Embeddings of clinical codes enable knowledge-grounded AI in medicine

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Informatics
• Clinical Data Science

Background:

• Standardization of electronic health records (EHRs) facilitates AI integration.
• Clinical codes within EHRs represent a rich source of medical information.
• Existing methods for representing clinical concepts may lack comprehensive integration.

Purpose of the Study:

• Introduce ClinVec, an embedding store for clinical codes and concepts.
• Develop ClinGraph, a knowledge graph tailored for clinical vocabularies.
• Demonstrate the utility of ClinVec embeddings in AI applications.

Main Methods:

• Generated embeddings for 153,166 clinical codes and concepts from eight vocabularies.
• Constructed ClinGraph with over 2 million edges from clinical vocabularies.
• Validated embedding similarity against an inter-institutional clinician panel and 3,767 clinical term pairs.

Main Results:

• Embedding similarity in ClinVec accurately reflects clinical relatedness across 11 disease areas.
• ClinVec embeddings were successfully used for knowledge injection in large language models.
• Demonstrated unsupervised patient stratification and risk prediction using ClinVec.

Conclusions:

• ClinVec offers a shared, knowledge-grounded representation of clinical concepts.
• The developed embeddings enhance AI systems for patient and population modeling.
• ClinVec supports the development of more sophisticated AI tools in healthcare.