Accurate physical activity monitoring is crucial for understanding its impact on health. However, progress in human activity recognition (HAR) algorithms has been hindered by limited labeled datasets. This study uses self-supervised learning and the UK Biobank's large accelerometer dataset (700,000 person-days) to build highly generalizable and accurate HAR models. These models outperform baselines across eight benchmark datasets, showing F1 relative improvements of 2.5–130.9% (median 24.4%). Unlike previous studies, generalization is achieved across diverse datasets, cohorts, environments, and sensor devices. The open-sourced pre-trained models are valuable for domains with limited labeled data.

Hang Yuan, Shing Chan, Andrew P. Creagh, Catherine Tong, Aidan Acquah, David A. Clifton, Aiden Doherty