An Explainable Machine Learning Model for Predicting Short-Term Haemodynamic Changes Post-TIPS With Prognostic...

Area of Science:

• Medical Imaging and Interventional Radiology
• Machine Learning in Medicine
• Hepatology and Liver Disease

Background:

• Transjugular intrahepatic portosystemic shunt (TIPS) is crucial for managing portal hypertension.
• Remeasuring portacaval pressure gradient (PPG) post-TIPS offers prognostic value but is invasive.
• Current methods lack efficient prediction of short-term PPG changes.

Purpose of the Study:

• To develop and validate an explainable machine learning (ML) model for predicting short-term PPG changes post-TIPS.
• To enhance the prognostic utility of PPG measurements in cirrhotic patients undergoing TIPS.
• To provide a non-invasive tool for clinical decision-making.

Main Methods:

• Developed and validated a Support Vector Regression (SVR_RBF) model using retrospective (n=328) and prospective (n=128) cohorts.
• Utilized hyperparameter tuning with 5-fold cross-validation for model optimization.
• Employed SHapley Additive exPlanation (SHAP) for model interpretability and identified key predictors.

Main Results:

• The 6-feature SVR_RBF model demonstrated robust performance in predicting PPG changes (R-squared up to 0.617).
• SHAP analysis revealed immediate PPG reduction as the main predictor of subsequent PPG changes.
• Relative PPG reduction, not absolute values, significantly correlated with 2-year decompensation, with optimal thresholds identified.

Conclusions:

• The developed ML model accurately predicts PPG changes and enhances prognostic value post-TIPS.
• This explainable AI approach supports clinical decision-making by offering reliable predictions without additional invasive procedures.
• The model holds potential for improving patient management in cirrhotic patients undergoing TIPS.