PAT: Pretrained Actigraphy Transformer
IEEE J Biomed Health Inform· 2026
Summary
PAT is a self-supervised masked autoencoder (MAE) foundation model for minute-level, week-long actigraphy sequences. Pretrained on 21,538 U.S. participants from the nationally representative NHANES cohort, PAT is the first open-source wearable-movement FM targeting psychiatric outcomes.