Spread of Covid-19 in the United States is controlled

Zixin Hu, Qiyang Ge, Shudi Li, Tao Xu, Eric Boerwinkle, Jin Li, Momiao XiongmedRxiv 2020 · DOI

Abstract

As of May 1, 2020, the number of confirmed COVID-19 cases in the United States surpassed 1,062,446. While interventions to slow the spread curtailed most social activities, they simultaneously triggered an economic crisis and rising resistance to strict measures. Traditional epidemiological models have suggested intermittent social distancing could extend the outbreak into 2022, raising urgent questions about the timing of relaxing quarantine measures. We developed a novel method integrating artificial intelligence and causal inference for the real-time prediction and control of nonlinear epidemic systems. Using an Intervention Auto-Encoder (IAE), we estimated that the US peak occurred on April 24, 2020, and projected the outbreak would conclude by the end of July 2020 with approximately 1,551,901 total cases. This framework provides a tool for balancing the risks to workers with the necessity of reopening the economy.

Introduction

The early stages of the COVID-19 pandemic presented policymakers with a grueling dilemma: maintain strict lockdowns to preserve public health or relax them to mitigate economic collapse. Standard mathematical models, such as the Susceptible-Exposed-Infected-Recovered (SEIR) model, often struggled to provide the precise, real-time guidance required for this decision-making.

We identified three primary limitations in traditional SEIR models:

  1. Homogeneity: They often assume populations mix uniformly, which contradicts real-world social dynamics.
  2. Unidentifiable Parameters: They rely on ordinary differential equations with static parameters that are difficult to estimate accurately from noisy, real-time data.
  3. Static Interventions: They frequently treat control measures as constants rather than dynamic factors that evolve in response to the state of the outbreak.

Methodology

To overcome these challenges, we developed the Intervention Auto-Encoder (IAE), a sequence-to-sequence deep learning architecture designed to reconstruct and forecast nonlinear, time-varying epidemic systems.

The Intervention Variable and Causal Inference

Central to our model is the quantification of public health measures (such as travel restrictions and school closures) into a single intervention variable ($A_t$), ranging from 0 (no restrictions) to 1 (maximum intervention). We utilized the Rubin Causal Model (Potential Outcomes Framework) to allow the IAE to estimate counterfactuals. This enables the model to simulate "what-if" scenarios, predicting the trajectory of infections if different intervention strategies had been chosen.

Architecture

The IAE consists of two stacked autoencoders. The encoder processes the historical data of cases and interventions to learn the underlying latent dynamics of the system. The decoder then uses these learned features to forecast future trajectories under specific intervention scenarios. The network was optimized with dimensions of 8 for input, 32 for hidden layers, and 4 for the latent space.

Data and Analysis

Our study utilized surveillance data of confirmed COVID-19 cases in the United States from January 22, 2020, to April 24, 2020, sourced from the Johns Hopkins Coronavirus Resource Center. The scope included the total US, all 50 states, and five additional regions. Data was segmented into 8-day windows and normalized for neural network training.

By applying k-means clustering to the predicted case curves, we grouped states into nine distinct categories. This analysis revealed that the pandemic's peak and end times varied significantly by geography—for instance, Montana was projected to see the end of its outbreak in May, while Massachusetts was expected to continue into August.

Key Insights

  • The Intervention Curve: The model identified that the degree of intervention was not linear. A notable dip occurred during late February (coinciding with Spring Break), which led to a subsequent surge before stricter policies were implemented.
  • System-Dependent Control: The IAE treats the intervention variable as system-dependent, meaning the model can automatically adjust the predicted control measures based on the current state of the epidemic.
  • Testing as an Offset: A significant finding was the quantification of testing as a control measure. Our results suggest that the risks associated with relaxing social distancing can be mathematically offset by increasing the volume of virus testing and contact tracing.

Results and Counterfactual Simulations

The IAE achieved high predictive accuracy, with a 10-step ahead forecasting error of only 0.0012, significantly outperforming traditional epidemiological models.

Forecasting the US Outbreak

The model correctly identified the peak of the US outbreak as April 24, 2020, with approximately 36,188 new cases reported that day. Assuming then-current intervention levels remained steady, the model predicted the US would reach a total of 1,551,901 cases and the outbreak would be largely controlled by the end of July 2020.

Economic Reopening Scenarios

We evaluated the impact of relaxing interventions to facilitate economic reopening:

  • Relaxation: A 40% reduction in intervention measures was predicted to result in an additional 317,284 cases (a 20.4% increase).
  • Enhanced Control: Conversely, increasing the intervention efficacy by 20%—potentially through massive testing—could reduce the total case count by 16.5%.

Limitations and Future Directions

While the IAE provides a powerful tool for pandemic management, it is limited by the confounded nature of early interventions; because multiple policies (e.g., masks and stay-at-home orders) were often introduced simultaneously, the model cannot isolate the specific impact of a single policy type. Furthermore, the model's accuracy is contingent on the quality of reported surveillance data. Future work will focus on incorporating individual-level intervention data to further refine the quantification of specific public health strategies.

Summary updated

Abstract

Abstract As of May 1, 2020, the number of cases of Covid-19 in the US passed 1,062,446, interventions to slow down the spread of Covid-19 curtailed most social activities. Meanwhile, an economic crisis and resistance to the strict intervention measures are rising. Some researchers proposed intermittent social distancing that may drive the outbreak of Covid-19 into 2022. Questions arise about whether we should maintain or relax quarantine measures. We developed novel artificial intelligence and causal inference integrated methods for real-time prediction and control of nonlinear epidemic systems. We estimated that the peak time of the Covid-19 in the US would be April 24, 2020 and its outbreak in the US will be over by the end of July and reach 1,551,901 cases. We evaluated the impact of relaxing the current interventions for reopening economy on the spread of Covid-19. We provide tools for balancing the risks of workers and reopening economy.
BibTeX
@misc{hu2020spread,
  author = {Zixin Hu and Qiyang Ge and Shudi Li and Tao Xu and Eric Boerwinkle and Jin Li and Momiao Xiong},
  title = {{Spread of Covid-19 in the United States is controlled}},
  year = {2020},
  doi = {10.1101/2020.05.04.20091272},
  howpublished = {medRxiv},
}
Figure