Researchers utilized a large-scale unlabeled dataset from the U.K. Biobank to train deep-learning models for accurately monitoring physical activity through wearables, demonstrating enhanced accuracy and generalizability across diverse environments and devices.