Comparing Step-Counting Algorithms for High-Resolution Wrist Accelerometry Data in Older Adults in the ARIC Study

Step counting from wrist accelerometry data is widely used in physical activity research and practice. While several open-source algorithms can estimate steps from high-resolution accelerometry data, there is a critical need to compare these algorithms and provide practical recommendations for their use in older adults.

A total of 1 282 Atherosclerosis Risk in Communities study participants (mean age 83.4, 60% female) wore ActiGraph GT9X wrist devices for 7 days, collecting 80-Hz triaxial accelerometry data. Five open-source step-counting algorithms (ADEPT, Oak, Step Detection Threshold [SDT], Verisense, and Stepcount) were applied to this data. Step count distributions and their cross-sectional associations with health outcomes were compared.

The estimated mean daily step counts varied widely across algorithms, ranging from 988 for ADEPT to 23 607 for SDT. Pearson correlations across methods ranged from moderate (r = 0.52) to very strong (r = 0.96). All step counts were highly associated with age, with an estimated decline of 119.0 to 142.8 steps/year (all p < .001) with comparable trends observed across demographic subgroups. After z-score standardization (subtracting the population mean and dividing by the population standard deviation), the estimated steps from each algorithm exhibited similar directionality and magnitude of association with various metabolic, cardiovascular, physical performance, and cognitive outcomes (all p < .001).

The estimated step counts algorithms are highly correlated, and, after z-scoring, have similar and highly significant associations with health outcomes. Because the total number of steps varies widely across algorithms, interpretation and translation of results for health monitoring and clinical use in older adults depend on the choice of step-counting algorithm.