Supplementary Material for: Machine learning models for Disease-free survival analysis after liver resection for hepatocellular carcinoma: a multicentric French collaborative study

Introduction Liver resection (LR) is a potentially curative treatment of hepatocellular carcinoma (HCC) but early recurrence rates remain high reaching as high as 70%. Accurate prediction of disease-free survival (DFS) after LR is crucial to optimize patients’ selection for clinical trials evaluating adjuvant strategies. This study assessed machine learning (ML) models for predicting DFS after LR for HCC. Methods A total of 663 patients who underwent LR between 2010 and 2020 in 3 French HPB referral centers were analyzed. Three ML models —random survival forest (RSF), gradient boosting survival (GBS), and fast survival support vector machine (FSSVM) —were compared with the Cox proportional hazards regression model. Model performance was assessed using C-index and time-dependent area under the curve (AUC). External validation was performed using an independent cohort from a fourth center. Statistical comparisons between models were conducted using paired t-tests. Results After a median follow up of 52 months, recurrence occurred in 43 % of patients and median DFS was 15 months. In the training cohort, RSF achieved the highest discrimination (C-index = 0.721 ± 0.027) and predictive accuracy (AUC = 0.768 ± 0.039) among the four evaluated models. Cox, GBS, and FSSVM showed comparable performance (C-index = 0.684–0.699; AUC = 0.718–0.745). Paired t-tests demonstrated that RSF significantly outperformed Cox for both C-index (p=0.023) and AUC (p = 0.039). In the external validation cohort, all ML models outperformed Cox regression, with RSF showing the highest performance (C-index 0.817, AUC 0.863), suggesting strong generalizability. Conclusions RSF significantly improved DFS prediction compared with Cox regression and yielded the best discriminatory performance in both training and external validation cohorts. These findings highlight the value of ML-based survival models, particularly RSF, for enhancing individualized postoperative prognostication and refining patient selection for clinical trials evaluating adjuvant treatment to surveillance after LR for HCC.