CDE-Mapper: Using retrieval-augmented language models for linking clinical data elements to controlled vocabularies.

Area of Science:

• Health Informatics
• Computational Linguistics
• Clinical Research Informatics

Background:

• Standardizing clinical data elements (CDEs) is crucial for consistent patient information across healthcare systems.
• Current methods struggle with CDEs of varying representation and complex structures, hindering data integration and interoperability.
• Automating CDE standardization is essential for advancing clinical research and decision support.

Purpose of the Study:

• To present CDE-Mapper, a novel framework for automating the alignment of CDEs with controlled vocabularies.
• To address the challenges of standardizing complex and variably represented CDEs.
• To improve data harmonization and support clinical decision support systems.

Main Methods:

• Utilized a retrieval-augmented generation strategy with large language models.
• Implemented query decomposition to handle CDE complexity.
• Integrated expert rules via prompt engineering and in-context learning with multiple retrievers.
• Developed a human-in-loop validated knowledge reservoir for concept linking.

Main Results:

• CDE-Mapper demonstrated an average of 7.2% higher accuracy improvement over baseline methods across four diverse datasets.
• Achieved accurate concept linking while minimizing computational costs.
• Successfully automated the alignment of CDEs with controlled vocabularies.

Conclusions:

• CDE-Mapper offers a powerful solution for automating CDE standardization, overcoming limitations of existing methods.
• The framework significantly enhances data harmonization, crucial for clinical research and decision support.
• Advanced language models show great potential in improving clinical data interoperability and research capabilities.