Development and Validation of a Machine Learning Model for Predicting 3-Year Overall Survival After Transjugular...

Predicting overall survival (OS) in cirrhotic patients undergoing transjugular intrahepatic portosystemic shunt (TIPS) remains challenging due to the complex interdependencies of clinical variables. This study aims to develop and validate a machine learning (ML)-based predictive model using preprocedural clinical variables to improve OS prediction for cirrhotic patients undergoing TIPS.

This multicenter, retrospective study included 347 cirrhotic patients undergoing TIPS from January 2017 to December 2023. Participants were randomly divided into training (n=243) and validation (n=104) cohorts. Key clinical data, including demographic, biochemical, and procedural variables, were collected. Several ML models, including gradient boosting machine (GBM), random survival forest (RSF), and others, were trained to predict 3-year OS after TIPS. Model performance was evaluated using time-dependent receiver operating characteristic (ROC) curves, area under the curve (AUC), and Harrell C-index. Kaplan-Meier survival analysis was performed to assess risk stratification.

Key prognostic factors identified included cirrhosis etiology, hemoglobin levels, creatinine, and prothrombin time. Among the 5 models, GBM demonstrated the best performance, with higher AUCs and C-indexes in both the training and validation cohorts. RSF also showed strong predictive performance but exhibited slightly inferior calibration compared with GBM. Lasso-Cox, CoxBoost, and SurvivalSVM showed lower predictive accuracy. Kaplan-Meier survival analysis confirmed that the GBM model effectively stratified patients into high-risk and low-risk groups, with significant differences in survival probabilities (P<0.001).

The GBM-based model outperforms other models, which effectively predict OS in cirrhotic patients undergoing TIPS, enabling improved risk stratification and personalized treatment strategies.