Estimating human metabolic rate at various activities using tri-axial accelerometers

Abstract

Accurate metabolic rate estimation is vital for predicting thermal states and regulating indoor thermal comfort. Linear regression models utilizing heart rate and accelerometers exhibit limited accuracy under low metabolic conditions, and the acquisition of heart rate data is often restricted in field experiments. This study aims to develop a non-invasive machine learning model that leverages three-axis accelerometer data and individual parameters to achieve higher accuracy in predicting short-term metabolic rates. Metabolic rates were measured in 151 healthy adults during various activities. Accelerometer data collected from the wrist, waist, and ankle, alongside individual-specific parameters, were employed. Evaluation of five machine learning algorithms—ANN, KNN, Random Forest, SVM, and XGBoost—was conducted using 10-fold cross-validation, and Bland-Altman analysis assessed prediction accuracy. The findings indicated that predictions based on the waist and ankle surpassed those from the wrist in accuracy when considering a single measurement placement. The XGBoost algorithm demonstrated the highest performance when integrating data from multiple accelerometer locations. Furthermore, incorporating individual parameters markedly improved the accuracy of metabolic rate predictions across all conditions, achieving an R² of up to 0.965 and an RMSE of 11.62 W/m². This approach addresses the inadequacies of existing prediction methods, particularly under low metabolic rate conditions. The proposed model bypassed the need for heart rate data, achieving high-precision metabolic rate predictions, and supported the application of wearable devices in health monitoring and indoor environmental control.