Graph Neural Network Drug-Drug Interaction Prediction Under Sparse African Pharmacovigilance Reporting: A Label-Noise-Robust Evaluation of Artificial Intelligence Prescribing Safety

This is the authors' abstract. We don't add key points for this paper.

Drug-drug interaction (DDI) prediction models trained on high-income pharmacovigilance corpora are increasingly proposed for prescribing safety in African health systems, yet the label-generating process in those systems differs in ways that standard evaluation ignores. the present analysis treats the problem as one of label noise rather than of model capacity. It formalises sparse African pharmacovigilance reporting as a non-random missingness mechanism over true interaction labels, derives the bias that this mechanism induces in the standard binary cross-entropy objective, and proposes a noise-robust graph neural network estimator that reweights edges by an estimated reporting propensity. Because no verified African DDI corpus was available for the present analysis, the paper is methodological: it characterises the operating characteristics of the proposed estimator under a declared variance structure rather than reporting empirical performance on observational data. The design is a before-after-control-impact evaluation with between-cluster standard deviation 6.0, residual standard deviation 4.5 and intraclass correlation 0.64, estimated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. Power analysis indicates that 52 clusters are required to detect a 5 percentage point effect at 80% power. The contribution is a formal account of how reporting sparsity propagates into DDI classifier error, together with a design-level evaluation protocol that African pharmacovigilance networks can use to size prospective validation studies before deploying DDI alerts.