Montoye Alexander H K, Pivarnik James M, Mudd Lanay M, Biswas Subir, Pfeiffer Karin A
Department of Integrative Physiology and Health Science, Alma College, United States.
Department of Kinesiology, Michigan State University, United States.
J Sci Med Sport. 2017 Nov;20(11):1003-1007. doi: 10.1016/j.jsams.2017.04.011. Epub 2017 Apr 21.
Evaluate accuracy of the activPAL and its proprietary software for prediction of time spent in physical activity (PA) intensities (sedentary, light, and moderate-to-vigorous) and energy expenditure (EE) and compare its accuracy to that of a machine learning model (ANN) developed from raw activPAL data.
Semi-structured accelerometer validation in a laboratory setting.
Participants (n=41 [20 male]; age=22.0±4.2) completed a 90-min protocol performing 13 activities for 3-10min each and choosing activity order, duration, and intensity. Participants wore an activPAL accelerometer (right thigh) and a portable metabolic analyzer. Criterion measures of time spent in sedentary, light, and moderate-to-vigorous PA were determined using measured MET values of ≤1.5, 1.6-2.9, and ≥3.0, respectively. Estimated times in each PA intensity from the activPAL software and ANN were compared with the criterion using repeated measures ANOVA. Window-by-window EE prediction was assessed using correlations and root mean square error.
activPAL software-estimated sedentary time was not different from the criterion, but light PA was overestimated (6.2min) and moderate- to vigorous PA was underestimated (4.3min). ANN-estimated sedentary time and light PA were not different from the criterion, but moderate- to vigorous PA was overestimated (1.8min). For EE estimation, the activPAL software had lower correlations (r=0.76 vs. r=0.89) and higher error (1.74 vs. 1.07 METs) than the ANN.
The ANN had higher accuracy for estimation of EE and PA than the activPAL software in this semi-structured laboratory setting, indicating potential for the ANN to be used in PA assessment.
评估activPAL及其专有软件预测身体活动（PA）强度（久坐、轻度和中度至剧烈）和能量消耗（EE）时间的准确性，并将其准确性与从原始activPAL数据开发的机器学习模型（人工神经网络，ANN）进行比较。
在实验室环境中进行半结构化加速度计验证。
参与者（n = 41 [20名男性]；年龄 = 22.0±4.2）完成了一个90分钟的方案，进行13项活动，每项活动持续3 - 10分钟，并可选择活动顺序、持续时间和强度。参与者佩戴activPAL加速度计（右大腿）和便携式代谢分析仪。使用测量的代谢当量（MET）值分别≤1.5、1.6 - 2.9和≥3.0来确定久坐、轻度和中度至剧烈PA所花费时间的标准测量值。使用重复测量方差分析将activPAL软件和人工神经网络在每个PA强度下估计的时间与标准值进行比较。逐窗口EE预测使用相关性和均方根误差进行评估。
activPAL软件估计的久坐时间与标准值无差异，但轻度PA被高估（6.2分钟），中度至剧烈PA被低估（4.3分钟）。人工神经网络估计的久坐时间和轻度PA与标准值无差异，但中度至剧烈PA被高估（1.8分钟）。对于EE估计，activPAL软件的相关性较低（r = 0.76 vs. r = 0.89），误差较高（1.74 vs. 1.07 METs），而人工神经网络则相反。
在这种半结构化实验室环境中，人工神经网络在估计EE和PA方面比activPAL软件具有更高的准确性，表明人工神经网络在PA评估中具有应用潜力。
