Enigma@ELCardioCC: Bridging NER and ICD-10 Entity Linking - A Hybrid Method for Greek Clinical Narratives

Boris Velichkov, Aleksis Datseris, Sylvia Vassileva, Svetla Boytcheva

Published: 01 Sept 2025, Last Modified: 28 Apr 2026CEUR Workshop ProceedingsEveryoneRevisionsCC BY-SA 4.0
Abstract: This paper presents an approach for the clinical term Named Entity Recognition (NER) and Entity Linking (EL) in Greek clinical texts. The approach was developed as part of the ELCardioCC shared task for clinical coding to the International Classification of Diseases, 10th edition (ICD-10). For the NER task, we used different BERT-based models, the monolingual Greek BERT and the multilingual XLM-RoBERTa. We adapted them to the biomedical domain by additional pretraining on biomedical texts in Greek. We further fine-tuned the models for token classification on the train set to determine the ICD-10 term mentions in the text. The best F1 score we achieved was 0.7167 on the test set. For the EL, we used a hybrid approach that combined two stages. The first stage was based on a gazetteer - exact match or statistical match to unambiguous terms in a gazetteer compiled from the train set, ICD-10 specification, and other public resources. The second stage was a fine-tuned bi-encoder model (BAAI/bge-m3), applied only to mentions that did not match anything in the first stage. Our best F1 score on this task was 0.6693.