Performance analyses of step-counting algorithms using wrist accelerometry

Abstract Step count is one of the most used real-world (RW) outcomes for understanding physical functioning, activity, and overall quality of life. In the current investigation, we systematically evaluated the performances of modern wrist-accelerometry-based algorithms based on peak detection, autocorrelation, template matching, movement frequency detection, and machine learning on a common dataset that included continuous walking trials of varying speeds and regularities. The accuracies were computed with respect to the ground truth step count derived using smartphone-based video recordings. On average, the movement frequency detection-based and ML-based algorithms outperformed the other algorithms showing the highest accuracies across all trials (95.3 ± 6% to 96.7 ± 6.41%). The other algorithms showed varied accuracies ranging from 59.8 ± 41% to 90.11 ± 10.3%. All algorithms showed relatively lower accuracies for 1-minute slower walks and showed relatively higher accuracies for the longest walking trials of 6-minute. Except for two algorithms (autocorrelation and template-based), all algorithms showed no significant effect of the device type (CentrePoint Insight Watch vs GT9X) as well as device placement (left wrist vs right wrist) on accuracies for all trials. The smartphone-based step detection algorithm showed the lowest accuracies and variability suggesting the need for fit-for-purpose algorithms in step count estimation using wrist accelerometry. The current investigation provides essential evidence to facilitate the application of wearable digital health technologies in clinical research and care.