Improving broad-coverage medical entity linking with semantic type prediction and large-scale datasets.

Area of Science:

• Biomedical Natural Language Processing (NLP)
• Computational Biology
• Medical Informatics

Background:

• Broad-coverage information extraction is crucial for analyzing scientific documents and clinical notes.
• Linking medical concepts to standardized vocabularies is challenging due to large concept inventories and multiple semantic types.
• Existing biomedical NLP tools face difficulties in accurately identifying and linking medical entities.

Purpose of the Study:

• To develop a novel semantic type prediction module for enhancing biomedical NLP pipelines.
• To create two large-scale, automatically constructed datasets (WikiMed and PubMedDS) for broad-coverage semantic type training.
• To improve the accuracy of medical entity linking in information extraction tasks.

Main Methods:

• Experimentation with five off-the-shelf biomedical NLP toolkits on four benchmark datasets.
• Introduction of a semantic type prediction module (MedType) to filter irrelevant candidate concepts during entity linking.
• Development of two novel datasets, WikiMed and PubMedDS, to address the lack of broad-coverage training data.

Main Results:

• Semantic type filtering significantly improved medical entity linking performance across all tested toolkits and datasets, with notable gains in F-1 scores.
• Pretraining the MedType model on the novel datasets achieved state-of-the-art performance in semantic type prediction for biomedical text.
• The proposed module and datasets enhance the accuracy of information extraction from biomedical literature and clinical notes.

Conclusions:

• Semantic type prediction is a vital component for accurate, broad-coverage information extraction in biomedical NLP.
• The developed MedType module and accompanying datasets offer valuable resources for advancing biomedical NLP research.
• Public availability of source code and datasets promotes reproducible research in the field.