|
Abstract:
|
Statistical methods for detection of trends in adverse-event (AE) data in clinical development can help limit and characterize drug safety risks to patients. However, modeling of AE data is complicated by the high-dimensional nature of AE coding (e.g. MedDRA includes 23,000+ Preferred Terms). It is advantageous to borrow strength between groupings of related events. To identify such groupings, we use an analogue to topic modeling, in which the topics of passages of text are inferred by analyzing the vocabulary. We applied the Bayesian topic model Latent Dirichlet Allocation (LDA) (Blei et al, 2003) to data from the FDA Adverse Event Reporting System (FAERS). LDA identified a few classes of coherent clinical concepts as topics, and improved out-of-sample prediction of AEs. We further improved topic coherence using the Latent-Feature LDA (LF-LDA) model (Nguyen et al, 2015) and a compressed metric of closeness in the MedDRA hierarchy. These models surpass the flexibility of System Organ Classes and Standardized MedDRA Queries (SMQs) for grouping events into related classes, with benefits for exploration and interpretation, and for downstream modeling of AE incidence and severity.
|
