Fast and Accurate Forecasting of COVID-19 Deaths Using the SIkJα Model

Ajitesh Srivastava, Tianjian Xu, Viktor K. PrasannaarXiv 2020 · arXiv

Introduction

Accurate forecasting during a pandemic is a notoriously difficult task due to the inherent "noise" in public health data. Factors such as unreported asymptomatic cases, varying testing rates, and significant reporting delays often obscure the true trajectory of an outbreak. Traditional epidemiological frameworks, like the standard Susceptible-Exposed-Infectious-Recovered (SEIR) model, typically rely on complex simulations or Bayesian inference to estimate numerous parameters simultaneously. While powerful, these approaches suffer from two major drawbacks: high computational costs and a tendency toward overfitting, where models become too "tuned" to historical noise to predict future trends accurately.

Researchers at the University of Southern California have addressed these challenges by developing an extension to the SIkJα model. This approach prioritizes speed and accuracy by mathematically simplifying complex epidemiological dynamics into a system of linear equations.

Methodology: The Power of Linearization

The core innovation of the SIkJα model lies in its use of linearization. While the underlying biological processes of an epidemic are non-linear, the authors demonstrate that these dynamics can be transformed into a mathematically equivalent system of simple linear models. This reduction in complexity significantly lowers the number of parameters that need to be estimated, thereby reducing the risk of overfitting and boosting computational efficiency.

The authors simplify a complex epidemiological network into a mathematically equivalent linear system to reduce parameters and boost speed.

Key technical features of the model include:

  • Sub-state Infection Modeling: The model incorporates $k$ sub-states over $J$ days to capture the evolution of infectiousness from the moment of contraction.
  • Weighted Least Squares (LSER): Instead of computationally intensive simulations, the model employs linear regression to learn parameters. A "forgetting factor" ($\alpha$) is used to assign greater weight to recent data, allowing the model to adapt rapidly to shifts in public health policy or social behavior.
  • Efficient Hyper-parameter Tuning: Parameters such as the fraction of the population that is immune or isolated ($\rho$) and reporting lags ($\lambda$) are optimized through a fast grid search on a validation set. This keeps the core training process strictly linear and exceptionally fast.

A significant hurdle in pandemic modeling is the lack of a single "ground truth." The study evaluates the SIkJα model against three major, often conflicting, US data sources: Johns Hopkins University (JHU), The New York Times (NYT), and US Facts (USF). As illustrated in the heatmaps below, these sources can disagree by over 150 deaths for the same state on the same day.

Heatmaps illustrating the significant discrepancies between JHU, NYT, and USF data sources across different states and dates.

By testing against multiple datasets, the authors ensure the model's robustness. Their findings suggest that in a fast-moving pandemic, prioritizing the most recent data points—facilitated by the $\alpha$ parameter—is crucial for reliable short-term (2-week) forecasts.

Key Results and Performance

The SIkJα model was benchmarked against seven prominent models used by the CDC, including UCLA’s SuEIR and Columbia University’s CU_SELECT. During the evaluation period from May 10 to June 28, 2020, the SIkJα model achieved a lower Root Mean Squared Error (RMSE) than all baseline models in the majority of test cases.

Beyond accuracy, the model's efficiency is remarkable. On a standard 2-core desktop machine, it takes only 3.18 seconds to tune hyper-parameters, learn parameters, and generate 100-day forecasts for all US states. This scalability extends to the local level, where the model can process data for over 3,000 US counties in just 101.03 seconds, enabling near real-time, granular forecasting that is often impossible with more complex mechanistic models.

Conclusion and Limitations

While the SIkJα model excels in prediction and speed, it involves trade-offs. The authors acknowledge that the linearized parameters may lose some specific biological interpretability compared to non-linear mechanistic models. Furthermore, while the model accounts for the impact of unreported cases, it does not explicitly differentiate between the transmission dynamics of asymptomatic and symptomatic individuals.

Despite these limitations, the SIkJα model represents a significant advancement in public health tools, providing a fast, accurate, and highly scalable framework for real-time pandemic response.

Summary updated

Abstract

Forecasting the effect of COVID-19 is essential to design policies that may prepare us to handle the pandemic. Many methods have already been proposed, particularly, to forecast reported cases and deaths at country-level and state-level. Many of these methods are based on traditional epidemiological model which rely on simulations or Bayesian inference to simultaneously learn many parameters at a time. This makes them prone to over-fitting and slow execution. We propose an extension to our model SIkJ$\alpha$ to forecast deaths and show that it can consider the effect of many complexities of the epidemic process and yet be simplified to a few parameters that are learned using fast linear regressions. We also present an evaluation of our method against seven approaches currently being used by the CDC, based on their two weeks forecast at various times during the pandemic. We demonstrate that our method achieves better root mean squared error compared to these seven approaches during majority of the evaluation period. Further, on a 2 core desktop machine, our approach takes only 3.18s to tune hyper-parameters, learn parameters and generate 100 days of forecasts of reported cases and deaths for all the states in the US. The total execution time for 184 countries is 11.83s and for all the US counties ($>$ 3000) is 101.03s.
BibTeX
@misc{srivastava2020fast,
  author = {Ajitesh Srivastava and Tianjian Xu and Viktor K. Prasanna},
  title = {{Fast and Accurate Forecasting of COVID-19 Deaths Using the SIkJα Model}},
  year = {2020},
  eprint = {2007.05180},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2007.05180},
}
Figure