Researchers at Severance Hospital have developed an artificial intelligence system that automatically writes emergency room discharge records for physicians, potentially reducing administrative burden and allowing doctors to spend more time with patients.
Severance Hospital announced Wednesday that a research team led by Professor Kim Ji-hoon of the Department of Emergency Medicine and Professor Yoo Seung-chan of the Department of Biomedical Systems Informatics at Yonsei University College of Medicine has developed "Y-Knot," an AI model based on large language models (LLM) that generates ER discharge records.
Under current medical law, emergency room physicians are required to complete emergency patient care records, also known as discharge records, after treating patients. These documents must contain the entire treatment process, including the reason for the visit, test results, treatment details, progress, transfer status, and discharge decision rationale. While necessary for patient safety management and continuity of care, this requirement has added to the workload of doctors who must rapidly treat a constant stream of emergency patients.
The research team developed the AI model to address this challenge. Once the AI generates a draft record, physicians only need to review and verify it. Previous LLM-based AI models existed but posed risks of exposing sensitive patient health information because they used networks that could communicate outside the emergency room.
The newly developed AI model is designed based on an "on-site large language model" and a "lightweight transformer model (Llama3-8B)." This approach minimizes risks such as personal information leaks by enabling use within the ER's internal network without connection to external networks.
When the research team had six emergency medicine physicians at a 2,400-bed tertiary hospital in Korea use the AI model, the time required to complete emergency patient care records decreased by more than 50 percent. While physicians took an average of 69.5 seconds to write records manually, using the AI model reduced this to 32.0 seconds.
Records created with AI assistance were also rated higher in quality than those written manually by physicians. When three emergency medicine physicians were shown AI-assisted and manually written records in random order and asked to evaluate them on completeness, accuracy, conciseness, and clinical usefulness, they rated the AI-assisted records superior in all four categories.
"Emergency patient care record generation using the AI model proved far superior to manual writing in both speed and quality," Professor Kim said. "With the added safety of patient information through internal network use, physicians will be able to spend more time treating patients."
"As we continue to refine the system, final review by specialists remains essential," Professor Yoo said, adding that "the system could be applied not only in emergency medicine but also in other medical departments."
The research findings were published in the latest issue of JAMA Network Open, an international academic journal published by the American Medical Association.
