Simultaneous prediction of early and delayed mortality in burn patients: a comparative machine learning analysis of feature importance in a single-center retrospective study

Simultaneous prediction of early (≤120 hours) and delayed ( > 120 hours) mortality in burn patients is essential for timely interventions and effective resource allocation. Early mortality often reflects the immediate shock period following severe burn, whereas delayed mortality is commonly associated with secondary complications. This dual prediction is clinically important because it enables tailored intervention timing facilitating rapid stabilization strategies for patients at risk of early death while guiding long-term monitoring and complication management for those at risk of delayed mortality. We retrospectively analyzed 1094 burn patients from a single center, incorporating demographic, clinical, and treatment-related variables. A multi-output classification framework was developed using Random Forest, Extra Trees, and CatBoost, with preprocessing, feature selection, and normalization. Model performance was evaluated on the independent hold-out test set using Accuracy, Precision, Recall, F1 score, log-loss, AUC, specificity, and negative predictive value. On the hold-out test set, Extra Trees yielded the highest accuracy (0.806) and F1 score (0.805), while CatBoost demonstrated higher probability calibration (lowest Log-Loss = 0.453) and Random Forest maintained the highest discrimination capacity (AUC = 0.882). Feature importance analysis consistently identified Total Burn Surface Area (TBSA), Weight, Debridement, Escharotomy, Age, and Perineogenital Area Burn as the most influential predictors across models. Collectively, these results indicate that three independently trained multi-output models can robustly stratify burn patients according to early (≤120 h) and delayed ( > 120 h) mortality risk, thereby enabling time-sensitive, personalized interventions and informed clinical resource allocation.