Very Early Recurrence as a Dynamic Prognostic Stratifier After Curative Hepatectomy for Hepatocellular Carcinoma: A Landmark Analysis with Internal Validation and Machine-Learning Benchmarking

Background: Very early recurrence (VER, defined as recurrence within 12 months) after hepatocellular carcinoma (HCC) resection is often linked to poor survival, but conventional survival analyses stratified by recurrence status are inherently susceptible to guarantee-time bias. The relation between VER and post-recurrence survival also remains uncertain. Methods: This retrospective cohort included 526 patients with histopathologically confirmed HCC who underwent curative hepatectomy at the First Affiliated Hospital of Xinjiang Medical University between January 2016 and July 2025. A 12-month landmark cohort included patients alive beyond 12 months after surgery. Patients were split 7:3 into training and internally held-out validation cohorts. Landmark Cox models, a time-varying Cox model, post-recurrence survival analyses, and machine-learning benchmarking were performed. Missing primary-model values were median-imputed; multiple imputation was not used. Results: The landmark cohort included 486 patients; 82 had VER. The proportional-hazards assumption was violated for VER; the conventional adjusted Cox model estimated an average HR of 3.86 (95% CI, 2.81– 5.29; P < 0.001) over post-landmark follow-up. The association varied over time, with model-estimated HRs of 10.06 at 1 month and 2.23 at 60 months after the landmark. In the internally held-out validation cohort, adding VER to the conventional Cox model improved the C-index from 0.591 (95% CI, 0.508– 0.675) to 0.716 (95% CI, 0.633– 0.797; C-index difference = 0.125, 95% CI, 0.051– 0.203; P = 0.002). VER was not independently associated with post-recurrence mortality (HR, 1.23; 95% CI, 0.87– 1.75; P = 0.25). Conclusions: Among patients alive at the 12-month landmark, VER was associated with higher subsequent mortality and improved discrimination in an internally held-out cohort. External validation is required before use for individual risk prediction or surveillance decisions.