Performance Evaluation of Device-Based Algorithms to Estimate Step Counts in Free-Living Adults Compared with 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 with video-recorded direct observation.

METHODS

Twenty adults (mean ± SD age, 36.1 ± 14.7 yr; 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, Verisense, and stepcount) were applied to the ActiGraph data. Participants completed two 3-h 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 Step Detection Threshold (231.5%) compared with direct observation. The variance explained was moderate-to-high (R2 = 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 (R2 = 0.87, mean absolute percentage error = 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.