Korean law requires emergency departments to produce discharge notes documenting each visit's reasons, test results, procedures, clinical course and disposition, a legal necessity that consumes physician time at the exact point in the hospital where time is scarcest. A Yonsei University College of Medicine team led by emergency physician Kim Ji-hoon and biomedical informatics professor You Seng-chan, with fourth-year medical student Song Ji-woo, built an AI to absorb that burden, and named it Y-KNOT, for Your Knowledgeable Navigator of Treatment.
The design choice that distinguishes the system is where it runs. Many large language model tools require external network access, an unacceptable exposure for health data, so Y-KNOT operates entirely within the hospital network, pairing an internal LLM on hospital servers with a lightweight Llama 3-8B transformer. The AI drafts each note and the physician reviews and finalises it, keeping a board-certified doctor as the last word.
Deployed with six emergency physicians at a 2,400-bed tertiary hospital, the system cut average note-writing time by more than half, from 69.5 seconds to 32.0 seconds, and in blind scoring by three emergency physicians the AI-assisted notes rated higher than manually written ones on completeness, correctness, conciseness and clinical utility. Published in JAMA Network Open and announced in December 2025, the work offers a template for privacy-preserving clinical AI that other departments, and other hospitals, can plausibly copy.