Presentation (CTI)

OBJECTIVES: As utility of real-world data increases, understanding the limitations and differences across data sources is important. In breast cancer, treatments associated with ER and HER2 in claims databases are often used as a proxy to ER and HER2 status. We evaluated concordance between ER and HER2 status derived from a claims-based algorithm and abstracted clinical notes.
METHODS: Patients were identified from the GuardantINFORM™ clinical-genomics database, which links cell-free circulating tumor DNA results to de-identified claims data. Clinical notes submitted at the time of a Guardant test were abstracted with Mendel.ai’s LLM and symbolic reasoning tools with human review. Adult mBC patients in the US who received at least one Guardant test from July 2014 to June 2024, after metastatic diagnosis and before first or second line (1L, 2L) treatments were included. From clinical notes, ER and HER2 results from FISH, IHC or tissue testing were utilized. Claims-based algorithm identified ER and HER2 status based on receipt of ER and HER2 therapies. Concordance was compared using Cohen’s Kappa statistic.
RESULTS: 812 patients had at least one Guardant test prior to 1L and 972 were tested prior to 2L. There was substantial agreement of ER status between claims and clinical notes (1L: k=0.68; 2L: k=0.64). Concordance of HER2 status was almost in perfect agreement in 1L (k=0.91), lower in 2L (k=0.78). Patients identified as ER+ve based on clinical notes but ER-ve based on claims primarily received chemotherapy or HER2 therapies.
CONCLUSIONS: We demonstrated that concordance of ER and HER2 status between claims-based algorithm and clinical notes was high, highlighting the value of treatment information for identifying ER or HER2 positivity. However, one should be cautious when using claims-based algorithm only to identify hormone status, especially in settings where neoadjuvant or adjuvant therapies are not well distinguished from metastatic treatment lines.