Precision Adverse Drug Reactions Prediction with Heterogeneous Graph Neural Network.

Area of Science:

• Computational pharmacology
• Biomedical informatics
• Machine learning in healthcare

Background:

• Patient-level Adverse Drug Reaction (ADR) prediction is crucial for safety and outcomes.
• Traditional methods struggle with individual patient demographics and ADR variations.
• Existing models often predict drug-level ADRs, not patient-specific risks.

Purpose of the Study:

• To propose a novel framework, Precise Adverse Drug Reaction (PreciseADR), for patient-level ADR prediction.
• To overcome limitations of traditional methods by integrating patient-specific data.
• To enhance the accuracy of identifying potential ADRs for individual patients.

Main Methods:

• Constructed a heterogeneous graph including patients, diseases, drugs, and ADRs.
• Utilized heterogeneous Graph Neural Networks (GNNs) to learn patient embeddings.
• Incorporated patient-disease and patient-drug relationships within the graph structure.

Main Results:

• PreciseADR demonstrated superior predictive performance on a large-scale real-world healthcare dataset (FAERS).
• Achieved a 3.2% higher AUC score and a 4.9% higher Hit@10 compared to the strongest baseline.
• Effectively captured local and global dependencies for subtle ADR pattern identification.

Conclusions:

• PreciseADR offers a powerful approach for accurate patient-level ADR prediction.
• The GNN-based framework effectively models complex interactions influencing ADRs.
• This advancement holds significant potential for improving patient safety and personalized medicine.