The Effect of Well-Known Burn-related Features on Machine Learning Algorithms in Burn Patients' Mortality Prediction

Introduction: Burns are one of the most common traumas worldwide. Severely injured burn patients have an increased risk for mortality and morbidity. This study aimed to evaluate well-known risk factors for burn mortality and comparison of six Machine Learning (ML) Algorithms' predictive performances. Methods: The medical records of patients had burn injuries treated at burn treatment center were examined retrospectively. Patients’ demographics such as age and gender, Total Burned Surface Area, Inhalation Injury, Full-Thickness Burns, Burn-Types were recorded and used as input features in ML models. Patients were analyzed under two groups: Survivors and Non-Survivors. Six ML algorithms including k-Nearest Neighbor, Decision Tree, Random Forest, Support Vector Machine, Multi-Layer Perceptron, and AdaBoost, were used for predicting mortality. Several different input feature combinations were evaluated for each algorithm. Results: The number of eligible patients was 363. All six parameters that were included in ML algorithms showed a significant difference (p<0.001). The results show that AdaBoost algorithm using all input features had the best prediction performance with an Accuracy of 90% and an AUC of 92%. Conclusion: ML algorithms showed strong predictive performance in burn mortality. The development of an ML algorithm with right input features could be useful in the clinical practice. Further investigations are needed on this topic.