A machine learning approach for automating review of a RxNorm medication mapping pipeline output

Area of Science:

• Health Informatics
• Artificial Intelligence in Healthcare
• Pharmacovigilance

Background:

• Standardized medication terminology is crucial for analyzing data across federated electronic health record (EHR) networks.
• TriNetX operates the world's largest federated EHR network, highlighting the need for efficient data harmonization.

Purpose of the Study:

• To introduce RxEmbed, a novel pipeline for mapping local EHR medication descriptions to RxNorm ingredient codes.
• To leverage Large Language Models (LLMs) and machine learning for automated mapping and review.

Main Methods:

• Developed RxEmbed, a pipeline utilizing LLMs to map local medication names to RxNorm ingredient codes.
• Incorporated machine learning for automated review of the generated mappings.
• Evaluated performance on a French public dataset and data from 6 US/Brazil healthcare organizations within the TriNetX network.

Main Results:

• RxEmbed demonstrated superior performance compared to existing LLM-based methods on a public dataset.
• In the TriNetX network, RxEmbed achieved high RxNorm mapping recalls (84%-93%) with exceptional precision (99.5%-100%).

Conclusions:

• A novel LLM-based pipeline, RxEmbed, effectively maps EHR medication data to RxNorm ingredient codes.
• The high precision of RxEmbed minimizes the requirement for manual intervention in the mapping process.
• This facilitates large-scale analytics on federated EHR data.