Speaker
Description
Effective reproduction numbers, $\mathcal{R}_E$ , provide a means of evaluating the effectiveness of interventions to control the transmission of pathogens in host populations. They can be estimated empirically from reported infections via a renewal equation or analytically via next generation matrices derived from mechanistic models. Both approaches have advantages and disadvantages, with one being more amenable to implementation via artificial and other natural intelligence. The main problems with empirical estimates are delayed, under- and biased reporting of infections and absence of heterogeneity, while analytical estimates depend on assumptions, which are too rarely evaluated. The main advantage of mechanistic models is their ability to assist in designing programs for deploying mitigation measures. To compare approaches, we used a metapopulation SEIR model of SARS-CoV-2 transmission and control in the United States that includes relevant biological characteristics of COVID-19 and whose parameters we either estimated from first principles or, with one exception, sourced from the literature. After ensuring that predictions matched accurate independent observations, we compared empirical estimates of $\mathcal{R}_E$ using simulated infections as input time-series to weekly estimates via the next generation matrix. We conclude not only that analytical results from reliable mechanistic models are more informative than empirical ones, but that artificial intelligence is not yet capable of such modeling.
Joint work with Troy Day (Queen's University) and Zhilan Feng (National Science Foundation).
