An Explainable Machine Learning Model for Predicting Short‐Term Haemodynamic Changes Post‐ TIPS With Prognostic Implications

BACKGROUND AND AIMS: While remeasuring portacaval pressure gradient (PPG) after transjugular intrahepatic portosystemic shunt (TIPS), it provides superior prognostic information, and its clinical utility is limited by invasiveness. We aimed to develop and validate an explainable machine learning (ML) model for predicting short-term PPG changes and improving prognostic value of PPG in cirrhotic patients undergoing TIPS. METHODS: We enrolled 328 and 128 patients (2018-2023) in retrospective training and prospective validation cohorts, respectively. PPG was measured pre-TIPS, immediately (imePPG) and 2-4 days post-TIPS (delPPG). All ML models were developed using hyperparameter tuning with 5-fold cross-validation. RESULTS: Support vector regression model (SVR_RBF) demonstrated superior performance across feature subsets, with its 6-feature model achieving optimal balance between robustness and simplicity. In the validation cohort, the final model significantly improved PPG agreement (R-squared: 0.265-0.617, mean square error: 16.64-4.90, mean absolute error: 3.45-1.65, intraclass correlation coefficient: 0.533-0.841). SHapley Additive exPlanation analysis identified imePPG reduction from baseline as the primary determinant of subsequent PPG changes. Restricted cubic spline analyses revealed significant nonlinear relationships between relative PPG reduction (but not absolute values) and 2-year further decompensation in the training cohort. Optimal reduction thresholds were defined as U-shaped curves with hazard ratio (HR) nodes at 1 (imePPG: 30%-70%, delPPG: 20%-55%), which were validated in the prospective cohort showing significant associations for both true (p = 0.003, HR = 0.45) and predicted (p = 0.044, HR = 0.57) delPPG, but not for imePPG (p = 0.196). CONCLUSIONS: The six-feature SVR_RBF model enhances the accuracy and prognostic utility of PPG measurements before TIPS completion, supporting clinical decision-making without additional invasive assessments.