Application of Augmented Intelligence for Pharmacovigilance Case Seriousness Determination

Area of Science:

• Pharmacovigilance and Patient Safety
• Artificial Intelligence in Healthcare
• Computational Linguistics

Background:

• Accurate identification of adverse event seriousness is crucial for patient safety and regulatory reporting timelines.
• Manual assessment of adverse event seriousness is time-consuming and struggles to keep pace with increasing report volumes.
• Scalable, automated solutions are needed to support pharmacovigilance experts.

Purpose of the Study:

• To develop an augmented intelligence methodology for automatic identification of adverse event seriousness.
• To evaluate deep learning models for accuracy and F1 score in classifying seriousness across different report types.
• To create a system that supports pharmacovigilance by automating seriousness determination.

Main Methods:

• Development of three neural networks: a binary seriousness classifier, a seriousness categorization classifier, and a seriousness criteria annotator.
• Utilized a stratified random sample of safety reports for model training and validation.
• Evaluated models against expert-labeled ground truth data.

Main Results:

• The seriousness classifier achieved high accuracy (83.0%-92.9%) across post-marketing, solicited, and medical literature reports.
• F1 scores for seriousness categorization ranged from 75.5% to 78.9% for death, hospitalization, and important medical events.
• The seriousness annotator demonstrated strong performance with F1 scores of 89.9% and 75.2% in solicited and medical literature reports, respectively.

Conclusions:

• Neural network approaches provide an accurate and scalable method for identifying adverse event seriousness.
• This technology can augment the capabilities of pharmacovigilance practitioners.
• The developed methodology supports timely and efficient safety report processing.