Abstract
Timely interventions and early preparedness of healthcare resources are crucial measures to tackle the COVID-19 disease. To aid these efforts, we developed the Mobility-Augmented SEIR model (MA-SEIR) that leverages Google’s aggregate and anonymized mobility data to augment classic compartmental models. We show in a retrospective analysis how this method can be applied at an early stage in the COVID-19 epidemic to forecast its subsequent spread and onset in different geographic regions, with minimal parameterization of the model. This provides insight into the role of near real-time aggregate mobility data in disease spread modeling by quantifying substantial changes in how populations move both locally and globally. These changes would be otherwise very hard to capture using less timely data.
Introduction
In the early stages of a pandemic, public health officials often operate under a "fog of war." Traditional epidemiological models, such as the classic Susceptible-Exposed-Infectious-Recovered (SEIR) framework, are effective at modeling disease dynamics within a closed population. However, they frequently struggle to predict the timing of disease transmission between different geographic regions.
The primary challenge is the lack of timely, high-resolution human mobility data. Standard models often rely on static census data or outdated flight logs, which fail to capture the rapid behavioral shifts that occur during a global health crisis. This research addresses the challenge of forecasting "onset"—the moment cases cross a specific threshold—by integrating aggregate, privacy-preserving mobility flows into a global metapopulation model.
Methods
The researchers introduced the Mobility-Augmented SEIR (MA-SEIR) model, which combines a metapopulation framework with real-time data. Instead of treating the world as a single population, the model treats each country or state as a "patch" with its own internal SEIR dynamics.
The core innovation is the integration of a time-varying mobility matrix $M(t)$, derived from Google’s aggregate and anonymized Location History data. This matrix defines the "flow" of individuals between patches. To align with the weekly cadence of mobility data, the authors formulated MA-SEIR as a discrete-time Markov process. To simulate a true early-stage scenario, the model used minimal parameterization, relying on fixed epidemiological parameters from early Wuhan data and fitting only a single multiplier factor based on early curves in Italy and South Korea.
Datasets
The study utilized several key datasets:
- Google COVID-19 Aggregated Mobility Research Dataset: Anonymized, differentially private mobility flows tracking region-to-region movement.
- NSSAC (University of Virginia): Global country-level COVID-19 daily case reports.
- The New York Times: US county and state-level confirmed case numbers.
- Proxy Data: International traffic from Hong Kong was used as a proxy for flows originating from Mainland China due to limited data coverage.
Key Results
The global simulation was remarkably seeded with just a single case in China on December 1, 2019. Despite this minimal starting point, the mobility data accurately routed the infection globally.
- Onset Prediction Accuracy: The model predicted onset (100 cases for countries, 30 for US states) with a Mean Absolute Error (MAE) of 19.9 days globally and 12.2 days for US states.
- Correlation: There was a strong positive correlation between predicted and actual onset dates, with a Pearson correlation of 0.51 globally and 0.61 for US states.
- Mobility as a Leading Indicator: The model captured drastic behavioral shifts, such as the 82% drop in international travel by April 2020. Counterfactual simulations showed that while an 85% reduction in global mobility helps, it only delayed the mean onset by approximately 5.16 days, highlighting the difficulty of stopping a highly transmissible virus once international seeding has occurred.
Discussion and Limitations
While MA-SEIR is a powerful tool for predicting when a disease will arrive, it has limitations. The mobility data is restricted to smartphone users who opted into Google's Location History, which may not perfectly represent all demographics. Additionally, the model assumes homogeneous mixing, where exposed individuals move at the same rate as susceptible ones.
Furthermore, while the model is effective at predicting onset timing, forecasting the magnitude (peak height) and duration of an outbreak remains more difficult, as these factors are heavily influenced by local policy interventions like social distancing and masking. Future work may involve integrating this mobility data into complex stochastic agent-based models to better capture these nuances.
Summary updated