Development and validation of an interpretable machine learning model integrating ctDNA methylation and clinical features for diagnosing hepatocellular carcinoma in HBV-infected patients with intrahepatic nodules

Distinguishing hepatocellular carcinoma (HCC) from non-HCC nodules remains challenging in patients with hepatitis B virus (HBV) infection. We developed an interpretable machine-learning model that integrates baseline circulating tumor DNA (ctDNA) methylation readouts with clinical variables for this specific diagnostic setting. The cohort included 586 patients, with 217 adjudicated HCC outcomes. The existing 411-patient training and 175-patient internal-validation partition was retained. Eleven baseline predictors previously selected by Boruta and LASSO were used. Ten algorithms underwent nested stratified cross-validation with 10 outer and five inner folds, regularized candidate settings, and a one-standard-error selection rule. Model selection and classification thresholds used training data. Frozen models were evaluated internally, and SHAP was used to explain CatBoost predictions. After feature selection using both Boruta and LASSO, eleven features were retained: age, cirrhosis, maximum intrahepatic nodule diameter, AFP-L3%, AFP, DCP, ALBI grade, AST, HIST1H3G/ACTB, OTX1/ACTB, and NLR. SHAP analysis was subsequently performed to quantify and visualize the contribution of each feature to the model predictions. Among the evaluated machine learning models, CatBoost achieved the highest observed area under the receiver operating characteristic curve in the internal validation set (AUC = 0.903). The sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), F1 score, and Youden’s index were 0.800, 0.818, 0.811, 0.722, 0.874, 0.759, and 0.618, respectively. The integrated CatBoost model demonstrated strong diagnostic performance for identifying HCC among HBV-infected patients with intrahepatic nodules by integrating clinical characteristics, laboratory indicators, and ctDNA methylation markers. The model provided an interpretable framework for distinguishing HCC from non-HCC nodules and showed potential as an auxiliary tool for individualized assessment of HCC likelihood after nodule detection. However, the use of historically derived feature selection and the reuse of the internal validation set may limit the robustness of the validation.