A Leakage-Free Survival-Modelling Benchmark for Hepatocellular Carcinoma Recurrence After Liver Transplantation:...

Background: Predicting recurrence after liver transplantation (LT) for hepatocellular carcinoma (HCC) remains important for post-transplant risk stratification and surveillance planning. The Milan criteria discriminate only moderately and some machine-learning re-analyses report overly optimistic results because of information leakage. Aim: The current study aimed to re-evaluate a previously published transplant cohort under a leakage-free survival-analysis framework and to benchmark post-transplant, explant-informed survival learners against the Milan criteria as a fixed pre-transplant reference. We hypothesised moderate rather than near-perfect discrimination, similar performance across learners of differing complexity, and better discrimination than the Milan criteria. Methods: This secondary analysis included 356 patients with HCC who underwent LT. The primary endpoint was recurrence-free survival, analysed from the observed event indicator and follow-up time rather than from a derived risk label. Seven survival learners were benchmarked with repeated nested cross-validation, using three repeats of a five-fold outer loop with a three-fold inner tuning loop. All data-dependent preprocessing, including robust multivariable outlier handling and imputation, was fitted within training folds only. Performance was assessed by the concordance indices of Harrell and Uno, the time-dependent area under the curve, the integrated Brier score, calibration, decision-curve analysis and descriptive competing-risk assessment. Results: Recurrence developed in 183 of the 356 patients over a median follow-up of 52 months. Discrimination was moderate rather than near-perfect and similar across learners; the random survival forest ranked highest and the Elastic-Net Cox model performed comparably. All learners showed higher descriptive concordance than the Milan criteria, and dependency-corrected comparisons supported higher concordance for the full-feature Cox model than for the Milan criteria, whereas the random survival forest and Cox did not differ materially. Out-of-fold calibration of the Elastic-Net Cox model at 36 months was acceptable, decision-curve analysis indicated positive net benefit across clinically relevant thresholds, and tumour size and alpha-fetoprotein were the leading contributors to prediction. Findings were stable in ablation and threshold-sensitivity analyses. Conclusions: Leakage-free survival modelling gave moderate but internally validated prediction of post-transplant recurrence and higher concordance than the Milan criteria in this cohort, supporting the stated hypotheses. Careful study design may matter more than architectural complexity in this setting, and leakage-free survival analysis is a practical standard for prognostic modelling in transplant oncology.