Standardizing in-hospital cause-of-death data using large language models and neural machine translation

This study aims to expand the Observational Medical Outcomes Partnership Common Data Model (CDM) by linking in-hospital mortality data with national mortality statistics. Specifically, we developed a pipeline utilizing large language models (LLMs) and neural machine translation (NMT) to standardize unstructured, bilingual cause-of-death text. From a cohort of 1,033,461 patients, 64,034 mortality records were identified. These records were converted into a standardized CDM format by mapping Korean Standard Classification of Diseases codes to the international standard, Systematized Nomenclature of Medicine—Clinical Terms. The pipeline structured 1,702 cases of unstructured, bilingual (Korean–English) free-text records containing medical abbreviations. The English translation performance of NMT- and LLM-based models was compared and subsequently validated against a ground truth reviewed by medical professionals. Finally, we employed retrieval-augmented generation-based LLM prompts to automatically assign International Classification of Diseases, Tenth Revision (ICD-10) codes to the translated text, identifying neoplasms (35.89%) and circulatory diseases (17.70%) as the primary causes of death. Technically, Nllb-200-1.3B achieved the highest translation accuracy, whereas SOLAR-10.7B and Llama 3.1-8B demonstrated superior performance in ICD-10 code matching. We successfully established a comprehensive mortality database and demonstrate that LLM-based modules are highly feasible for standardizing clinical big data, offering a scalable solution for enhancing real-world medical data interoperability.