Methods for classifying physical activities using accelerometer data: a scoping review

Accurate classification of physical activity from accelerometer data is essential for health research and population-scale studies, yet the wide range of computational approaches has created inconsistencies in implementation, validation, and reproducibility. This scoping review identified and categorised methods used to classify physical activities from accelerometer data, emphasising implementation, simplicity, validation, and feasibility for large datasets such as the All of Us Research Programme. We searched PubMed, Web of Science, and SPORTDiscus (2015-2025) for studies using accelerometer data to classify activities or activity levels with a reported validation strategy. Of 1851 records screened, 158 met the inclusion criteria. Machine learning was most common (n = 73), followed by deep learning (n = 38), hybrid models (n = 27), rule-based methods (n = 5), and unsupervised or other novel approaches (n = 3). Walking (n = 97), sitting (n = 79), and standing (n = 68) were most frequently studied. Most studies used lab-based protocols with k-fold or leave-one-subject-out validation. Only 16 studies provided public code, and just two examined seasonality, a limitation for generalisability. Substantial variation exists in classification methods and reporting practices. Open-source tools and real-world validation remain limited. Simpler, validated, and reproducible approaches are needed for population-scale datasets such as All of Us and the UK Biobank.