Machine learning-based hepatocellular carcinoma risk prediction model for patients with HBV-related compensated advanced chronic liver disease.

purposePatients with hepatitis B virus (HBV)-related compensated advanced chronic liver disease (cACLD) demonstrate significant liver fibrosis and portal hypertension, further increasing their hepatocellular carcinoma (HCC) risk. This study aimed to develop and validate machine learning-based HCC risk prediction models.

methodsWe retrospectively enrolled 1051 patients with HBV-related cACLD, randomly allocated patients to training (n = 736) and validation (n = 315) cohorts. Feature selection was performed using least absolute shrinkage and selection operator regression, random forest (RF), and support vector machine (SVM). Based on the selected key features, five machine learning models were constructed: SVM, RF, logistic regression, extreme gradient boosting, and Naive Bayes. Model performance was evaluated using area under the curve (AUC), accuracy, sensitivity, and specificity, etc. The Shapley additive explanations (SHAP) method was employed for model interpretability analysis.

resultsDuring a median follow-up time of 35 (20–55) months, 103 patients (9.8%) developed HCC. Feature selection analysis identified five key predictors: liver stiffness measurement (LSM), age, platelet, bile acid, and white blood cell count. The RF model demonstrated superior performance with an AUC of 0.979, an accuracy of 0.977, and a sensitivity of 0.808. SHAP interpretability analysis identified LSM as the most influential predictor (mean SHAP value 1.2), followed by age and other indicators. Feature interaction analysis revealed significant synergistic effects between LSM, platelet, and bile acid.

conclusionMachine learning-based HCC risk prediction models, particularly the RF algorithm, demonstrated excellent predictive performance in patients with HBV-related cACLD. LSM emerged as the most critical predictive factor.