Applications of deep learning in forecasting COVID-19 pandemic and county-level risk warning

In COVID-19 pandemic, it is important to construct a reliable prediction scheme and a real-time risk warning system. Rapid-developing deep learning (DL) techniques, in particular, provide toolboxes for forecasting situations from considerable amounts of real-time data that are currently collected and characterized by unprecedented spatial and temporal coverage. In this chapter, we first review the recent efforts that adopted diverse DL approaches to predict the COVID-19 pandemic after the early burst. After that we introduce different analyses of spatial motions at multiscale. Some new paradigms are discussed on spatial-temporal datasets. In the end, we demonstrate our epidemiological model-driven DL framework on county-level predictions, in which spatial-temporal evolutions of infection cases are driven by a cellular automata sensitive-undiagnosed-infected-removed (CA-SUIR) model. Training on the dataset generated from CA-SUIR model, this new toolbox is used to predict the prevalence of multiscale Covid-19 in all 412 counties in Germany. It can also be naturally extended to multinational or transnational analyses. Based on this framework, we also discuss the possible extensions of introducing vaccination rates and virus variants into DL methods.