Artificial intelligence–enabled analysis of unstructured EHR data for identification of actionable HER2-low breast cancer phenotypes.

Abstract

e13029 Background: Antibody-drug conjugates (ADCs) targeting HER2 have shown clinical effectiveness in breast cancer, even at low levels of HER2 expression, thereby widening patient eligibility. Detailed HER2 results from immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH) is often available only in unstructured clinical notes and pathology reports, leading to underestimation of HER2-low disease prevalence when relying on structured electronic health record (EHR) data alone. Artificial intelligence (AI) and machine learning methods can enable the identification of HER2-low disease by permitting the curation of data from unstructured notes. We compared HER2 status derived from 1) structured EHR fields and 2) AI-curated data from unstructured clinical documentation. Methods: ConcertAI Precision360 is a US nationwide, validated, AI-powered platform that utilizes advanced language models to deliver insights from structured and unstructured machine-curated EHR clinical data. This study focused on a subset of Precision360 breast cancer patients diagnosed on or after 1/1/2014 with HER2 results within 180 days of initial diagnosis from both 1) EHR-derived, native, structured data and 2) AI-curated data from unstructured EHR clinical notes. HER2 results as defined by the American Society of Clinical Oncology/College of American Pathologists guidelines within 180 days of initial diagnosis were included and classified as positive (IHC 3+ or IHC2+/FISH positive), low (IHC 2+/FISH negative or IHC 1+) negative, equivocal, or unknown. Results: HER2 status was available in both structured and unstructured/AI-curated data sources for 27,126 patients. HER2-low rates for structured, unstructured/AI-curated, and the combined data sources of both were 10%, 28%, and 34%, respectively. HER2-negative rates were 73%, 56%, and 52%, while HER2-positive status did not vary between the three groups (13%). Conclusions: AI-based curation of unstructured clinical notes reclassified a subset of patients, increasing the number of HER2-low patients and reducing those labeled as HER2-negative. Unstructured clinical documentation contained substantially more detailed and higher-resolution HER2 information than structured EHR fields alone. This approach can increase the identification of patients eligible for HER2-targeted therapies HER-2 status distribution from structured EHR, AI-abstracted unstructured EHR, and both sources for breast cancer patients diagnosed from 2014-2025. Structured HER(n-27,126) Unstructured/AI Curated EHR (n-27,126) Structured and unstructured EHR (n-27,126) HER2 status N % N % N % Positive 3454 12.7 3498 12.9 3575 13.2 Low 2803 10.3 7703 28.4 9328 34.4 Equivocal 583 2.1 290 1.1 121 0.4 Negative 19848 73.2 15084 55.6 13984 51.6 Unknown 438 1.6 551 2.0 118 0.4