Translating UMLS Concepts to Improve Medical Entity Linking in French: A SapBERT-Based Approach

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Medical Entity Linking (MEL) is crucial for analyzing unstructured clinical data.
• SapBERT is an effective English MEL tool, but its French version lacks performance due to limited training data.
• Existing cross-lingual methods struggle with French medical concept identification.

Purpose of the Study:

• To investigate training SapBERT on machine-translated French medical terminologies.
• To improve the accuracy of French Medical Entity Linking.
• To overcome the challenge of insufficient French medical training data for SapBERT.

Main Methods:

• Utilized machine-generated French translations of the UMLS Metathesaurus.
• Trained the SapBERT model using these translated French medical terms.
• Evaluated the model's performance on the QuaeroFrenchMed benchmark dataset.

Main Results:

• The proposed training strategy significantly outperformed the standard Cross-lingual SapBERT.
• Achieved superior performance in SNOMED-CT Medical Entity Linking for French.
• Demonstrated the effectiveness of using translated terminologies for low-resource languages.

Conclusions:

• Training SapBERT on translated UMLS Metathesaurus is a viable strategy to enhance French MEL.
• This method effectively addresses the data scarcity issue for French medical concept codification.
• The findings offer a practical solution for improving cross-lingual medical NLP applications.