Machine learning-based prediction of time-dependent recurrence risk after ablation for hepatocellular carcinoma integrating glypican-3 and clinical features: a retrospective study

Hepatocellular carcinoma (HCC) frequently recurs after curative ablation, and current surveillance strategies lack time-specific risk stratification. We aimed to develop a machine learning model integrating glypican-3 (GPC-3) expression with routine clinical features to predict recurrence across distinct post-ablation time windows. This retrospective single-center study analyzed 447 patients with treatment-naïve, solitary HCC who underwent ablation. An ensemble feature selection combining logistic regression, minimum Redundancy Maximum Relevance (mRMR), and Mutual Information (MI) identified five key predictors: tumor size, alpha-fetoprotein (AFP), gender, GPC-3, and BCLC stage. Multiple models were built, and the Extreme Gradient Boosting (XGB) algorithm was selected as the best performer. Model performance was evaluated using area under the curve (AUC) for overall and time-specific recurrence (0–1 year, 1–2 years, > 2 years) in training, validation, and test sets (3:1:1 split). The XGB model achieved AUCs of 0.819, 0.813, and 0.790 in the training, validation, and test sets, respectively, for overall recurrence. For the 0–1 year window, the test set AUC was 0.809; for 1–2 years and > 2 years, AUCs were 0.776 and 0.781, respectively. Subgroup analyses confirmed stable performance across gender, age, and tumor size categories (interaction p > 0.7 for all). The model effectively separated high- and low-risk patients in Kaplan–Meier analysis. The GPC-3-integrated XGB model shows potential clinical utility for individualized, time-dynamic recurrence risk assessment after HCC ablation. It may facilitate tailored surveillance intervals and adjuvant therapy decisions. Prospective multi-center validation is warranted to confirm generalizability.