Developing an Artificial Intelligence-Guided Signal Detection in the Food and Drug Administration Adverse Event Reporting System (FAERS): A Proof-of-Concept Study Using Galcanezumab and Simulated Data

Abstract

Introduction: Time- and resource-demanding activities related to processing individual case safety reports (ICSRs) include manual procedures to evaluate individual causality with the final goal of dismissing false-positive safety signals. Eminent experts and a representative from pharmaceutical industries and regulatory agencies have highlighted the need to automatize time- and resource-demanding procedures in signal detection and validation. However, to date there is a sparse availability of automatized tools for such purposes.

Objectives: ICSRs recorded in spontaneous reporting databases have been and continue to be the cornerstone and the most important data source in signal detection. Despite the richness of this data source, the incessantly increased amount of ICSRs recorded in spontaneous reporting databases has generated problems in signal detection and validation due to the increase in resources and time needed to process cases. This study aimed to develop a new artificial intelligence (AI)-based framework to automate resource- and time-consuming steps of signal detection and signal validation, such as (1) the selection of control groups in disproportionality analyses and (2) the identification of co-reported drugs serving as alternative causes, to look to dismiss false-positive disproportionality signals and therefore reduce the burden of case-by-case validation.

Methods: The Summary of Product Characteristics (SmPC) and the Anatomical Therapeutic Chemical (ATC) classification system were used to automatically identify control groups within and outside the chemical subgroup of the proof-of-concept drug under investigation, galcanezumab. Machine learning, specifically conditional inference trees, has been used to identify alternative causes in disproportionality signals.

Results: By using conditional inference trees, the framework was able to dismiss 20.00% of erenumab, 14.29% of topiramate, and 13.33% of amitriptyline disproportionality signals on the basis of purely alternative causes identified in cases. Furthermore, of the disproportionality signals that could not be dismissed purely on the basis of the alternative causes identified, we estimated a 15.32%, 25.39%, and 26.41% reduction in the number of galcanezumab cases to undergo manual validation in comparison with erenumab, topiramate, and amitriptyline, respectively.

Conclusion: AI could significantly ease some of the most time-consuming and labor-intensive steps of signal detection and validation. The AI-based approach showed promising results, however, future work is needed to validate the framework.