Researchers use LLMs to standardize bilingual mortality data

Researchers processing 1,033,461 patient records identified 64,034 mortality entries to test a new data standardization pipeline linking in-hospital deaths with national statistics.

Key Clinical Takeaways:

• A pipeline using large language models and neural machine translation successfully structured 1,702 bilingual free-text cause-of-death records containing abbreviations.
• Nllb-200-1.3B achieved the highest translation accuracy, while SOLAR-10.7B and Llama 3.1-8B excelled at ICD-10 code matching.
• Neoplasms accounted for 35.89% of identified deaths, and circulatory diseases represented 17.70% within the standardized database.

Expanding the Common Data Model for Mortality Tracking

The study aims to expand the Observational Medical Outcomes Partnership Common Data Model by linking in-hospital mortality data with national mortality statistics. Investigators developed a pipeline utilizing large language models and neural machine translation to standardize unstructured, bilingual cause-of-death text. From the total cohort of 1,033,461 patients yielding 64,034 mortality records, the system converted the data into a standardized format by mapping Korean Standard Classification of Diseases codes to the international standard, Systematized Nomenclature of Medicine—Clinical Terms.

Validating Neural Translation and Code Mapping Performance

The pipeline structured 1,702 cases of unstructured, bilingual Korean-English free-text records containing medical abbreviations. Researchers compared the English translation performance of neural machine translation and large language model approaches, subsequently validating the output against a ground truth reviewed by medical professionals. Retrieval-augmented generation-based prompts then automatically assigned International Classification of Diseases, Tenth Revision codes to the translated text.

Identifying Primary Causes of Death in the Cohort

Through automated coding and database integration, neoplasms emerged as the primary cause of death at 35.89%, followed by circulatory diseases at 17.70%. Technical evaluation showed that Nllb-200-1.3B achieved the highest translation accuracy among tested models. Meanwhile, SOLAR-10.7B and Llama 3.1-8B demonstrated superior performance in matching the resulting text to standard diagnostic codes. The project established a comprehensive mortality database, demonstrating that LLM-based modules offer a scalable solution for enhancing real-world medical data interoperability.

Disclaimer: The information provided in this article is for educational and scientific communication purposes only and does not constitute medical advice.