RAPID: Fast COVID-19 Scenario Projections in Presence of Vaccines and Competing Variants

After more than a year, COVID-19 remains a concern worldwide. While the United States is moving towards a fast reopening of economic activities, it is crucial to ensure that it can be done without an increased burden on the healthcare system. As a result, there is an urgency to produce reliable long-term scenario projections of cases, hospitalizations, and deaths to inform policymakers. New challenges in modeling and estimations are emerging due to emerging competing variants with transmissibility advantage, possible waning immunity and immune escape, vaccine hesitancy, and changes in non-pharmaceutical interventions. The project will address these emerging challenges in scenario projections at the state-level in the US. A key advantage of the modeling technique is that it can incorporate various complexities and learn from a changing epidemiological and social environment, and yet it can produce fast projections. The techniques developed in the project will not only be applicable to the US locations but also locations around the world where COVID-19 is still a severe disaster. The scenario modeling framework developed during the project will also set the foundations for quick scenario generation for better preparedness during future epidemics. This project provides training opportunities for a graduate student.

The proposed project develops a discrete-time heterogeneous rate model that can incorporate various complexities of COVID-19 and yet produce long-term scenario projections quickly on commodity hardware (2-3 mins/scenario for all US states). The fast projections of cases, hospitalizations, and deaths are enabled by decoupling of the parameter estimations so that they can be learned independently using simple regression techniques. This also results in the elimination of over-fitting arising from simultaneously learning complex interdependent parameters and from high-dimensional machine learning approaches. Projections are generated as probabilistic quantiles for a given scenario, health outcome, week, and location; the quantiles are predicted based on an ensemble of projections resulting from the uncertainties in data inputs and estimations. The project will also develop a novel constrained optimization-based learning approach to estimate the temporal dynamics of competing variants from genomics data.

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.