Mining adverse drug events from patients' disease histories via a GNN-based subgraph prediction method

Adverse drug events have been a significant concern in medication development as this issue has resulted in comparatively high risks of hospital admissions and death rates worldwide. Traditionally, these risks should have been identified through clinical trials, which could be time-consuming and cost-inefficient. At the same time, it still leaves some adverse drug events unknown due to limited samples in labs during the early phases of drug development. Therefore, various machine-learning techniques have been used in this field of study to support the discovery of adverse drug events as early as possible. This paper presents a state-of-the-art network-based approach to model each patient as a subgraph that consists of nodes of ICD-10 codes and directed edges showing the progression of their diseases. With the help of four Graph Neural Network variants, we could address three research questions 1) whether, 2) when, and 3) which ADEs would occur for a particular patient. In this short paper, we analysed the first question - how this method could be used in identifying cohorts associated with adverse drug events. The experiment showed that the GraphSage method employed on our suggested graph provided the highest accuracy - 0.8863 and the highest recall - 0.9128 for this research question. The same GNN-based framework could also be applied to the remaining research questions.