Abstract
BACKGROUND: Accurate stratification of recurrence risk after curative resection remains a critical challenge in the management of hepatocellular carcinoma (HCC). Dysregulated ceramide (CER) metabolism has been implicated in HCC progression and relapse. This paper evaluates whether preoperative plasma CER profiling coupled with machine learning (ML) enhances the risk prediction of HCC recurrence. METHODS: In this retrospective study, 257 HCC patients undergoing curative resection participated. Preoperative plasma CERs were quantified by targeted Lipidomics. Independent predictors were identified via multivariate Cox regression and incorporated into ten ML models. Using an internal 20% validation cohort, model performance was assessed by the area under the curve (AUC), concordance index (C-index), calibration, and decision curve analysis. Model interpretability employed Shapley additive explanations (SHAP), correlation analysis, and Bayesian network-based causal inference. The model's risk stratification capability was evaluated. This study was registered on clinicaltrials.gov (NCT06623474). RESULTS: Preoperative plasma CERs exhibited significant prognostic value in patients with HCC after curative resections. Multivariate analyses revealed that serum α-fetoprotein (AFP), tumor size, CER(d18:1/20:1), and CER(d18:1/22:1) independently predicted recurrence, and these variables were incorporated into ten ML models. Among them, the gradient boosting machine (GBM) algorithm demonstrated the best predictive performance (AUC: 0.959 at 1 year, 0.954 at 2 years, 0.931 at 3 years; C-index ≈ 0.93), outperforming all the other approaches. Model interpretability analysis (SHAP) highlighted tumor burden as the major determinant, with CER (d18:1/20:1) acting as a recurrence-promoting factor, and CER (d18:1/22:1) exerting a protective effect. Correlation analyses further revealed that CER(d18:1/20:1) was positively related to tumor multiplicity, systemic inflammation, and shorter recurrence-free survival (RFS), whereas CER(d18:1/22:1) was linked to better liver function and longer RFS. Bayesian causal inference indicated that elevated CER(d18:1/20:1) directly accounted for approximately 26% of the recurrence risk through effects on AFP and tumor size, whereas reduced CER(d18:1/22:1) conferred approximately 30% causal protection by modulating RFS, AFP, Liver function, and inflammation. Notably, the GBM model successfully identified 54 of 56 recurrent cases as high risk, enabling clear stratification of patients for precision surveillance. CONCLUSIONS: Preoperative plasma CER profiling, integrated with clinical parameters in a GBM framework, provides a highly accurate and interpretable strategy for predicting postoperative HCC recurrence, which paves the way for precise risk stratification and targeted management. This study provides insights that may enhance liver health and reduce disease burden in patients with HCC.