Purpose: Accelerometry is commonly used to objectively measure physical activity (PA), however, differential data collection methods and analysis techniques yield dissimilar outcomes. The aims of this research were to (1) understand how accelerometer output varies among accelerometers worn on the non-dominant wrist (NDW), dominant wrist (DW), and hip; (2) develop site-specific algorithms to predict activity type classification, activity intensity classification, and estimates of metabolic intensity; and (3) compare the algorithms in a free-living setting.
Methods: Forty participants (16.8 – 64.2 yr) completed a sequence of sedentary and physical activities in a laboratory while wearing accelerometers on the NDW, DW, and hip. Participants also wore a portable metabolic analyzer to objectively measure oxygen consumption (VO2). One-second accelerometer output was compared across wear locations by activity type and intensity classifications (Aim 1). Accelerometer output data were transformed into variables related to the magnitude (ϒ), horizontal angle (φ), and inclination (θ) of acceleration, and used to develop algorithms for the NDW, DW, and hip. Random forest algorithms were developed to predict activity type classification (i.e., sedentary, lifestyle, and ambulatory) and activity intensity classification (i.e., sedentary, light, moderate, and vigorous), and regression models were built to predict VO2 (Aim 2). Following the laboratory visit, participants simultaneously wore an accelerometer at each of the three locations for three days of free-living data collection. The site-specific algorithms developed in Aim 2 were compared for equivalence (Aim 3).
Aim 1 Results: Analysis of variance indicated that accelerometer output differed between the NDW, DW, and hip for all activities completed, except for lying supine. Differences were expected; thus, Pearson correlation coefficients were calculated between the NDW, DW, and hip, and compared across activity type and intensity classifications. For activity type, the relationships between all wear locations were different for all activity types (i.e., sedentary, lifestyle PA, and ambulatory PA). For activity intensity, the relationships between the wrists were significantly different between sedentary and light activities. Additionally, relationships between the NDW, DW, and hip differed between light and moderate, and light and vigorous PA for all wear locations. The disparate correlations indicated that accelerometer signals do not just increase in magnitude as intensity increases; rather they increase differentially by wear location and activity type.
Aim 2 Results: Site-specific random forest algorithms were developed to predict activity type and intensity classification. The algorithms utilized 10-15 features of the accelerometer signal related to variability, location, and central tendency. The hip had prediction accuracies of 84.9% for activity type classification and 80.2% for activity intensity classification. The dominant wrist had activity type prediction accuracy of 83.6% and intensity prediction accuracy of 78.9%. The non-dominant wrist had prediction accuracies of 83.1% and 78.0% for activity type and intensity, respectively. The VO2 prediction algorithms had Mean Absolute Errors of 2.96 ml/kg/min for the hip, 3.34 ml/kg/min for the NDW, and 3.49 ml/kg/min for the DW. This equates to an average error of 0.93 metabolic equivalents (METs); algorithms currently used in practice yield errors of 0.89 to 2.00 METs.
Aim 3 Results: The site-specific prediction algorithms were applied to free-living data. Using the random forest algorithms, activity type classification estimates differed by 2 to 82 minutes/day, and activity intensity classification estimates differed by 0 to 83 minutes/day; however, these differences were not significantly different. The VO2 prediction models provided estimates of PA within 0 to 57 minutes/day of one another. The hip provided the lowest estimates of MVPA, while the NDW provided the highest estimates, however the VO2 estimates from all wear locations were statistically equivalent to one another.
Conclusion: The differential relationships among accelerometer outputs from the NDW, DW, and hip indicate that output differs based on activity type and intensity. This non-systematic error prevents scaling or comparing data collected at different wear locations and supports the need for site-specific analysis methods. Site-specific prediction algorithms provided comparable to improved performance over currently-utilized analysis methods in PA research, and the PA estimates were equivalent across wear locations. This research provides a more nuanced understanding of the impact of wear location on accelerometer output and alternative methods for analysis. Importantly, the algorithms created allow for comparisons to be made among data collected at the NDW, DW, and hip, which has not previously been possible.