A Pilot, Predictive Surveillance Model in Pharmacovigilance Using Machine Learning Approaches.

Area of Science:

• Pharmacovigilance and Drug Safety
• Computational Toxicology
• Pharmaceutical Data Science

Background:

• Ensuring drug product safety through adverse event (AE) identification is critical in the pharmaceutical industry.
• Machine learning (ML) offers potential to enhance traditional pharmacovigilance (PV) surveillance and signal detection.
• This pilot study evaluated ML for detecting safety signals and assessing early detection capabilities.

Purpose of the Study:

• To assess the capability of ML models in detecting potential safety signals for two pharmaceutical products.
• To compare the early detection of safety signals by ML models against human review.
• To evaluate the accuracy and performance of ML in a real-world pharmacovigilance setting.

Main Methods:

• Two drugs, Drug X (mature product) and Drug Y (recently approved), were selected for analysis.
• Gradient boosting ML algorithms, specifically XGBoost, were employed as the primary modeling strategy.
• The models were trained and tested on post-marketing data to identify potential safety signals.

Main Results:

• For Drug X, the model identified four true signals out of 12 potential new signals (50.0% sensitivity, 33.3% PPV). One signal was detected six months earlier than human review.
• For Drug Y, the model identified five true signals out of 13 potential new signals (55.6% sensitivity, 38.5% PPV). No new true signals were confirmed by human review for Drug Y.
• The ML model demonstrated potential for earlier safety signal detection compared to traditional human assessment.

Conclusions:

• The developed ML model showed acceptable accuracy for safety signal detection in pharmacovigilance.
• The study highlights the potential of ML to supplement human review and enable earlier identification of adverse events.
• Human expertise remains crucial for the final assessment and interpretation of safety signals.