Flexible Empirical Bayesian Approaches to Pharmacovigilance for Simultaneous Signal Detection and Signal Strength Estimation in Spontaneous Reporting Systems Data

TL;DR

This paper develops flexible, scalable empirical Bayesian methods for pharmacovigilance that improve signal detection and estimation of adverse event strengths in spontaneous reporting data, with better accuracy and efficiency.

Contribution

It introduces novel non-parametric empirical Bayes approaches that are computationally efficient and more adaptable than existing methods for adverse event signal detection and estimation.

Findings

Methods outperform existing approaches in simulation studies.

Achieve comparable or better signal detection rates.

Provide accurate uncertainty quantification for AE signals.

Abstract

Inferring adverse events (AEs) of medical products from Spontaneous Reporting Systems (SRS) databases is a core challenge in contemporary pharmacovigilance. Bayesian methods for pharmacovigilance are attractive for their rigorous ability to simultaneously detect potential AE signals and estimate their strengths/degrees of relevance. However, existing Bayesian and empirical Bayesian methods impose restrictive parametric assumptions and/or demand substantial computational resources, limiting their practical utility. This paper introduces a suite of novel, scalable empirical Bayes methods for pharmacovigilance that utilize flexible non-parametric priors and custom, efficient data-driven estimation techniques to enhance signal detection and signal strength estimation at a low computational cost. Our highly flexible methods accommodate a broader range of data and achieve signal detection performance comparable to or better than existing Bayesian and empirical Bayesian approaches. More importantly, they provide coherent and high-fidelity estimation and uncertainty quantification for potential AE signal strengths, offering deeper insights into the comparative importance and relevance of AEs. Extensive simulation experiments across diverse data-generating scenarios demonstrate the superiority of our methods in terms of accurate signal strength estimation, as measured by replication root mean squared errors. Additionally, our methods maintain or exceed the signal detection performance of state-of-the-art techniques, as evaluated by frequentist false discovery rates and sensitivity metrics. Applications on FDA FAERS data for the statin group of drugs reveal interesting insights through Bayesian posterior probabilities.