Abstract / Summary
Burn injury is one of the leading causes of mortality worldwide, and dynamic prediction of in-hospital mortality during the patient’s hospital course can help identify patients at risk and support improved clinical management. While variables such as ICU length of stay and overall hospital length of stay are not known at admission, our models are intended to provide risk estimates that can be updated as these data become available, rather than serving as a static admission-time score. In this study, machine learning (ML) and deep learning (DL) models were implemented to predict mortality using a burn dataset comprising six key variables, including age, burn percentage, number of hospitalization days, number of ICU stay days, infection status, and gender. Multiple supervised learning algorithms, including ensemble tree-based methods and fully connected neural networks, were trained and evaluated using standard classification and discrimination metrics. In addition, to enhance model transparency, several explainable artificial intelligence techniques, including Shapley Additive explanations (SHAP), permutation importance, and Local Interpretable Model-agnostic Explanations (LIME), were applied. These analyses enabled the assessment of the importance, nonlinear risk patterns, and patient-level predictive behavior. Among all evaluated models, XGBoost achieved the highest discriminative performance, with a receiver operating characteristic area under the curve (ROC-AUC) of 0.9896 and a precision-recall area under the curve (PR-AUC) of 0.9502, and was selected as the reference model for interpretability analyses. Several other ensemble tree-based methods, including Gradient Boosting Machine, Light Gradient Boosting Machine (LightGBM), and Random Forest, also performed strongly, with ROC-AUC values exceeding 0.98. However, these results are based on internal validation from a single center; external validation is essential before clinical application.