Improvement of Predictive Scores in Burn Medicine through Different Machine Learning Approaches

The mortality of severely burned patients can be predicted by multiple scores which have been created over the last decades. As the treatment of burn injuries and intensive care management have improved immensely over the last years, former prediction scores seem to be losing accuracy in predicting survival. Therefore, various modifications of existing scores have been established and innovative scores have been introduced. In this study, we used data from the German Burn Registry and analyzed them regarding patient mortality using different methods of machine learning. We used Classification and Regression Trees (CARTs), random forests, XGBoost, and logistic regression regarding predictive features for patient mortality. Analyzing the data of 1401 patients via machine learning, the factors of full-thickness burns, patient's age, and total burned surface area could be identified as the most important features regarding the prediction of patient mortality following burn trauma. Although the different methods identified similar aspects, application of machine learning shows that more data are necessary for a valid analysis. In the future, the usage of machine learning can contribute to the development of an innovative and precise predictive score in burn medicine and even to further interpretations of relevant data regarding different forms of outcome from the German Burn registry.