Graph neural network-based subgraph analysis for predicting adverse drug events.

Area of Science:

• Computational medicine
• Pharmacovigilance
• Machine learning in healthcare

Background:

• Adverse drug events (ADEs) pose significant global health risks, leading to hospitalizations and mortality.
• Current methods often identify ADEs late, after widespread drug use, challenging patient safety.
• There is a critical need for computational tools to predict ADEs earlier, before or during clinical trials.

Purpose of the Study:

• To develop and evaluate a network-based computational approach for the early identification of ADEs.
• To predict the occurrence, timing, and type of ADEs based on patient diagnostic history.
• To leverage Graph Neural Networks (GNNs) for enhanced ADE prediction models.

Main Methods:

• A network-based approach modeling patients as subgraphs of International Classification of Diseases (ICD) codes.
• Utilized four Graph Neural Network (GNN) variants (e.g., GraphSage, GAT) for predictive modeling.
• Employed binary classification for predicting occurrence/timing and multi-label classification for predicting ADE type.

Main Results:

• The network-based approach showed superior performance in predicting ADEs.
• GraphSage achieved the highest accuracy for predicting ADE occurrence (0.8863) and type (0.9367).
• Graph Attention Networks (GAT) performed best in predicting ADE timing (0.8769), with specific advantages for certain ADE categories.

Conclusions:

• Graph Neural Networks (GNNs) show significant potential for early ADE detection and prevention.
• Accurate ADE prediction can inform clinical decisions, preventive actions, and medication adjustments.
• This predictive method can optimize healthcare resource allocation by preventing hospital admissions and reducing the burden of ADEs.