Interpretable machine learning model for predicting post-hepatectomy liver failure in hepatocellular carcinoma.
Tianzhi Tang1, Tianyu Guo2, Bo Zhu3
1Department of Hepatobiliary and Pancreatic Surgery, Cancer Hospital of China Medical University/Liaoning Cancer Hospital & Institute, Shenyang, Liaoning Province, People's Republic of China.
Scientific Reports
|May 2, 2025
Summary
A new machine learning model accurately predicts post-hepatectomy liver failure (PHLF) in liver cancer patients. The XGBoost model, using total bilirubin, MELD score, and ICG-R15, offers reliable preoperative risk assessment.
Area of Science:
- Hepatobiliary surgery
- Machine learning in medicine
- Oncology
Background:
- Post-hepatectomy liver failure (PHLF) is a critical complication after liver surgery.
- Accurate prediction of PHLF is essential for patient management and surgical planning.
Purpose of the Study:
- To develop and validate an interpretable machine learning (ML) model for predicting PHLF in hepatocellular carcinoma (HCC) patients.
- To identify key predictors of PHLF using advanced ML techniques.
Main Methods:
- Enrolled 312 HCC patients undergoing hepatectomy, with internal validation on 30% of samples.
- Utilized least absolute shrinkage and selection operator regression, random forest, and recursive feature elimination (RF-RFE) for variable selection.
- Evaluated 12 ML algorithms, with SHapley Additive exPlanations (SHAP) used to interpret the optimal model (XGBoost).
- Conducted an independent prospective validation with 62 patients.
Main Results:
- The XGBoost model achieved high predictive accuracy (AUCs of 0.983 training, 0.981 validation) among 12 ML models.
- Calibration curves and decision curve analysis (DCA) confirmed clinical applicability.
- Prospective validation (AUC = 0.942) demonstrated strong generalization.
- SHAP analysis identified total bilirubin (TBIL), MELD score, and ICG-R15 as critical predictors.
Conclusions:
- The interpretable XGBoost model accurately predicts PHLF in patients with resectable HCC.
- This ML approach surpasses traditional models in predictive performance.
- Key variables identified can guide preoperative risk stratification and clinical decision-making.
