Machine learning-driven risk stratification to guide variceal embolization in TIPS-treated cirrhotic patients with acute variceal bleeding.

BACKGROUND: Transjugular intrahepatic portosystemic shunt (TIPS) combined with variceal embolization (VE) is a therapeutic strategy for cirrhotic patients with acute variceal bleeding (AVB). However, the efficacy of this combination remains controversial. This study aimed to develop and validate prediction models to guide treatment decisions for TIPS combined with VE in AVB patients, focusing on predicting rebleeding risk. METHODS: This retrospective study included 1,336 cirrhotic patients with AVB undergoing TIPS with (941 cases) or without VE (395 cases) between January 2010 and June 2020. Patients were divided into training (n = 338), internal validation (n = 146), and external validation (n = 852) cohorts. Data were collected on baseline characteristics, clinical variables, and procedural details. Several machine-learning algorithms were evaluated alongside a multivariable logistic regression (LR) model, with LR demonstrating the best predictive performance. Model performance was assessed using area under the curve (AUC), sensitivity, specificity, and decision curve analysis (DCA). The SHapley Additive exPlanations (SHAP) method was used to identify key predictors of rebleeding risk. RESULTS: Rebleeding occurred in 17.74% of patients, with no significant difference between the TIPS alone and TIPS + VE groups (p = 0.946). Shunt dysfunction was more frequent in the TIPS + VE group (p < 0.001). The multivariable logistic regression model had the highest AUC for predicting rebleeding in both internal (AUC = 0.762) and external validation cohorts (AUC = 0.762). It showed a balanced sensitivity and specificity across validation cohorts. SHAP analysis identified key predictors, including creatinine, Child-Pugh class, and international normalized ratio. High-risk patients identified by the LR model benefited from TIPS + VE (p = 0.018), while low-risk patients saw no additional benefit. CONCLUSION: The LR model effectively stratified TIPS patients into high- and low-risk groups for post-procedural rebleeding, providing a useful tool to guide personalized treatment decisions for cirrhotic patients undergoing TIPS with or without VE. PRACTICE IMPLICATIONS: The LR-based model enables individualized treatment decisions by identifying AVB patients most likely to benefit from TIPS combined with variceal embolization, thereby optimizing therapeutic outcomes and avoiding unnecessary interventions.