[Paper Review] Dynamic communicability and epidemic spread: a case study on an empirical dynamic contact network

Isabel Chen, Michele Benzi+2 more|arXiv (Cornell University)|Jan 27, 2016

Complex Network Analysis Techniques34 references3 citations

TL;DR

This study evaluates dynamic communicability and other temporal centrality measures on an empirical emergency department contact network to assess their ability to predict epidemic spread. While temporal centrality identifies distinct top-spreaders compared to static network measures, it does not consistently outperform simpler metrics like duration observed and time of first appearance in predicting epidemic outcomes.

ABSTRACT

We analyze a recently proposed temporal centrality measure applied to an empirical network based on person-to-person contacts in an emergency department of a busy urban hospital. We show that temporal centrality identifies a distinct set of top-spreaders than centrality based on the time-aggregated binarized contact matrix, so that taken together, the accuracy of capturing top-spreaders improves significantly. However, with respect to predicting epidemic outcome, the temporal measure does not necessarily outperform less complex measures. Our results also show that other temporal markers such as duration observed and the time of first appearance in the the network can be used in a simple predictive model to generate predictions that capture the trend of the observed data remarkably well.

Motivation & Objective

• To assess whether temporal centrality measures improve prediction of epidemic outcomes compared to static network centrality.
• To investigate the role of dynamic communicability and other time-resolved network metrics in identifying key spreaders in real-world contact networks.
• To compare the predictive power of temporal centrality with simpler temporal markers such as duration observed and time of first appearance.
• To evaluate whether incorporating temporal network structure enhances the accuracy of epidemic outcome prediction in a linear regression framework.
• To explore the impact of aggregation time scales on centrality measure differentiation and predictive performance.

Proposed method

• Applied dynamic communicability, a temporal generalization of Katz centrality, to a time-resolved empirical contact network from an urban emergency department.
• Simulated epidemic spread using contact duration as a key parameter in the infection process.
• Used linear regression models to assess the relationship between centrality scores and epidemic outcomes, adjusting for confounding variables like duration observed and time of first appearance.
• Computed prediction accuracy using RMSE and BIAS metrics across all individuals, comparing models with and without centrality predictors.
• Explored aggregation of contact data into 20-minute time frames to assess sensitivity of centrality measures to time-scale choice.
• Evaluated normalization techniques and alternative centrality measures to address skewness and improve predictive performance.

Experimental results

Research questions

• RQ1 Does temporal centrality, such as dynamic communicability, improve the prediction of epidemic outcomes compared to static network centrality?
• RQ2 How do temporal markers like duration observed and time of first appearance compare to complex temporal centrality measures in predicting epidemic trends?
• RQ3 To what extent does the aggregation time scale affect the differentiation of top nodes by centrality measures?
• RQ4 Can linear regression models incorporating temporal centrality better predict epidemic outcomes than models without them?
• RQ5 Do more nuanced temporal centrality measures (e.g., BC) provide substantial improvement over simpler measures (e.g., AD) in predictive performance?

Key findings

• Temporal centrality measures like dynamic communicability identify a distinct set of top-spreaders compared to time-aggregated network centrality, improving identification accuracy when used together.
• Despite better top-spreader identification, temporal centrality does not consistently outperform simpler measures like duration observed and time of first appearance in predicting overall epidemic outcomes.
• The model including centrality measures reduced prediction error, with the aggregated duration (AD) measure outperforming both BC and RC in predictive performance.
• The time of first appearance and duration observed were essential in reproducing realistic epidemic trends, suggesting their strong predictive power.
• Aggregation into 20-minute time frames improved differentiation among top RC nodes, likely due to reduced normalization effects despite loss of fine-grained temporal detail.
• The linear regression framework showed marginal improvement over the null model, indicating that while centrality helps, the added computational cost may not be justified without more sophisticated modeling.