Evaluating AI's Role in Discharge Summaries With: Dr. Ben Rosner
Dr. Ben Rosner, a hospitalist and digital health researcher at UCSF, explores the promise and pitfalls of AI-generated discharge summaries. In this wide-ranging discussion, Dr. Rosner explains how LLMs can reduce administrative burden, improve communication at discharge, and potentially enhance patient safety—if implemented thoughtfully. He shares findings from his JAMA-published study evaluating LLM-drafted summaries and outlines how these tools perform against physician-written counterparts. The conversation expands into the risks of de-skilling, challenges of AI trust, and the need for systems like "LLMs as juries" to monitor AI-generated clinical documentation. Rosner also reflects on AI’s broader impact on medical education and the role of emerging roles like Chief Health AI Officers.
Episode Contents:
About the Guest
Dr. Ben Rosner is a hospitalist at UCSF and a published researcher focused on digital health, AI in healthcare, and patient engagement. View his full profile: https://www.linkedin.com/in/benrosnermdphd/
Key Takeaways
- AI-generated summaries can match or exceed physician-written ones
- Clinical LLMs require safety frameworks and performance monitoring
- Medical education must address AI-driven de-skilling risks
Transcript Summary
What inspired your study on AI-generated discharge summaries?
As a hospitalist, Dr. Rosner saw firsthand how labor-intensive it is to write summaries—especially for long hospitalizations. He and his team asked whether LLMs could generate drafts, freeing up physician time.
How did the LLM-generated summaries compare to those written by physicians?
Reviewers found LLM summaries just as good, sometimes better—more concise and coherent. While LLMs made more errors, they were generally low-risk. Importantly, human-written summaries also had hallucinations.
What are some concerns with deploying LLMs in hospitals?
Trust without oversight. Rosner warns of "cognitive laziness" where clinicians may stop reviewing AI outputs. Tools must be evaluated and monitored post-deployment.
What's the solution?
He proposes "LLM as jury"—using multiple models to critique each other's output. It's more scalable than relying solely on human reviewers.
Can AI enhance or hinder medical training?
Both. Rosner sees use cases like automatic procedure logs and clinical reasoning feedback. But over-reliance may lead to de-skilling.
How can educators balance benefit and risk?
Some schools are withholding AI tools from early trainees to ensure foundational skills develop before layering in AI assistance.
More Topics
- AI in Patient Care
- AI in the Healthcare Industry
- AI and Medical Innovation
- Healthcare Ethics and Policy
Keep Exploring
About the Series
AI and Healthcare—with Mika Newton and Dr. Sanjay Juneja is an engaging interview series featuring world-renowned leaders shaping the intersection of artificial intelligence and medicine.
Dr. Sanjay Juneja, a hematologist and medical oncologist widely recognized as “TheOncDoc,” is a trailblazer in healthcare innovation and a rising authority on the transformative role of AI in medicine.
Mika Newton is an expert in healthcare data management, with a focus on data completeness and universality. Mika is on the editorial board of AI in Precision Oncology and is no stranger to bringing transformative technologies to market and fostering innovation.