Overview of MedProcNER task on medical procedure detection and entity linking at BioASQ 2023
Abstract
Recent advances in NLP techniques, the use of large language models and Transformers are showing promising results for processing clinical content. The development of tools for automatic recognition of medical concepts, variables, and clinical expressions is key for the semantic analysis of clinical records, semantic search engines and the generation of structured data representations. Despite the importance of medical procedures for management, diagnosis prevention and prognosis, there are few comprehensive resources for medical procedure extraction and normalization. In order to foster the development of procedure mention detection and entity linking systems, we have released the MedProcNER (Medical Procedures Name Entity Recognition) corpus, a high quality, manually annotated collection of 1000 clinical case reports written in Spanish. The corpus has been exhaustively labeled by physicians following detailed annotation guidelines and quality control measurements. Additionally, a multilingual Silver Standard corpus has also been generated for English, Italian, French, Portuguese, Romanian, Dutch, Swedish and Czech, to provide a clinical NLP resource for research in these languages. A total of 9 teams from 8 different countries have participated in the MedProcNER track of BioASQ 2023 (part of CLEF 2023), using mostly Transformers architectures and models like RoBERTA, BioMBERT, ALBERT, Longformers or SapBERT. MedProcNER was structured into three sub-tracks: a) Clinical Procedure Entity Recognition task, b) Clinical Procedure Normalization task to SNOMED CT and c) Clinical Procedure-based Document Indexing task. The MedProcNER corpus, guidelines, and resources (including cross-mappings to MeSH and ICD-10) are freely available at: https://zenodo. org/record/7929830
