Clinical Validation of RlapsRisk BC in an international multi-cohorts setting

ABSTRACT Purpose This study evaluated the prognostic performance of RlapsRisk BC, a multimodal deep learning tool designed to predict distant recurrence-free interval (dRFI) in early-stage, ER-positive, HER2-negative breast cancer using routine H&E-stained whole-slide images (WSIs) and standard clinicopathologic features. Methods RlapsRisk BC was developed and internally validated on seven retrospective cohorts totaling 6,039 patients. Its ability to stratify patients into high- and low-risk groups for dRFI was then assessed in a blinded, retrospective validation across three independent international cohorts (UK, France, USA), including 591 patients with non-metastatic ER+/HER2− breast cancer treated with adjuvant endocrine therapy alone. Results Across all validation cohorts, RlapsRisk BC showed strong prognostic performance, stratifying patients into distinct low- and high-risk groups with hazard ratios from 3.93 to 9.05. At 5 years, distant recurrence ranged from 0.85%–4.7% in low-risk vs. 6.39%–34.8% in high-risk groups. This separation remained robust across subgroups, including grade 2 tumors, menopausal status, and nodal involvement. RlapsRisk BC was also an independent prognostic factor and improved performance when combined with clinical variables (age, tumor size, nodal status). It increased the c-index by 0.08, 0.19, and 0.20 across the three cohorts, with significant improvement in two. Compared to genomic assays, RlapsRisk BC showed complementary—and sometimes superior—performance, particularly for identifying low-risk patients. At matched specificity, it achieved higher sensitivity: 0.85 vs. 0.33 (Oncotype DX) and 0.74 vs. 0.49 (EndoPredict). Conclusion RlapsRisk BC demonstrates strong, independent prognostic value and may offer a scalable, accessible alternative to genomic assays. Further studies are needed to confirm clinical utility and support integration into treatment decisions. Context Key Objective This study aimed to assess whether RlapsRisk BC, a digital pathology-based AI model, can provide clinically meaningful prognostic stratification in ER-positive, HER2-negative early breast cancer. The AI model combines standard histology exhibited in whole slide images of hematoxylin and eosin stained tissue sections and clinical data. The prognostic performance of RlapsRisk BC was evaluated across multiple independent patient cohorts. Knowledge Generated The study demonstrates that RlapsRisk BC offers independent prognostic value beyond established clinical variables and genomic assays. It consistently stratifies patients into high- and low-risk groups for distant recurrence, with reproducible performance across diverse and independent cohorts, supporting its potential integration into routine clinical decision-making. Relevance RlapsRisk BC may serve as a scalable alternative or adjunct to molecular assays, supporting more personalized and accessible treatment decisions in breast cancer, particularly in settings where genomic testing is unavailable, limited, or yields intermediate-risk results.