Evaluating machine learning approaches to predict the energy expenditure of cross-national preschool children: a study of preprocessing and feature effects.

Abstract

Purpose: This study aimed to examine the impact of preprocessing and inclusion of various features on predicting the energy expenditure (EE) of preschool children (3.0-6.99 years).

Methods: The internal Canadian sample consisted of 36 children, equipped with accelerometers on their wrists (OPAL) and right hip (ActiGraph GT9X). The external German sample consisted of 41 children, equipped with accelerometers on their wrists (GENEActiv) and right hip (GENEActiv; ActiGraph GT3X +). Both datasets used portable metabolic units to record EE. The effects of filtering, rectifying, adding a time delay, frequency domain (FD) features, and participant features on EE prediction across linear regression, random forest (RF), and fully connected neural network models. The Canadian sample was split into training (2/3) and validation (1/3) sets, and the German sample served as an external validation set.

Results: Consistently it was found that the RF with filtered, not rectified data with FD, participant features, and a time delay resulted in improved performance compared to approaches used previously. The models also performed similarly in the holdout sample but resulted in higher error when applied in the external validation dataset.

Conclusions:  Results attest that filtering, not rectifying, FD features and participant features result in improved model performance to predict the EE of preschool children.