Physicians were asked to provide feedback on the AI-generated summaries — namely, whether there were inaccuracies, hallucinations or omissions in the summaries — and their perceived time savings, if applicable.
Feedback on 100 AI-enabled summaries found some omissions (25%) and inaccuracies (20%), but hallucinations were rare (2%). Physicians rated 88 unedited summaries as having no harm potential and 21 as having mild harm potential.
One summary was rated by a physician as likely to cause moderate harm. In that case, the summary noted that a patient was discontinued from intravenous antibiotics and transitioned to oral antibiotics. But it failed to mention that a nine-day course of treatment antibiotics had already been completed, and the oral antibiotics were prophylactic due to an unrelated condition. Independent reviewers determined the summary posed no risk.
No severe harm was reported.
Less exhaustion
While physicians presumed they had a time savings of more than 10 minutes per discharge summary with MedAgentBrief, the researchers determined, from checking time logs in the electronic medical record, that the time savings were modest (three minutes at best) and variable for every user. The discrepancy is likely because revising and checking the AI-generated summaries felt to physicians like less effort than creating the summary from scratch, the researchers said.
Physicians participating in the pilot reported significant reduction in their burnout scores.
Liang said she found the summaries especially helpful — and time saving — for complex patients with long hospital stays whom she met shortly before discharge. In one case, the AI-generated summary highlighted a patient’s anemia, which wasn’t central to their hospitalization but was important to convey to their primary care team.
“I didn’t save too much time,” she said, “but it definitely was very helpful and important that it highlighted these things that I wasn’t fully aware of.”
The pilot ended on Oct. 11, but the MedAgentBrief remained in routine clinical use for another six months. It was turned off in April in advance of the rollout of a similar tool by Stanford Medicine’s electronic medical record supplier.
Grolleau said the researchers plan to develop a system to evaluate AI tools from vendors.