Wearable units file physiological and behavioral alerts that may enhance well being predictions. Whereas basis fashions are more and more used for such predictions, they’ve been primarily utilized to low-level sensor information, regardless of behavioral information usually being extra informative because of their alignment with physiologically related timescales and portions. We develop basis fashions of such behavioral alerts utilizing over 2.5B hours of wearable information from 162K people, systematically optimizing architectures and tokenization methods for this distinctive dataset. Evaluated on 57 health-related duties, our mannequin exhibits robust efficiency throughout numerous real-world purposes together with individual-level classification and time-varying well being state prediction. The mannequin excels in behavior-driven duties like sleep prediction, and improves additional when mixed with representations of uncooked sensor information. These outcomes underscore the significance of tailoring basis mannequin design to wearables and display the potential to allow new well being purposes.
- * Equal contribution
- † Work finished whereas at Apple
- ‡ College of Southern California
