Development and validation of machine learning models to predict esophagogastric variceal rebleeding risk in HBV-related cirrhosis after endoscopic treatment: a prospective multicenter study.

Background: Rebleeding after initial endoscopic therapy is associated with high mortality in patients with hepatitis B virus (HBV)-related liver cirrhosis complicated by esophagogastric variceal bleeding (EGVB), imposing a substantial public health burden. Spontaneous portosystemic shunts (SPSS), a compensatory mechanism for portal hypertension, are closely associated with disease progression. This study aimed to develop and validate machine learning (ML) models incorporating clinical and imaging features to predict the risk and frequency of rebleeding following initial endoscopic treatment. Methods: This multicenter prospective study enrolled patients with HBV-related cirrhosis and EGVB treated at Zhongshan Hospital, Fudan University (the development cohort). External validation was completed in five tertiary centers in China. The trial was registered at ClinicalTrials.gov, NCT03277651. Data were collected between January 2017 and January 2022. Five classic ML algorithms, Hierarchical Gradient Boosting (HGB), Multilayer Perceptron (MLP), Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting with classification trees (XGB), were utilized to predict rebleeding. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1 score. Time-dependent ML was further applied, with predictive performance compared between conventional and time-dependent models using the concordance index (C-index). The optimal model was interpreted via Shapley Additive Explanations (SHAP) and externally validated. Additionally, key predictors were integrated into a Support Vector Regression (SVR) model to estimate rebleeding frequency. Findings: Among 295 patients in the development cohort and 190 in the external cohort, rebleeding occurred in 77 and 68 patients with SPSS, respectively. The XGB model demonstrated the best discrimination (AUCs: 0.814 internal, 0.776 external), significantly outperforming the other models ( Interpretation: The ML-based model offers a noninvasive, accurate tool for individualized risk stratification and follow-up planning in patients with HBV-related cirrhosis and SPSS after initial endoscopic therapy. Funding: The work was supported by National Natural Science Foundation of China (82370622); Fujian Provincial Medical Innovation Project (2022CXB020); and Xiamen Key Medical and Health Project (3502Z20234006).