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

Area of Science:

• Biomedical Engineering
• Sports Science
• Human Physiology

Background:

• Accurate energy expenditure (EE) measurement is crucial for health research.
• Existing accelerometer-based EE prediction methods lack consensus on optimal sensor placement.
• Artificial neural networks (ANNs) offer a potential advancement in EE prediction accuracy.

Purpose of the Study:

• To validate and compare ANNs for predicting EE using accelerometers at various body locations.
• To assess the accuracy of ANNs with different accelerometer combinations.
• To compare ANN performance against established linear regression equations.

Main Methods:

• Twenty-seven children underwent treadmill tests and exergaming sessions.
• Nine ActiGraph accelerometers were worn on multiple body sites (chest, wrists, hips, knees, ankles).
• ANNs were developed using accelerometer data (mean and variance of axes) and compared to measured EE via correlation and RMSE.

Main Results:

• ANNs utilizing single accelerometers showed comparable accuracy (r=0.77-0.82) to multi-accelerometer ANNs.
• Single-accelerometer ANNs outperformed a 9-accelerometer ANN (r=0.69) and the Freedson equation (r=0.75).
• Root mean square error (RMSE) was lowest for single and few-accelerometer ANNs, outperforming the 9-accelerometer ANN and Freedson equation.

Conclusions:

• ANNs developed from single accelerometers demonstrate equivalent accuracy in predicting EE in semi-structured settings.
• ANNs show improved EE prediction accuracy compared to traditional linear regression.
• These findings support the use of ANNs with strategically placed single accelerometers for reliable EE assessment.