Cross-Validation and Comparison of Energy Expenditure Prediction Models Using Count-Based and Raw Accelerometer Data in Youth

Abstract

Background: Machine learning may improve energy expenditure (EE) prediction from body-worn accelerometers. However, machine learning models are rarely cross-validated in an independent sample, and the use of machine learning raises additional questions including the effect of accelerometer placement and data type (count vs. raw) for optimal EE prediction. Purpose: To assess the accuracy of artificial neural network (ANN) models for EE prediction in youth using count-based or raw data from accelerometers worn on the hip, wrist, or in combination, and compare these to count-based, EE regression equations. Methods: Data were collected in two settings; one (n = 27) to calibrate the EE prediction models, and the other (n = 34) for model cross-validation. Participants wore a portable metabolic analyzer (EE criterion) and accelerometers on the left wrist and right hip while completing 30 minutes of exergames (calibration, cross-validation) and a maximal exercise test (calibration only). Six ANNs were created from the calibration data, separately by accelerometer placement (hip, wrist, combination) and data format (count-based, raw) to predict EE (15-second epochs). Three count-based linear regression equations were also developed for comparison to the ANNs. Results: The count-based, hip ANN demonstrated lower error (RMSE: 1.2 METs) than all other ANNs (RMSE: 1.7–3.6 METs) and EE regression equations (RMSE: 1.5–3.2 METs). However, all models showed bias toward the mean. Conclusion: An ANN developed for hip-worn accelerometers had higher accuracy for EE prediction during an exergame session than wrist or combination ANNs, and ANNs developed using count-based data had higher accuracy than ANNs developed using raw data.