Seoul National University Hospital has developed what could become South Korea’s first large-scale medical language model.
The project to develop medical LLM began last March. The SNUH research team began by collecting 38 million clinical texts, including hospitalizations, outpatients, surgery, prescriptions, and nursing records, and dismissed and anonymized them to create a basic corpus for model learning.
Then, at the start of 2025, it developed a department-specific knowledge base integrated into LLM through later retrieved generations (RAG). These knowledge bases consist of local Korean medical laws, paper summary, treatment guidelines, and medical terminology standards and abbreviations dictionary.
LLM was tested on questions from the Korean Medical License Survey over the past three years, earning 86.2%, with an average Taker score above 79.7%. Snuh also said that the model exhibits high translation performance and processes 50,000 words of text simultaneously.
SNUH research teams can verify the performance and safety of LLM for a year and then apply to aid clinical work and research. It also plans to expand applications across a variety of healthcare sectors and further enhance the model’s medical data processing capabilities.
Why is it important?
According to Snuh, existing medical LLMs developed by high-tech giants, including Google. Med-Palm 2 and Microsoft’s Llava-Med are limited to understanding Korean medical textbooks, laws, and treatment guidelines. These models are optimized in Western contexts, which proves to be unreliable for Korean health professionals. Therefore, hospital researchers have begun the development of Korean medical LLMs to meet the needs of local clinicians.
Bigger trends
Late last year, the startup, backed by the conglomerate SK Group, claimed to introduce the first LLM-based search platform unique to South Korea’s healthcare. Phynx Lab’s Cheiron helps pharmacists and pharmaceutical industry researchers obtain relevant information by picking up user intent through natural language processing. It also allows you to self-check information and quickly generate high-quality answers.
Anam Hospital, Korea University It is also reportedly developing its own LLM, which is scheduled to be piloted this year.
Asan Medical Center We also used LLM to create voice-based clinical scribes. It is currently used in 16 departments.
Meanwhile, Snuh is currently working on other AI projects, releasing clinicians from management work. These include multimodal AI, which can automatically generate patient summary, AI for enhancing claims processing, and AI, which curates the latest papers for researchers.
record
“Through this Korean-style medical LLM development, we have opened up a new chapter in medical innovation by maximizing work efficiency for medical staff and providing faster and more accurate medical services to patients.”
