Predictive and prognostic value of MRI radiomics in patients with hormone-receptor–positive, HER2-negative breast cancer: A multicenter study.

e12559 Background: Treatment decisions for patients (pts) with early-stage hormone receptor-positive, human epidermal growth factor receptor 2-negative (HR+/HER2-) breast cancer (BC) routinely rely on the use of a genomic-based (21-gene) biomarker. Insufficient samples can preclude its use. We sought to evaluate whether a new breast MRI image-based artificial intelligence (AI) tool, with or without standard clinicopathologic (CP) factors, can (1) predict high 21-gene score (≥26) for guiding adjuvant chemotherapy and (2) prognosticate 5-year recurrence risk better than the 21-gene assay. Methods: In this multicenter, retrospective study, we built the MRI-based AI predictive/prognostic tool on a single-institution development cohort (DC), then validated its performance on an independent validation cohort (VC) comprised of three institutions. The AI tool used 1,688 MRI-derived radiomic features extracted from each pre-treatment breast MRI, with and without CP factors. The AI classification tool for predicting the high 21-gene score optimized discrimination between the binary endpoints of ≥26 versus < 26 using elastic-net logistic regression with nested cross-validation. Discrimination ability was quantified by area under the receiver operating characteristic curve (AUC) with performance metrics and reported at a derived Youden threshold. We assessed prognostic performance of 5-year recurrence risk using the IPCW concordance index (C-index). Results: 1,277 pts (DC n = 516; VC n = 761) with HR+/HER2- BC diagnosed between January 2010 and December 2024 with pre-treatment breast MRI were included from four institutions. The Kaplan–Meier 5-year recurrence incidence was 9.6% (95% CI, 7.9%–11.7%). For predicting 21-gene score ≥26, combining MRI-derived radiomic features with CP predictors improved classification in the DC compared with CP predictors alone (AUC 0.81 vs 0.77) and increased balanced accuracy at the Youden operating point (0.76 vs 0.74), driven by higher sensitivity (0.78 vs 0.58) with similar specificity (0.78 vs 0.79). In the VC, the combined model also improved AUC (0.80 vs 0.78) and achieved higher balanced accuracy at the Youden threshold (0.72 vs 0.70), with sensitivity (0.76 vs 0.60) and specificity (0.69 vs 0.70), respectively. For prognostication, radiomic features demonstrated higher discrimination ability for 5-year recurrence risk than a model combining the 21-gene score with CP factors, improving the C-index from 0.62 to 0.71 in the DC (Δ = +0.09) and from 0.55 to 0.61 in VC (Δ = +0.06; p < 0.0001). Conclusions: In early-stage HR+/HER2− BC, an AI tool comprised of MRI-based and CP features predicted high 21-gene score (≥26) (AUC 0.8), as well as prognosticated 5-year recurrence better than the 21-gene score combined with CP factors. Given the ubiquity of breast MRI, this tool may serve as a proxy for the 21-gene score when not available.