Abstract PS3-04-01: Prediction of pathologic complete response from histopathology images of HER2+ breast cancer using an AI foundation model

Abstract Background: Neoadjuvant taxane-based chemotherapy in combination with anti-HER2 therapy is standard in patients with locally advanced HER2-overexpressing (HER2+) breast cancer. Pathologic complete response (pCR) after combination therapy is associated with improved long-term outcomes and occurs less frequently in patients whose disease is also estrogen and/or progesterone receptor positive (HR+) compared to patients with estrogen and progesterone receptor negative (HR-) disease. In patients with HER2+, HR+ disease, improved patient selection for standard neoadjuvant therapy is needed to optimize outcomes and minimize toxicity in patients unlikely to achieve pCR. Here we sought to utilize digital pathology and artificial intelligence/machine learning (AI/ML) algorithms on baseline biopsy tissue of newly diagnosed patients with operable HER2+ breast cancer to predict likelihood of pCR after standard neoadjuvant therapy, with a particular focus on HR status-specific differences. Methods: Digitized H&E-stained baseline biopsy slides from patients with HER2+ operable breast cancer treated neoadjuvantly with taxane-based chemotherapy plus anti-HER2 therapy with pathologic response data available were identified from 2 publicly available datasets and an institutional cohort. Utilizing CanvOI, a pre-trained computational pathology foundation model, we developed an AI/ML model for predicting pCR from whole-slide images (WSIs). For processing, WSIs were divided into 380x380 pixel tiles. CanvOI’s tile encoder converted tile image data into numeric vectors, and subsequently, CanvOI’s slide encoder converted tile vectors into an aggregated slide vector. The slide vectors were used as predictors of pCR for HR+ and HR- (ER and PR ≤ 10%) cohorts independently, using logistic regression as a vector-data classification method. Model performance and generalizability were evaluated using 4x4 cross-validation, consisting of 4-fold cross-validation repeated 4 times with randomly mixed data. In each iteration, 75% of the data was used for model training and 25% for evaluation. Multivariable logistic regression analysis was conducted in each cohort to evaluate the model and pathological features associated with pCR. To address bias due to the naturally occurring imbalance of pCR rates within the HR+ and HR- cohorts, we implemented sample trimming in the dominant class within each cohort. Results: A total of 231 patients had adequate baseline tissue for analysis, N=119 HR+ and N=112 HR-. The rates of pCR for these patients was 40.4% for HR+ disease and 64.1% for HR- disease. For prediction of pCR status, the model achieved a mean Area Under the Curve (AUC) of 0.69 (95% CI: 0.66-0.73) and Matthews Correlation Coefficient (MCC) of 0.304 in HER2+/HR+ patients, and a mean AUC of 0.65 (95% CI: 0.61-0.69) and MCC of 0.213 in HER2+/HR- patients. After multivariable analysis evaluating the ML model, tumor grade (1-2 vs 3) and HER2 immunohistochemistry status (2+ vs 3+) for patients in whom the pathological data was available (N=70 HR+ and N=73 HR-), the ML model remained a significant predictor of pCR in both HR+ and HR- cohorts. Applying data balancing techniques improved model performance, with a mean AUC of 0.70 (95% CI: 0.67-0.73) and MCC of 0.318 in HER2+/HR+ patients, and a mean AUC of 0.67 (95% CI: 0.63-0.71) and MCC of 0.181 in HER2+/HR-. Conclusions: These findings demonstrate the potential of a digital pathology-based AI/ML model to capture information about tumor sensitivity to therapeutic agents that are not provided by current clinical parameters. In patients with HER2+, HR+ operable breast cancer, where rates of pCR remain low, such a tool has the potential to guide the evaluation of novel therapeutic strategies in appropriately selected patients. Larger datasets are needed to build upon these findings. Citation Format: R. Basho, E. Heredia, H. McArthur, Z. Fang, V. Narumi, A. LeVee, T. Jonsson, F. Dadmanesh, S. Bose, S. Patre, T. Keller, J. Paine, B. Tran, B. Simental, O. Castellanos, J. Lee, S. Shiao, D. Agus, N. Matasci.. Prediction of pathologic complete response from histopathology images of HER2+ breast cancer using an AI foundation model [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 nr PS3-04-01.