Deep learning application for real-time prediction of COVID-19 outbreak with susceptible-infected-recovered-deceased model

Abstract

Due to the complex nature of a pandemic such as COVID-19, forecasting how it would behave is difficult, but it is indeed of utmost necessity. Furthermore, adapting predictive models to different data sets obtained from different countries and areas is necessary, as it can provide a wider view of the global pandemic situation and more information on how models can be improved. Therefore, we combine here the long-short-term memory (LSTM) model and the traditional susceptible-infected-recovered-deceased (SIRD) model for the COVID-19 prediction task in Ho Chi Minh City, Vietnam. In particular, LSTM shows its strength in processing and making accurate numerical predictions on a large set of historical input. Following the SIRD model, the whole population is divided into 4 states (S), (I), (R), and (D), and the changes from one state to another are governed by a parameter set. By assessing the numerical output and the corresponding parameter set, we could reveal more insights about the root causes of the changes. The predictive model updates every 10 days to produce an output that is closest to reality. In general, such a combination delivers transparent, accurate, and up-to-date predictions for human experts, which is important for research on COVID-19.