Explainable artificial intelligence for pharmacovigilance: What features are important when predicting adverse...

Area of Science:

• Pharmacovigilance
• Machine Learning
• Explainable Artificial Intelligence

Background:

• Machine Learning (ML) models use health data to predict adverse outcomes.
• Explainable Artificial Intelligence (XAI) determines feature importance in ML predictions.
• Acute Coronary Syndrome (ACS) is a critical adverse outcome for monitoring.

Purpose of the Study:

• To develop an XAI-based pharmacovigilance technique.
• To quantify the contribution of specific drugs to ACS predictions.
• To utilize ML and XAI for adverse drug event detection.

Main Methods:

• Trained ML models using health data (drug history, comorbidities) to predict ACS adverse outcomes.
• Applied XAI techniques (LIME, SHAP) to quantify feature importance, specifically drug contributions.
• Analyzed linked Western Australian health datasets for individuals aged over 65.

Main Results:

• ML models predicted ACS adverse outcomes with 72% accuracy.
• XAI successfully identified rofecoxib and celecoxib as contributing to ACS predictions.
• SHAP and LIME accurately distinguished important from unimportant features, with SHAP showing a slight advantage.

Conclusions:

• ML and XAI algorithms effectively quantify feature importance from health datasets.
• This approach shows potential for pharmacovigilance monitoring of adverse outcomes.
• Further development could lead to robust, automated pharmacovigilance systems.