ESPWA: a deep learning tool to inform precision-based use of endocrine therapy in resource-limited settings

Abstract

Immunohistochemistry for estrogen receptor (ER) expression is often unavailable in low-and-middle-income countries (LMICs), leading to empiric use of endocrine therapy (ET) and unnecessary toxicity in ER-negative patients. To address this unmet need, we developed ESPWA, a deep-learning model trained on 3448 H&E slides and tissue-matched ER status from breast cancer patients treated at Zanmi Lasante (ZL), Haiti. A model trained on The Cancer Genome Atlas (TCGA) exhibited substantial domain shift when applied to the ZL cohort, with AUROCs dropping from 0.846 on TCGA cross-validation to 0.671 on the ZL cohort. In contrast, ESPWA demonstrated improved performance on ZL cross-validation (AUROC=0.790; p=0.005). In an independent test set of 134 Haitian patients with parallel slides prepared and scanned in Mirebalais Hospital (Haiti) and Brigham and Womens Hospital, ESPWA was robust to variations in slide preparation, quality, and scanners, achieving AUROCs of 0.794 on BWH-prepared WSIs and 0.805 on Mirebalais-prepared WSIs. Prospective studies using ESPWA are underway in sub-Saharan Africa to evaluate its utility in informing precision-based use of ET.