Objective: To develop and evaluate a machine learning-based CT radiomics model for the prediction of hepatic encephalopathy (HE) after transjugular intrahepatic portosystemic shunt (TIPS).
Methods: A total of 106 patients who underwent TIPS placement were consecutively enrolled in this retrospective study. Regions of interest (ROIs) were drawn on unenhanced, arterial phase, and portal venous phase CT images, and radiomics features were extracted, respectively. A radiomics model was established to predict the occurrence of HE after TIPS by using random forest algorithm and 10-fold cross-validation. Receiver operating characteristic (ROC) curves were performed to validate the capability of the radiomics model and clinical model on the training, test and original data sets, respectively.
Results: The radiomics model showed favorable discriminatory ability in the training cohort with an area under the curve (AUC) of 0.899 (95% CI, 0.848 to 0.951), while in the test cohort, it was confirmed with an AUC of 0.887 (95% CI, 0.760 to 1.00). After applying this model to original data set, it had an AUC of 0.955 (95% CI, 0.896 to 1.00). A clinical model was also built with an AUC of 0.649 (95% CI, 0.530 to 0.767) in the original data set, and a Delong test demonstrated its relative lower efficiency when compared with the radiomics model (p < 0.05).
Conclusion: Machine learning-based CT radiomics model performed better than traditional clinical parameter-based models in the prediction of post-TIPS HE.
Advances in knowledge: Radiomics model for the prediction of post-TIPS HE was built based on feature extraction from routine acquired pre-operative CT images and feature selection by random forest algorithm, which showed satisfied performance and proved the advantages of machine learning in this field.