H&E-based AI model assessed for recurrence score estimation in early breast cancer

Evidence before this study

We searched Google Scholar for studies published between Jan 1, 2016, and March 1, 2025, using the terms (“Hematoxylin” OR “Haematoxylin” OR “Eosin”) AND “Breast cancer” AND “Deep learning” AND (“Oncotype” OR “Oncotype DX”), without language restrictions. The search identified 416 papers; 11 peer-reviewed studies specifically addressed the prediction of the Oncotype DX recurrence score from haematoxylin and eosin (H&E)-stained breast cancer slides. Review of reference lists found no additional relevant reports. These studies showed increasing interest in using machine-learning models to predict genomic recurrence scores, but all studies were based on retrospective observational data, where treatment assignment confounds prognostic and predictive evaluation. None were validated in randomised clinical trials, preventing assessment of chemotherapy benefit, the main clinical purpose of the Oncotype DX assay. These studies also lacked large-scale external validation across diverse populations and laboratory settings. A separate body of work has aimed to infer patient outcomes directly from H&E images, including regulatory-approved commercial tools (PreciseDX, Stratipath, Ataraxis, and RlapsRisk). Although these methods can outperform existing genomic assays in some datasets, their clinical impact remains uncertain because they were validated on observational data, where nearly all high-risk patients receive chemotherapy, precluding evaluation of treatment benefit. By contrast, predicting the Oncotype DX score itself allows AI models to inherit the assay's established clinical validation from randomised trials such as TAILORx. Across published studies, the risk of bias was high to moderate, driven by retrospective design, small sample size, and absence of external patient-level validation.

Added value of this study

To our knowledge, this study presents the first AI model based on H&E pathology images and clinicopathological variables to be assessed retrospectively using data from a randomised clinical trial. Leveraging this trial enabled assessment of the model's ability to predict chemotherapy benefit derived from randomised treatment assignment, rather than only correlate with recurrence scores or outcomes. The integration of a large-scale foundation model with a simple calibration method that does not rely on local genomic data shows that such an AI model can generalise across diverse populations and laboratory conditions. Extensive external validation across multiple independent cohorts, comprising one of the largest evaluations of an AI pathology model to date, provides first-of-its-kind evidence that combines retrospective assessment using randomised trial data with extensive real-world validation of recurrence score estimation and prognostication.

Implications of all the available evidence

This study presents a digital pathology AI model that advances towards real-world deployment to guide chemotherapy decisions in hormone receptor-positive, HER2-negative early breast cancer. The model's potential impact is particularly significant in low-income and middle-income countries, where genomic testing reaches fewer than 5% of patients. These findings show the potential for democratising precision oncology through widely accessible histopathological analysis, extending personalised treatment decisions to resource-limited settings globally. Future prospective and randomised validations across node-positive patients and populations from low-income and middle-income countries will further establish clinical utility and support global implementation.

Source: The Lancet Oncology