Predicting survival of patients with deep burns using supervised machine learning algorithms

Deep burns carry a high risk of death and early reliable mortality prediction can support triage, resource allocation, and clinical decision-making. We aimed to develop and evaluate a leakage-resistant Machine Learning (ML) model to predict in-hospital mortality in patients with deep burns using only information available within the first 24 h of admission. In this retrospective, single-center study of 528 patients with deep (third-degree) burns (240 in-hospital deaths, 45.5%), twenty-four first-24-hour predictors were analyzed. Ten supervised ML algorithms were compared using nested 10 × 5 cross-validation, with all preprocessing performed within each fold to prevent leakage. A final untouched 20% hold-out set, derived from the same center and period, was used for evaluation. Class imbalance was addressed with balanced class weights, and missingness not at random was modeled explicitly. All models showed high discrimination. We ultimately selected a regularized logistic regression model, as it combined the highest cross-validated discrimination (area under the receiver-operating-characteristic curve (AUROC) 0.982 ± 0.020) with excellent calibration and full interpretability. On the hold-out set, the model remained robust, achieving an AUROC of 0.992 (95% CI: 0.980–0.999) and a Brier score of 0.04. Notably, while only five of 106 patients were misclassified, the single false negative(FN) highlights a critical need for clinical oversight to prevent missed high-risk cases. Compared to traditional scoring, our model outperformed the Baux (AUROC = 0.943; p = 0.009) and revised Baux (AUROC = 0.950; p = 0.011) scores, while performing on par with the Abbreviated Burn Severity Index (ABSI) (AUROC = 0.982; p = 0.22). Even after Bonferroni correction, the improvements over Baux scores remained statistically significant. The strongest predictor was clearly the percentage of total body surface area burned (%TBSA), followed by early organ-stress markers, inhalation injury, and diabetes. Performance was statistically significant in a permutation test and showed no evidence of bias by sex. The proposed model demonstrates high predictive performance for early mortality among deep-burn patients. Despite being an internal validation study in which burn extent was recorded in ordinal 10% bands rather than exact %TBSA, a limitation that constrains direct clinical translation, the model offers a promising framework for supporting early clinical triage and stratification, provided it undergoes rigorous external multi-center evaluation.