Adverse Drug Reaction Prediction: Graph Neural Networks and Causal Inference Techniques

This paper presents an innovative and integrated approach for predicting and understanding Adverse Drug Reactions (ADRs) by combining Graph Neural Networks (GNNs) with causal inference techniques. Utilizing the SIDER dataset, comprising information on 1430 drugs, 5880 ADRs, and 140,064 drug–ADR pairs, our study employed GNNs to analyze chemical structures and predict ADRs with an accuracy of 87%. Concurrently, causal inference methodologies were applied to discern causal relationships between drug treatments and adverse reactions, revealing a 12% increase in ADR likelihood associated with specific drug treatments. The integration of GNN and causal inference models provided a nuanced understanding of ADRs, addressing limitations inherent in individual methodologies. The doubly robust estimator in causal inference enhanced robustness in estimating causal effects, while cross-validation experiments confirmed the generalizability of our models. The comprehensive insights gained from this dual approach contribute to advancing pharmacovigilance systems, offering a valuable tool for proactive healthcare measures and pharmaceutical safety assessment.