AI-quantified high-branching tumor vasculature (H&E) stratifies recurrence risk beyond TNM in adjuvant RCC

In a post-hoc biomarker analysis of CheckMate 914 (CM914; NCT03138512), an adjuvant randomized Phase 3 trial after nephrectomy in localized/high-risk RCC, investigators used a deep-learning model on H&E whole-slide images to quantify the tumor-wide fraction of “high-branching” (HB) vessels in 1,098/1,641 evaluable patients (67%). HB fraction was continuously prognostic: each +10% HB reduced recurrence/death risk (DFS HR 0.92, 95% CI 0.89–0.95) and mortality (OS HR 0.89, 95% CI 0.83–0.95; both p<0.005), with best outcomes in the top quartile (OS HR 0.39; DFS HR 0.47); a data-driven cutoff (≥74% HB; n=275) identified an “intermediate” group vs “high-risk” (<74% HB; n=823) with worse DFS (HR 2.11; C-index 0.57) and OS (HR 2.57; C-index 0.56). TNM-only risk performed slightly worse (DFS C-index 0.53; OS C-index 0.55), while combining HB+TNM improved discrimination (C-index 0.59); clinically, this aims to refine current adjuvant selection where pembrolizumab is the US standard option based on DFS benefit vs placebo in KEYNOTE-564 (DFS HR ~0.68–0.72), whereas nivolumab+ipilimumab in CheckMate 914 Part A did not improve DFS vs placebo.