Incremental prognostic value of an AI-derived histology signature beyond the 21-gene recurrence score: A prospective-retrospective validation in TAILORx.

Abstract

554 Background: The 21-gene recurrence score (RS) is a foundational tool for risk-stratifying HR+/HER2- early breast cancer (EBC). However, clinical outcomes vary within RS categories. RlapsRisk BC (RR), an AI pathology-based test, integrates features from H&E-stained whole-slide images with clinical data (age, tumor size, nodal status) and was developed using 7 retrospective cohorts totaling 6,039 patients. We evaluated the clinical validity of RR and its histology-only component (RR-H) beyond RS and standard clinicopathologic factors. Methods: The clinical validity of RR was established through a validation program of 4 cohorts and over 8,521 patients across diverse geographic regions and laboratory settings. This included 3 international cohorts (n=933) and a prospective-retrospective analysis of the TAILORx trial, where RR-H was evaluable in 7,585 (97.5% of analyzable patients). The primary endpoint was distant recurrence-free interval (DRFI). In TAILORX, we assessed the additive value of the RR-H score to a base model (composed of age, tumor size, histological grade and RS) using Cox proportional hazards models and C-index comparison. Results: Across international validation cohorts (median follow-up 7.5 years, 9,8% DRFI events), RR successfully stratified patients (HR=4.91; 95% CI: 3.13-7.71; p < 0.00001), and identified a low-risk population with a 97.5% (95% CI: 95.8%-98.5%) 5-year DRFI rate. In the TAILORx population (n=7,584), the RR-H score was a highly significant independent predictor of DRFI. Incorporating RR-H as a continuous variable to the base model (including RS) significantly increased the C-index from 0.6730 to 0.7007 (p < 0.0001). The estimated HR for a 1-point difference in RR-H was 1.108 (95% CI: 1.078-1.138; p < 0.0001). When analyzed by quartiles, patients in the highest risk group (Q4) exhibited a significantly higher risk of distant recurrence compared to the lowest risk group (Q1) HR= 2.656 (95% CI: 2.003-3.522). Concordance analysis revealed that RR-H is independent of stromal TILs (R 2 =0.01). While RR-H correlated with increasing histological grade, substantial distribution overlap confirms intra-grade prognostic granularity. Furthermore, RR-H distributions were consistent across ILC and non-NLC. Conclusions: RR demonstrated robust and reproducible prognostic value across 8520 patients from diverse populations and settings establishing its clinical validity. Additional findings on TAILORx demonstrate that RR-H provides significant, independent prognostic value that complements existing prognostic tools (including genomic assays and clinicopathological factors), suggesting that integrating AI-pathology can refine precision risk assessment, and thus optimize adjuvant treatment in HR+ HER2- EBC.