Time series forecasting of COVID-19 infections and deaths in Alpha and Delta variants using LSTM networks.

Abstract

Since the beginning of the rapidly spreading COVID-19 pandemic, several mutations have occurred in the genetic sequence of the virus, resulting in emerging different variants of concern. These variants vary in transmissibility, severity of infections, and mortality rate. Designing models that are capable of predicting the future behavior of these variants in the societies can help decision makers and the healthcare system to design efficient health policies, and to be prepared with the sufficient medical devices and an adequate number of personnel to fight against this virus and the similar ones. Among variants of COVID-19, Alpha and Delta variants differ noticeably in the virus structures. In this paper, we study these variants in the geographical regions with different size, population densities, and social life styles. These regions include the country of Iran, the continent of Asia, and the whole world. We propose four deep learning models based on Long Short-Term Memory (LSTM), and examine their predictive power in forecasting the number of infections and deaths for the next three, next five, and next seven days in each variant. These models include Encoder Decoder LSTM (ED-LSTM), Bidirectional LSTM (Bi-LSTM), Convolutional LSTM (Conv-LSTM), and Gated Recurrent Unit (GRU). Performance of these models in predictions are evaluated using the root mean square error, mean absolute error, and mean absolute percentage error. Then, the Friedman test is applied to find the leading model for predictions in all conditions. The results show that ED-LSTM is generally the leading model for predicting the number of infections and deaths for both variants of Alpha and Delta, with the ability to forecast long time intervals ahead.