Lundell, Sydney

Wearable sensors are increasingly used to monitor physical activity, yet low-power devices often rely on data aggregation to conserve battery life, potentially impacting measurement accuracy. This study evaluates the performance of a new low-power wearable (LPW), designed for monitoring steps across multiple months in a free-living environment, compared to a research-grade sensor (RGS) that collects raw acceleration data, with a focus on how different aggregation intervals impact step count accuracy. Thirty-two participants wore both sensors over two days, with LPW data collected in 10 min, 1 min, or 10 s aggregation periods (APs). Sensitivity and specificity of wear time detection were high across all APs (0.96 and 0.98, respectively). While total daily step count error did not differ significantly between APs, the 10 min AP exhibited greater undercounting and wider limits of agreement, especially in APs containing more than 40 steps. These findings suggest that although AP does not affect total daily step count, it influences the accuracy and variability of more granular data windows. Aggregating step counts over longer intervals may obscure short, fragmented bouts common in daily activity, leading to underestimation of steps. Optimizing APs and sensor settings is critical for improving accuracy in low-power wearables used outside laboratory settings.

With the increased push for personalized medicine, researchers and clinicians have begun exploring the use of wearable sensors to track patient activity. These sensors typically prioritize device life over robust onboard analysis, which results in lower accuracies in step count, particularly at lower cadences. To optimize the accuracy of activity-monitoring devices, particularly at slower walking speeds, proven methods must be established to identify suitable settings in a controlled and repeatable manner prior to human validation trials. Currently, there are no methods for optimizing these low-power wearable sensor settings prior to human validation, which requires manual counting for in-laboratory participants and is limited by time and the cadences that can be tested. This article proposes a novel method for determining sensor step counting accuracy prior to human validation trials by using a mechanical camshaft actuator that produces continuous steps. Sensor error was identified across a representative subspace of possible sensor setting combinations at cadences ranging from 30 steps/min to 110 steps/min. These true errors were then used to train a multivariate polynomial regression to model errors across all possible setting combinations and cadences. The resulting model predicted errors with an R2 of 0.8 and root-mean-square error (RMSE) of 0.044 across all setting combinations. An optimization algorithm was then used to determine the combinations of settings that produced the lowest RMSE and median error for three ranges of cadence that represent disabled low-mobility ambulators, disabled high-mobility ambulators, and healthy ambulators (30–60, 20–90, and 30–110 steps/min, respectively). The model identified six setting combinations for each range of interest that achieved a ±10% error in cadence prior to human validation. The anticipated range of errors from the optimized settings at lower walking speeds are lower than the reported errors of wearable sensors (±30%), suggesting that pre-human-validation optimization of sensors may decrease errors at lower cadences. This method provides a novel and efficient approach to optimizing the accuracy of wearable activity monitors prior to human validation trials.