Enhancing post-TIPS hepatic encephalopathy risk stratification: a hybrid TabPFN model leveraging radiomics, deep transfer learning features, and MELD score.
Authors
Affiliations (4)
Affiliations (4)
- Department of Interventional Therapy, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, 17 Panjiayuan Nanli, Chaoyang District, Beijing, 100021, China.
- Department of Gastroenterology and Hepatology, Sichuan University-University of Oxford Huaxi Joint Centre for Gastrointestinal Cancer, West China Hospital, Sichuan University, 37 Guoxue Ln, Chengdu, 610041, China.
- State Key Laboratory of Holistic Integrative Management of Gastrointestinal Cancers and National Clinical Research Center for Digestive Diseases, Xijing Hospital of Digestive Diseases, Air Force Medical University, Xi'an, China. [email protected].
- Department of Interventional Therapy, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, 17 Panjiayuan Nanli, Chaoyang District, Beijing, 100021, China. [email protected].
Abstract
Predicting hepatic encephalopathy (HE) after transjugular intrahepatic portosystemic shunt (TIPS) is critical for guiding portal hypertension treatment strategies and enabling early intervention. This study aims to employ the Tabular Prior-data Fitted Network (TabPFN) algorithm to develop a machine learning (ML) model that predicts post-TIPS HE. This study retrospectively enrolled 218 patients who underwent TIPS across three hospitals. Preoperative contrast enhanced CT (CECT) scans were used to delineate the volumetric region of interest (VOI) for the liver, spleen, abdominal fat, and abdominal muscle. Radiomics and deep transfer learning (DTL) features were extracted from each VOI. Overt HE occurrence during follow-up was divided into two groups. 171 patients (two hospitals) were randomly split (7:3) into training and validation set, 47 patients (third hospital) formed an external test set. After feature selection, we trained and compared multiple ML models. Shapley additive explanation (SHAP) was performed for model interpretability. The overall incidence of overt HE in the study cohort was 20.6%. The combined TabPFN model with the best predictive performance achieved AUCs of 0.953 (training set), 0.870 (validation set), and 0.942 (external test set), with accuracies of 0.933, 0.846, and 0.872, respectively. SHAP analysis identified the liver radiomics signature as a dominant predictors. Time-dependent AUCs at 90, 180, 365, and 730 days exceeded 0.88 in all cohorts, and high-risk patients had significantly higher HE occurrence (p < 0.01). A TabPFN-based ML model integrating CECT radiomics, DTL features, and MELD score enables accurate, externally validated prediction of post-TIPS HE, supporting personalized risk stratification and clinical decision-making.