Abstract 7109: Understanding race-specific differences in the predictive performance of computational outcome models for breast cancer

Abstract Background: Neoadjuvant chemotherapy (NACT) is the standard of care for locally advanced breast tumors and pathologic complete response (pCR) is the most prominent determinant of favorable outcomes. We built computational predictors of pCR for breast cancer patients (pts) treated with NACT using different data modalities acquired at baseline (clinical, genomic, pathology and radiology) and investigated race-specific differences in their predictive value when pts were stratified into hormone receptor subtypes. Methods: We analyzed data from 1, 135 breast cancer pts who self-identified as Asian (n=151, 13%), Black (168, 15%), Hispanic (154, 14%) or White (662, 58%). Patients were treated with NACT at Memorial Sloan Kettering (MSK) between 2014 and 2021. The cohort included HR-/HER2+(n=183, 16%), HR+/HER2- (353, 31%), HR+/HER2+(339, 30%) and TNBC (260, 23%) pts. Digital pathology whole-slide images from H&E stained pre-NACT tumor biopsies were analyzed for 863 pts. Pre-NACT MRIs were analyzed for 608 pts. Targeted DNA sequencing from MSK-IMPACT was analyzed for 302 pts. Predictors of pCR were built using penalized regression models for the clinical and genomic modalities, and deep learning (DL) architectures for pathology and radiology. Genetic ancestry was computed for pts with sequencing information using published methods. Model performance was quantified as the area under the receiver operating characteristic curve (AUC). Results: Genetic ancestry labels and self-reported race were highly concordant (297/302, 98% match). The overall pCR rate was 38.5% and varied strongly by receptor subtype, being lowest for HR+/HER2- (13%) and highest for HR-/HER2+ (75.4%). pCR rate also varied across races, being highest for Asians (45%) and lowest for Black pts (34.5%), although this can be at least partially explained by the different representation of receptor subtypes across races (e.g., Black pts had significantly higher rates of TNBC tumors, 37.5% vs. 20.3%, p < 0.0001). After controlling for subtype, Hispanics tended to be diagnosed at a younger age (median 46 vs. 50 years, p=0.02). We identified several combinations of subtype and data modalities where AUCs varied by race: for example, AUCs for the clinical predictor were higher for Asian pts in the TNBC subtype (0.64 AUC vs. 0.49 mean AUC across all races), while the predictor based on the same modality overperformed among Hispanics of the HR-/HER2+ subtype (0.84 AUC vs. 0.65 mean AUC across all races). Training subtype-specific models instead of a pan-subtype model only improved overall cross-type AUC for Black pts (mean AUC increase of 0.07). Conclusions: Racial background can impact the predictive performance of computational outcome models in breast cancer and should be factored for method design. The influence of race-specific differences varies across data modalities and receptor subtypes. Citation Format: William Asinger, Sarah Eskreis-Winkler, Armaan Kohli, Enrico Moiso, Anika Begum, Mehnaj Ahmed, Andrew Aukerman, Jianjiong Gao, Jian Carrot-Zhang, Arfath Pasha, Hong Zhang, Elizabeth Sutton, Sohrab Shah, Pedram Razavi, Francisco Sanchez-Vega. Understanding race-specific differences in the predictive performance of computational outcome models for breast cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 7109.