PAIR Center at American Medical Informatics Association 2025 Annual Symposium

Privacy-Preserving, Asynchronous, Federated Training and Validation of a Clinical Decision Support System for Ventilator Management (Poster)

Antonia Angeli Gazola

Managing patients receiving invasive mechanical ventilation settings is complex and no effective clinical decision support systems exist that are generalizable and comprehensive to support such management. We trained the AI Vent Assistant (AVA) using an asynchronous, privacy-preserving, federated approach across seven geographically diverse health systems to make simultaneous recommendations for twelve interrelated ventilator settings. The resultant, aggregated model required no data sharing or additional training and outperformed most local models in held-out, external validation.

View poster here.

How Should We Oversee AI in Medicine To Ensure Safety and Facilitate Innovation? (Panel)

Moderated by Gary Weissman

The use of artificial intelligence (AI) in clinical care is rapidly accelerating. While AI tools show promise for improving patient care and streamlining clinical workflows, understanding of the requirements for overseeing the implementation and ensuring safety of such tools is lacking. Traditional medical device oversight frameworks are inadequate for many AI tools, because patient populations, clinical practice, documentation, and information systems vary considerably across hospitals and clinics and vary over time. In this panel, we will bring together a diverse set of experts to describe the gaps in the current oversight of clinical AI systems and discuss opportunities to solidify emerging policy and practice efforts for optimal oversight.

View slides here.

Unregulated Large Language Models Produce Medical Device-Like Output (Oral Presentation)

Gary Weissman

Large language models (LLMs) show considerable promise for clinical decision support (CDS) but none is currently authorized by the Food and Drug Administration (FDA) as a CDS device. We evaluated whether two popular LLMs could be induced to provide device-like CDS output. We found that LLM output readily produced device-like decision support across a range of scenarios, suggesting a need for regulation if LLMs are formally deployed for clinical use in the future.

View slides here.