Dynamic parameterization of a modified SEIRD model to analyze and forecast the outbreak evolution of COVID-19 in the United States



Journal Title

Journal ISSN

Volume Title



The rapid spread of the SARS-CoV-2 (Severe Acute Respiratory Syndrome Coronavirus 2) virus since its first detection in Wuhan, China, in December 2019 caused the COVID-19 (Coronavirus Disease 2019) epidemic to quickly turn into a pandemic. Numerous outbreaks across continents fueled the interest in mathematical models to understand and predict infectious disease dynamics with the ultimate goal of contributing to the decision-making of public health authorities. These mathematical epidemiology models can be classified into two main groups. Statistical models rely heavily on data series to exploit observed trends in contagion dynamics for the prediction of their future evolution, but they do not provide insight into the underlying mechanisms of COVID-19 spread. Conversely, mechanistic models incorporate underlying mechanisms of disease contagion and transmission but usually assume a constant parameterization for separate waves of the pandemic. Some of these mechanistic models are also not widely deployable as they require specific data that are not universally collected. In this work, we propose a mechanistic approach that leverages widely available epidemiological data to provide a smooth time-resolved parameterization of the temporal dynamics of COVID-19 spread. We employ a modified SEIRD (Susceptible-Exposed-Infected-Recovered-Deceased) model that is initially fit to successive one-week intervals of cumulative infection and deaths over a given timeframe. Then, we apply a mean filter to obtain daily estimates of the model parameters from the weekly parameterizations. Finally, we fit quadratic splines to the daily parameter estimates to render a smooth parameterization for analysis and forecasting. We tested our methodology in five of the most heavily impacted states of the US (by cases) between March 2020 and August 2020, when the nationwide seroprevalence (i.e., presence of SARS-CoV-2 antibodies in the blood) estimates were first available. Our calibration framework achieves a median [range] NRMSE (normalized root mean squared error) of cumulative cases and deaths of 3.64% [1.68%, 7.96%] and 4.22% [1.38%, 6.20%], respectively. Then 2-week forecasts with the state-calibrated model result in a median [range] NRMSE of cumulative cases and deaths of 3.48% [1.30%, 12.16%] and 4.70% [0.86% 8.91%], respectively, while 4-week forecasts show a corresponding median NRMSE of 3.95% [0.91%, 43.90%], and 11.03% [0.96%, 35.26%]. Assimilation of the 2 subsequent weeks, decreases the median NRMSE of the initial forecast of cumulative cases and deaths by 8.79% and 1.96% respectively in the first week, while the corresponding median NRMSE in the forecast of the second week is reduced by 9.5% and 1.23%. As dynamic parameters can enable a consistent model-based explanation of the progression of the pandemic, we believe that our methodology can be further developed into a predictive technology enabling the analysis and forecast of the outbreak dynamics of infectious diseases, which can contribute to the design of effective pandemic-arresting public health policies.


LCSH Subject Headings