2B-04 - Leveraging natural language processing (NLP) for the detection of drug-induced anaphylaxis in clinical notes

Background: Drug-induced anaphylaxis is a life-threatening hypersensitivity reaction and requires timely detection to improve patient safety. However, incomplete and uncoded documentation in electronic health records hinders identification. Natural language processing (NLP) offers a solution by extracting meaningful information from unstructured clinical notes and standardizing it for integration into clinical decision support systems.

Objectives: To identify drug-induced anaphylaxis using NLP, compare its detection in unstructured clinical notes with structured data from a coded hypersensitivity tool, and map free-text findings to SNOMED CT codes to enable clinical decision support.

Methods: We applied the open-source iKnow NLP tool to analyze clinical notes from Universitair Ziekenhuis Brussel between May 2022 and October 2023, across all patients and departments. iKnow extracts semantic concepts and relationships from unstructured text without requiring predefined dictionaries or ontologies. This is the first application of iKnow for identifying drug-induced anaphylaxis. We began with exploring, selecting and validating relevant NLP concepts. We then identified relationships between anaphylaxis and medication terms using a bottom-up approach. Findings from unstructured text were compared to structured data from the coded hypersensitivity module (implemented in May 2022), and extracted terms were mapped to SNOMED CT codes.

Results: The iKnow NLP tool identified 2,865 anaphylaxis-related concepts, of which 378 were excluded after validation. After filtering for the study period, anaphylaxis-related terms appeared in 600 clinical documents from 921 patients. The analysis found 1,877 potential causative NLP concepts, with 352 confirmed as medication-related. Ultimately, 567 documents contained drug-induced anaphylaxis events, exceeding the 70 cases recorded in the structured hypersensitivity tool. All selected medication and anaphylaxis concepts were successfully mapped to SNOMED CT codes.

Conclusions: NLP enhances the identification and coding of drug-induced anaphylaxis from unstructured clinical data, improving patient safety and ADE prevention. With iKnow's ability to analyze over 30 years of clinical notes and expand to other hypersensitivity reactions, future research can leverage this approach to develop a comprehensive dataset for prediction modeling, risk factor analysis, and clinical decision support.