Prediction of Internal Body Temperature using Machine Learning Models

Context: The only validated methods for assessment of deep body temperature during exercise in the heat are invasive or logistically difficult to implement. Non-invasive prediction of deep body temperature has the potential to provide critical information to individuals who exercise in environmental extremes. Objective: To examine the use of machine learning methods for the prediction of deep body temperature using non-invasive measures. Setting: Research laboratory. Participants: Twenty-five recreationally active participants (mean±SD; male, n=19; female, n=6, age, 24±4 y; height, 177±10 cm; body mass, 75.94±12.45 kg; body fat, 15.31±6.55%). Interventions: We pooled data from two studies wherein participants walked and ran on a motorized treadmill in an environmental chamber (ambient temperature, 39.8±1.7°C; relative humidity, 33.4±10.7%). 7-site skin temperature (chest, abdomen, back, upper arm, neck, thigh and calf), heart rate, speed, incline and rectal temperature were collected regularly. Main Outcome Measures: Data were split into a 70%/30% partition for the purposes of model development and evaluation. Skin temperature, heart rate, speed, incline, environmental conditions and demographic information were selected as predictors. Multivariate linear regression, recursive partitioning, M5’ modeling and multivariate adaptive regression splines analyses were performed to develop prediction models. K-nearest neighbor and C5.0 model tree analyses were performed to develop classification models for individuals becoming hyperthermic (\u3e39°C). Results: Standard stepwise linear regression accounted for 61% of the variability in rectal temperature (SEE=0.52). A Multivariate adaptive regression spline model accounted for 77.6% of the variance in rectal temperature (RMSE=0.428). A C5.0 decision tree was able to identify cases where an individual was hyperthermic with a sensitivity of 0.625 and a specificity of 0.906. This yielded a positive likelihood ratio of 6.58. Conclusions: Machine learning techniques improved upon traditional regression analyses for the prediction of rectal temperature. Additionally, decision tree models were able to identify individuals who were hyperthermic with moderate shift in diagnostic probability. These techniques may be useful for refinement and implementation of future models to predict deep body temperature in an athletic setting