Applying machine learning to pharmacovigilance data: A proof-of-concept study.

Area of Science:

• Pharmacovigilance and Drug Safety
• Computational Toxicology
• Health Informatics

Background:

• Machine learning (ML) applications in pharmacovigilance are limited.
• Assessing ML feasibility for drug-adverse reaction classification is crucial.

Purpose of the Study:

• To evaluate the feasibility of using eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP) for classifying warfarin-associated gastrointestinal bleeding.
• To identify risk factors associated with reporting this adverse drug reaction using French National Pharmacovigilance Database (FNPV) data.

Main Methods:

• Extracted 2025 individual case safety reports (ICSRs) involving warfarin from the FNPV.
• Trained an XGBoost model on 1045 preprocessed ICSRs to classify gastrointestinal bleeding.
• Interpreted model predictions using SHAP values to identify influential features.

Main Results:

• The XGBoost model achieved 0.74 F1-score and 0.716 AUC-ROC on the test set.
• Key predictors for gastrointestinal bleeding reporting included age, gastrointestinal bleeding risk, antiplatelet use, and history of renal dysfunction.
• SHAP analysis provided insights into feature importance and directional effects.

Conclusions:

• Demonstrated the feasibility of applying XGBoost and SHAP analysis to pharmacovigilance data.
• Successfully classified a specific drug-adverse reaction pair (warfarin-GI bleeding).
• Identified key predictive features for adverse drug reaction reporting, aiding pharmacovigilance efforts.