Performance Evaluation of Device-Based Algorithms to Estimate Step Counts in Free-Living Adults Compared to Direct Observation

Purpose: Step counts are a widely used indicator of activity in studies of physical activity and health, but direct evidence of the accuracy and precision of these measures in free-living adults is limited. We evaluated the performance of six methods to estimate step counts using research-oriented devices compared to video-recorded direct observation. Methods: Twenty adults (mean ± SD age, 36.1 ± 14.7 years; 50% female) were affixed with the thigh-worn activPAL and wrist-worn ActiGraph GT3X+ devices for 7 days. The activPAL was processed using standard software and five algorithms (ActiLife, Oak, Step Detection Threshold [SDT], Verisense, and stepcount) were applied to the ActiGraph data. Participants completed two, 3-hour sessions during which they were recorded using a GoPro camera. Camera data were annotated where step counts, activity type, posture and whole-body movements were labeled. Linear mixed effects regression and equivalence testing were used to compare each algorithm’s step counts to direct observation. Results: Among the wrist algorithms, the mean absolute percent error was lowest for stepcount (17.1%) and highest for SDT (231.5%) compared to direct observation. The variance explained was moderate-to-high (R 2 = 0.64–0.90). The stepcount algorithm was similar to activPAL estimates; both were statistically equivalent to direct observation at a 15% level and provided similar estimates over a 7-day period (R 2 =0.87, MAPE=12.8%). Accuracy was highest during walking/running for all algorithms, and was lower and highly variable for biking, modified walking, and mixed movements, such as pushing a stroller, working at a screen, and food preparation. Conclusions: Step count estimates differed among algorithms, however those measured from the thigh-worn activPAL and the wrist algorithm stepcount were the most accurate.