Data assimilation method for improving the global spatiotemporal predictions of epidemic dynamics yielded by an ensemble Kalman filter and Metropolis–Hastings sampling

Abstract

AbstractAssimilating the latest epidemic data can improve the predictions of epidemic dynamics compared with those using only dynamic models. However, capturing the nonlinear spatiotemporal heterogeneity remains challenging. We propose a data assimilation method to simultaneously update the parameters and states with respect to their spatiotemporal variation intervals by (1) developing a susceptible-infected-removed-vaccinated model by considering vaccination strategy and quarantine periods and (2) assimilating real-time epidemic data using an ensemble Kalman filter for daily updates of the state variables and Metropolis–Hastings sampling for weekly parameter estimation. Synthetic experiments and a WebGIS-based global prediction system demonstrate the sufficient nowcasting accuracy of this method. An analysis of the system outcomes shows that modeling vaccination details, embedding reasonable model and observation errors, using up-to-date parameters, and avoiding the prediction of sporadic cases can increase the correlation coefficient and coefficient of determination by more than 31.35% and 161.19%, respectively, and decrease the root mean square error by more than 54.17%. Our prediction system has been working well for more than 700 days. Its worldwide nowcasting accuracies have been continuously improved, where the overall correlation coefficients, coefficient of determination, and threat percent score exceed 0.7, 0.5 and 65%, respectively. The proposed method lays promising groundwork for the real-time spatiotemporal prediction of infectious diseases.