SapBERT-Based Medical Concept Normalization Using SNOMED CT.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Word embeddings are fundamental in Natural Language Processing (NLP).
• Contextualized embeddings have recently shown high performance in various NLP tasks.
• Medical concept normalization is essential for standardizing clinical terminology.

Purpose of the Study:

• To evaluate the effectiveness of contextualized versus non-contextualized word embeddings for medical concept normalization.
• To compare the performance of different embedding types in mapping clinical terms to SNOMED CT.

Main Methods:

• Utilized k-Nearest Neighbors (k-NN) approach for concept mapping.
• Compared performance using both non-contextualized and contextualized word embeddings.
• Assessed performance via F1-score for mapping clinical terms to SNOMED CT.

Main Results:

• Non-contextualized embeddings achieved a significantly higher F1-score (0.853).
• Contextualized embeddings yielded a substantially lower F1-score (0.322).
• Non-contextualized embeddings demonstrated superior performance in this specific task.

Conclusions:

• Non-contextualized word embeddings are more effective for medical concept normalization than contextualized ones.
• The choice of embedding type critically impacts the accuracy of clinical term mapping to SNOMED CT.
• Future research should consider non-contextualized embeddings for similar medical NLP applications.