COVID-Net: A deep learning based and interpretable predication model for the county-wise trajectories of COVID-19 in the United States

ABSTRACT

The cases of COVID-19 have been reported in the United States since January 2020. We propose a COVINet by combining the architecture of both Long Short-Term Memory and Gated Recurrent Unit. First, we use the 10-fold cross-validation to train and assess different prediction models for which all counties serve alternatively as the training and test counties. Then, we focus on the prediction for the 10 severest counties. We employ the Mean Relative Errors (MREs) to measure the performance of the COVINet in predicting confirmed cases and deaths. Two COVINet models with 26 and 19 input variables, respectively, are trained. We estimate their respective MREs in the last 30 days before January 23, 2021, by the 10-fold CV, which are 0.0898 and 0.1068 for the number of confirmed cases, and 0.0694 and 0.0724 for the number of deaths. The MREs are also small for all predictions of the events in the last 7 or 30 days before January 23, 2021. The COVINet uses features including workforce driving alone to work, traffic volume, income inequality, and longitude and latitude of infected counties to predict the trajectories of COVID-19 in counties of the United States. The increasing awareness of how predictors affect the pandemic helps policymakers develop plans to mitigate the spread of COVID-19.