Interpretative machine learning for predicting 60-day mortality in burn patients with suspected infection
- Author:
Haitao Ren
1
Author Information
- Publication Type:Journal Article
- Keywords: Burns; Machine learning; Mortality prediction; Feature selection
- From: World Journal of Emergency Medicine 2026;17(3):244-249
- CountryChina
- Language:English
- Abstract: BACKGROUND: Traditional burn severity scores have limited accuracy in predicting mortality in burn patients with infection. This study aimed to develop an interpretative machine learning (ML) model to predict 60-day mortality in burn patients with suspected infection. METHODS: Data on burn patients with suspected infection were extracted from the Dryad database and divided into a training cohort (70%) and a test cohort (30%). Feature selection was conducted by combining the Boruta algorithm and least absolute shrinkage and selection operator (LASSO) regression. Twelve ML models were developed to predict 60-day mortality. Model robustness was evaluated in the training cohort, and the discrimination capacity was assessed in the test cohort. DeLong’s test was performed to compare the area under the curve (AUC) between the optimal model and the traditional scores (abbreviated burn severity index [ABSI] and revised Baux [rBaux]). SHapley Additive exPlanations (SHAP) analysis was used for model interpretation. RESULTS: A total of 1,391 adult burn patients with suspected infections were included: training cohort (n=973), test cohort (n=418). The overall mortality was 23.7% (n=329). The percentage of total body surface area (%TBSA), Acute Physiology and Chronic Health Evaluation IV (APACHE IV) score, and age were identified as significant predictors of 60-day mortality among burn patients with suspected infections. CatBoost achieved a well-balanced performance and better ability than the ABSI and rBaux did. CONCLUSION: The ML model incorporating the APACHE IV score improved the predicting performance of 60-day mortality in burn patients with infection. Its high interpretability may facilitates its clinical application for In the future.
