Enhancing Post-Marketing Pharmacovigilance: An AI-Driven System Approach for Safety Signal Detection of Adverse Drug Events: A Case Study of Glucagon-Like Peptide-1 Receptor Agonists

Author

Ali Alsuhibani, Virginia Commonwealth UniversityFollow

DOI

https://doi.org/10.25772/BTP3-ZE96

Defense Date

2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Pharmaceutical Sciences

First Advisor

Dr. Dayanjan S Wijesinghe

Abstract

Pharmacovigilance is paramount for safeguarding patient welfare by detecting, assessing, and preventing adverse drug reactions (ADRs). Traditional monitoring methods—such as spontaneous reporting—face growing challenges from increasingly complex drug therapies and ever-expanding datasets. To address these gaps, we developed an AI-driven framework that automatically retrieves and cross-references FDA-approved drug labels from DailyMed with real-world adverse event data from the FAERS. Our approach uses a streamlined pipeline involving (1) automated label extraction and text-processing to capture critical safety data from PDFs, (2) advanced preprocessing of FAERS data—filtering, merging, and deduplicating millions of records across multiple quarters, and (3) integrated signal detection algorithms (PRR, ROR, and IC) to identify safety concerns with robust statistical validation.

We showcased this framework using Glucagon-Like Peptide-1 receptor agonists (GLP-1 RAs)—a medication class widely used for type 2 diabetes and increasingly for obesity. Despite proven therapeutic benefits, GLP-1 RAs present complex safety profiles, including gastrointestinal disturbances, rare neoplastic concerns, and various device-related issues (for injection-based formulations). By unifying data from drug labels and FAERS, our system identified multiple potential safety signals: for instance, semaglutide showed pronounced gastrointestinal signals (e.g., cholecystitis, ileus, pancreatitis), while liraglutide indicated strong neoplasm-related signals (e.g., thyroid and pancreatic malignancies). Exenatide data prominently revealed device-handling complications, underscoring the importance of injection technique. The system also automatically mapped adverse events to System Organ Classes (SOCs) via large language models (LLMs), reducing manual curation and improving consistency.

Validation showed substantial agreement (Cohen’s kappa ≥ 0.73) between LLM-based SOC mappings and established literature classifications. Additionally, our disproportionality analyses (PRR, ROR, IC) demonstrated excellent reliability (intraclass correlation coefficient = 1.0) compared to reference implementations, confirming the technical soundness of our approach. These validation results underscore the promise of AI for harmonizing disparate safety data sources and accelerating ADR detection. By bridging FDA labels (pre-market insights) with spontaneous reports (post-market data), our method allows more proactive, accurate, and near-real-time pharmacovigilance.

In conclusion, this study illustrates a novel, automated system that advances pharmacovigilance through AI-powered data extraction, integration, and signal detection. The successful GLP-1 RA case study highlights the transformative potential of this approach for broad application across therapeutic classes—ultimately fostering faster, more reliable safety monitoring and better patient outcomes.

Rights

© Ali Alsuhibani

Is Part Of

VCU University Archives

Is Part Of

VCU Theses and Dissertations

Date of Submission

4-24-2025