AI Test Predicts Breast Cancer Recurrence Using Standard Pathology Slides

Researchers at New York University developed a multimodal artificial intelligence (AI) test that may provide a faster and less expensive method for predicting breast cancer recurrence than currently available genomic assays. Published in Nature Communications, the AI-based tool analyzes routine pathology slides alongside standard clinical information to estimate recurrence risk across multiple breast cancer subtypes, including those for which no guideline-recommended genomic prognostic tests currently exist.¹

Current genomic assays, such as Oncotype DX, play an important role in guiding adjuvant treatment decisions for patients with hormone receptor (HR)-positive, HER2-negative breast cancer. However, these tests are costly, require specialized laboratory processing, may take several weeks to return results, and consume tissue samples that could otherwise be preserved for future molecular analyses.¹

"Our AI test can read the same tumor slides pathologists already examine and, combined with basic clinical details, accurately estimate how likely a patient's cancer is to return," senior author Krzysztof J. Geras, PhD, visiting scholar at NYU's Center for Data Science and adjunct assistant professor at NYU Grossman School of Medicine, said in a news release.²

Strong Performance Across Diverse Patient Populations

To develop the model, investigators trained and validated the AI system using data from 8161 patients across 15 cohorts spanning 7 countries. The algorithm integrates digitized hematoxylin and eosin-stained pathology slides with routinely collected clinical variables, including tumor stage, patient age, estrogen receptor, progesterone receptor (PR), HER2 status, and histologic subtype.¹

For external validation, researchers evaluated the model in 3502 patients from 5 independent cohorts. Overall, the AI test demonstrated strong prognostic performance, achieving a pooled concordance index (C-index) of about 0.71 for predicting disease-free interval while effectively distinguishing patients at higher versus lower risk of recurrence. Patients classified as high risk experienced a significantly greater likelihood of recurrence over time, with a hazard ratio of 3.63 compared with those identified as low risk.¹

Importantly, investigators found that the model maintained predictive accuracy across several clinically meaningful subgroups, including triple-negative breast cancer and HER2-positive disease. These subtypes currently lack National Comprehensive Cancer Network-recommended genomic assays for recurrence risk assessment, highlighting a potential gap the AI model could help address if validated prospectively.^1,3

Comparison With Current Genomic Testing

The investigators also directly compared the AI test with the widely used 21-gene Oncotype DX assay in 858 patients who had undergone both evaluations. Although confidence intervals overlapped, the AI model demonstrated numerically higher prognostic discrimination overall, with a pooled C-index of 0.67 compared with 0.61 for Oncotype DX.¹

Additionally, the AI system reclassified many patients who received intermediate-risk Oncotype DX scores into either lower- or higher-risk categories. Among those with intermediate genomic scores, approximately 80% were categorized as low risk by the AI model, whereas nearly 20% were identified as high risk, suggesting the technology could potentially improve treatment stratification in this challenging population.¹

The model's foundation is self-supervised learning, enabling it to identify meaningful morphologic patterns from millions of pathology image patches without relying solely on manually labeled datasets.

Potential Implications for Oncology Practice

Although additional validation in randomized clinical trials will be necessary before clinical implementation, investigators believe the technology could streamline recurrence-risk assessment while reducing costs and preserving valuable tissue samples.

Because the model relies on digital pathology slides already generated during routine diagnosis, results could potentially be available within hours rather than weeks. The investigators also estimated that digital pathology-based testing may be substantially less expensive than current genomic assays, which can further expand prognostic assessment to patient populations not currently served by available molecular tests.^1,2

For pharmacists involved in breast cancer care, more rapid and accessible recurrence-risk information could eventually support earlier multidisciplinary treatment discussions and more individualized decisions regarding adjuvant systemic therapy. As AI continues to expand within oncology, tools that complement traditional pathology and clinical assessment may help improve precision medicine while increasing access to personalized cancer care.

REFERENCES

1. Witowski J, Zeng KG, Cappadona J, et al. Multi-modal AI for comprehensive breast cancer prognostication. Nat Commun. 2026;17(1):5879. Nat Commun. 2026;17:5879. doi:10.1038/s41467-026-73088-y

2. New York University. Researchers develop AI test to predict recurrence of breast cancer. News Release. EurekAlert! July 6, 2026. Accessed July 9, 2026. https://www.eurekalert.org/news-releases/1134931

3. National Comprehensive Cancer Network. NCCN Clinical Practice Guidelines in Oncology 2026: Invasive Breast Cancer. Accessed July 9, 2026. https://www.nccn.org/patients/guidelines/content/PDF/breast-invasive-patient.pdf