Utilizing Deep Learning for Detecting Adverse Drug Events in Structured and Unstructured Regulatory Drug Data Sets

Area of Science:

• Pharmacovigilance
• Regulatory Science
• Computational Linguistics

Background:

• The US Food and Drug Administration (FDA) collects extensive post-market drug data, including adverse event (AE) reports and structured drug product labels (SPLs).
• The FDA Adverse Event Reporting System (FAERS) contains millions of public submissions detailing suspected medication-related AEs, requiring standardized coding for analysis.
• Manufacturers are not currently mandated to code drug labels with associated AEs, presenting a data gap.

Purpose of the Study:

• To assess the suitability of manually annotated FDA FAERS and SPL datasets for predictive modeling.
• To enhance regulatory efficiency through automated classification of adverse event reports by preferred terminology.

Main Methods:

• A proof-of-concept recurrent neural network (RNN) was developed for automated extraction of preferred AE terminology.
• Two separate RNN models were trained and cross-validated: one on 325 annotated FAERS patient narratives and another on 100 SPLs.

Main Results:

• The RNN model for product labels performed comparably to conventional models on most selected AE terms, based on F1-score.
• Model performance on the FAERS dataset yielded mixed results, indicating potential limitations.

Conclusions:

• A machine learning approach was successfully demonstrated as a proof-of-concept for automated AE detection in regulatory datasets.
• Limited instances of specific AE classes may have hindered model generalization; additional data could improve validation robustness.