Facilitating the spread prediction of public health emergencies based on spatio-temporal neural network.

To overcome the shortcomings of traditional models in capturing the spatiotemporal dynamics of epidemic spread, this study introduces an innovative model, hybrid spatio-temporal neural network architecture (MI-GCN-LSTM), which integrates mutual information (MI), long short term memory (LSTM), and graph convolutional network (GCN) to collaboratively predict epidemic trends across multiple regions during public health emergencies. First, an adjacency matrix is constructed using mutual information theory to quantify the influence of complex behaviors such as interregional population movement, thus capturing potential spatial dependencies more accurately. Second, GCN are applied to identify transmission patterns across regions and extract spatial features. Ultimately, LSTM are employed to model long-term temporal dependencies and dynamic trends in the development of epidemics. To evaluate the model's effectiveness, a case study uses COVID-19 data from several European countries provided by the World Health Organization (WHO), as well as city-level data from Hubei Province released by its Health Commission. Results show that the proposed model achieves superior performance compared to the suboptimal LSTM-GCN model on datasets from several European countries, reducing MAE and RMSE by 9.22% and 4.24%, respectively. On the Hubei dataset, the improvements are more substantial, MAE decreased by approximately 27.74%. This study provides decision making support for emergency response and lays the foundation for future forecasting and resource allocation.