Online Program

Return to main conference page
Tuesday, September 24
Tue, Sep 24, 2:45 PM - 4:00 PM
Thurgood Marshall South
Machine Learning and Artificial Intelligence: Dynamic Landscape in Pharmaceutical Development and Clinical Decision-Making

An Evolving Perspective on AI and Machine Learning at CBER FDA (301024)

Steven A Anderson, FDA/CBER 
Mark Owen Walderhaug, FDA/CBER 
*Alan E. Williams, FDA 

Keywords: AI, ML, NLP, Hemovigilance, safety

At present, applications of AI and ML in FDA’s Center for Biologics Evaluation and Research (CBER) have been in the post-market area. A major effort AI effort in CBER is a part of a program called Biologics Effectiveness and Safety (BEST) Initiative which uses Natural Language Processing (NLP) and ML on electronic health records (EHR) in conjunction with other surveillance data to find insights in biologics safety. A project in BEST using NLP was to examine nursing notes that included observations on patient transfusions. After selected representative set of nursing notes were identified, this set was used to train the NLP algorithm to extract vital signs and to derive information such as elapsed time of transfusion based on recorded start and stop times. After running the algorithm on 34,000 transfusion notes, a random sample of 100 NLP extracted notes were compared with expert review of the notes. The NLP extracted information was highly accurate for vitals and accurate for derived values in nursing notes that had complete information. Other areas in which NLP may be applied are real world evidence, automated high-level adverse event reporting from unstructured data in EHRs, and semi-automated chart review. An example of the latter is a BEST project combining AI, ML, and NLP in the analysis of EHR for adverse events related to transfusions. After training an AI algorithm on 727 transfusion adverse events from the National Healthcare Safety Network Hemovigilance reports, the algorithm was applied to EHRs reporting 334,485 transfusion exposures. Incident rates were determined for febrile non-hemolytic transfusion reaction (0.099%), allergic reaction (0.049%), transfusion-associated circulatory overload (0.018%), delayed serologic transfusion reaction (0.014%), and hypotensive transfusion reaction (0.011%). AI, ML, and NLP show promise in enhancing CBER’s mission to improve the safety and effectiveness of biologics.