Stony Brook highlights study finding doctors preferred AI discharge summaries in 95% of cases

Stony Brook University is considering a clinical pilot of AI-written hospital discharge summaries after physician reviewers preferred them in 95% of cases in a 60-hospitalization study. In a November 2 announcement, the university described better-rated drafts with fewer missing details—but stressed that doctors must verify them before they enter a patient’s record.

The announcement highlights research published October 2 in npj Digital Medicine. It was a retrospective comparison of existing hospital cases, not a trial of AI-directed care. The possible pilot would evaluate the tool in everyday practice at Stony Brook Medicine; it remains under consideration, not a confirmed rollout.

Summarizing a week—or three—of hospital care

Discharge summaries carry information from a hospital stay into the next stage of care. They describe treatment, medication needs and changes, unresolved medical issues, and outpatient follow-up. Primary care doctors rely on them to continue treatment, while patients and caregivers need to understand what happens after discharge.

The researchers focused on 60 internal medicine hospitalizations at Stony Brook University Hospital, each lasting seven to 21 days. Longer, complex stays create more material to condense and can increase documentation work. The paper notes that earlier evaluations largely involved shorter stays and rarely examined receiving doctors’ priorities or the reporting of incidental findings.

Clinicians helped design the approach around the information physicians need after discharge. A large language model—AI that processes and generates text—organized medical records and other patient information into a summary within minutes, according to the university. That generation time is distinct from the actual physician time savings researchers still intend to assess.

Hospital-based physicians and primary care doctors assessed the AI and human summaries in paired evaluations. They rated the AI versions higher for quality, readability, factual accuracy and completeness. Primary care reviewers also found them more useful for follow-up care and easier to use when explaining hospital treatment to patients and caregivers.

93.5%Findings rated factually correct

Among 31 AI-identified incidental radiology findings, reviewers rated 93.5% factually correct and 87.1% appropriate for reporting. These ratings concern that subset of findings, not the accuracy of entire discharge summaries.

Fewer omissions, not a demonstrated safety advantage

The AI summaries received fewer reviewer annotations for factual inaccuracies or clinically relevant omissions, mainly because they left out fewer details. Both versions nevertheless contained errors. Reviewers found no statistically significant difference in their estimates of those errors’ potential severity or likelihood of harm—a narrower result than showing that AI summaries were safer.

Who reviewed the summaries also mattered. Primary care physicians marked more omissions and assigned a higher likelihood of harm than hospital-based reviewers. That difference sits alongside their favorable usefulness ratings: the doctors receiving patients after discharge valued the AI drafts, while also identifying more missing information than their hospital counterparts.

Co-corresponding author Fusheng Wang said further development will focus on questions the comparison leaves open. The planned work covers:

• Accuracy and safety as the team continues developing the discharge-summary approach.
• Actual time savings for physicians, rather than simply how quickly the model produces text.
• How the tool would fit into clinical workflows, with an everyday-practice pilot still being considered.