Abstract PS3-04-28: Espwa: a deep learning-enabled computational pathology tool that facilitates precision oncology for haitian breast cancer patients

Background: Global disparities in breast cancer outcomes are a public health crisis. In Haiti, one of the most resource-limited health care settings in the world, the challenge is especially dire. With only five practicing pathologists serving the entire country, timely and complete pathologic assessment of tumor tissue is nearly impossible. As estrogen receptor (ER) assessment with immunohistochemistry (IHC) cannot be performed, all Haitian breast cancer patients are empirically treated with endocrine therapy (ET). This “one-size-fits-all” strategy results in increased morbidity and mortality. To address this unmet need, we present ESPWA: ER Status assessment using deep learning-enabled histoPathology Whole slide imaging (WSI) Analysis, a low-cost tool that informs precision-based use of ET for Haitian patients. Methods: To develop ESPWA, we curated a dataset of H&E WSIs and tissue-matched ER status for 3448 breast cancer patients from Zanmi Lasante (ZL), the largest health care provider in Haiti. This dataset encompasses primary tumors (n = 2933) and lymph node metastases (n = 515). Using the ZL and The Cancer Genome Atlas (TCGA)-breast cancer datasets (n = 1349), we used weakly-supervised attention-based multiple instance learning to train models, using H&E as input, that predict ER status for each patient. Binary classification (output: ER-positive vs negative) and regression (output: quantitative expression of ER– 0-100%) models were trained. Label-preserving image augmentation methods (e.g., color augmentation, rotations, flips) were applied to increase diversity of training data. To assess performance, both the TCGA and ZL datasets were split into training (70%), validation (15%), and testing (15%) sets, and the results were compared across both patient populations. Results: First, we trained an ER classification model using the TCGA dataset. Performance of this TCGA-trained model was sensitive to the domain shift between the Black and Caucasian populations, with a performance of an area under receiver operating characteristic (AUROC) of 0.846 [95% confidence interval: 0.831-0.876] on the “held-out” TCGA test set, and 0.671 [0.655-0.683] on the ZL test set. Next, using 10-fold cross validation and bootstrapping analyses, we trained ER classification and regression models with the ZL dataset. These models demonstrated an AUROC of 0.790 [0.774-0.805] and an intraclass correlation coefficient of 0.524 [0.471, 0.551] on the ZL test sets. Subgroup analyses revealed that ESPWA performed less well for Haitian patients harboring ER-weak positive tumors (compared to ER-negative and strong positive tumors) and poorly differentiated tumors (compared to well-differentiated tumors).A head-to-head comparison between ESPWA and a breast pathologist with 22 years of experience (J.E.B.) illustrated ESPWA’s potential to outperform human experts. ESPWA (accuracy: 0.726 [0.711, 0.740]) displayed superior performance in determining ER status from an H&E compared to the pathologist (accuracy: 0.639 [0.627, 0.651]) for the entire dataset and across nearly all subgroups. Conclusions: In collaboration with the American Society of Clinical Pathology, we are planning clinical trials in ZL and other African countries to deploy ESPWA (“Hope” in Haitian Creole) to inform precision-based use of endocrine therapy for prospective patients. Citation Format: D. Pulido-Arias, R. Henderson, J. Lormil, C. Millien, M. Djenane Jose, G. Flambert, M. Mathelier, M. Corrielus, T. Goncalves, J. Kalpathy-Cramer, E. Gerstner, K. Landgraf, A. Brown, P. Castle, J. Jeronimo, S. Sirintrapun, S. Wander, D. Milner, J. Brock, C. Bridge, A. E. Kim. Espwa: a deep learning-enabled computational pathology tool that facilitates precision oncology for haitian breast cancer patients [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 nrPS3-04-28.