Epidemic Model Guided Machine Learning for COVID-19 Forecasts in the United States

Difan Zou, Lingxiao Wang, Pan Xu, Jinghui Chen, Weitong Zhang, Quanquan GumedRxiv 2020 · DOI

Abstract

We propose a new epidemic model (SuEIR) for forecasting the spread of COVID-19, including numbers of confirmed and fatality cases at national and state levels in the United States. Specifically, the SuEIR model is a variant of the SEIR model by taking into account the untested/unreported cases of COVID-19, and trained by machine learning algorithms based on the reported historical data. Besides providing basic projections for confirmed and fatality cases, the proposed SuEIR model is also able to predict the peak date of active cases, and estimate the basic reproduction number ($R_0$). In particular, the forecasts based on our model suggest that the peak date of the US, New York state, and California state are 06/01/2020, 05/10/2020, and 07/01/2020 respectively. In addition, the estimated $R_0$ of the US, New York state, and California state are 2.5, 3.6 and 2.2 respectively. The prediction results for all states in the US can be found on our project website: https://covid19.uclaml.org, which are updated on a weekly basis, and have been adopted by the Centers for Disease Control and Prevention (CDC) for COVID-19 death forecasts.

Introduction: Addressing the "Dark Figure" of COVID-19

Early in the COVID-19 pandemic, researchers faced a significant "data gap." Classical epidemic models like SIR (Susceptible-Infectious-Removed) and SEIR (Susceptible-Exposed-Infectious-Removed) rely heavily on reported data. However, due to limited testing capacity and a high volume of asymptomatic "silent spreaders," the number of publicly reported cases was often an order of magnitude lower than the actual number of infections. Standard models that failed to account for this "dark figure" of unreported cases often produced unreasonable predictions. Furthermore, most existing work focused on national trends, leaving a need for granular, state-level forecasts to assist local policy makers in resource allocation.

Methodology: The SuEIR Framework

The authors introduce the SuEIR model, a variant of the SEIR model designed to bridge the gap between reported data and actual disease dynamics. The core innovation lies in the treatment of the "Exposed" ($E$) compartment and the introduction of an "Unreported" pathway.

The SuEIR model structure, explicitly modeling the path to unreported recovery.

Unlike standard models where the "Exposed" ($E$) category is a passive waiting room for symptoms, the SuEIR model treats the $E$ compartment as individuals who are already infectious but haven't been tested. A key novel parameter, the Discovery Rate ($\mu$), represents the fraction of exposed cases that eventually get confirmed and reported ($I$). The remaining $(1-\mu)$ fraction transitions directly to an "Unreported Recovered" state. As shown in the model diagram, both the "Exposed" (untested) and "Infectious" (reported) groups contribute to the infection of the "Susceptible" population, reflecting the reality of asymptomatic transmission.

Machine Learning Integration and Parameter Estimation

Instead of manually tuning parameters, the authors use a gradient-based optimizer (BFGS) to minimize a logarithmic-type mean square error (MSE) loss function. This allows the model to "learn" the contact rate ($\beta$), discovery rate ($\mu$), and transition rates ($\sigma, \gamma$) directly from historical data.

One significant challenge was that many jurisdictions only reported "Total Confirmed" cases, lacking a breakdown of active versus removed cases. To address this, the authors developed an exponential function to estimate the ratio between daily fatalities and total removed cases. This method was validated using international datasets where more complete data was available.

Comparison of estimated and true removed cases across four countries, validating the model's robust parameter estimation.

Key Insights: Silent Spreaders and Regional Variance

The model revealed that the discovery rate ($\mu$) was remarkably low—often less than 0.1 for most states. This suggests that the actual number of infected individuals in the US was likely 10 times higher than official counts, a finding later corroborated by seroprevalence studies.

Furthermore, the model captured significant regional variance in the basic reproduction number ($R_0$). The estimated $R_0$ in New York (3.6) and New Jersey (4.5) was significantly higher than in California (2.2), reflecting the higher population density and earlier outbreak intensity in the Northeast.

Results and Real-World Impact

At the time of its initial deployment in May 2020, the model provided critical foresight:

  • Peak Predictions: It accurately predicted peak active cases for the US on 06/01/2020 and New York on 05/10/2020.
  • Forecasting Accuracy: Short-term predictions for total deaths were remarkably close to ground truth. For instance, on 05/18/2020, the model predicted 91.13k US deaths; the actual reported number was 90.35k.
  • Long-term Outlook: The model forewarned that by the end of June 2020, the US would reach approximately 2 million confirmed cases and 120k deaths.

Due to its robust performance, the SuEIR model was officially adopted by the CDC for its COVID-19 death forecasting ensemble, contributing to national public health planning.

Conclusion and Limitations

The primary limitation of the current SuEIR model is the assumption of a constant contact rate. Because it was trained on data from a period of strict "stay-at-home" orders, it does not automatically account for sudden changes in human behavior or policy, such as phased reopenings. Future work involves incorporating time-varying parameters to account for social distancing fatigue, policy shifts, and the eventual introduction of vaccines.

Summary updated

Abstract

Abstract We propose a new epidemic model (SuEIR) for forecasting the spread of COVID-19, including numbers of confirmed and fatality cases at national and state levels in the United States. Specifically, the SuEIR model is a variant of the SEIR model by taking into account the untested/unreported cases of COVID-19, and trained by machine learning algorithms based on the reported historical data. Besides providing basic projections for confirmed and fatality cases, the proposed SuEIR model is also able to predict the peak date of active cases, and estimate the basic reproduction number ( ). In particular, the forecasts based on our model suggest that the peak date of the US, New York state, and California state are 06/01/2020, 05/10/2020, and 07/01/2020 respectively. In addition, the estimated of the US, New York state, and California state are 2.5, 3.6 and 2.2 respectively. The prediction results for all states in the US can be found on our project website: https://covid19.uclaml.org , which are updated on a weekly basis, and have been adopted by the Centers for Disease Control and Prevention (CDC) for COVID-19 death forecasts ( https://www.cdc.gov/coronavirus/2019-ncov/covid-data/forecasting-us.html ).
BibTeX
@misc{zou2020epidemic,
  author = {Difan Zou and Lingxiao Wang and Pan Xu and Jinghui Chen and Weitong Zhang and Quanquan Gu},
  title = {{Epidemic Model Guided Machine Learning for COVID-19 Forecasts in the United States}},
  year = {2020},
  doi = {10.1101/2020.05.24.20111989},
  howpublished = {medRxiv},
}
Figure