Understanding and predicting spatio-temporal spreading of COVID-19 via diffusive graph embedding and dual SIR

Keywords: COVID-19; spread; transmission path; graph embedding; SIR
Session Type: Virtual Paper
Day: Friday
Session Start / End Time: 4/9/2021 11:10 AM (Pacific Time (US & Canada)) - 4/9/2021 12:25 PM (Pacific Time (US & Canada))
Room: Virtual 7

Abstract

In order to find useful intervention strategies of the novel coronavirus (COVID-19), it is vital to understand how the disease spread In this study, we address the modeling of COVID-19 spread across space and time, which facilitates the understanding of the pandemic and the prediction of future transmission. The number of patient cases is divided into two components: one for imported cases from other areas, and the other for local transmission cases. We leverage human mobility data to estimate the number of imported cases and initialize a transmission matrix via diffusive convolution for each areal unit. The total case data are fed into a modified compartment epidemiological model to derive time-varying infectious rates, which are used to compute both imported and local transmission case numbers. Two graph embedding-based representation schemes are used to integrate local socio-economic, demographic, and traffic flow information, resulting in node embeddings for areal units and diffusion embeddings for potential transmission routes. These embeddings, which encode potential local and imported transmission risks, are used to predict future infected numbers. The transmission matrix is updated on a daily bases and can be visualized to reveal historical transmission paths. The embeddings can also be used to underpin many visual analytics tasks, such as community detection and role analysis. A case study in state of Colorado, USA is performed to demonstrate the effectiveness of the proposed transmission modeling approach in terms of understanding and predicting spatio-temporal spreading of COVID-19 at the neighbo