A supervised machine learning approach to predict core temperature for workers with high-level personal protective equipment

This study assesses machine-learning approaches to predict core temperature (Tc) in workers wearing high-level personal protective equipment (PPE), hypothesizing improved accuracy over conventional methods. Methods Fifty participants (48 male, 2 female) performed exercise protocols while wearing high-level PPE in one of two conditions (NORM [n = 14]: 25-26°C, 45% humidity, HYPR [n = 36]: 35-36°C, 45-50% humidity). The variables collected were skin temperature (Tsk), heart rate (HR), time, respiratory rate (RR), and rate of skin temperature acquisition per minute (Tsk/min). A total of 4063 data points were utilized in two separate, complementary supervised machine learning analyses (random forest [RF] and generalized additive models [GAM]). Model performance was evaluated using both leave-one-subject-out (LOSO) cross-validation and an observation-level 80/20 split (80/20). The two models’ performance was measured using adjusted R 2 , bias, limits of agreement (LoA), root mean squared error (RMSE), mean absolute error (MAE), and standard error of the estimate (SEE). Results LOSO approach : GAM model (R 2 of 0.84, a bias of 0.00, a 95% LoA [-0.50°C, 0.50°C], an RMSE of 0.26°C, an MAE of 0.21°C, and a SEE of 0.26°C), RF model (R 2 of 0.84, a bias of 0.01, a 95% LoA [-0.50°C, 0.50°C], an RMSE of 0.26°C, an MAE of 0.20°C, and a SEE of 0.26°C). 80/20 approach : GAM model (R 2 of 0.89, a bias of 0.00, a 95% LoA [-0.41°C, 0.41°C], an RMSE of 0.21°C, an MAE of 0.17°C, and a SEE of 0.21°C), RF model (R 2 of 0.96, a bias of 0.00, a 95% LoA [-0.26°C, 0.26°C], an RMSE of 0.13°C, an MAE of 0.09°C, and a SEE of 0.21°C). Conclusion Supervised machine learning approaches can effectively predict Tc among workers wearing high-level PPE.