Evaluation of Optimal Epoch Lengths for Real-Time Physical Activity Measurement for mHealth Applications: Cross-Sectional Study (Preprint)

BACKGROUND Wearable accelerometers have become integral to mobile health (mHealth) research, particularly for delivering real-time physical activity (PA) monitoring and applications in interventions such as Just-in-Time Adaptive Interventions (JITAIs). One critical yet underexplored factor in real-time PA monitoring is epoch length, which is the time interval over which raw accelerometry data are aggregated to classify activity intensities and levels. Shorter epochs (e.g., 1 second) enhance precision but increase computational and battery demands, while longer epochs (e.g., 60 seconds) reduce data burden but may miss brief activity bouts. Although previous studies have examined epoch effects using post-processed data, limited evidence exists regarding their influence on real-time, wrist-based PA estimates, especially for moderate-to-vigorous PA (MVPA). Identifying an optimal epoch length for real-time PA measurement remains a critical gap in supporting scalable and efficient mHealth interventions. OBJECTIVE This study determined the impact of varying epoch lengths on real-time MVPA estimates derived from a wrist-worn accelerometer to identify an optimal epoch that balances measurement accuracy with practical feasibility for mHealth applications. METHODS Twenty adults (Age: 32.5 ± 15.1 years) completed a series of carefully selected simulated free-living activities in a controlled laboratory setting. Participants wore the MotionSense HRV wristband, which computed real-time Euclidean Norm Minus One (ENMO) values, and a COSMED K5 indirect calorimetry for metabolic reference. ENMO values were aggregated into 5-, 10-, 15-, 30-, and 60-second epochs. MVPA was classified using validated ENMO cut-points. Epoch-level MVPA estimates were compared against the 1-second reference using mean absolute percent error (MAPE), Pearson’s correlations, Bland-Altman (BA) plots, and equivalence testing with a ±10% equivalence zone. RESULTS MVPA estimates from all epoch lengths were statistically equivalent to the 1-second standard. The 15-second epoch demonstrated the best trade-off between accuracy and efficiency, with minimal bias (0.05 min), low MAPE (6.3%), and strong correlation (r = 0.97). However, indicators of individual-level error increased with longer epochs; MAPE increased to 9.5% at 60 seconds, and the limits of agreement widened (from ± 2.9 min at 15s to ± 4.9 min at 60s), suggesting greater potential misclassifications in estimating MVPA with longer epochs. CONCLUSIONS Although MVPA estimates using the MotionSense HRV wristband were robust across all epoch lengths, findings from this study suggest that a 15-second epoch provides an optimal balance between measurement precision and processing efficiency, making it well-suited for mHealth interventions, such as JITAIs that rely on timely activity detection.