Design and Integration of a Relation Extraction Component for Adverse Drug Events Detection in Swedish Clinical Notes

Medications are essential for health care but can cause adverse drug events (ADEs), which are harmful and sometimes fatal. ADEs are often not documented in structured data of electronic health record (EHR) or explicitly written in clinical notes. Detecting them requires advanced Natural Language Processing (NLP) techniques, including Named Entity Recognition (NER) and Relation Extraction (RE). This is particularly challenging in Swedish clinical text, where compound words, non-contiguous phrases representing a single concept are common and not well-supported by standard models. This study aims to fine-tune the SwedClin-BERT model for relation extraction which will work with the compound words (split entities), integrate it with an existing medical named entity recognition pipeline, and build an end-to-end ADE detection pipeline capable of operating at both the sentence and document level. Finally, the study intends to evaluate the performance of the RE component, and integrated NER-RE pipeline through controlled experiments on annotated clinical data from the Geriatric Clinic in Kalmar. Three experiments were conducted: (1) 5-fold cross-validation of the RE component using annotated relation data with stratified splits for split entities on 850 clinical notes; (2) document-level ADE classification using the integrated NER-RE pipeline on 255 annotated notes; and (3) generalization of NER-RE pipeline testing on 1200 unannotated labeled The RE model achieved a macro-avergae F1-score of 0.91 in cross-validation. The full pipeline reached 100% agreement with the ground truth on the annotated dataset but showed limited generalizability on the unannotated dataset (F1: 77.47%, specificity: 41.45%), primarily due to error propagation from NER model. The proposed NER-RE pipeline enhances ADE detection performance in structured evaluation but suffers from low specificity on broader datasets due to biased training data. Future work should focus on diversifying data sources and revising the annotation rules to enhance model robustness.