Optimal hyperparameters selected for each model.

IntroductionBurn injuries are a major global health challenge. Predicting mortality in burn patients using data mining can enhance therapeutic decision-making, but the long-term reliability of such models is a critical concern for safe clinical deployment. This study aimed to develop a high-performing mortality prediction model and rigorously evaluate its temporal stability.MethodsThis retrospective cohort study utilized data from 651 patients, comprising 93 predictive features, from the burn registry of the Injuries and Burn Subspecialized Teaching Hospital, affiliated with Isfahan University of Medical Sciences. After data preprocessing, including robust imputation using a Generalized Linear Model (GLM), five tree-based models were developed. Model performance was evaluated using a 10-fold stratified cross-validation and a rigorous temporal validation, with models trained on an initial nine-month cohort (n = 435) and tested on a subsequent nine-month cohort (n = 216).ResultsThe Gradient Boosted Trees (GBT) model demonstrated the best overall performance in cross-validation, achieving an accuracy of 93.1% (± 2.1%), an Area Under the ROC Curve (AUC) of 0.966 (± 0.010), and the lowest Brier Score of 0.060 (± 0.017), indicating superior discrimination and calibration. The most influential predictors were the Abbreviated Burn Severity Index (ABSI), Total Burn Surface Area (TBSA), and the percentage of third-degree burns. Crucially, the temporal validation revealed strong model stability, with the GBT model’s AUC only decreasing slightly to 0.948. However, a clinically significant drop in sensitivity was observed (from 78.1% to 68.3%).ConclusionTree-based models, particularly GBT, are powerful and accurate tools for predicting burn mortality. While our model demonstrated strong temporal stability, the observed decrease in sensitivity highlights that even robust models are subject to performance shifts over time. This underscores the vital importance of implementing a governance framework, including continuous monitoring and periodic recalibration, to ensure the sustained safety and efficacy of predictive models in clinical practice.