Energy expenditure prediction in preschool children: a machine learning approach using accelerometry and external...

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Pediatric Health

Background:

• Accurate estimation of energy expenditure (EE) in children is crucial for understanding physical activity and metabolic health.
• Wearable accelerometers are widely used to monitor physical activity, but their accuracy in predicting EE varies.
• Previous studies have explored various machine learning models for EE prediction, with limited external validation.

Purpose of the Study:

• To develop and validate convolutional neural networks (CNNs) for predicting children's energy expenditure (EE) using raw accelerometer data.
• To compare the performance of CNN models against established models like linear regression (LM), random forest (RF), and fully connected neural networks (FcNN).
• To assess the generalizability of these models across different datasets, age groups, and accelerometer devices.

Main Methods:

• Developed CNN models using raw accelerometer data from 41 German children (3-7 years) for internal validation.
• Externally validated CNN models alongside LM, RF, and FcNN using data from 39 Canadian children (3-6 years).
• Utilized portable metabolic units for simultaneous EE measurement during semi-structured activities of varying intensities.
• Evaluated model performance using root mean square error (RMSE) values.

Main Results:

• CNN models significantly outperformed LM, FcNN, and RF models in the internal validation dataset, showing lower mean RMSE values.
• In contrast, CNN models exhibited consistently higher RMSE values compared to LM, FcNN, and RF when applied to the external validation dataset.
• The performance differences highlight potential issues with model generalizability across different populations and accelerometer hardware.

Conclusions:

• While CNNs demonstrate potential for accurate energy expenditure prediction, their generalization capabilities are limited compared to simpler models.
• Linear regression, random forest, and fully connected neural networks show better adaptability to new datasets and accelerometer types.
• Further research is needed to improve the robustness and generalizability of deep learning models for EE prediction in pediatric populations.