Background: Hematoma expansion or rebleeding after decompressive craniectomy (DC) is a critical determinant of poor prognosis in traumatic brain injury (TBI). However, reliable and interpretable tools for early risk prediction remain limited. Methods: We retrospectively analyzed a bicentric cohort (training/internal validation, n = 880; external validation, n = 302). Preoperative clinical and computed tomography variables were collected. Missing data were addressed by multiple imputation with chained equations, and features were selected through LASSO and collinearity screening. Five machine-learning algorithms (logistic regression, elastic net, support vector machine, random forest, XGBoost) were optimized using Bayesian tuning and compared. Rubin’s rules integrated performance estimates across imputations. Shapley Additive Explanations (SHAP) was employed for model interpretability, and the best-performing model was implemented as an online predictive tool. Results: The XGBoost model achieved the best discrimination and calibration pooled area under the receiver operating characteristic curve (AUC) 0.868, 95% confidence interval (CI) 0.794–0.943; area under the receiver operating characteristic curve precision–recall curves (AUPRC) 0.769 and outperformed other methods in decision-curve analysis. External validation confirmed robust generalizability (AUC 0.847, 95% CI 0.793–0.900; AUPRC 0.758). At the predefined threshold, accuracy reached 83.4%, sensitivity 75.6%, and specificity 86.6%. Eleven preoperative predictors were retained, with age, admission Glasgow Coma Scale, anticoagulant/antiplatelet use, hypertension, and basal cistern status as the most influential factors. SHAP visualizations enhanced transparency at both the population and individual levels. Conclusion: We developed and externally validated an interpretable XGBoost-based model for the early prediction of hematoma expansion or rebleeding after DC in patients with TBI. This tool offers practical clinical value for perioperative decision-making and targeted monitoring.
Ding et al. (Thu,) studied this question.