A Neural Embedding Approach to Mapping Health Concepts to Concept Unique Identifiers

Understanding health concepts in free text is an important task in biomedical NLP. Being able to map the extracted concepts to unique concept identifiers can facilitate integration of and interoperability across biomedical informatics applications. Advancement of pretrained large language models made it possible to identify health concepts in free text with a high degree of accuracy. However, they lack the ability to map the concepts to unique identifiers correctly. In this study we investigated a neural embedding approach to mapping health concepts to the Unified Medical Language System (UMLS) Metathesaurus' Concept Unique Identifier (CUIs). A vector store containing the embeddings of 57819 unique concepts and corresponding CUIs was created, and a collection of annotated COVID-19 signs and symptoms was tested on 3 combinations of neural embeddings and vector stores. The results show that the neural embedding approach does significantly outperform the baseline string match method by >200%, which is very encouraging. However, its performance will need to be further improved for integration with large language models (LLMs).