P240: Emerging Patient Safety Considerations for LLM Based Adverse Event Detection
Poster Presenter
Adam Clarke
Innovation Manager
Bristol Myers Squibb United States
Objectives
Design an evidence-based approach to detect Adverse Event language in company Text sources via LLMs using centralized prompt libraries and propose how this new capability should be employed across the company.
Method
Over the last year, we reviewed literature and internal case examples on using LLMs to detect adverse events in enterprise text. Patient Safety and partner function use cases were examined, including new opportunities, new risks, and approaches for responsible deployment.
Results
The study produced a draft framework for identifying Adverse Event–related language in text generated by other functions using a vetted LLM prompt strategy. Analysis identified both opportunities for AE awareness and risks, including potential for under-reporting due to machinal bypass and over-reporting driven by duplication. A structured library of representative use cases was generated for inclusion in the Pharmacovigilance System Master File (PSMF), supporting emerging regulatory expectations for AI-enabled safety processes.
Conclusion
Widespread and accessible ability to look for AEs in text should be approached with caution as it caries new risks, but also potential benefits.
