Predicting Oral Temperature from Infrared Thermography and Environmental Data using Ensemble Machine Learning

Infrared thermography (IRT) is a non-contact imaging technology that enables the assessment of surface temperature patterns and has potential applications in healthcare, including temperature monitoring and physiological assessment. However, the relationship between facial skin temperature and oral temperature is influenced by multiple physiological, environmental, and measurement-related factors, making accurate prediction a challenging regression problem. This study presents a comparative evaluation of four machine learning regression algorithms—Random Forest, K-Nearest Neighbors (KNN), Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGboost) for predicting oral temperature from infrared thermography data. The study utilized a dataset containing 816 observations and 33 predictive features representing demographic characteristics, environmental conditions, and thermographic measurements obtained from facial regions. Model performance was assessed using the Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and coefficient of determination (R²). Among the evaluated approaches, the Random Forest Regressor achieved the strongest overall performance, obtaining an R² of 0.5608, an RMSE of 0.2299, and an MAE of 0.1722. LightGBM produced a comparable MAE of 0.1721 and an R² of 0.5476, whereas KNN and XGBoost achieved lower predictive performance. Feature importance analysis indicated that thermographic measurements associated with the periocular and facial regions were among the most influential predictors of oral temperature. The findings demonstrate the potential of machine learning-assisted infrared thermography for non-contact temperature estimation while also indicating that substantial unexplained variability remains. Future research should investigate systematic hyperparameter optimization, robust feature selection, external validation, larger and more diverse datasets, and clinically relevant validation protocols before deployment in real-world healthcare environments.