Dual foundation models for accelerometry predict future health

Summary

This study asks what one week of passively collected wrist accelerometry actually predicts about a person's future health, across the whole phenome rather than one disease at a time. The authors encode UK Biobank accelerometry from 97,696 participants using two frozen self-supervised foundation models, then train a single multilabel survival model on those embeddings plus age and sex to predict 390 disease outcomes simultaneously. In 5,253 held-out participants, mean concordance across outcomes is 0.688.