Abstract
Understanding the spread of SARS-CoV-2 provides important insights for control policies such as social-distancing interventions and vaccine delivery in the post-pandemic era. In this work, we take the advantage of action tracking reports of confirmed COVID-19 patients, which contain the mobility trajectory of patients. We analyzed reports of patients from April 2020 to January 2021 in China, a country where the residents are well-prepared for the “new normal” world following COVID-19 spread. We developed natural language processing (NLP) tools to transform the unstructured text of action-tracking reports to a structured network of social contacts. An epidemiology model was built on top of the network. Our analysis provides important insights for the development of control policies. Under the “new normal” conditions, we find that restaurants, locations less protected by mask-wearing, have a greater risk than any other location categories, including locations where people are present at higher densities (e.g., flight). We find that discouraging railway transports is crucial to avoid another wave of breakout during the Chunyun season (a period of travel in China with extremely high traffic load around the Chinese New Year). By formalizing the challenge of finding the optimal vaccine delivery among various different population groups as an optimization problem, our analysis helps to maximize the efficiency of vaccine delivery under the general situation of vaccine supply shortage. We are able to reduce the numbers of infections and deaths by 7.4% and 10.5% respectively with vaccine supply for only 1% of the population. Furthermore, with 10% vaccination rate, the numbers of infections and deaths further decrease by 52.6% and 78.1% respectively. Our work will be helpful in the design of effective policies regarding interventions, reopening, contact tracing and vaccine delivery in the “new normal” world following COVID-19 spread.
Introduction
Traditional epidemiological models often rely on coarse, city-level mobility data, which can overlook the specific transmission dynamics of different social venues. In the "new normal" phase of the COVID-19 pandemic—characterized by low community spread but frequent localized outbreaks—understanding these nuances is critical. This study leverages unstructured action tracking reports from confirmed patients in China to reconstruct high-resolution social contact networks and inform targeted intervention strategies.
Methodology
The researchers developed a multi-stage computational pipeline to transform text-based public health reports into a structured mobility network. As shown in Figure 1, the process begins with a BERT-based NLP model that extracts entities (people, locations, times) and predicts their relationships. This patient-location network is then augmented with aggregated mobility data from SmartSteps to estimate the total population "denominator." Finally, a Susceptible-Exposed-Infectious-Removed (SEIR) model is applied, with transmission rates tailored to 11 location categories and various demographic factors such as age, sex, and city tier.
Key Insights and Results
The analysis revealed several critical findings regarding transmission risks in a society with high mask compliance. Notably, restaurants were identified as the highest-risk locations, surpassing even high-density environments like airplanes. This "mask paradox" stems from the necessity of removing masks while eating, whereas mask mandates are strictly enforced during flights. As shown in Figure 6a, the reproduction rate for restaurants is significantly higher than for other categories.
Furthermore, the study highlighted the disproportionate role of railway transport in spreading the virus compared to air travel. As illustrated in Figure 6g, discouraging railway transport is significantly more effective at preventing outbreaks during peak travel seasons like "Chunyun." Transmission rates were also found to be higher in lower-tier cities (Tiers 2-5) compared to Tier 1 metropolises, likely due to differences in public health enforcement strictness.
The model also addressed vaccine allocation under supply shortages. By formalizing delivery as an optimization problem, the researchers found that even a 1% vaccination rate could reduce infections by 7.4% and deaths by 10.5% if prioritized correctly. The most effective strategy involved targeting mobile adults to stop spread versus seniors to prevent mortality, creating a mathematical trade-off based on policy goals.
Conclusion
By converting unstructured tracking reports into actionable data, this work provides a blueprint for more precise public health interventions. The findings emphasize that reopening policies should prioritize mask-off environments like restaurants for stricter control and that transportation-based interventions should focus on high-volume modes like railways. This NLP-to-network approach offers a scalable method for real-time risk assessment and vaccine distribution in future public health crises.
Summary updated