Foundation Model for Wearable Health Data Trained on Trillions of Minutes
A new foundation model for wearable health data has been proposed, pretrained on over one trillion minutes of unlabeled sensor signals from five million participants. The model aims to convert low-level physiological and behavioral data into personalized health insights, addressing challenges such as high phenotypic diversity and scarcity of labeled data. The research demonstrates that scaling model capacity and pretraining data volume jointly improves performance. The work is detailed in a preprint on arXiv (2605.22759).
Key facts
- Proposes a foundation model for wearable health data
- Pretrained on more than one trillion minutes of unlabeled sensor signals
- Data drawn from a cohort of five million participants
- Aims to convert low-level sensor data into higher-level health insights
- Addresses challenges of phenotypic diversity and label scarcity
- Demonstrates joint scaling of model capacity and pretraining data volume
- Published as arXiv preprint 2605.22759
Entities
Institutions
- arXiv