A Novel Multi-scale Spatiotemporal Graph Neural Network for Epidemic Prediction

Predicting epidemics is of vital significance for safeguarding human life, health, and safety. Spatio-temporal graph neural networks have been successfully employed in epidemic forecasting, as they can extract information from both the temporal and spatial dimensions of the transmission process. However, the current approaches of using graph neural networks for epidemic forecasting only consider single variate infectious disease data and overlook the relationships between diffusion patterns at the micro level and macro level in terms of spatial information. Incorporating the relationships between diffusion patterns at different spatial scales is crucial for accurate epidemic forecasting, as it captures the complex dynamics of disease transmission across various levels of granularity. These limitations significantly affect the accuracy and rationality of prediction results. In this paper, we propose a Multi-Scale Spatio-Temporal graph neural network (MSST) that incorporates multivariate infectious disease data for epidemic prediction. This multi-scale structure aligns the predictive model more closely with the characteristics of infectious disease transmission and can better capture the hierarchical nature of disease transmission, from local clusters to regional and global spread. The experimental results demonstrate that our model effectively extracts predictive information and integrates it across multiple scales, leading to improved epidemic forecasting accuracy.