Enhancing City-Level Influenza Nowcasting on Island Terrain with Graph Neural Networks: Spatial Feature Insights

Accurate prediction of influenza outbreaks is crucial for implementing effective preventive measures. Many deep learning-based models have been proposed by incorporating temporal and spatial information, showing promising performance in influenza prediction. However, these models predominantly focus on exploring spatial information at the geographical aspect, such as gridding geographical geometry to depict geographical proximity, lacking the flexibility to incorporate spatial connections from other patterns among cities. To tackle this limitation, in this study, we propose a novel graph neural networks (GNN)-based approach with multi-graph processing, which contains different graphs reflecting different spatial patterns beyond geographical connectivity. Our proposed approach demonstrates an advantage in mining the interdependence of influenza outbreaks among distinct cities by incorporating different types of spatial patterns flexibly. The experiments on real-world influenza-like illness (ILI) dataset show that our approach achieves competitive performance, surpassing existing well established baselines. Notably, in exploring geographical relationships among nodes (cities), we innovatively incorporate background information as a new node. This new node is endowed with features distinct from the target nodes based on actual conditions. Following this strategy aids and deepens the feature interaction among nodes, resulting in a significant enhancement in predictive performance.