Term Candidate Generation to Enrich Clinical Terminologies with Large Language Models

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Terminology

Background:

• Clinical routine documentation offers valuable data for secondary use.
• Identifying comprehensive terms for standardized codes like ICD-10 remains a challenge.

Purpose of the Study:

• To leverage annotated clinical data for identifying additional term candidates for ICD-10 codes.
• To bridge the lexical gap between real-world clinical jargon and standardized terminologies.

Main Methods:

• Log-likelihood analysis of co-occurring de-identified ICD-10 codes and problem list entries in German (approx. 1.9 million).
• Statistical significance threshold set at p < 0.01 to identify seed terms.
• Interfacing with a large language model to harvest additional term candidates.

Main Results:

• The approach successfully identified additional term candidates.
• Performance metrics included MAP@5 for hypernyms (0.801), synonyms (0.723), and hyponyms (0.507).

Conclusions:

• Re-using annotated clinical datasets with large language models is an effective strategy.
• This method enhances the completeness of clinical terminologies by incorporating real-world language.