Machine learning for in-hospital mortality prediction in postoperative neurosurgical ICU patients: model comparison with SHAP-based interpretation

Abstract

Postoperative mortality in neurosurgical ICUs remains high. Using the Jinhua NSICU database, we developed and internally validated machine learning models with SHAP-based interpretation to predict in-hospital mortality. This retrospective cohort study included 7,841 postoperative neurosurgical patients partitioned into training (70%, n  = 5,488) and test (30%, n  = 2,353) sets. Seven machine learning algorithms were trained on 21 LASSO-selected features without class weighting. Model selection was based on composite performance including discrimination, calibration, and precision-recall metrics. SHAP analysis was applied for feature importance ranking. Among 7,841 patients (in-hospital mortality 6.1%), XGBoost achieved test AUC 0.877 (95% CI 0.849–0.905) with acceptable calibration (Brier score 0.0424, slope 1.149). Performance was equivalent to Logistic Regression in discrimination but showed lower Brier score. SHAP analysis identified mean neutrophil count, mean glucose, and mean albumin as top predictors. False-negative cases showed relatively preserved laboratory markers, suggesting mortality potentially driven by acute neurological events. XGBoost showed favorable performance for postoperative neurosurgical ICU mortality prediction in this single-center cohort. SHAP analysis identified inflammatory, metabolic, and nutritional markers as important predictors. External validation in independent cohorts is required before clinical application. Clinical trial number: not applicable.