Finding the Goldilocks zone for toddler accelerometry: how many days are needed for a reliable estimate of physical activity using machine learning?

Abstract Accelerometers are used to measure sedentary time (SED) and physical activity (PA) in toddlers, but they may struggle to wear them for extended periods of time (e.g., weeks). Previous studies have investigated the minimum number of days needed to reliably estimate SED and PA using count-based methods. Machine learning (ML) methods use raw data which is more variable, thus potentially requiring more days for a reliable estimate. The objective of this study is to understand how many days and hours per day of accelerometer wear are needed for a reliable estimation of SED and PA using ML. Methods 109 toddlers wore an accelerometer on the right hip at home for 7 days. Time in SED, light PA (LPA), moderate-to-vigorous PA (MVPA), and total PA (TPA) were assessed using a validated ML model for toddlers. Single day intraclass coefficients (ICCs) were calculated for each minimum hours per day of wear time and each outcome. These ICCs were passed to the Spearman-Brown prophecy equation to determine the reliability of each hour per day and days combination (3-12 hours, 1-10 days). Results Predicted reliabilities ranged from 0.32 to 0.98, increasing as both numbers of hours per day and number of days increased. Discussion Our findings support the recommended 6 hours per day of wear for at least 4 days as it balances acceptable reliability with participant retention. This recommendation is valid for ML methods and we anticipate that it can be used to further explore SED and PA in toddlers using ML advances. Key Highlights This study calculates reliability of estimates of toddlers’ physical activity and sedentary time using machine learning method for a range of days (1-10) and hours per day (3-12). Our findings support the recommended 6 hours per day of wear for at least 4 days as it balances acceptable reliability with participant retention. We hope that this recommendation can be used to further explore physical activity and sedentary time in toddlers using machine learning advances.