RAPID: Collaborative Research: Operational COVID-19 Forecasting with Multi-Source Information

Mathematical and Physical Sciences - This project aims to develop a new deep learning predictive platform for COVID-19 transmission, integrating multi-source information under model and data uncertainties. In contrast to other viruses such as influenza, SARS, and MERS, COVID-19 differs in a number of ways, including uncertainties in response to weather conditions, history of the disease, as well as the effectiveness of responses from public health officials or from the general public. An important aspect is to integrate multi-source data such as official reports, atmospheric variables, and social media data into operational biosurveillance and real-time prediction of COVID-19. The proposed biosurveillance framework will be used to forecast COVID-19 dynamics and to enhance mitigation strategies. In addition, it could also be applicable to tracking many other infectious diseases, thereby contributing to security of our society as a whole. Furthermore, the project will build innovative connections within and across mathematical biology, statistics, and deep learning, with a strong focus on interdisciplinary graduate research training.

As the main forecasting framework, the widely used Susceptible-Exposed-Infected-Recovered (SEIR) dynamic models can accurately describe the disease dynamics, but only with precise knowledge of disease parameters, which can take a long time to accurately estimate. Deep learning algorithms can potentially have superior predictive ability, but they require extensive training. Another key challenge in the statistical modeling of these events is how to timely and systematically integrate multiple sources of surveillance, anecdotal, and other health-related information under uncertainty. The proposed new predictive approach is based on the interaction between multiple data sources, dynamical SEIR models, and deep learning algorithms. The key idea is to view simulation SEIR models as ?surrogate? pre-trainers for the deep learning models, resulting in less real data needed to retrain the predictive model to reflect ?real world? COVID-19 progression. Deep learning predictive models can then be used for making predictions about the future COVID-19 dynamics, which can be compared to the predictions made by the original SEIR model. Depending on which mathematical model makes better predictions, another model can be updated with the better prediction as inputs, thereby representing reinforcement learning from both data and the best mathematical model. As a result, the new predictive framework will allow one to assess impacts of the immediate responses such as declaration of a national emergency, a school closing, or a quarantine, and can be considered as a step toward interpretable AI for COVID-19 biosurveillance.

This grant is being awarded using funds made available by the Coronavirus Aid, Relief, and Economic Security (CARES) Act supplemental funds allocated to MPS.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.