Drug Saf. 2025 Dec 6. doi: 10.1007/s40264-025-01632-8. Online ahead of print.
ABSTRACT
BACKGROUND: Databases for safety monitoring of medicinal products contain records of a huge number of pairings of drugs and adverse events (AEs). Existing disproportionality methods for safety monitoring in such databases estimate background rates of AE occurrence in ways that may be susceptible to masking effects that can hinder signal detection, particularly in the context of large overall counts of AE or drug occurrence in the data.
OBJECTIVES: To develop a new statistical model for determining the background rate against which individual drug-AE pairs are to be evaluated, which is robust against masking effects, and to incorporate this into an algorithm which simultaneously estimates background rates and detects drug-AE pair counts that deviate significantly from these rates.
METHODS: We constructed a hierarchical Bayesian model for background rates, and background rate samples were drawn from the model parameters using an iterative Markov Chain Monte Carlo (MCMC) method. At each iteration, any counts whose probability is low given current background rate estimation were removed from the computation that sampled the next set of background rates. The algorithm, called Markov Chain Signal Generation (MCSG), was implemented using a combination of Python and the probabilistic programming language Stan.
RESULTS: The MCSG algorithm outperformed routinely used quantitative approaches for signal detection on both synthetic data designed to include a drug-AE pair with very strong masking effects and a reference set featuring 69 unique active substances and 792 unique AEs. On a synthetic dataset where selected pairs occurred at rates deviating from a constant background, MCSG accurately identified these pairs in the presence of strong masking signals. On a subset of some real data from the FDA Adverse Event Report System (FAERS), it effectively identified a reference set of positive and negative controls and was able to identify drug-AE pairs suggested in the literature.
CONCLUSION: We have demonstrated a new approach to signal generation, which avoids the confounding effect of masking more effectively than currently used methods. The algorithm is best used in a setting of multiple drug-AE pairs, the majority of which are expected to have counts at background rate, although with substantial datasets the algorithm can take minutes or hours to run. It is therefore particularly suitable for infrequent, large-scale analysis (for example, quarterly analysis of the entirety of a pharmacovigilance database).
PMID:41351764 | DOI:10.1007/s40264-025-01632-8
