Practical use case of natural language processing for comprehensive atrial fibrillation cohort characterization from electronic health records in Poland

Abstract

Abstract Background Electronic health records (EHR) contain data valuable for clinical research but mostly in textual format. Natural language processing (NLP) is a computational technique that allows text analysis. Purpose To demonstrate a practical use case of NLP in Polish EHR for a large retrospective study cohort characterization and compare it to a standard, manual retrieval by a human. Material and methods The dataset consisting of anonymized discharge documentation of 10314 patients from the cardiology tertiary care department in Poland was analyzed for inclusion in the registry of patients with oral anticoagulation for atrial fibrillation (AF). Extensive clinical characteristics regarding: 73 concomitant diseases, medications from 21 classes, and 15 numeric echocardiography parameters were collected manually and through NLP. Results There were 2624 and 2601 patients fulfilling study inclusion criteria according to human and NLP-based approaches, respectively, reflecting 99.5% accuracy of NLP in identifying correct study participants. The calculated CHA2DS2VASc and HASBLED scores based on both methods did not differ (human vs NLP; median, IQR, p-value): 3 (2-4) vs 3 (2-4) p=0.74 and 1 (0-1) vs 1(0-1) p=0.98. Despite numeric differences, no statistically significant differences in any of the analyzed clinical characteristics entities were found. Therefore, similar conclusions on cohort characteristics would be made in a shorter time. Conclusions NLP utilization in Polish EHR may accelerate acquisition and provide accurate data for a retrospective study.Patient flow.