STE-COVIDNet: A Multi-channel Model with Attention Mechanism for Time Series Prediction of COVID-19 Infection

ABSTRACT

The outbreak of COVID-19 has had a significant impact on the world. The prediction of COVID-19 can conduct the distribution of medical supplies and prevent further transmission. However, the spread of COVID-19 is affected by various factors, so the prediction results of previous studies are limited in practical application. A deep learning model with multi-channel combined multiple factors including space, time, and environment (STE-COVIDNet) is proposed to predict COVID-19 infection accurately in this paper. The attention mechanism is applied to score each environment to reflect its impact on the spread of COVID-19 and obtain environmental features. The experiments on the data of 48 states in the United States prove that STE-COVIDNet is superior to other advanced prediction models in performance. In addition, we analyze the attention weights of the environment of the 48 states obtained by STE-COVIDNet. It is found that the same environmental factors have inconsistent effects on COVID-19 transmission in different regions and times, which explains why researchers have varying results when studying the impact of environmental factors on transmission of COVID-19 based on data from different areas. STE-COVIDNet has a certain explainability and can adapt to the environmental changes, which ultimately improves our predictive performance. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.