Description
Contemporary approaches to identification of potential on-market adverse drug effects (ADEs) depend upon aggregation of many data sources, including spontaneous reporting systems (SRSs) and the biomedical literature. Due to inherent scale and complexity, information from these data sources require robust and scalable methods. This research leverages relational information extracted from the biomedical literature in direct complement to observational signal extracted from the FDA Adverse Event Reporting System (FAERS, a SRS) using representation learning and composition with downstream supervised learning for drug safety monitoring. By combining these data, performance on a publicly-available drug safety reference standard improves ~20% over using SRS-based data alone, and several points of improvement over utilizing the literature alone. In an ensemble configuration, the basis for classification can be modulated to favor either data source, with best performance occurring at equal contribution. These results demonstrate utility in leveraging these complementary data stores for drug safety monitoring.
Learning Objective: Learn and be able to discuss the challenges and opportunities that multimodal data integration for drug safety monitoring presents, especially in regards to automated systems that have the potential to impact regulatory action concerning drug products and patient health.
Authors:
Justin Mower (Presenter)
Rice University
Trevor Cohen, University of Washington Medicine
Devika Subramanian, Rice University
Presentation Materials: