Investigating optimal accelerometer placement for energy expenditure prediction in children using a machine learning approach

Accurate measurement of energy expenditure (EE) is imperative for identifying and targeting health-associated implications. Whilst numerous accelerometer-based regression equations to predict EE have been developed, there remains little consensus regarding optimal accelerometer placement. Therefore, the purpose of the present study was to validate and compare artificial neural networks (ANNs) developed from accelerometers worn on various anatomical positions, and combinations thereof, to predict EE. Twenty-seven children (15 boys; 10.8 +/- 1.1 years) participated in an incremental treadmill test and 30 min exergaming session wearing a portable gas analyser and nine ActiGraph GT3X+ accelerometers (chest and left and right wrists, hips, knees, and ankles). Age and sex-specific resting EE equations (Schofield) were used to estimate METs from the oxygen uptake measures. Using all the data from both exergames, incremental treadmill test and the transition period in between, ANNs were created and tested separately for each accelerometer and for combinations of two or more using a leave one- out approach to predict EE compared to measured EE. Six features (mean and variance of the three accelerometer axes) were extracted within each 15 s window as inputs in the ANN. Correlations and root mean square error (RMSE) were calculated to evaluate prediction accuracy of each ANN, and repeated measures ANOVA was used to statistically compare accuracy of the ANNs. All single-accelerometer ANNs and combinations of two-, three-, and four-accelerometers performed equally (r = 0.77-0.82), demonstrating higher correlations than the 9-accelerometer ANN (r = 0.69) or the Freedson linear regression equation (r = 0.75). RMSE did not differ between single-accelerometer ANNs or combinations of two, three, or four accelerometers (1.21-1.31 METs), demonstrating lower RMSEs than the 9-accelerometer ANN (1.46 METs) or Freedson equation (1.74 METs). These findings provide preliminary evidence that ANNs developed from single accelerometers mounted on various anatomical positions demonstrate equivalency in the accuracy to predict EE in a semi-structured setting, supporting the use of ANNs in improving EE prediction accuracy compared with linear regression.