Quantitative Study on American COVID-19 Epidemic Predictions and Scenario Simulations

Jingtao Sun, Jin Qi, Zhen Yan, Yadong Li, Jie Liang, Sensen WuIsprs Int. J. geo inf 2024 · DOI

TL;DR

This paper introduces a novel spatiotemporal heterogeneity analysis (GTNNWR) model for predicting COVID-19 deaths in the US, demonstrating superior accuracy to traditional SEIR models. It also quantifies the effectiveness of various prevention measures, finding that combined strategies are significantly more impactful than individual ones, especially during winter.

What Problem Does This Paper Address?

The COVID-19 pandemic has had a profound global impact, necessitating accurate prediction of epidemic trends to inform public health policies and resource allocation. While traditional epidemiological models like SEIR have been used, they often struggle with the complex spatiotemporal heterogeneity of disease spread and can be highly sensitive to parameter errors. Existing geographical models, while accounting for spatial variations, sometimes lack temporal dynamics or efficient handling of large datasets. This study aims to bridge this gap by developing a model that captures both spatial and temporal complexities for more accurate COVID-19 death predictions and to quantitatively assess the impact of various prevention strategies.

Methods

The core of this research is the Geographically and Temporally Neural Network Weighted Regression (GTNNWR) model. This model builds upon the Geographically Neural Network Weighted Regression (GNNWR) by incorporating a spatiotemporal proximity neural network (STPNN) to account for both spatial and temporal dependencies. This allows it to estimate spatiotemporally non-stationary relationships, which are crucial for understanding disease spread. The GTNNWR model uses ordinary linear regression (OLR) to calculate initial coefficients, which are then refined using the neural network to capture complex, non-linear interactions.

As a benchmark, the study also employs the traditional SEIR (Susceptible-Exposed-Infected-Recovered) model, a standard epidemiological model that divides the population into compartments and uses differential equations to simulate disease progression.

For scenario simulations, five distinct scenarios were designed, varying the intensity of three key epidemic prevention measures: mask-wearing, vaccination, and social distancing. These scenarios were then fed into the trained GTNNWR model to predict future death tolls.

Datasets

The study utilizes monthly COVID-19 death data from April 2020 to June 2023 for 49 states in the continental United States. The independent variables include:

  • Epidemic prevention factors: Outdoor mask usage rate, social distancing index, and total vaccination coverage rate.
  • Natural environmental factors: Temperature, wind speed, surface pressure, and precipitation.
  • Socioeconomic factors: Percentage of population with a college degree, poverty rate, unemployment rate, median household income, percentage of population that did not complete high school, and proportion of the elderly population.
  • Seroprevalence factor: Percentage of the population that had been infected by COVID-19.

Data for mask usage and vaccination rates were sourced from the Institute for Health Metrics and Evaluation (IHME). Social distancing index data was obtained from the University of Maryland's COVID-19 dataset.

Key Insights and Ideas

  • GTNNWR's Superiority: The GTNNWR model significantly outperforms the SEIR model in predicting COVID-19 deaths, especially for longer-term predictions. This highlights the importance of capturing spatiotemporal heterogeneity in disease modeling.
  • Combined Measures are Key: Individual prevention measures have a noticeable impact, but their combined implementation is exponentially more effective. This suggests a holistic approach is crucial for robust epidemic control.
  • Winter is Critical: The effectiveness of prevention measures, particularly combined strategies, is amplified during winter months, indicating a need for intensified efforts during this period.
  • Masks as a Primary Tool: Among individual measures, mask-wearing shows the most significant impact in reducing deaths, underscoring its importance as a fundamental public health intervention.
  • "Herd Immunity" Nuances: The study's findings, particularly the effectiveness of combined measures and the potential for deaths to exceed standard levels even with strict individual measures in certain contexts, question the sole reliance on "herd immunity" as a strategy.

Key Results

  • The GTNNWR model achieved approximately 9% higher predictive accuracy than the SEIR model.
  • Individual preventive measures showed the following reductions in deaths:
  • Mask-wearing: 5.38%
    • Vaccination: 3.59%
    • Social distancing: 2.69%
  • The combined implementation of mask-wearing, vaccination, and social distancing reduced deaths by 27.2%, proving 5 to 10-fold more effective than any single measure.
  • During winter, the combined implementation of these three measures could reduce deaths by approximately 45%, which is 1.5-3 times higher than in other seasons.
  • Without stringent measures (Scenario 1), an estimated 78,000 deaths are predicted for the continental US from July 2023 to February 2025.
  • The most effective scenario (Scenario 5, combined measures) predicted 56,800 deaths for the same period.

Limitations and Future Work

The authors acknowledge that using months as the minimum unit for analysis is a limitation. This could lead to a "loophole" if a sudden outbreak occurs within a month, making the model unable to capture it in time.

Future work suggested includes:

  • Daily factor analysis: Interpolating daily driving factors and matching them with dependent variables to construct a daily prediction model.
  • Further exploration of the complex interactions between various driving factors and their impact on epidemic dynamics.
  • Investigating the effectiveness of these models and strategies in other geographical regions.

Summary updated

BibTeX
@article{sun2024quantitative,
  author = {Jingtao Sun and Jin Qi and Zhen Yan and Yadong Li and Jie Liang and Sensen Wu},
  title = {{Quantitative Study on American COVID-19 Epidemic Predictions and Scenario Simulations}},
  journal = {Isprs Int. J. geo inf},
  year = {2024},
  doi = {10.3390/ijgi13010031},
}
Figure