Thigh Accelerometer Placement Accuracy for Physical Activity and Energy Expenditure Prediction in a Semi-Structured Setting

Abstract

An inclinometer-based accelerometer (PAL) worn on the thigh, coupled with proprietary software, may provide accurate estimates of sedentary time. However, its accuracy for estimating other physical activity (PA) intensities and energy expenditure (EE) is unknown. PURPOSE: To evaluate accuracy of the PAL software for prediction of time spent in PA intensities [sedentary, light, and moderate-to-vigorous (MVPA)] and EE and compare its accuracy to that of a machine learning model (ANN) developed from raw PAL data. METHODS: Participants (n=39 [19 male]; age=22.1±4.3) completed a 90-min, semi-structured protocol in which they performed 13 activities for 3-10 min each, choosing activity order, duration, and intensity. Participants wore a PAL accelerometer on the right thigh and a portable metabolic analyzer (OXY). Time spent in sedentary, light, and MVPA was determined using MET values of ≤1.5, 1.6-2.9, and ≥3.0, respectively, calculated from OXY. Estimated times in each PA intensity from the PAL software and ANN were compared with OXY using difference scores and 95% confidence intervals; non-overlap of confidence intervals with zero indicated significant differences from OXY. Window-by-window EE prediction was assessed using correlations and root mean square error (RMSE). RESULTS: PAL software-predicted sedentary time was not different from OXY (-1.6 min; 95% CI: -3.6 - 0.4 min), but light PA was over-predicted (6.2 min; 95% CI: 4.1 - 8.3 min) and MVPA was under-predicted (-4.6 min; 95% CI: -6.5 - -2.7 min). ANN-predicted sedentary time and light PA were not different from OXY, (-0.7 min, 95% CI: -2.3 - 0.8 min; -0.9 min; 95% CI: -2.8 - 1.1 min, respectively), but MVPA was over-predicted (1.6 min; 95% CI: 0.1 - 3.2 min). For EE prediction, the PAL software had lower correlations (r=0.76 vs. r =0.89) and higher RMSE (1.74 vs. 1.07 METs) than the ANN. Under-prediction by the PAL software was more pronounced at intensities >5.0 METs. CONCLUSIONS: The PAL software distinguished between sedentary and non-sedentary activities but had high error for EE prediction, especially for higher-intensity PA. The ANN had high accuracy for prediction of sedentary and light PA and EE prediction, indicating strong potential for raw data from a thigh-worn PAL to be used in PA assessment. Supported by Blue Cross Blue Shield of Michigan Foundation.