A Machine Learning Model for Predicting Posthepatectomy Liver Failure After Hepatectomy With Extrahepatic Bile Duct Resection for Perihilar Cholangiocarcinoma: With and Without Indocyanine Green

ABSTRACT Aim Posthepatectomy liver failure (PHLF) is a serious complication of major hepatectomy and is closely related to perioperative mortality. Whether or not indocyanine green (ICG) is used to measure liver function varies by country. Using machine learning, we aimed to develop two highly accurate, user‐friendly models for predicting PHLF depending on whether ICG data were used. Methods This was a retrospective, three‐center study. SHapley Additive exPlanations (SHAP) evaluated the feature importance of Random Forest (RF) analysis. SHAP quantitatively assessed the impact of each feature on model predictions and identified the three most important features in both ICG‐used and ICG‐unused models. A Decision Tree (DT) model was constructed using two of these three key features to enhance clinical interpretability. PHLF was defined as Grade B or C according to the International Study Group of Liver Surgery. Results Feature importance was calculated using the SHAP analysis, in the ICG‐used model the ICG clearance rate (ICGK) multiplied by the percentage of remaining liver measured by CT volumetry (ICGK‐F), CRP, and operative procedure were identified as the three highest factors, whereas the predicted liver resection rate, CRP, and total bilirubin were identified as the most important features in the ICG‐unused model. In the constructed DT model, we categorized the cases into negative and positive, and the negative cases were further classified into three categories based on their risks. Conclusion These models may offer a simple and practical approach for predicting the risk of PHLF and hold promise for future clinical application.