Setting the basis for a natural language processing tool to improve the reporting of notifiable diseases: a descriptive study

Timely reporting of notifiable diseases is crucial for public health, but it is often delayed by semimanual International Statistical Classification of Diseases and Related Health Problems Tenth Revision (ICD-10) coding of electronic health records (EHRs). Natural language processing (NLP) could enhance timely reporting by identifying potential cases of patients in unstructured clinical documents. This study aimed to describe the clinical documents containing key terms that suggest notifiable diseases for informing the future development of an NLP pipeline in a tertiary hospital, and this study also aimed to evaluate the current semimanual hospital notification system. This retrospective observational study included all the notifiable disease episodes according to Catalonian regulations reported between May and July 2023 from a tertiary hospital in Barcelona. Information on age, sex, notification-related variables, and the healthcare area where suspicion was determined was obtained from the hospital surveillance database. The adequacy of disease suspicion on the reporting date was assessed based on the regional public health criteria. For correct and timely notifications, the frequency of the presence of clinical documents and the frequency of their inclusion of terms that raise suspicion were described. Notification delay since the first date of evidence of suspicion in the EHR was calculated. Results were presented globally and stratified by 10 disease groups, with eight being communicable disease groups. A total of 284 episodes were notified. The most frequent group was sexually transmitted diseases (25.8%), with the outpatient clinic being the healthcare area where most notifications were found (145, 51.2%). Adequate notifications accounted for 210 (74.2%) episodes, including 25 diseases. Incorrect notifications were obtained mainly due to the reporting of chronic conditions as onset diseases (34.2%). Up to 25% of notifications were delayed for 3 days or more. Terms that raise suspicion were mostly found in medical notes (93.0%-78.7%). The unreleased clinical draft, the clinical course, and the microbiology reports were available for over 50% of the episodes of six or more disease groups, with a high inclusion of such suspicion-raising terms (81.1%-94.8%). The results revealed suboptimal reliability and faster reporting in the current reporting process. As the goal was to identify notifiable diseases before the confirmation of diagnosis, suspicion-raising terms were frequently found in unstructured clinical notes. Hence, the integration of the described documents into an NLP-based workflow could improve the reporting process.