Speaker
Eric Karl Oermann, MD
Associate Professor of Neurosurgery, Radiology, and Data Science
Director, OLAB – Health AI Research Lab of NYU Langone Health
Associate Program Director, Neurosurgery Residency program
Investigator, Neuroscience Institute
NYU Langone Medical Center
New York University
Keywords
Artificial Intelligence; Large Language Models; AI in Medicine; Clinical Benefit
Key Points
- While clinicians often treat large language models (LLMs) like textbooks, books are peer-reviewed, static, and universal. In contrast, LLMs are stochastic and probabilistic, meaning different users will see different responses. AI should be treated as a clinical tool.
- There is a gap between the amount of work and publishing being done in this field and its impact on clinical practice. Since 2020, roughly 5,000 papers have been published on the topic of LLMs in medicine. Of those, only 21 generated legitimate medical evidence.
- Process benefit does not imply clinical benefit. The success of medical AI lies in how the tool is paired with a clinical workflow. Dr. Oermann demonstrated this with examples from 4 real-world clinical trials. Future trials should test both models and workflows; how clinicians utilize a model is as important as the model they use.
Discussion Themes
Traditional randomized controlled trials can take years to plan and execute, during which time an AI model may go through several generations of updates. To match the speed of technology, Dr. Oermann advocated for adaptive trial designs and rapid learning health systems where institutions can run silent deployments, AB test AI tools in-house, and observe how they affect local care delivery.
Health systems are ultimately responsible for patient outcomes, placing the accountability for monitoring and validating AI performance over time onto the individual institutions rather than relying solely on third-party vendors.