AI-Driven Polypharmacy Risk Prediction Using Graph Neural Networks

Abstract

Polypharmacy, which is the usage of several drugs simultaneously, is becoming a common phenomenon, especially among elderly patients and those with long-term illnesses. Although it may be essential to the successful treatment, polypharmacy is a leading contributor to the occurrence of adverse drug reactions (ADRs), drug-drug interactions (DDIs), medication errors, and decreased therapeutic efficacy. Classical rule-based clinical decision support systems are primarily concerned with pairwise drug interactions and are based on predefined medical knowledge. But in treatment settings where a combination of drugs is available, these methods tend to fail to represent the complicated dynamics that arise when multiple drugs interact at the same time. In this project, we propose an artificial intelligence-based system to predict polypharmacy risk using Graph Neural Networks (GNNs). The model proposed is that each drug is a node in a graph, and drug-drug interactions and pharmacological relationships are edges between nodes. This multi-drug interaction is an effective way in which this graph-based representation enables the system to model the complex interaction patterns between drugs. The model is trained using deep learning algorithms on graph-structured biomedical data to predict important representations of drug interactions and the probability of an adverse reaction developing with medication combinations. The given system combines drug interaction datasets and leverages GNNs' ability to spread information across interconnected nodes to identify known and unseen interaction patterns. The ex-perimental assessments prove the given method to be more accurate in making predictions, generalizing, and being more robust in comparison with the classical machine learning and rule-based methods.