Title: Daily Predictions of COVID-19 Cases

Abstract:

Emerging infectious diseases present unique challenges to policy makers and disease modelers. Such novel pathogens are poorly described, and epidemiological surveillance systems may not be able to capture and disseminate complete case data. What is more, policy makers may take drastically different actions in response to the same public health threat and populations adherence to those policies will vary. To account for the paucity of information on the emergent pathogen and high variation in the efficacy of control strategies, we developed the seirDynamic package. The package contains epidemiological models that require minimal data inputs to produce reliable predictions of reported cases in a given location.

The primary model in seirDynamic is an SEIR compartmental model that incorporates an estimated time varying reproduction number (Rt) to generate a dynamic effective contact rate (beta). By estimating Rt, we are able to capture local variation in a population’s behavior using case count data and the serial interval for the pathogen. Other parameters in the SEIR model are static and readily described in scientific literature early in an outbreak. Using the probability of recovery (gamma) and incubation period (sigma), it is possible to establish initial conditions for the model from case count data as well.

As SARS-COV-2, the virus that causes COVID-19, spread across the United States of America, the impacts of the virus and public health responses to it were extremely heterogenous. To address the need for short term COVID-19 forecasts, we employed our model using case data from the Johns Hopkins Center for Systems Science and Engineering. In that time, we have been able to provide timely and reliable short-term predictions of COVID 19 cases at state and county levels. Under certain conditions, our model provided unrealistic estimates at early points in a local outbreak or when daily case counts are very low. To overcome this limitation, we used the United States Department of Agriculture Economic Research Services county classification to generate state level Rt values form metropolitan counties and non-metropolitan counties. By applying these aggregated Rt values, we were able to recover adequate model performance as compared to a general state-level Rt estimate.

In conclusion, with seirDynamic, we are able to provide reliable predictions for all counties in the United States using a lightweight and transparent model. We feel this transparency is especially important when working with policy makers who must justify their decisions. We recognize that our model is unable to forecast the effects of planned policy changes, and would caution against using it for that type of analysis. However, our model intrinsically captures behavioral changes once they are reflected in the epidemiological data, making it a reliable predictor of epidemiological trends under current conditions. Given the flexibility and minimal data requirements, the seirDynamic modeling approach should be able to inform policy makers during other emerging non-vector borne communicable disease events, allowing them to respond quickly to the situation on the ground.