Data-Driven Models Informed by Spatiotemporal Mobility Patterns for Understanding Infectious Disease Dynamics

Die Zhang, Yong Ge, Xilin Wu, Haiyan Liu, Wen-Bin Zhang, Shengjie LaiISPRS Int. J. Geo Inf 2023 · DOI

TL;DR

This paper introduces a novel approach to predict infectious disease outbreaks by integrating human mobility data with traditional epidemiological models. By developing "mobility-informed risk indices," the researchers significantly improved the accuracy of identifying high-risk areas and forecasting disease spread, demonstrating a more robust and physically consistent method for disease surveillance and containment.

What Problem Does This Paper Address?

Predicting the spread of infectious diseases is critical for effective public health interventions. Traditional epidemiological models, like SEIR, often rely on simplifying assumptions about disease spread that don't fully capture the complex, heterogeneous nature of real-world transmission. While data-driven approaches using machine learning have shown promise, they often struggle to incorporate the nuanced dynamics of human movement, which is a primary driver of disease transmission. Existing methods may not fully account for how people move between and within regions, limiting their predictive accuracy and consistency, especially in the early stages of an outbreak or for widespread, long-standing diseases.

Methods

The core innovation of this paper lies in the development of two "mobility-informed risk indices": Case Flow Intensity (CFI) and Case Transmission Intensity (CTI).

  • CFI: This index quantifies regional infection risk by counting the cumulative number of initial cases that have visited a region. It leverages data on time-varying population flow and the locations of initial confirmed cases.
  • CTI: This index builds on CFI to represent the risk introduced by both inter-regional movements and intra-regional contact with initial cases. It estimates the number of potential new infections resulting from the activity of initial cases.

These indices are then used as auxiliary predictive variables in statistical regression (logistic regression) and machine learning (random forest) models. The study also employed a travel network-based SEIR model to simulate outbreak scenarios for generating training data, particularly for the early stages of an outbreak. The models were evaluated using metrics like sensitivity, specificity, Mean Absolute Error (MAE), and R-squared.

Datasets

The study utilized several real-world datasets for training and evaluation:

  • COVID-19 outbreak data:
    • Beijing and Guangzhou, China: Data on actual COVID-19 outbreaks, including affected subdistricts and case numbers.
  • Contiguous United States: County-level daily COVID-19 cases.
  • Mobility data:
    • China Mobile: Anonymized population movement flows (hourly data) between subdistricts in Beijing and Guangzhou.
  • SafeGraph: Daily origin-destination (OD) mobility data at the county level for the contiguous United States.
  • Point of Interest (POI) data:
    • AMap Services (China): POI data for Beijing and Guangzhou.
  • SafeGraph (USA): Spatial distribution of POIs across the contiguous United States.
  • Population data:
    • WorldPop: Population data at a 100-meter resolution for Beijing and Guangzhou.
  • U.S. Census (2018 estimates): Used as baseline demographic variables.
  • Epidemiological data:
    • USA Facts: County-level daily COVID-19 cases for the United States.

The paper also mentions using simulated data from an SEIR model for training and validation in the initial outbreak stage scenario.

Key Insights and Ideas

  • Human mobility is a critical, often underutilized, predictor: The study strongly emphasizes that understanding how people move is not just supplementary but fundamental to accurately modeling disease spread.
  • "Physics-informed" data improves models: By incorporating the physical reality of human movement (CFI and CTI), the models become more robust and consistent, moving beyond purely statistical correlations.
  • Early outbreak detection is enhanced: The mobility-informed indices proved particularly effective in identifying high-risk subdistricts in the early stages of an outbreak, enabling targeted interventions.
  • Forecasting accuracy improves across horizons: While the impact was more pronounced for shorter forecast horizons (1-2 weeks), the inclusion of mobility data consistently improved prediction accuracy for up to four weeks ahead.

Key Results

  • Case detection in China: The proposed models successfully identified 87% of affected subdistricts in Beijing and Guangzhou during real-world COVID-19 outbreaks.
  • Forecasting in the US:
    • The proposed method led to a decrease in Mean Absolute Error (MAE) for most forecast dates compared to reference studies.
  • The R-squared (R²) values were improved, with the proposed method achieving R² higher than 0.5 for one- to four-weeks-ahead forecasts on 39 dates in one scenario and higher than 0.8 on 11 dates in another.
  • Variable importance: The CFI risk index was identified as the most dominant variable in estimating potential subdistricts with cases, highlighting its significance.

Limitations and Future Work

  • Limited accounting for infection events during movement: The current CFI and CTI indices capture risk associated with visiting regions but don't fully account for the risk of infection events occurring while an infected person moves between two regions. Future work could optimize these indices to address this.
  • Reliance on SEIR simulations for early detection: The method for identifying high-risk areas at the early stage of an outbreak currently requires training data generated by SEIR model simulations. The authors suggest that using available historical real-world epidemiological data could further improve prediction accuracy.
  • Potential for multicollinearity: Despite using variable selection methods, the authors acknowledge that multicollinearity among extensive variables might have influenced permutation importance rankings. They suggest variable filtering could be explored for future case forecasts.
  • Generalizability: While validated on COVID-19 in China and the US, the authors suggest obtaining more data on various infectious diseases and from other countries to comprehensively understand disease dynamics in diverse contexts.

Summary updated

BibTeX
@article{zhang2023data,
  author = {Die Zhang and Yong Ge and Xilin Wu and Haiyan Liu and Wen-Bin Zhang and Shengjie Lai},
  title = {{Data-Driven Models Informed by Spatiotemporal Mobility Patterns for Understanding Infectious Disease Dynamics}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2023},
  doi = {10.3390/ijgi12070266},
}
Figure