Machine learning-augmented Cox regression for evaluating adjuvant transarterial chemoembolization efficacy and predicting survival in hepatocellular carcinoma with satellite nodules following radical hepatectomy

The five-year post-resection recurrence rate of hepatocellular carcinoma (HCC) remains 50–70%, and satellite nodules constitute a well-established independent high-risk factor for early recurrence. The therapeutic benefit of adjuvant transarterial chemoembolization (TACE) in this specific subpopulation has not been rigorously evaluated in large cohorts. We therefore aimed to quantify the impact of postoperative adjuvant TACE and to develop a machine learning-based nomogram for individualized survival prediction in HCC patients with pathologically confirmed satellite nodules following radical hepatectomy. This retrospective, non-randomized, multi-center cohort enrolled 273 HCC patients with pathologically confirmed satellite nodules who underwent radical hepatectomy between January 2015 and March 2024. Assignment to the adjuvant TACE group versus the radical-resection-only control group was determined by shared decision-making between the treating multidisciplinary team and the patient (or family) at the first postoperative outpatient review , rather than by randomization. Recurrence-free survival (RFS) and overall survival (OS) were compared by Kaplan–Meier analysis with log-rank tests. A 1:1 propensity-score-matched sensitivity analysis (caliper 0.2 SD; covariates: sex, age, BCLC stage, cirrhosis, HBV infection, ALBI grade, MVI, Edmondson grade, satellite multiplicity, resection type, and surgical margin) was performed to adjust for residual baseline imbalance. Independent prognostic factors were identified by univariable and multivariable Cox proportional hazards regression. A peritumoral delta-radiomics score (peritumoral_delta_radscore), derived from contrast-enhanced CT by three-dimensional volumetric segmentation, was generated. Three machine learning-based Cox models—XGBoost-Cox, Cox-LASSO, and CoxBoost—were constructed and evaluated by time-dependent AUC, decision curve analysis (DCA), calibration, and concordance index (C-index) under ten-fold cross-validation. SHAP (SHapley Additive exPlanations) values were applied to quantify feature contributions, and a clinical nomogram was constructed. Adjuvant TACE markedly improved both RFS and OS. In the full cohort, Median RFS was 18.0 months in the TACE group versus 10.6 months in the control group (log-rank P < 0.001), with 1- and 2-year RFS rates of 72.8% vs. 44.6% and 39.2% vs. 8.8%, respectively. Median OS was 36.5 months in the TACE group versus 19.5 months in controls (log-rank P < 0.001), with a 3-year OS of 53.6% vs. 13.5%. After 1:1 propensity-score matching ( n = 248; 124 matched pairs), the survival advantage of adjuvant TACE remained statistically significant for both RFS (median 18.0 vs. 9.8 months; log-rank P < 0.001; 1-year RFS 72.6% vs. 42.7%; 2-year RFS 39.5% vs. 9.7%) and OS (median 36.9 vs. 22.9 months; log-rank P < 0.001; 1-year OS 100% vs. 100%; 3-year OS 54.0% vs. 16.1%). Multivariable Cox analysis identified treatment group, female sex, cirrhosis, BCLC stage, and peritumoral_delta_radscore as independent prognostic factors for both RFS and OS. Among the three machine learning models, XGBoost-Cox achieved the highest predictive performance, with time-dependent AUCs of 0.977/0.987/0.984 (1/3/5-year) in the training set ( n = 163, events = 155) and 0.900/0.880/0.900 in the independent test set ( n = 110, events = 105), consistently outperforming Cox-LASSO and CoxBoost. SHAP analysis identified treatment assignment, sex, BCLC stage C, and peritumoral_delta_radscore as the four dominant contributors. A 4-variable nomogram integrating these features demonstrated excellent calibration and net clinical benefit for 1-, 3-, and 5-year OS prediction. Adjuvant TACE might significantly prolong both RFS and OS in HCC patients with satellite nodules after radical hepatectomy. The XGBoost-Cox nomogram provides accurate individualized 1-, 3-, and 5-year survival prediction and may serve as an evidence-based decision-support tool for this high-risk subpopulation. As a retrospective , non-randomized study , the present findings require confirmation by prospective multicenter validation , ideally incorporating a randomized or propensity-score-stratified design to mitigate residual selection bias.