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

swarna kamal paul; Saikat Jana; Parama Bhaumik

is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10.1101/2020.04. 17.20069898 doi: medRxiv preprint longitude with each section containing 9x15 grids. A set of 4 heterogenous ensembles are trained for each of the 4 sections. The configuration of networks is same as that of Italy except they contain 2 hidden layers and each network is trained for 20 epochs. The implementation has been done on Python and code is available at https://github.com/swarna-kpaul/covid19spatiotemporal. Table I shows the performance of the models in terms of KL divergence and MAPE. For both USA and Italy, low KL divergences states that the predicted geospatial probability distribution of total infection cases nearly matches with the actual probability distribution. The pixel level MAPE for Italy stays below 30%. For USA in 8-day forecasting period MAPE is 44% as there are many pixels in USA with low total patient count. A slight deviation in the prediction for these pixels shoots up the MAPE. Country level MAPE is low for both Italy and USA. Fig. 3a and 4a shows predicted vs actual total Covid-19 cases for a period of 8 and 10 days in USA and Italy respectively. For Italy the prediction follows closely with the actual whereas for USA it is little underestimated. Fig. 3b and 4b shows predicted vs actual daily new Covid-19 cases for a period of 8 and 10 days in USA and Italy respectively. Fig. 5a and 5b shows the distribution of total predicted vs actual infection cases in each pixel after 10 day and 8 day in Italy and USA respectively. The predicted distribution closely follows with actual with residuals distributed both on negative and positive side.