Interpretable prediction model for short- and long-term hepatic encephalopathy and variceal rebleeding after TIPS: a...

Transjugular intrahepatic portosystemic shunt (TIPS) relieves portal hypertension, but post-TIPS overt hepatic encephalopathy (OHE) and variceal rebleeding (VRB) remain outcomes that significantly impair survival. We aimed to identify risk factors and develop an interpretable machine learning model predicting 2- and 6-year OHE and VRB.

In this multicenter retrospective study, 636 patients undergoing TIPS (2014-2023) with 2-6 years follow-up were analyzed. A detailed perioperative dataset was curated. LASSO selected key predictors for model training in an integrated machine learning voting ensemble (ML-VE) framework. Efficacy was assessed by AUC and compared with conventional clinical scores (Child-Pugh, MELD, MELD-Na). Subgroup analyses were conducted, and SHapley Additive exPlanations (SHAP) were used for interpretation.

The cumulative incidence of OHE and VRB was 22.2% and 8.2% at 2 years, increasing to 24.8% and 10.1% at 6 years, respectively. Internally, ML-VE achieved AUCs of 0.960 (OHE) and 0.966 (VRB) at 2 years, and 0.849(OHE) and 0.828(VRB) at 6 years; external validation for OHE yielded AUCs of 0.914 at 2 years and 0.924 at 6 years, outperforming conventional scores (AUC 0.528-0.563 for OHE; 0.528-0.555 for VRB). All prespecified subgroups had 2-year AUCs≥0.93 (OHE prediction in portal vein thrombosis: 0.993; VRB prediction in portal pressure gradient 9-12 mmHg: 0.987). SHAP identified "diabetes" as the strongest predictor for OHE and "recent bleeding" for VRB at 2-years.

The interpretable ML-VE model showed good discrimination for 2- and 6-year post-TIPS OHE and VRB risk, suggesting potential utility for short- and long-term risk stratification after TIPS.