Comparing cadence-based and machine learning based estimates for physical activity intensity classification: The UK Biobank

Abstract

ObjectivesCadence thresholds have been widely used to categorize physical activity intensity in health-related research. We examined the convergent validity of two cadence-based intensity classification approaches against a machine-learning-based intensity schema in 84,315 participants (≥40 years) with wrist-worn accelerometers. DesignValidity study. MethodsBoth cadence-based methods (one-level cadence, two-level cadence) calculated time for each intensity based on cadence thresholds while the two-level cadence identified stepping behaviors first. We used an overlapping plot, mean absolute error, and Spearman's correlation coefficient to examine agreements between the cadence-based and machine learning methods. We also evaluated agreements between methods based on practically-important-difference (moderate-to-vigorous-physical activity: ±20 min/day, moderate-physical activity: ±15, vigorous-physical activity: ±2.5, light-physical activity: ±30). ResultsThe group-level (median) moderate-to-vigorous-physical activity and moderate-physical activity minutes derived by one-level cadence were within the range of practically-important-difference compared to the machine learning method (bias of median: moderate-to-vigorous-physical activity, −3.5, interquartile range [−15.8, 12.2]; moderate-physical activity, −6.0 [−17.2, 4.1]). The group-level vigorous-physical activity and light-physical activity minutes derived by two-level cadence were within practically-important-difference range (vigorous-physical activity: −0.9 [−3.1, 0.5]; light-physical activity, −1.3 [−28.2, 28.9]). The individual-level differences between the cadence-based and machine learning methods were high across intensities (e.g., moderate-to-vigorous-physical activity: mean absolute error [one-level cadence: 24.2 min/day; two-level cadence: 26.2]), with the proportion of participants within the practically-important-difference ranging from 8.4 % to 61.6 % and the correlation ranging from 0.35 (95 % confidence interval: 0.34, 0.35) to 0.77 (0.77, 0.77). ConclusionsOne-level cadence showed acceptable group-level estimates in moderate-to-vigorous-physical activity and moderate-physical activity while two-level cadence showed acceptable group-level estimates in vigorous-physical activity and light-physical activity. The cadence-based methods (one-level cadence and two-level cadence) might not be appropriate for individual-level time estimation at each intensity category. The 60-second cadence windows may be too high to detect the short bursts of moderate-to-vigorous-physical activity.