Exploring Feasibility of Multivariate Deep Learning Models in Predicting COVID-19 Epidemic

Background: Mathematical models are powerful tools to study COVID-19. However, one fundamental challenge in current modeling approaches is the lack of accurate and comprehensive data. Complex epidemiological systems such as COVID-19 are especially challenging to the commonly used mechanistic model when our understanding of this pandemic rapidly refreshes. Objective: We aim to develop a data-driven workflow to extract, process, and develop deep learning (DL) methods to model the COVID-19 epidemic. We provide an alternative modeling approach to complement the current mechanistic modeling paradigm. Method: We extensively searched, extracted, and annotated relevant datasets from over 60 official press releases in Hubei, China, in 2020. Multivariate long short-term memory (LSTM) models were developed with different architectures to track and predict multivariate COVID-19 time series for 1, 2, and 3 days ahead. As a comparison, univariate LSTMs were also developed to track new cases, total cases, and new deaths. Results: A comprehensive dataset with 10 variables was retrieved and processed for 125 days in Hubei. Multivariate LSTM had reasonably good predictability on new deaths, hospitalization of both severe and critical patients, total discharges, and total monitored in hospital. Multivariate LSTM showed better results for new and total cases, and new deaths for 1-day-ahead prediction than univariate counterparts, but not for 2-day and 3-day-ahead predictions. Besides, more complex LSTM architecture seemed not to increase overall predictability in this study. Conclusion: This study demonstrates the feasibility of DL models to complement current mechanistic approaches when the exact epidemiological mechanisms are still under investigation.