Machine Learning Approaches to Predicting Energy Expenditure in Preschool Children - Insights from Accelerometry, Gyroscope Data, and Cross-National Validation

Aug 5, 2026

Upload by Aurelia

Machine learning models for predicting energy expenditure in preschool children address physical inactivity as a public health crisis. Wearable sensors provide indirect estimates of energy expenditure, enabling analysis of complex behavior relationships. Across four studies, models were developed and evaluated using Canadian and German calibration datasets with accelerometers, gyroscopes, and portable metabolic units during semi-structured protocols. Results showed deep learning lowest training error, while feature-based models performed best in external validation. Frequency-based filtering, frequency-domain features, participant characteristics, and dual-sensor integration improved accuracy. After selecting optimal features, resting-period measurement and MET definitions affected sedentary and activity estimates, advancing validated EE estimation methods.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages230

Reading Time9 h 40 min

AccessFree to download

Abstract

Acknowledgements

Machine Learning Approaches to Predicting Energy Expenditure in Preschool Children: Insights from Accelerometry, Gyroscope Data, and Cross-National Validation by Hannah Joyce Coyle-Asbil A Thesis presented to The University of Guelph In partial fulfilment of requirements for the degree of Doctor of Philosophy in Human Health and Nutritional Sciences and The University of Bremen In partial fulfilment of requirements for the degree of Dr. rer. nat. in Faculty 3: Mathematics and Computer Science Guelph, Ontario, Canada © Hannah Joyce Coyle-Asbil, August, 2025 Abstract Machine Learning Approaches to Predicting Energy Expenditure in Preschool Children: Insights from Accelerometry, Gyroscope Data, and Cross-National Validation Hannah Joyce Coyle-Asbil University of Guelph, 2025 University of Bremen, 2025 Advisor(s): Lori Ann Vallis Marvin N. Wright As highlighted by the World Health Organization, physical inactivity has been recognized as a public health crisis affecting not only adults, but also children and adolescents. To address this alarming trend, it is essential to establish a reliable and robust measure of physical activity (PA) to better understand its underlying determinants. For this purpose, wearable sensors are often used, offering an indirect measure to predict/estimate the energy expenditure (EE) of PA. With the adoption of wearable sensors, numerous researchers are implementing more sophisticated machine learning approaches in their analyses that are better equipped to model complex relationships. The overarching aim of this doctoral research was to develop and refine machine learning models to predict the EE of preschool children. Across four studies, key aspects of the modeling process were explored, including model selection, preprocessing strategies, feature selection, sensor integration, the influence of metabolic equivalent (METs) definitions, and external validation. Two calibration datasets, one consisting of Canadian preschool children and the other of German preschool children, were used to develop and evaluate models using accelerometers, gyroscopes, and portable metabolic units during semi-structured activity protocols. The findings indicated that while deep learning models achieved the lowest error on the training datasets, feature-based models demonstrated superior performance in external validation. Furthermore, preprocessing techniques, specifically frequency- based filtering, and the inclusion of frequency-domain features and participant characteristics (age, sex, height, and weight) contributed to reduced prediction error. When comparing models built using gyroscope data, accelerometer data, and a combination of both, the dual-sensor models consistently outperformed single-sensor models, yielding lower error rates. Finally, after identifying the optimal feature set, the models were applied to a large cohort of Canadian children to generate and compare PA estimates based on different METs definitions. Notably, it was found that measuring the resting period, rather than estimating it using predictive approaches, resulted in higher estimates of sedentary time and lower estimates of overall PA. Collectively, this thesis advances the field of movement behaviour research by contributing validated machine learning models for estimating EE in preschool children and addressing key methodological questions relevant to this domain. Acknowledgements “The most important thing is to recognize that, along our journey, we’ll encounter certain people who are placed in our path for a reason. Our task is simply to listen and learn the lesson they’re meant to teach us.” This was the advice my grandfather gave me when I asked him how he had achieved success in his life. I have tried to live by this advice throughout my PhD, learning from every person who crossed my path. I would like to begin by sincerely thanking the participants and their families who took

FAQs

What was the main goal of the doctoral research?

To develop and refine machine learning models that predict energy expenditure in preschool children, improving methodological choices for movement behavior research.

What data sources and calibration datasets were used?

Models were built and evaluated using accelerometer and gyroscope signals, along with portable metabolic units, using Canadian and German preschool calibration datasets collected during semi-structured activity protocols.

Which modeling choices improved external validation performance?

Feature-based models performed better in external validation, and preprocessing with frequency-based filtering plus frequency-domain features and participant characteristics reduced prediction error.

Aug 5, 2026

Upload by Aurelia