Accuracy in Step Counting Is Not Accuracy in Quantifying Walking Activity: A Criterion Validity Study of Wearable Devices Under Free-Living Conditions

Purpose: To determine the criterion validity of research- and consumer-grade wearables (18 step-based methods) for quantifying step counts and step-derived walking metrics in free-living conditions. Methods: Sixty healthy adults (20–80 years, balanced by sex and decade) wore six monitors for seven days: StepWatch (ankle), activPAL (thigh), ActiGraph (hip and wrist), Fitbit Inspire 2 (hip), Fitbit Charge 5 (wrist), and Garmin Vívoactive 4S (wrist), while wearing a waist-mounted video camera at their discretion during free-living sessions. Step counts from all devices were used to derive metrics of bout duration, frequency, and intensity. Raw ActiGraph data were also processed using five wrist-based machine-learning models for step detection. Video-identified steps labeled by trained researchers served as the criterion. Equivalence testing was performed using a ±10% equivalence margin. Results: Fifty-six participants provided 150 hours of valid video and 10,194 walking bouts. Only StepWatch fell within the equivalence margin for step counts with a mean absolute percent error (MAPE) <10%, with seven additional methods at 10–20%. For total walking bout duration, only activPAL and ActiGraph wrist-StepCount met equivalence criteria; activPAL achieved MAPE <10% and five others at 10–20%. No method met equivalence for the number of walking bouts, with MAPE >20% and balanced accuracy in bout detection <50%. For time spent in moderate-to-vigorous physical activity, activPAL (18.6%) and StepWatch (16.2%) achieved MAPE <20%, but no methods met equivalence. Conclusions: The level of step-count accuracy obtained from wearable devices did not translate into a comparable level of accuracy in quantifying real-world walking activity. Overall, activPAL, StepWatch, ActiGraph hip/1s and ActiGraph wrist-StepCount showed the most consistent performance. These findings can help researchers select step-based methods that more accurately capture real-world walking patterns.