Benchmarking open-source step counting algorithms for wrist-worn devices

Smartwatches and fitness trackers are widely used to monitor physiology and physical activity, for example, by counting the number of steps over time. Step count has been associated with health outcomes and it is therefore important that its measurement is reliable and that the algorithms used to derive it from inertial sensors are transparent.We reproduced 7 open-source algorithms for step counting for wrist-worn devices: 3 peak detectors (Bangle simple, Espruino, Oxford) 3 periodicity detectors (Block-autocorrelation, Windowed-autocorrelation, Windowed-FFT) and one Dummy algorithm that assumes a constant step rate when movement is detected. The algorithms were benchmarked against a dataset collected from 20 healthy participants who wore the open source Bangle.js smartwatch on the wrist, and a custom inertial sensor unit on the right foot as reference while performing 4 activities: resting, low intensity activity with no walking, indoor walking on a treadmill with 3 speeds and outdoor walking over a fixed path with stops.We compared the step count computed from the accelerometry collected through the smartwatch with the one collected on the foot on 30 s segments. Results show that the most accurate algorithm is the one based on windowed autocorrelation with a 22±30% mean absolute percentage error over segments where walking is present. Interestingly, the Dummy algorithm had higher accuracy than the peak detectors, which highlights the importance of motion detection strategies in these algorithms.