A Multimodel Superensemble Approach for Epidemiology with Application to COVID-19 in the United States

Abstract:

In the late 1980s and early 1990s, meteorologists developed approaches that blend initial-condition and multiple-model forecasts from numerical weather prediction models. Currently, climate scientists routinely integrate climate or earth system simulations with different initializations from multiple models to understand the past and project into the future. These insights and approaches from meteorology and climate science have informed epidemiological model development over the years. This has been observed with the method of analogues, which was originally developed in meteorology and has since been successfully adapted across disciplines to epidemiology. Specifically, methods developed in our Sustainability and Data Sciences Laboratory at Northeastern University in climate risk analysis and critical infrastructure resilience have informed risk and recovery planning in the US and in India. Here we examine the hypothesis that epidemiological disease transmission predictions, from multiple mechanistic and statistical models each with different parameter estimates and initial conditions, can be effectively blended to develop credible projections with uncertainty quantification, along with characterizations of extreme values. Furthermore, we examine the hypothesis that insights from these approaches, which we call epidemiological multimodel superensemble (EMS), have the potential to inform policy with a particular emphasis on non-pharmaceutical interventions including, but not limited to, testing or sampling strategies and recovery policies. Finally, we use the EMS approach to explore the hypothesis that environmental variables, including hydrometeorology and air pollution, can significantly impact the spread of disease in diverse ways, depending on the stage of a pandemic as well as policy intervention scenarios. We present a summary of the literature, preliminary results based on existing model runs and our own simulations, as well as possible next steps.