearable health data, pre-trained on more than one trillion minutes of multimodal sensor signals from five million consented participants. the model ingests 34 one-minute features from photoplethysmography, accelerometry, electrodermal activity, skin temperature, and altimetry, capturing heart rate, sleep, motion, and more over 24-hour windows. it uses self-supervised reconstruction with adaptive and inherited masking, which handles real-world missing data without imputation or discarding incomplete windows, making the representation missingness-aware by design.
scaling experiments show that increasing both model size and pre-training data yields predictable gains in reconstruction loss and downstream performance, with no sign of saturation. the largest model, sensorfm-b, reduces reconstruction loss by 31% over the smallest variant and improves downstream classification auc by 9% and regression pearson coefficient by 21% on average. evaluated across 35 health tasks from three prospective studies with 13,985 participants, linear probes on frozen sensorfm embeddings outperform feature-engineered baselines on 34 of 35 tasks, with particular strength in hard-to-measure conditions like depression and anxiety, and show strong label efficiency.
to automate adaptation, an agentic classroom of llms iteratively generates and refines prediction heads on sensorfm embeddings, beating simple linear probes on most tasks. sensorfm also grounds a personal health agent: clinician-rated summaries using sensorfm predictions matched those using ground-truth measurements, significantly improving over a baseline without sensorfm on all rubric dimensions. this suggests a shift toward a single generalist representation for wearable health, adaptable across cardiovascular, metabolic, sleep, and mental health domains.
why it matters: sensorfm shows that a single model trained on unlabeled wearable data can generalize across many health tasks, reducing the need for bespoke models and expensive labels, and can serve as a reliable grounding tool for ai health agents.
source: google research: sensorfm: towards a general intelligence and interface for wearable health data