Nutritional indicator-based prediction model for esophagogastric variceal bleeding in cirrhosis using XGBoost and SHAP.

Purpose: Malnutrition is common in cirrhotic patients and associated with poor prognosis, yet current models have suboptimal predictive performance for esophagogastric variceal bleeding (EGVB). This study aimed to construct a noninvasive prediction model integrating nutritional indicators. Methods: We retrospectively analyzed 416 cirrhotic patients (135 with EGVB, 281 without EGVB) at the First Hospital of Shanxi Medical University between March 2023 and March 2025. Eight nutritional screening tools were compared by univariate logistic regression and receiver operating characteristic (ROC) analysis to identify the optimal tool. An XGBoost prediction model was then constructed with permutation-importance-based feature selection. Model performance was assessed by AUC, accuracy, sensitivity, specificity, and F1-score, with SHAP for interpretation. Results: Among the eight nutritional assessment tools, univariate logistic regression analysis revealed that MUST (OR = 1.44, 95%CI: 1.22-1.71, Conclusion: The XGBoost-SHAP model integrating nutritional indicators accurately predicts EGVB risk in cirrhosis, outperforming single traditional nutritional screening tools. This explainable, noninvasive machine learning approach provides a reliable tool for risk stratification and individualized nutritional intervention, potentially improving outcomes in this high-risk population.