Comparative Analysis of ChatGPT-4 for Automated Mapping of Local Medical Terminologies to SNOMED CT

Standardizing medical terminology is critical for healthcare informatics, particularly for improving data interoperability and patient management systems. This study evaluated four distinct GPT-4-based approaches for mapping local medical terminologies to SNOMED CT: baseline, prompt-engineered, fine-tuned, and Retrieval-Augmented Generation (RAG). Using 1,200 diagnostic terms from a Korean hospital, we assessed the models' accuracy and error rates. The RAG model achieved the highest performance with a 96.2% valid SNOMED CT term match rate and a 57.6% overall exact match rate, surpassing the fine-tuned model (69.2% valid term match, 47.2% exact match). Error analysis showed that the RAG model also minimized structural errors to 14%, significantly lower than other models. While the fine-tuned and RAG models struggled with specificity, they showed promise for improving automated mapping and Named Entity Recognition (NER) tasks in clinical settings. This study highlights the potential of AI-human collaboration for enhancing autocoding and data standardization in healthcare. Further research is needed to refine specificity and validate these systems for clinical use.