Artificial Intelligence for Personalized Prediction of Post-TIPS Outcomes: Integrating Clinical, Biochemical, and Radiomics Data—A Narrative Review

Transjugular intrahepatic portosystemic shunt (TIPS) is an established treatment for complications of portal hypertension, but hepatic encephalopathy (HE), liver dysfunction, rebleeding, and mortality remain difficult to predict in otherwise eligible candidates. However, predicting post-TIPS outcomes remains challenging using conventional risk scores such as MELD 3.0 and Child–Turcotte–Pugh. This narrative review critically evaluates how clinical, biochemical, procedural, conventional imaging, handcrafted radiomics, and deep-learning features can be integrated for personalized post-TIPS risk prediction, supplemented by backward and forward reference checking. Thirty-two original post-TIPS studies met the core inclusion criteria: 18 focused primarily on clinical, biochemical, hemodynamic, microbiome, or procedural predictors and 14 on imaging, body composition, radiomics, or multimodal models. AI, ML, and radiomics models, by the aim of logistic regression, tree-based ensembles, support vector machines, artificial neural networks, and hybrid deep-learning models, able to capture non-linear interactions, have consistently demonstrated improved predictive performance compared with conventional scores, particularly for HE, with reported incidences of approximately 20–47%, mortality, and liver dysfunction. Their advantage lies in the ability to model complex, non-linear relationships and integrate heterogeneous data sources, including laboratory parameters, ammonia levels, hemodynamic variables, and imaging-derived features. Radiomics and deep learning approaches further enhance predictive accuracy. CT is currently the principal imaging substrate, while direct post-TIPS radiomics evidence for MRI and ultrasound remains sparse. Studies of liver and spleen morphology, portal-vein geometry, muscle and adipose tissue, and radiomic texture suggest incremental information beyond conventional scores, particularly when clinical and imaging features are combined; however, negative volumetric findings show that additional image features do not automatically improve prediction. However, most studies are retrospective, single-center, and lack external validation and no validated transformer-based or other sequence model has yet been established for post-TIPS outcomes. Standardization issues in radiomics and limited model interpretability remain significant barriers. Future directions should lead to prospective multicenter cohort validation, increasing sample sizes, harmonized imaging and endpoint definitions, locked external validation with recalibration, and the development of clinically interpretable tools to make it easier to identify those patients suitable for TIPS and their post-procedural management.