퇴원 요약 작성을 위한 AI

A discharge summary tells the next clinician what happened and what to do now. Core content includes: the reason for admission and final diagnoses; a hospital course organized by problem; procedures and complications; significant results; discharge medications, marked as new, changed or stopped, with reasons; allergies; condition at discharge; code status; follow-up appointments; and tests still pending. AI tools draft this by feeding a large language model the whole encounter, or the most relevant parts of it, then asking for a structured summary. That is hard. A multi-day stay produces many notes full of copy-forward text, contradictions and working hypotheses that were later dropped. The main accuracy risks are: Omission. The model leaves out a complication, a medication change or a pending result. This is the most dangerous error because nothing on the page looks wrong; Hallucination. The model states a finding or plan that appears nowhere in the record; Stale information. An early differential diagnosis appears as the final diagnosis because it kept being copied forward; Medication errors. The model uses the home medication list from admission instead of the discharge orders; and Lost nuance. Details such as 'patient declined' or 'per family preference' disappear. Studies comparing model-drafted and physician-written summaries have often found the drafts readable and concise, but reviewers still find omissions and inaccuracies. That is why a physician review stays mandatory. A common misconception is that the summary must be complete because the AI 'read everything'. It cannot include what was never documented, and it may drop items to stay concise. A practical review checklist: check each discharge medication against the orders, including held and stopped drugs; Confirm the diagnoses are final, not working hypotheses; List every pending result and who owns it; Confirm follow-up appointments; Verify allergies and code status; and Remove any statement you cannot trace to a source.