Ensemble machine learning for early mortality risk stratification in septic orthopedic trauma: an international cohort study.

Abstract

Sepsis represents a life-threatening complication in severe orthopedic trauma, significantly increasing short-term mortality risk. Despite the clinical urgency for early prognosis assessment, current predictive tools remain inadequate. To address this gap, this study used a machine learning (ML)-based framework for mortality risk stratification in this high-risk population. This retrospective cohort study established ML models to predict 30-day all-cause mortality in critically ill patients with orthopedic trauma and sepsis. Data from 2,060 eligible patients were extracted from the intensive care unit (ICU) of Beth Israel Deaconess Medical Center (2008-2019) in the United State and randomly split into training (80 %) and internal validation (20 %) sets. After handling missing data and addressing class imbalance, seven ML algorithms (including CatBoost [Categorical Boosting], RF [Random Forest], and SVM [Support Vector Machine]) were trained and optimized using 10-fold cross-validation. Model performance was assessed based on discrimination (AUC [Area Under the Curve], accuracy, F1-score), calibration (Brier score, calibration slope), and clinical utility. The top-performing models were further validated on an independent external Chinese cohort (n = 273, 2020-2024). The study cohort had a mean age of 62.8 years and a 30-day mortality rate of 19.9 % (410/2060). Non-survivors were significantly older, had a higher comorbidity burden, and more severe physiological derangements. The LASSO analysis identified 16 prognostic variables, with age, hematologic parameters (RDW, WBC), SOFA scores, hemodynamic measures (SBP), and antihypertensive therapy emerging as significant predictors. Among all models, the CatBoost algorithm demonstrated superior performance in the internal validation set, achieving the highest AUC (0.955), accuracy (0.884), and F1-score (0.878), along with excellent calibration (Brier score: 0.081). A soft voting ensemble model, integrating the top three algorithms (CatBoost, RF, SVM), was subsequently constructed. In external validation, this ensemble model generalized robustly, maintaining strong discrimination (AUC: 0.842, Accuracy: 0.737) and calibration (Brier score: 0.173), outperforming the standalone CatBoost model. SHapley Additive exPlanations analysis provided interpretable, individualized risk assessments. This study trains, optimizes, and evaluates a high-performing ML-based prediction model for 30-day mortality in patients with critical orthopedic trauma and sepsis. The CatBoost model and the soft voting ensemble, particularly the latter, demonstrates strong generalizability and clinical utility, offering a potential tool for early risk stratification and personalized management in this vulnerable population.