Interpretable machine learning for predicting 30-day mortality following intracranial hemorrhage surgery

Objective: This study aims to utilize interpretable machine learning models based on perioperative data to forecast the 30-day mortality risk following intracranial hemorrhage (ICH) surgery. By employing SHapley Additive exPlanations (SHAP) to interpret the Extreme Gradient Boosting (XGBoost) model, we sought to identify modifiable prognostic factors to improve clinical decision-making. Methods: A retrospective analysis was conducted on perioperative data from 1,271 ICH patients. After applying exclusion criteria, 992 patients were included. The dataset was randomly partitioned into training and validation cohorts (7:3 ratio). Multiple machine learning algorithms, including logistic regression, SVM, Random Forest, and XGBoost were developed. Model performance was rigorously assessed via ROC curves, calibration curves, and decision curve analysis (DCA), with hyperparameters optimized using 5-fold cross-validation. Results: The observed 30-day postoperative mortality rate was 13%. The XGBoost model achieved an AUC of 0.931 (95% CI 0.91-0.96) in the training cohort and 0.937 (95% CI 0.90-0.97) in the validation cohort, outperforming the logistic regression model (AUC 0.669). Decision curve analysis indicated that the XGBoost model provided the greatest net benefit within a threshold probability range of 5.79 to 33.52%. SHAP analysis identified postoperative pH, lactate, APTT, and CRP as the primary predictive factors. Conclusion: This study establishes an interpretable XGBoost model that leverages perioperative data to accurately predict short-term mortality after ICH surgery. By highlighting the prognostic value of these modifiable biomarkers, the model serves as a practical tool for early risk stratification, assisting in the optimization of perioperative management in critical care settings.