Deep learning forecasting using time-varying parameters of the SIRD model for Covid-19

Abstract

Abstract Accurate epidemiological models are necessary for governments, organizations, and individuals to respond appropriately to the ongoing novel coronavirus pandemic. One informative metric epidemiological models provide is the basic reproduction number ( $$R_0$$ R 0 ), which can describe if the infected population is growing ( $$R_0 > 1$$ R 0 > 1 ) or shrinking ( $$R_0 < 1$$ R 0 < 1 ). We introduce a novel algorithm that incorporates the susceptible-infected-recovered-dead model (SIRD model) with the long short-term memory (LSTM) neural network that allows for real-time forecasting and time-dependent parameter estimates, including the contact rate, $$\beta$$ β , and deceased rate, $$\mu$$ μ . With an accurate prediction of $$\beta$$ β and $$\mu$$ μ , we can directly derive $$R_0$$ R 0 , and find a numerical solution of compartmental models, such as the SIR-type models. Incorporating the epidemiological model dynamics of the SIRD model into the LSTM network, the new algorithm improves forecasting accuracy. Furthermore, we utilize mobility data from cellphones and positive test rate in our prediction model, and we also present a vaccination model. Leveraging mobility and vaccination schedule is important for capturing behavioral changes by individuals in response to the pandemic as well as policymakers.