Deep-learning-based automated terminology mapping in OMOP-CDM.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Interinstitutional medical data access is hindered by diverse vocabularies.
• Standardization efforts like the common data model require costly human oversight.
• Automating semantic code mapping is crucial for efficient data integration.

Purpose of the Study:

• To develop a trainable system for automated semantic interinstitutional code mapping.
• To overcome the limitations of manual standardization in healthcare data.
• To enhance the accuracy and efficiency of medical data harmonization.

Main Methods:

• Computed embedding-based semantic similarity between descriptive sentences for code mapping.
• Implemented a systematic approach for preparing training data for similarity computation.
• Compared deep learning-based semantic matching against traditional word-based mappings and the Usagi system.

Main Results:

• The proposed semantic matching method significantly outperformed the Usagi system.
• Achieved at least 10% greater matching accuracy compared to Usagi, consistent across top-k measurements.
• Demonstrated that incorporating contextual and semantic information improves mapping accuracy.

Conclusions:

• Deep learning-based semantic mapping surpasses traditional word-level algorithms by leveraging contextual information.
• The methodology for selecting negative training samples critically impacts system performance.
• The developed approach offers a more accurate and efficient solution for interinstitutional code mapping in medical data.