A multivariate spatiotemporal spread model of COVID-19 using ensemble of ConvLSTM networks

swarna kamal paul; Saikat Jana; Parama Bhaumik

Disease spread is a complex dynamical system and numerous factors contribute to the dynamics of spread making it non-stationary. Covid19 is no different. Geographical location, weather conditions [7] , human mobility [8] , population statistics might be some of the impacting factors changing the dynamics of the spread. Epidemic spread is correlated in time as well as spatial dimensions. There is high chance of spreading the infection in spatially co-located geographical regions specially during community spread. Thus, a spatiotemporal model can more accurately capture the dynamics of the spread. However, it may be spatially autocorrelated in a small localized region but not across wide regions. Thus, we divided a large geographic region into relatively smaller grids and trained our model with samples drawn from local distribution of infection cases. The goal of the model is to forecast new cases of infection on daily basis in different regions across a country which can be added up to calculate total cases of infection. The model can be trained with historical time series data of new cases of infection in different regions along with some external spatial and/or spatiotemporal features. During prediction, the . CC-BY 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.