Predicting Polypharmacy Effects Using Explainable AI

Polypharmacy - the concurrent use of multiple medications, common among elderly and chronically ill patients - significantly increases the risk of adverse drug-drug interactions that are difficult to predict due to the combinatorial explosion of possible drug combinations. This work proposes an explainable AI framework for pairwise drug-drug interaction prediction as a foundation for polypharmacy risk assessment, combining graph-based drug-interaction modelling with explainability techniques so that predictions are both accurate and clinically interpretable. A three-layer relational graph neural network with a relation-specific bilinear decoder is trained on TWOSIDES adverse-event edges under negative sampling, and compared against logistic-regression, MLP and graph baselines on AUROC, AUPRC, F1 and expected calibration error. A subgraph-explanation layer is evaluated quantitatively on fidelity and sparsity rather than presented illustratively, together with an error-mode analysis separating knowledge-graph coverage failures from model-capacity failures. Reported values are point estimates from single runs without per-seed dispersion or significance testing. Modelling of higher-order (three or more drug) interactions is left to future work.