Predicting Metabolic Rate for Firefighting Activities with Worn Loads using a Heart Rate Sensor and Machine Learning

Monitoring the metabolic rate of occupational workers who often perform physically demanding tasks is of significance in maintaining their performance and safety. We investigate the viability of accurate metabolic rate estimation from heart rate measurements in physically demanding occupational activities, with data collected from simulated firefighter activities. Various regression methods including linear regression, tree-based methods, kernel-based methods, support vector machine (SVM), and neural networks are employed to predict breath-by-breath metabolic rates for firefighting activities under three different loading conditions: firefighting gear, gear + self-contained breathing apparatus (SCBA), and gear + SCBA + 10 lb. With both heart rate and activity types as predictors, the best-performing machine learning method (Coarse Gaussian SVM) is able to estimate metabolic rate with R2 = 0.90 and RMSE = 0.375 for activities under the two SCBA conditions, and the method is robust against differences in the subjects' heart rates and metabolic rates from cross-validation. Without activity types as predictors, the prediction accuracy is significantly lower (decreases by 34% on average). Future research to incorporate IMU sensors and/or force insoles as additional predictors for metabolic rate could eliminate the reliance on activity types, thus enhancing the generality and applicability of the method for a broader range of occupational and daily activities.