NeighBERT: Medical Entity Linking Using Relation-Induced Dense Retrieval.

Area of Science:

• Clinical Natural Language Processing
• Artificial Intelligence in Healthcare
• Bioinformatics

Background:

• Medical Entity Linking (MEL) is crucial for clinical NLP but is challenged by ambiguous text in Electronic Health Records (EHRs).
• Existing transformer models show promise but struggle with the inherent ambiguity of clinical language.
• Resolving ambiguity is key to improving the accuracy of information extraction from clinical notes.

Purpose of the Study:

• To introduce NeighBERT, a novel pre-training technique for transformer models designed to address ambiguity in clinical text.
• To enhance the performance of medical entity linking and named entity recognition.
• To improve the extraction of relevant medical information from electronic health records.

Main Methods:

• Developed NeighBERT, a custom pre-training technique extending BERT by encoding entity relationships from a knowledge graph.
• Integrated relational context into the transformer architecture to better resolve textual ambiguity.
• Evaluated NeighBERT on two standard clinical datasets for named entity recognition and medical entity linking tasks.

Main Results:

• NeighBERT demonstrated significant improvements in Named Entity Recognition (NER) precision, recall, and F1-score by 1-3 points.
• Medical Entity Linking (MEL) performance saw substantial gains, with improvements of 10-15 points in F1-score.
• The relational context encoding effectively reduced ambiguity in clinical text processing.

Conclusions:

• NeighBERT offers a powerful new approach to enhance transformer-based clinical NLP tasks, particularly MEL and NER.
• The method successfully incorporates relational knowledge, overcoming limitations of standard BERT in handling ambiguous clinical text.
• This advancement holds potential for more accurate and reliable information extraction from electronic health records.