Abstract P2-10-14: HER2 evolution in breast cancer: Evidence from an AI-assisted real-world data analysis of 3 million women in a diverse urban population

Abstract Introduction: HER2 positive status gives breast cancer patients a wide range of treatment options. Previous HER2-targeted therapies did not show benefit in patients with low HER2 expression who were treated in the same way as HER2 negative cancers. Regulatory approval of trastuzumab deruxtecan for patients with low HER2 expression addresses a significant unmet medical need and impacts HER2 status reporting. As part of a qualitative exploratory analysis we utilized a novel AI-based approach to analyze a large unstructured real-world dataset, providing insights into biomarker testing and real-world prescribing practices. We found a number of patients with multiple HER2 statuses reported at different points in their patient journey. Methods: A retrospective analysis of Real-World Data (RWD) was performed on 3 million females aged over 18, using data from a health information exchange (HIE) that includes six urban health systems in the US, varying significantly in size and activity profile. A total of 25 million clinical documents - principally visit and discharge summaries - were analyzed using Natural Language Processing (NLP) specifically optimized for medical language. This analysis generated a curated cohort of 17,125 female breast cancer patients diagnosed and actively treated between August 2021 and August 2023 at three of the six health systems, which contributed most of the cohort. Patients were included in the curated cohort if their data showed evidence of at least two relevant biomarker tests, such as ER/PR, HER2, BRCA1/2, PIK3CA, AKT1, or PTEN. Patient records demonstrating discordant results where the treating physician had recorded HER2 low, HER2 positive and HER2 negative findings either at the same time or at different times were manually examined. Patients who had been described as: “HER2 low”; “HER2: 1+”; “HER2: 2+ with no amplification by FISH”; “HER2: 2+, negative by FISH” or variations of these statements with the same meaning were considered HER2 low. Patients with “HER2: 0” or “HER2 negative” were considered HER2 negative. Patients with “HER2: 2+” but no qualifying reference to the result of a FISH test were considered HER2 positive. Results: Distribution by age group and stage in the curated cohort of 17,125 patients closely matched national statistics. Evidence of biomarker testing was strongest for ER/PR, then HER2, then BRCA1/2, and finally PIK3CA, AKT1, or PTEN in descending order of frequency, consistent with the evolution of breast cancer care. A total of 55 patients with discordant results showing all three possible values of HER2 status were found. Transitions from HER2 positive to either HER2 low then HER2 negative or to HER2 negative then HER2 low occurred at a similar rate to transitions to HER2 positive from either of the other statuses. Conclusions: AI-supported analyses of very large RWD datasets can be conducted rapidly, revealing patterns of evolving disease and practice across entire populations and within subgroups with complex inclusion/exclusion criteria. Our study shows that reporting of HER2 positive, HER2 low and HER2 negative status in tumor samples from the same patient is more common and dynamic than previously thought. Further research is needed to understand the evolution of HER2 status, including its relationship to disease progression and treatment with potentially significant implications for clinical practice. Citation Format: Maylis Larroque, Shunmugam Mohan, Jean-Patrick Tsang, Tielman Van Vleck, Hadrian Green, Tessa Williams, Rob Hale, Otis Jackson, Scott Briercheck, Chris Tackaberry, Erin Gardner, Stefan Santos, Francisco Esteva. HER2 evolution in breast cancer: Evidence from an AI-assisted real-world data analysis of 3 million women in a diverse urban population [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2024; 2024 Dec 10-13; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(12 Suppl):Abstract nr P2-10-14.