Large-scale neural biomedical entity linking with layer overwriting.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Entity linking is crucial for standardizing biomedical information, but challenges arise from the vast number of concepts and variations in entity mentions.
• Existing neural methods for entity linking require extensive training data, which is difficult to obtain for the millions of biomedical concepts.

Purpose of the Study:

• To develop a novel neural model for biomedical entity linking that addresses the limitations of sparse training data.
• To improve the accuracy and efficiency of linking biomedical entity mentions to database entries.

Main Methods:

• A pure neural classification model was developed, incorporating layer overwriting to surpass performance ceilings.
• Training data augmentation using database entries was employed to compensate for insufficient data.
• A cosine similarity-based loss function was utilized to effectively distinguish between millions of biomedical concepts.

Main Results:

• The system achieved first place in the n2c2 2019 Track 3 challenge, linking clinical entity mentions to 434,056 Concept Unique Identifier (CUI) entries.
• The model demonstrated effectiveness on the MedMentions dataset (3.2M concepts) and achieved new state-of-the-art performance on the NLM-CHEM corpus (350K concepts).

Conclusions:

• The proposed neural model effectively handles sparse training data in biomedical entity linking.
• The method offers a robust solution for accurately classifying and linking a vast number of biomedical concepts.