OBJECTIVES: French National Health Insurance Database (SNDS) is a claims database commonly used to generate Real-World Evidence. One of the main limitations is patient selection as it may rely on partial or inconsistent information. This research aims to validate an approach using natural language processing (NLP)-powered solution on EHR to evaluate the SNDS algorithm and improve patient selection with metastatic breast cancer and HER2+ status : metastasis may especially be inconsistently coded, having no impact on the hospital’s remuneration.

METHODS: Data is obtained from EHR from patient hospitalized in 3 participating sites (CHU Reims, CH Troyes, GHNE) between 2017 and 2021 with or for breast Cancer. A cohort comparison is made between the results obtained using the SNDS algorithm, and the NLP-powered solution (using semantic elements and AI). A manual quality control (verification of positives and negatives) was then carried out by doctors to legitimize and interpret the results.

RESULTS: With an initial population of 3079 at the participating site, the NLP solution allowed to identify 999 patients with non ganglionary metastases while the SNDS-only algorithms allowed to identify 878. Manual analysis implied a sensibility/specificity ratio of 92%/99% (SNDS) and 99%/95% (NLP). Moreover, metastasis are usually coded in a timely manner (88% within 2 months). The combination of metastasis coding and anti-HER2+ drug dispensation allows to identify the population with HER2 gene’s presence with a good accuracy (sensibility/specificity 78%/98%), because these drugs are part of the en-sus list (allowing full financial support by the national health insurance). Few HER2+ patients are not found (no anti-HER2+ treatment : patient refusal, age, contra-indication, or lack of metastasis coding).

CONCLUSIONS: The ICD-10 coding of metastases is accurate, validating the usability of SNDS-algorithm for identification of metastatic breast cancer, however, the use of NLP-powered solution allowed to quickly improve patient selection with minimal reduction in specificity and much richer data.