Long-Term COVID-19 Outbreak Prediction using Time Difference Data

The massive damage caused by COVID-19 worldwide over the past two years has highlighted the importance of predicting the spread of infectious diseases. Therefore, with advances in deep learning, numerous and diverse methods have been considered for predicting the spread of infectious diseases. However, these studies have shown that the long-term prediction abilities of deep learning models are insufficient to predict the course and propagation of COVID-19 outbreaks. In this study, we consider the hypothesis that COVID-19 data of a specific region often have a strong correlation with that of other regions and related web data, such as Google search queries, with a time difference. We propose a long-term COVID-19 prediction methodology using time-difference data, in which the correlation is calculated by applying the time difference between COVID-19 propagation data of a specific region and of other regions or related web data. We subsequently processed the data with a sequence-to-sequence attention model. The results of an experimental evaluation showed that the proposition was inductively proven. Our model performed better than existing models for long-term predictions. In addition, our proposed model showed better performance in predicting data on the spread of COVID-19 in nine US states than previous state-of-the-art methods. Our code is available at https://github.com/lurker18/COVID-19\_TDD