Abstract
BACKGROUND: The 5-year survival rate for hepatocellular carcinoma (HCC) is stage-dependent, yet existing models lack accuracy in predicting hepatitis B virus-associated HCC (HBV-HCC). We therefore aimed to develop and validate an interpretable machine learning (ML) model integrating multidimensional biomarkers for HBV-HCC risk stratification. METHODS: This retrospective multicenter study included 3568 participants (1872 HBV-infected and 1696 HBV-HCC). Patients from Mengchao Hepatobiliary Hospital were divided into training and validation sets (3:1 ratio), while those from Eastern Hepatobiliary Surgery Hospital and the First Affiliated Hospital of Xiamen University formed the external validation set. Five key predictors were identified through random forest, LASSO regression, and XGBoost methods. Seven ML models were evaluated using area under the curve (AUC), sensitivity, specificity, accuracy, and F1-score, with the top model compared against previous models (GALAD, C-GALAD, C-GALAD II, and ASAP). RESULTS: Key predictors were log(10)DCP (mean SHAP value 1.784), log(10)HBVDNA (1.063), log(10)ALT (0.753), AFP-L3% (0.444), and log(10)AFP (0.392). The XGBoost model achieved AUCs of 0.985 (95% CI: 0.981-0.989) in the training set, 0.978 (0.969-0.987) in the validation set, and 0.942 (0.911-0.973) in the external validation set. XGBoost significantly outperformed previous models in both the training and validation sets (DeLong test; p < 0.001). In the external validation set, XGBoost demonstrated superior individualized risk prediction accuracy (IDI = 0.228), net clinical benefit, calibration, and high-risk patient identification compared to the ASAP model. An interactive web tool was developed to facilitate clinical implementation. CONCLUSIONS: We developed a novel diagnostic model for HBV-HCC that demonstrates higher accuracy in identifying HBV-HCC compared to existing models.