The work focused on Hospital Clínic de Barcelona, a predominantly adult tertiary hospital with limited pediatric services. Because notification rules in Spain and Catalonia include both communicable and some noncommunicable conditions, the investigators included every notifiable-disease episode reported at the hospital during the three-month period, rather than selecting diseases in advance. The sample covered 25 conditions grouped into 10 categories, including sexually transmitted infections, mosquito-borne diseases, respiratory infections, hepatitis, zoonoses and asthma. Of the analyzed episodes, 173 involved males, 110 females and one person whose sex was recorded as unknown. Median age was 42 years, with an interquartile range of 25 years. One episode with no information in the clinical record was excluded from subsequent analyses, leaving 283 for assessment by healthcare setting and document type.

Sexually transmitted diseases formed the largest notified group, accounting for 73 episodes, or 25.8%. Most notifications originated in the outpatient clinic, which accounted for 145 episodes, or 51.2%; 96, or 33.9%, came from the emergency room and 42, or 14.8%, from hospital care. Medical doctors reviewed whether sufficient information was available on the reporting date to justify notification under Catalonia’s public health criteria. They classified 210 episodes, representing 74.2%, as adequately suspected and 73, or 25.8%, as inadequate. The investigators also calculated the interval between the first appearance of suspicion in the electronic health record and the actual report.

Among the adequately notified episodes for which relevant records could be identified, 170 were subsequently confirmed as having the reported condition. Thirty-three episodes, equal to 15% of the adequately notified group, had been reported before suspicion was recorded in the electronic health record and were excluded from the delay calculation. Among the remaining 177 episodes, the median delay was zero days and the interquartile range was three days. Sixty-two percent were reported on the same day that the first suspicion appeared, but 10% were reported more than two weeks later. The authors said explanations for reports preceding the electronic record, such as telephone calls or internal messages, remained hypotheses because the original notification source had not been registered.

The leading cause of inadequate reporting was the treatment of a chronic condition as a new reportable event, which explained 34.2% of the 73 inadequate notifications. References to a disease in a patient’s medical history accounted for 26.0%, while disease-related language whose clinical context did not support current suspicion accounted for 23.3%. A further 12.3% lacked supporting information, and 4.1% involved a mismatch between the disease reported and the condition suggested by the available records. Importantly, 83% of inadequate notifications contained at least one term from the disease-specific thesaurus. The results therefore show that the mere presence of a disease word or clinical concept is not enough to establish that an episode warrants notification.

Manual review of six types of electronic health record documents revealed where suspicion-related terms were most likely to appear. When the corresponding document was available, investigators found such terms in 93.0% of hospital discharge notes, 89.0% of unreleased clinical drafts, 84.7% of emergency-room discharge notes, 78.7% of clinical course reports, 63.3% of microbiology reports and 45.4% of laboratory reports. These percentages describe the content of available documents, however, and do not mean that every document type was equally accessible across episodes. Availability differed substantially by healthcare setting and disease group, making the importance of a source depend on where and how a patient was being treated.

The pattern changed markedly across care environments. In emergency-room episodes, emergency-room discharge notes contained suspicion-related information in 85.4% of cases, while unreleased clinical drafts did so in 83.8%. Clinical course reports were less frequent in that setting, and hospital discharge notes and microbiology reports were often unavailable. For hospitalized patients, clinical course reports contained relevant information in every episode in which they were assessed, with additional contributions from hospital discharge notes and unreleased drafts. Outpatient records had a different profile: clinical course reports contained relevant terms in 96.8% of episodes, and laboratory and microbiology reports also contributed, whereas the three other document types were not frequently available. These differences argue against a surveillance tool based on one fixed electronic record source.

Disease group also altered the contribution of each document. Clinical course reports were the most consistently available source, appearing in more than 80% of episodes in every disease group; when available, they contained suspicion-related terms in 81.2% of episodes. Unreleased clinical drafts were present in more than half the episodes of six disease groups and contained relevant terms in 94.8% of those available documents. Emergency-room discharge notes were representative for asthma, exanthematous viral diseases and mumps, and heatstroke, while hospital discharge notes were representative only for respiratory infectious diseases. Microbiology reports were available in at least half the episodes of six groups and contained relevant terms in 81.1% of available reports. Their contribution was particularly prominent for exanthematous viral diseases and mumps, foodborne and waterborne diseases, mosquito-borne diseases, and sexually transmitted infections.

The authors say the findings point toward a future system that selects document sources according to disease and care pathway, maps variations in language to standardized clinical concepts, and interprets those concepts in context rather than treating them as isolated keywords. A system would need to distinguish active suspicion from chronic disease, past history, negation and contextual mentions that do not support notification. The analysis was performed on raw clinical documentation, retaining abbreviations, misspellings, typographical errors, idiomatic expressions and colloquialisms; investigators manually identified and then standardized concepts for classification. A future system would also need expert review in a human-in-the-loop workflow, particularly because the study found overnotification as well as delays.

This work does not demonstrate that natural language processing or large language models can improve disease surveillance. No automated system was tested, and the researchers could not calculate the sensitivity, specificity or diagnostic performance of a future pipeline. The analysis included only episodes that had already been notified, rather than independently identifying cases that escaped reporting, and it did not include prior notification records needed to verify repeated reports. Its three-month summer sample may underrepresent conditions with different seasonal patterns, while the single hospital setting limits generalizability. Future research would need larger and more representative datasets, automated testing on unprocessed text, and direct measurement of accuracy and reporting time. Even with those limitations, the study offers a concrete starting point: the documents most likely to contain early signals, the conditions that generate false alerts, and the clinical pathways any future surveillance tool would have to understand.

Subject of Research: Evaluation of notifiable-disease reporting and clinical documents for future NLP-assisted hospital surveillance

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

Article References: Cano-Escobar, C., Barón-Miras, L. E., Gualda-Gea, J. J., Frid, S., & Bertran, M. J. (2026). Setting the basis for a natural language processing tool to improve the reporting of notifiable diseases: a descriptive study. Frontiers in Digital Health, 8, Article 1882377. https://doi.org/10.3389/fdgth.2026.1882377

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1882377

Keywords: natural language processing, notifiable diseases, public health surveillance, electronic health records, clinical informatics, disease reporting, hospital epidemiology, text mining, Barcelona, Setting, basis, natural