Stony Brook University 研究發現，醫生在 95% 的住院個案中偏好 AI 生成的出院摘要

Physician reviewers preferred AI-generated discharge summaries over their human-authored counterparts in 95% of the hospitalizations evaluated in a Stony Brook University study. The retrospective analysis focused on 60 Internal Medicine admissions lasting seven to 21 days, where extended treatment and large volumes of clinical notes make the handoff to outpatient care especially demanding.

The findings, published in npj Digital Medicine, suggest that large language models could help clinicians produce more complete and usable discharge documents. They also show why human verification remains essential: both AI-generated and human-authored summaries contained errors, and higher ratings did not establish better patient outcomes.

Why Discharge Summaries Are a Difficult AI Task

A discharge summary connects care delivered in a hospital with the clinicians responsible for what happens afterward. It needs to explain the reason for admission, the course of treatment, medication changes, unresolved problems, and the follow-up plan. For a primary care physician, a missing explanation or unclear next step can mean reconstructing the hospitalization from a much larger record.

Longer admissions compound that problem. Information accumulates across progress notes, consultations, imaging reports, and other documentation, often with repetition and changes in the treatment plan. A useful summary must preserve the final clinical picture while explaining how decisions evolved.

Stony Brook’s approach was designed with clinicians around the information receiving physicians need. The university says the system organized patient records into discharge drafts within minutes. That describes generation time; the effect on physicians’ total documentation and verification time still needs testing in clinical practice.

What the Physicians Rated More Highly

The study paired assessments by hospitalists and primary care physicians. Reviewers gave the AI summaries higher ratings for overall quality, readability and conciseness, factual accuracy, and completeness. They also identified fewer missing clinical details.

On the supplied five-point ratings scale, the AI summaries averaged 4.71 for overall quality compared with 3.16 for human-authored summaries. Factual accuracy averaged 4.62 versus 3.73, while completeness averaged 4.69 versus 3.42. These are physician ratings, rather than percentages of statements proven correct.

The graphic reports 120 ratings per summary type for the combined hospitalist and primary care measures, and 60 per type for the primary care-specific questions. Primary care reviewers rated the AI summaries more highly for understanding the hospitalization, explaining it to patients and caregivers, and supporting follow-up care.

Those distinctions matter because the person writing a discharge summary and the person acting on it have different perspectives. A document can recount inpatient events thoroughly yet still leave the receiving clinician uncertain about pending issues or the rationale for a medication change.

How the Summarization Pipeline Handles Long Records

The researchers’ public code release adds context about the technology. Its exported LangChain/LangGraph pipeline groups notes chronologically, builds a structured summary, and refines that draft as it processes additional chunks. A separate stage extracts incidental findings from flagged notes before a final narrative is formatted.

This staged approach addresses a practical constraint of language models: a long clinical record may exceed the amount of text a model can process in one pass. Carrying a structured draft forward allows successive portions of the record to contribute to the result.

The public export also records source identifiers and produces citation metadata alongside structured and narrative outputs. It includes synthetic demonstration data, while the repository provides de-identified rating and error-annotation results for analysis. That supports examination of the workflow without publicly releasing the original patient records.

Fewer Omissions Do Not Eliminate Clinical Risk

The published abstract attributes the lower number of annotated inaccuracies or omissions primarily to fewer omissions. It reports no statistically significant difference between the two summary types in reviewers’ estimates of error-related harm severity or likelihood.

That finding does not demonstrate that the approaches have identical safety, nor that an AI draft is ready for a patient record without review. The evaluation concerned documents and physician judgments; it did not demonstrate reductions in readmissions, medication-related harm, or other patient outcomes.

The authors stress physician verification before use. Medication instructions, unresolved medical issues, and follow-up recommendations are central to the document’s purpose and remain consequential places for an error or omission.

What Comes Before Routine Clinical Use

“These results provide evidence that artificial intelligence can support discharge documentation for complex hospital stays,” said Rachel Wong, MD, a co-corresponding author and clinical associate professor at Stony Brook, in the university’s announcement.

Co-corresponding author Fusheng Wang, PhD, said the next research steps include assessing accuracy, safety, actual physician time savings, and fit within clinical workflows. The university says a pilot in everyday Stony Brook Medicine practice is being considered.

Led by computer science PhD candidate Tyler Osborne, the interdisciplinary project brings together physicians, computer scientists, biomedical informaticists, and communications expertise. The study provides encouraging evidence for clinician-supervised drafting of complex discharge summaries. Establishing how that translates into reliable daily use requires evaluating the review process alongside the generated text.