Evaluation Of Wearable Activity Monitor Algorithms For Estimating Physical Activity Metrics In Adults

PURPOSE: To examine the validity of physical activity intensity (PAI) estimates derived from different processing methods for hip and wrist worn wearable devices compared to criterion measures as assessed by an extended stay in a whole-room calorimetry chamber. METHODS: Twenty-one adults (mean age 42.5 ± 18.6y, 57.1% female) had their resting metabolic rate (RMR) measured and participated in a 12-hr metabolic chamber stay while engaging in semi-structured activities. Participants wore Actigraph GT3X-BT accelerometers (ACC) on the hip (HIP) and wrist (WRIST). Metabolic energy cost (dividing min-by-min measured energy cost by RMR) data from the metabolic chamber were collapsed into 3 categories: Sedentary (SED <1.5 METs), Light (LGT >1.5 METs) and Moderate-Vigorous Physical Activity (MVPA >3.0 METs). ACC data was processed using 2 HIP and 5 WRIST published algorithms to estimate time in different PAI. Mean differences between PAI estimates and the criterion were calculated and evaluated with equivalency testing set at a 15% threshold (TOST). Confusion matrices were used to calculate overall accuracy, and F1 statistics. RESULTS: Criterion mean time spent in different PAI were 451.6 min in SED, 213.5 min in LGT, and 52.8 min in MVPA. HIP ACC methods overestimated SED (77.8 to 135.8 min), underestimated LGT (-58.5 to -139.1 min), and over- and under-estimated MVPA (3.4 to -19.3 min). No HIP ACC estimate was equivalent to the criterion, with overall percent accuracy and F1 scores for HIP ACC estimates ranging from 70.9-71.1% and F1 0.32-0.83. WRIST ACC methods over- and under-estimated time spent in SED, LGT, and MVPA by 22.4 to -267.9 min, 449.1 to -155.5 min, and 108.9 to -44.0 min, respectively. No WRIST ACC method estimate was equivalent to the criterion, with overall percent accuracy and F1 score for WRIST estimates ranging from 31.0-60.1% and F1 0.18-0.79. CONCLUSIONS: HIP ACC data processing methods to estimate different PAI were generally more accurate than WRIST ACC data processing method estimates. Substantial differences across HIP and WRIST wearable monitor estimates for time spent in SED, LGT and MVPA were noted. In accordance with best practice calibration practices, continued work refining wearable device data processing algorithms are necessary. Supported by: NIH 5R01CA215318