Cross-modality AI modeling of histopathology images to stratify outcomes among patients treated with hormonal or chemo-hormonal therapy in ER+/HER2− breast cancer.

Abstract

1028 Background: Oncotype DX (ODX) is widely used to guide adjuvant treatment decisions in early-stage ER+/HER2− breast cancer. However, ODX provides limited stratification of residual metastatic risk within patients receiving hormonal therapy (HT) alone or hormonal therapy following chemotherapy (HT+CT). We developed a cross-modality AI approach that derives risk scores directly from routine H&E-stained whole-slide images (WSIs) to stratify metastatic outcomes among patients receiving standard adjuvant therapies. Methods: Using a cross-modality transfer learning framework, we first developed genomic biomarkers by mapping RNA-seq data to distant metastatic outcomes by integrating thousands of genes and signaling pathway activities in the publicly available ScanB dataset. Deep learning models were then trained to infer these biomarkers directly from WSI of breast tumors using the TCGA dataset. External validation was performed in an independent cohort of 287 early-stage ER+/HER2−, node-negative patients from MD Anderson Cancer Center. Among these, 147 patients with ODX<25 received HT alone, while 140 patients with ODX>25 were included, 121 of whom received HT+CT. Kaplan–Meier survival analysis, log-rank tests, and Cox proportional hazards models were used to evaluate outcome stratification within each treatment group. Results: Among patients treated with HT alone, AI score stratified metastatic outcomes (HR=2.75; p <0.01), identifying patients who developed distant recurrence despite low ODX scores. Among patients treated with HT+CT, AI score also stratified outcomes (HR=3.94; p <0.03), identifying patients who experienced metastasis despite combination therapy. Across treatment groups, AI score provided prognostic information beyond ODX and clinicopathologic variables. In multivariable analyses adjusting for ODX and clinical covariates, AI score remained independently associated with metastatic risk and demonstrated consistent stratification across clinically relevant subgroups. Validation of these findings in an expanded cohort is ongoing. Conclusions: This AI-based approach further stratifies outcomes among patients receiving HT or HT+CT as guided by ODX, directly from routine histopathology images in early-stage ER+/HER2− breast cancer. Thus, integration of AI-derived outcome stratification with ODX scores may improve individualized treatment decision-making and support consideration of alternative therapies.