(087) Assessment of the Long-Term Stability of Drug Safety Signals between Frequency and Bayesian-based methods for Glucagon-like Peptide-1 Receptor Agonists

Background: The accumulation of Individual Case Safety Reports (ICSRs) in the spontaneous reporting system over time can affect which drug-reaction pairs are detected as safety signals. However, there has been little research comparing the performance of signal detection methods in generating stable drug safety signals over time.

Objectives: This study aimed to compare the long-term stability between frequency-based and Bayesian-based methods in detecting drug safety signals for glucagon-like peptide-1 receptor agonists (GLP-1 RAs), specifically dulaglutide and albiglutide, which were first approved in 2014.

Methods: The ICSRs data consisted of quarterly FDA Adverse Event Reporting System (FAERS) data files on GLP-1 RAs, spanning from Q3 2014 to Q4 2023. Frequency-based methods included the proportional reporting ratio (PRR) and reporting odds ratio (ROR). Bayesian-based methods included the information component (IC) and empirical Bayes geometric mean (EBGM). Long-term stability was assessed by comparing the positive predictive value (PPV) and negative predictive value (NPV) of drug safety signals as additional years of FAERS data were consecutively added to the initial data from Q3 2014 to Q4 2015. Drug safety signals detected from the entire dataset were considered the gold standard. There were eight comparisons for PPV and NPV between the entire data (gold standard) and the test data accumulated one year at a time for eight years.

Results: The number of total drug-reaction pairs (N) was 415,528 from the entire dataset. Initially, PRR and ROR generated more signals compared to Bayesian methods, but this difference diminished over time. For PPV, Bayesian-based methods showed values greater than 0.7 in almost all subsets. In contrast, frequency-based methods had a PPV of only 0.42 in the first limited period subset, exceeding 0.7 only after seven years of data accumulation. In terms of NPV, EBGM had the highest values in all subsets, while other methods also performed well.

Conclusions: Frequency-based methods detected more signals for dulaglutide and albiglutide shortly after drug approval compared to Bayesian methods. However, the PPV and NPV of these signals were lower than those of Bayesian methods. This suggests that Bayesian methods may identify more robust and stable signals in situations with limited data availability, such as performing pharmacovigilance for new drugs.