Abstract RF3-07: A Multimodal-Multitask Deep Learning Model Trained in NSABP B-42 and Validated in TAILORx for Late Distant Recurrence Risk in HR+ Early Breast Cancer

Abstract Background: Clarity BCR is a multimodal-multitask (MMMT) deep learning algorithm developed to estimate late distant recurrence (DR) risk in hormone receptor–positive (HR+) breast cancer and to identify patients likely to benefit from extended endocrine therapy (EET). The algorithm was initially developed and validated in the NSABP B-42 trial, a pivotal randomized phase III study evaluating EET in postmenopausal women with HR+ early breast cancer. In B-42, Clarity BCR stratified patients into high- and low-risk groups with strong prognostic performance (HR = 5.71; 95% CI: 3.50-9.32, p < 0.001). It identified a clinically meaningful differential EET benefit for the 49% of patients in the high-risk group (10-year absolute benefit 4.09%). In contrast, among the 51% classified as low risk, the 10-year DR risk was 1.93% without EET and 1.44% with EET, yielding a minimal absolute benefit of 0.49%, supporting its clinical utility in sparing EET. Building on these findings, we externally validated Clarity BCR in the TAILORx translational cohort. TAILORx was a large prospective trial designed to optimize adjuvant therapy for patients with ER-positive, HER2-negative early breast cancer. Methods: We included 6,516 TAILORx patients with digitized hematoxylin and eosin (H&E) slides and relevant clinical data. Slides were scanned using the Pramana SpectralHT scanner. Clarity BCR integrates image features from whole-slide images (WSIs) and clinical variables (age, surgery type, pathological node status) to produce a continuous risk score and a dichotomized risk group (high vs. low), using a prespecified threshold. A subset of 4,469 patients who completed at least 4.5 years of endocrine therapy (ET) and remained disease-free at 5 years post-randomization was included in the late DR analysis. The primary endpoint was distant recurrence–free interval (DRFI). The primary objective was to validate Clarity BCR for late DR prognostication. The secondary objective was overall DR risk estimation across the full follow-up period. Univariable and multivariable Cox models were used. Results: In the overall cohort, 1,134 (17.4%) patients were classified as Clarity BCR high risk and 5,382 (82.6%) as low risk. In the late DR analysis cohort (n = 4,469), 770 were high risk and 3,699 were low risk. High-risk patients exhibited significantly worse late DR outcomes (HR = 1.88; 95% CI: 1.43-2.48; p < 0.001). The estimated 15-year DRFI was 86.4% (95% CI: 83.1-89.9%) in the high-risk group vs. 93.0% (95% CI: 92.0-94.0%) in the low-risk group. The prognostic discrimination C-index was 0.59 for Clarity BCR and 0.54 for Oncotype DX.In a multivariable analysis adjusting for age, tumor size, grade, Oncotype DX recurrence score, surgery type, therapy type, and menopausal status (n = 4,368), Clarity BCR remained independently prognostic (HR = 1.54; 95% CI: 1.11-2.13; p = 0.010).Across the entire follow-up period (DRFI from randomization), patients in the high-risk experienced significant worse DR outcomes (HR = 1.73; 95% CI: 1.40-2.15; p < 0.001), with 15-year DRFI of 83.7% in the high-risk group vs. 90.8% in the low-risk group. Multivariable analysis (n = 6,354) confirmed Clarity BCR’s independent prognostic value (HR = 1.32; 95% CI: 1.02-1.72; p = 0.037). The C-index for DRFI was 0.59 for Clarity BCR and 0.60 for Oncotype DX. Conclusions: Clarity BCR demonstrated robust, independent prognostic performance for late and overall DR up to 15 years in TAILORx. Together with prior NSABP B-42 findings, these results support its potential clinical utility to guide long-term treatment decisions in HR+ breast cancer. Research Support: Supported by NIH/NCI U10CA180820, U10CA180794. NSABP B-42 trial funded by NIH (U10CA180868, U10CA180822, UG1CA189867, U24CA196067) and Novartis. Citation Format: E. Mamounas, V. Wang, M. Chen, J. A. Sparano, R. J. Gray, M. Rahman, Y. Cheng, P. Rastogi, E. Amidi, C. E. Geyer Jr., T. Boucher, T. J. Freeman, M. Ramzanpour, M. Varma, H. Ghani, C. Cheng, C. Bales, J. R. Ribeiro, N. Stransky, M. R. Miglarese, M. Oberley, D. Spetzler, M. Radovich, G. W. Sledge, N. Wolmark. A Multimodal-Multitask Deep Learning Model Trained in NSABP B-42 and Validated in TAILORx for Late Distant Recurrence Risk in HR+ Early Breast Cancer [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr RF3-07.