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

The high R-naught factor of SARS-CoV-2 has created a race against time for mankind and it necessitates rapid containment actions to control the spread. In such scenario short term accurate spatiotemporal predictions can help understanding the dynamics of the spread in a geographic region and identify hotspots. However due to the novelty of the disease there is very little disease specific data generated yet. This poses a difficult problem for machine learning methods to learn a model of the epidemic spread from data. A proposed ensemble of convolutional LSTM based spatiotemporal model can forecast the spread of the epidemic with high resolution and accuracy in a large geographic region. A data preparation method is proposed to convert available spatial causal features into set of 2D images with or without temporal component. The model has been trained with available data for USA and Italy. It achieved 5.57% and 0.3% mean absolute percent error for total number of predicted infection cases in a 5day prediction period for USA and Italy respectively.