Two Decades of Physical Activity Classification Algorithms: An Artificial Intelligence-Assisted Systematic Review by the LABDA Consortium

Accelerometer-based human activity recognition (HAR) has advanced rapidly over the past two decades, yet its capacity to generalize beyond scripted locomotion remains uncertain. Prior reviews have emphasized algorithmic performance but lacked a systematic synthesis of activity coverage, data set use, and methodological reproducibility. We conducted an artificial intelligence-assisted systematic review of accelerometer-based HAR algorithms published from 2004 to 2024, following Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines. A custom pipeline, built entirely in Python with open-source libraries, processed 1,290 full texts, extracted structured answers to 34 predefined questions, and applied a standardized activity taxonomy (Harmonized European Time Use Surveys). Eligibility was determined after extraction, ensuring inclusion/exclusion decisions were grounded in verified methodological details rather than underspecified abstracts. Of the 681 compliant studies, machine learning (41%) and deep learning (34%) dominated algorithmic contributions, with a shift toward deep learning models after 2019. Research remains concentrated on a narrow activity set: 10 Harmonized European Time Use Surveys Level 3 categories accounted for 73% of all instances, dominated by walking, sitting, and running. Nearly 60% of studies relied on external scripted data sets; for instance, the PAMAP2 data set, which was used 114 times in the literature, covered only nine subjects with 12 activities. Only 11% of publications provided open-source implementations, limiting reproducibility. HAR has achieved remarkable methodological innovation but remains focused on a limited set of activities, has high dependence on scripted data sets, and has limited code sharing. To enable real-world application, the field must invest in open, ecologically valid data sets, adopt standardized taxonomies, and require open-sourced code with clear documentation. Our artificial intelligence-assisted workflow also provides a scalable blueprint for conducting large-scale evidence synthesis.