Evaluating Open-Source Wrist-Worn Accelerometer Models for Sedentary Time Detection Against Thigh-Worn Accelerometer Data

Abstract Objective Wrist-worn accelerometers are common in large-scale epidemiological studies, but their ability to measure sedentary behaviour in free-living environments is unknown. We therefore aimed to evaluate the accuracy of openly-available methods to infer sedentary time from wrist-worn accelerometers. Methods We analysed data from 662 working-age adults in the SMART Work & Life study (20-70 years; mean age 45 years; 72% female) who concurrently wore wrist- and thigh-worn accelerometers for up to eight free-living days. Reference measurements of sedentary time were derived from the thigh accelerometer data using proprietary algorithms. Wrist accelerometer data were processed using widely used, publicly available activity recognition models. Performance was evaluated at 30-second epochs to generate per-participant metrics, alongside comparisons of mean daily sedentary time, mean daily number of prolonged sedentary bouts ([≥] 30 minutes) and proportion of sedentary time in prolonged bouts. Model performance was examined across subgroups defined by age, sex, body mass index, season, recruitment centre, and in sensitivity analyses restricted to daytime hours (08:00-22:00). Results The best performing machine learning model (Actinet) accurately classified sedentary time from wrist-worn accelerometer data with a mean per-participant accuracy of 0.87 and F1 score of 0.85. Cut point-based approaches demonstrated lower accuracy of 0.80 (F1 score of 0.79). The ActiNet machine learning model showed strong agreement in daily sedentary time, daily number of prolonged sedentary bouts and proportion of sedentary time in prolonged bouts, all within 10% of the free-living thigh reference. Findings were consistent across subgroups and in analyses restricted to daytime hours. Conclusion Wrist-worn accelerometers can provide accurate measurements of sedentary behaviour in free-living settings, when assessed using current machine learning models, particularly ActiNet. This work provides confidence in future epidemiological research to examine sedentary behaviour patterns from wrist-worn accelerometers and their associations with health outcomes.