To investigate the clinical predictive value of machine learning in patients with cirrhosis undergoing transjugular intrahepatic portal shunt(TIPS).

A total of 53 cirrhosis patients admitted to Shandong Provincial Hospital Affiliated to Shandong University from January 2016 to December 2019 were included in this study.Sixty-eight preoperative and postoperative clinical variables were collected, and postoperative bleeding, hepatic encephalopathy, liver disease-related death and shunt disfunction were used as independent outcomes, respectively.Logistic regression was used to select clinical variables that had significant influence on the outcome.Corresponding support vector machine(SVM)models were constructed and Shapley Additive explanation(SHAP)model was constructed for interpretation and analysis.

The accuracy, recall rate and area under the curve(AUC)of the SVM model for postoperative bleeding were 0.75, 1.00 and 0.81, respectively.The three parameters for hepatic encephalopathy were 0.75, 0.67 and 0.77, respectively.The three parameters for liver disease-related death were 0.88, 1.00 and 0.87, respectively.The three parameters for shunt disfunction were 0.94, 0.67 and 0.87, respectively.Among the four constructed models, the highest SHAP values of included variables were diuretics, probiotics, preoperative portal venous pressure and end-stage liver disease model(MELD)score.

Machine learning has good practical value in predicting different clinical outcomes after TIPS for liver cirrhosis patients, which can assist clinicians to predict the postoperative status of such patients and carry out effective intervention at an early stage.