The Pulse
Stony Brook study finds clinicians favor AI discharge summaries 95%
A Stony Brook University study found physician reviewers preferred AI-generated discharge summaries in 95% of 60 hospitalizations lasting seven to 21 days. The retrospective evaluation also found fewer annotated omissions, but no significan
AI.info Team · Written with AI · fact-checked ·
Stony Brook University said on Monday, November 2, that physician reviewers preferred AI-generated discharge summaries in 95% of the hospitalizations examined in a new study. The research, published online in npj Digital Medicine on October 2, compared model-generated drafts with summaries written by clinicians for 60 internal medicine hospital stays.
The cases involved stays of seven to 21 days, where clinicians had a larger volume of records to condense for the handoff from hospital to outpatient care. Reviewers rated the AI summaries higher for quality, readability, factuality and completeness. But the results measure clinicians’ assessments of documents—not whether AI improved patient outcomes or reduced the time doctors spent working.
Longer stays put the handoff under scrutiny
Discharge summaries tell patients and the clinicians who will care for them what happened in the hospital, including treatments, medication changes, unresolved issues and follow-up needs. Missing or unclear information can complicate that transition. The study focused on longer, complex stays, rather than treating every hospital discharge as the same documentation task.
Researchers compared each AI-generated summary with the human-authored version for the same case, with paired assessments from hospitalists and primary-care physicians. That design let the reviewers consider the documents against the same underlying hospitalization. Primary-care physicians also rated whether the AI summaries helped them understand the hospital care, explain it to patients or caregivers, and plan follow-up.
Reviewers favored the AI draft, but errors remained
The 95% preference result is striking, but it does not mean the model produced error-free records. Reviewers annotated fewer factual inaccuracies or clinically relevant omissions in AI-generated summaries, with the difference driven mainly by fewer omissions. Yet the study found no significant difference between the two summary types in reviewers’ estimates of the errors’ potential severity or likelihood of harm.
“In this retrospective clinician evaluation, LLM-generated summaries for longer, complex hospitalizations were rated more favorably and had fewer annotated omissions than human-authored summaries, with no significant differences in estimated harm potential or likelihood.”
Tyler Osborne et al., “Clinician-centered evaluation of large language model-generated discharge summaries for longer hospitalizations”
The paper also reports that primary-care reviewers rated the model’s summaries more favorably for understanding and communicating hospital care and providing follow-up. The study’s authors examined 31 incidental findings identified by the model in radiology reports: reviewers judged 93.5% factually correct and 87.1% appropriate to include. Those figures concern this selected set of findings, not every statement in the discharge summaries.
One hospital’s retrospective test is not a deployment trial
The evaluation used historical hospitalizations at Stony Brook University Hospital. It therefore cannot establish how the approach performs prospectively across other hospitals, patient groups or clinical workflows. Nor does the reported summary-generation time establish real-world time savings for physicians; Stony Brook said further work will assess time saved and how the tool fits into clinical practice.
Stony Brook’s announcement stresses that both human-written and AI-generated summaries contained errors and says clinicians must review and verify any model-generated draft before it enters a patient record. The preference result supports testing AI as a drafting aid for complex discharges, not handing over responsibility for the final clinical document.