Comparison of linear and non-linear models for predicting energy expenditure from raw accelerometer data.

Area of Science:

• Biomedical Engineering
• Sports Science
• Wearable Technology

Background:

• Accurate estimation of energy expenditure (EE) is crucial for health and performance monitoring.
• Body-worn accelerometers are widely used for physical activity tracking, but their accuracy in predicting EE varies by placement and modeling technique.
• Evaluating different modeling approaches (linear vs. machine learning) and sensor locations is essential for optimizing EE prediction.

Purpose of the Study:

• To compare the accuracy of linear regression, linear mixed models, and artificial neural network (ANN) models for predicting energy expenditure (EE).
• To assess the impact of accelerometer placement (hip, thigh, wrist) on EE prediction accuracy.
• To determine the optimal modeling approach and sensor location for reliable EE estimation.

Main Methods:

• Forty participants engaged in 13 activities during a 90-minute laboratory protocol.
• Accelerometers were worn on the hip, thigh, and wrists, with a portable metabolic analyzer serving as the criterion measure for EE.
• Four EE prediction models (linear regression, linear mixed, two ANN models) were developed and compared using correlation, RMSE, and bias.

Main Results:

• Linear regression and linear mixed models showed similar EE prediction accuracy across all accelerometer placements (correlations: 0.71-0.88, RMSE: 1.11-1.61 METs).
• For thigh-worn accelerometers, linear and ANN models yielded comparable EE prediction accuracy (correlations: ~0.88-0.89, RMSE: ~1.07-1.11 METs).
• ANN models significantly outperformed linear models for wrist-worn accelerometers (higher correlations, lower RMSE), while showing a slight advantage for hip-worn accelerometers.

Conclusions:

• Artificial neural network models offer significant improvements in energy expenditure prediction accuracy when using wrist-worn accelerometers.
• Linear models provide comparable accuracy to machine learning models for hip- and thigh-worn accelerometers, presenting a viable alternative.
• Accelerometer placement and the choice of modeling technique critically influence the accuracy of energy expenditure estimation.