Machine Learning-Based Models for the Prediction of Postoperative Recurrence Risk in MVI-Negative HCC

Area of Science:

• Hepatology
• Machine Learning in Medicine
• Oncology

Background:

• Hepatocellular carcinoma (HCC) patients without microvascular invasion (MVI) have a high risk of early recurrence (ER) after surgery.
• Prognostic factors for ER in this group are not well understood, and existing models have limitations.

Purpose of the Study:

• To develop and validate an interpretable machine learning (ML) model to predict ER in MVI-negative HCC patients.
• To utilize routine clinical parameters for accurate risk stratification.

Main Methods:

• Retrospective analysis of 578 MVI-negative HCC patients undergoing radical resection.
• Benchmarking of seven ML algorithms using clinical, laboratory, and imaging features, optimized with recursive feature elimination (RFE) and hyperparameter tuning.
• SHapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• The CatBoost model achieved superior performance (AUC: 0.7957, Accuracy: 0.7290).
• Key predictors for ER included absence of tumor capsule, elevated HBV-DNA and CA125 levels, larger tumor diameter, and lower body weight.
• SHAP analysis provided individualized risk insights.

Conclusions:

• The CatBoost model offers robust prediction of ER in MVI-negative HCC patients.
• This interpretable ML tool can personalize risk stratification and optimize postoperative management for HCC patients.