Presentation (CTI)

OBJECTIVES: The SNDS is one of the leading claim databases in Europe, covering the entire French population and patient pathways. Despite its advantages, evidence generation using the SNDS is constrained by the limited ability to select specific populations due to the lack of detailed ICD-10 coding, and potential coding delay. The use of NLP to analyze patients' EHR to identify these gaps could significantly improve SNDS studies without requiring time-consuming and costly human review. The aim of this study was to validate REALLI compared to manual review by physicians of EHR from patients with HER2+ metastatic breast cancer (BC).
METHODS: Retrospective analysis on EHR from BC patients treated in 3 general hospitals and 1 university hospital between January 1st, 2017, and December 31st, 2021. The performance of REALLI was tested on the classification of metastatic and non-metastatic patients, the diagnostic date of metastatic status, and on HER2+ status. Specificity and sensibility of REALLI was calculated by comparing it to a gold standard based on a manual control performed by physicians on a random sample size of 200 patients.
RESULTS: Between January 2017 and December 2021, 3214 BC patients were included. The sensibility [IC95] /specificity [IC95] for metastatic status and HER2+ status were 100% [94%;100%] / 95% [90%;98%] and 100% [66%;100%] / 100% [90%;100%] respectively. A concordance of 90% on the metastatic diagnosis dates was observed between REALLI and the manual review. Results were similar in each hospital.
CONCLUSIONS: REALLI was able to identify metastatic status and HER2+ status with high sensitivity and specificity. It also accurately detected the date of metastatic diagnosis. This method is a timesaving, cost-effective, and reliable alternative to manual review. Further studies involving various medical centers are necessary to validate the REALLI solution considering variability in coding practices.