Improving Clinical Decisions in IR: Interpretable Machine Learning Models for Predicting Ascites Improvement after
Okan İnce1, Hakan Önder2, Mehmet Gençtürk3
1Department of Radiology, Rush University Medical College, Chicago, Illinois.
Journal of Vascular and Interventional Radiology : JVIR
|October 10, 2024
Summary
Machine learning models show promise in predicting ascites improvement after transjugular intrahepatic portosystemic shunt (TIPS) placement. These interpretable models can aid in selecting patients likely to benefit from TIPS for refractory ascites.
Area of Science:
- Hepatology
- Interventional Radiology
- Artificial Intelligence
Background:
- Refractory ascites significantly impacts liver disease prognosis.
- Transjugular intrahepatic portosystemic shunt (TIPS) placement is a key intervention for managing refractory ascites.
- Predicting treatment response is crucial for optimizing patient selection and outcomes.
Purpose of the Study:
- To evaluate the predictive potential of interpretable machine learning (ML) models for ascites improvement post-TIPS.
- To identify key features influencing treatment response in patients with refractory ascites.
Main Methods:
- Retrospective analysis of 218 patients undergoing TIPS for refractory ascites.
- Utilized 29 demographic, clinical, and procedural features.
- Developed and validated Support Vector Machine (SVM) and CatBoost ML models with Shapley additive explanations for interpretability.
Main Results:
- 77% of patients experienced ascites improvement.
- Higher sodium and albumin levels, lower creatinine, and lower MELD/MELD-Na scores were associated with improvement.
- SVM and CatBoost models achieved high accuracy (83% and 87%) and AUC (0.83 and 0.87).
Conclusions:
- Interpretable ML models demonstrate significant potential in predicting ascites improvement after TIPS.
- These models can assist in patient selection for TIPS procedures.
- Further validation is warranted to integrate ML into clinical decision-making for refractory ascites.
