Graph neural ordinary differential equations for epidemic forecasting

• Published in: CCF transactions on pervasive computing and interaction (Online) Vol. 6; no. 3; pp. 281 - 295
• Main Authors: Yanqin, Xiong, Huandong, Wang, Guanghua, Liu, Yong, Li, Tao, Jiang
• Format: Journal Article
• Language: English
• Published: Singapore Springer Nature Singapore 01.09.2024; Springer Nature B.V
• Subjects: Computer Science; COVID-19; Differential equations; Disease control; Disease transmission; Epidemics; Forecasting; Graph neural networks; Infectious diseases; Medical research; Neural networks; Ordinary differential equations; Predictive control; Regions; Regular Paper; Time series; Traffic flow; User Interfaces and Human Computer Interaction; Epidemic; Neural ODE; Spatial temporal forecasting; Graph neural network
• ISSN: 2524-521X; 2524-5228
• DOI: 10.1007/s42486-024-00161-0

Developing a practical model for predicting the dynamics of epidemic spread is critical, the results of which can be used to evaluate the effectiveness of prevention and control measures as well as the allocation of medical and health resources. However, accurate prediction is challenging because epidemic spread is closely associated with population mobility, which is nonlinear and complex, making reliable prediction difficult. Furthermore, the epidemic observed data is sparse and irregularly sampled, rendering the traditional time series models ineffective. Under these circumstances, this paper designs a graph neural ordinary differential equations approach, which combines Ordinary Differential Equation Networks (ODENet) and Graph Neural Networks (GNNs). This approach adopts a new attention mechanism taking into account the interaction between regional epidemic information and interaction between regions, achieving the precise continuous-time epidemic prediction based on non-Euclidean data. In addition, we use Transformer to deduce the value of the initial hidden state via future observable data in an innovative way, reconstructing the hidden state successfully. We conduct a lot of experiments based on the contact matrix and simulated epidemic data in San Francisco from March 2020 to May 2020, results show that our method can not only forecast the dynamics of epidemic spread, but also mine hidden patterns in observable data and extract hidden states.