Unsupervised ensemble methods for non-wear detection from triaxial accelerometer data in children following continuous monitoring

Accelerometer signals from non-wear periods are often misclassified as stationary time or sleep, potentially biasing interpretations of physical activity’s impact on health. Non-wear detection algorithms exist but their performance varies depending on the population and activity patterns. To address this, we developed and evaluated unsupervised ensemble algorithms using data from the Stanford Sleep Accelerometer Study (15 children, aged 4–18) and a simulated dataset (100 subjects) constructed from real-world pediatric data. We compared two unsupervised ensemble methods (SUMMA and BI) and a simple majority-voting ensemble against nine base classifiers derived from three algorithms: Choi, NHANES and a Hidden Markov Model (HMM). Non-wear was treated as the positive class throughout. Performance was evaluated using sensitivity, specificity, and balanced accuracy, and the variability of error across subjects. Individual algorithms exhibited performance trade-offs: Choi performed well for long stationary and sleep periods but often missed short non-wear intervals, while HMM identified short non-wear intervals but underestimated stationary time. Ensemble methods provided comparable balanced accuracy with a better balance between sensitivity and specificity. In the simulation dataset, SUMMA achieved 67.7% sensitivity, 87.1% specificity, and 77.5% balanced accuracy, compared to the best individual baseline algorithm, NHANES_VM (63.4%, 90.3%, 77.1%). Overall, the ensemble approach showed more stable performance across subjects and provides a useful alternative to relying on a single non-wear detection algorithm in settings with heterogeneous activity patterns and short non-wear periods.