606 - Utilizing Machine Learning to Predict 30-Day Mortality from Acute Variceal Hemorrhage After TIPS

Purpose: Traditional clinical scores such as MELD, Child-Pugh, and AIMS-65 rely on laboratory and demographic variables but do not incorporate imaging or procedure-specific factors that directly influence outcomes after transjugular intrahepatic portosystemic shunt (TIPS) for variceal hemorrhage. We evaluated whether a multimodal score incorporating variceal morphology, portal hemodynamics, and procedural complexity, analyzed with machine learning, can improve prediction of 30-day mortality compared with logistic regression (LR) and established clinical scores, by leveraging imaging and procedural data not captured in prior models.

Materials and Methods: We retrospectively analyzed 305 patients undergoing TIPS between August 2011 and August 2023 at a single institution, filtering for patients with acute variceal hemorrhage with known variceal location. Variables include variceal morphology, portal hemodynamics through pre-TIPS hepatic venous pressure gradient (HVPG) and portosystemic gradient changes, laboratory indices, transfusion and vasopressor use, and procedural characteristics. We trained and evaluated a spectrum of machine-learning algorithms spanning linear and non-linear function classes, including Gradient Boosting (GB), Random Forest (RF), and Support Vector Machine (SVM), and compared them with LR (control). Performance was quantified using AUC values under the ROC curve and concise operating-point metrics to reflect clinical trade-offs; model interpretability was stratified using SHAP values to identify the most clinically significant variables.

Results: Among 90 patients (11% 30-day mortality), tree-based ensembles outperformed LR. GB showed the highest discrimination (AUC ≈ 0.90), RF was next (AUC ≈ 0.84), and SVM was comparable to LR (AUC ≈ 0.75). At clinically chosen thresholds, GB produced the fewest false positives with the highest overall accuracy, while RF provided the best sensitivity and specificity; both significantly improved practical utility over LR.


Conclusion: A multimodal ML framework substantially improves 30-day mortality prediction after TIPS versus logistic regression, thereby moving risk assessment from one-size-fits-all scores toward individualized, patient-specific probabilities. By integrating imaging, hemodynamics, labs, and procedural complexity, the model delivers higher discrimination with clinically useful actionable operating points, supporting more actionable post-TIPS risk stratification and decision making that can be prospectively validated and calibrated for bedside use.