Graph neural networks for the enhanced prediction of new atrial fibrillation episodes after stroke

Traditionally, statistical and machine learning (ML) algorithms have been used to develop risk prediction models for adverse clinical events, such as atrial fibrillation (AF) after stroke. However, these algorithms often fail to encapsulate or exploit possible connections between patients, assuming each patient is fully independent. This study builds a graph of patients interlinked by their medical histories to create a graph-based risk prediction model for AF. We investigate the ability of Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) to predict AF risk in critically ill stroke patients. We introduce a novel, patient-specific approach for computing similarities between GNN nodes and explore several methods for GNN explainability, including node-specific Shapley value analysis and node relationships based on the attention coefficients of the GAT model. Our findings show that GCN and GAT models, with AUCs of 0.81[0.78-0.84] and 0.84[0.81-0.87], consistently outperform traditional algorithms such as Random Forest, XGBoost, and Logistic Regression, which had the best AUC of 0.78 [0.74-0.82]. This superior performance is observed when our proposed custom similarity metric is used to construct the graph, highlighting the importance of task-specific graph design in enhancing model effectiveness. The attention mechanisms in GAT models likely contributed to this improved performance. This study highlights the strength of GNNs in capturing complex relationships and provides insights into model predictions, demonstrating the generalisability of our methodological approach to other risk prediction models.