Abstract
Epidemics of certain viruses in a population can have major impact effects, as is the case in the recent global pandemic caused by the COVID-19 virus. Identifying infected individuals during the course of an epidemic is extremely important for measuring spread and designing more effective control measures. However, in some epidemics infected individuals do not exhibit clear symptoms despite being infected and contributing to the contagion of others (called asymptomatic). This work addresses the problem of identifying asymptomatic individuals in network epidemics based on the observation of infected (symptomatic) individuals. The main contribution of this work is the evaluation of different centrality measures to identify asymptomatic individuals when a fraction of the infected nodes in a network epidemic is observed at a given moment in time. In particular, a variation of the betweenness centrality measure is proposed in this work. An evaluation using different network models and different asymptomatic rates shows that the proposed centrality measure outperforms other centrality measures in many scenarios. Furthermore, the performance of centrality measures increases as the fraction of asymptomatic decreases, showing an interesting trade-off.
Introduction
Asymptomatic carriers represent a significant challenge in public health surveillance. These individuals are infected and contagious but do not exhibit symptoms, making them "invisible" to standard observation. During the COVID-19 pandemic, these carriers played a crucial role in silent transmission. While mass testing can identify these cases, the logistical and financial costs are often prohibitive.
This paper proposes a network-theoretic approach to identify these hidden spreaders. By analyzing the structure of social contact networks and the locations of known symptomatic individuals, we can rank "healthy-looking" nodes by their likelihood of being infected.
Methodology
The authors model the epidemic using the Susceptible-Infected (SI) model on various network topologies. To identify asymptomatic nodes, they compare three centrality measures:
- Degree Centrality: Ranks nodes by their total number of connections, assuming highly connected "hubs" are more likely to be infected.
- Contact Method: Ranks nodes based on the proportion of their immediate neighbors who are symptomatic.
- Modified Betweenness Centrality: This is the core proposal. Traditional betweenness centrality measures how often a node lies on the shortest path between any two nodes. The proposed variation restricts this to shortest paths between observed symptomatic nodes.
The intuition is that infection typically spreads via shortest paths. If a node bridges two known symptomatic individuals, it is a prime candidate for being an asymptomatic link in the infection chain.
Experimental Setup
The study utilizes synthetic networks of 3,000 nodes with an average degree of 8 to test different social structures:
- Watts-Strogatz (WS): Captures "small-world" properties with high local clustering.
- Barabási-Albert (BA): Represents "scale-free" networks dominated by a few highly connected hubs.
- Erdős-Rényi (ER): Represents purely random connectivity.
Simulations were run until infection reached 20% and 60% of the population, with observation rates varying from 5% to 90% to simulate different levels of surveillance visibility.
Results and Insights
The performance was measured using the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC).
- Superiority of Betweenness: The modified betweenness measure consistently outperformed Degree and Contact methods across most scenarios. In Watts-Strogatz networks with 20% infection and 50% observation, it achieved an AUC of 0.83.
- Robustness to Sparse Data: A key finding is that betweenness remains effective even when only 10% of infected nodes are observed, which is critical for early-stage outbreak management.
- Network Topology Impact: In scale-free (BA) networks, while hubs are naturally central, the betweenness method still provided better discrimination by accounting for the actual state of neighbors rather than just potential connectivity.
- Saturation Effect: Performance tends to decrease as the epidemic spreads (e.g., at 60% infection). In a highly infected network, the "noise" increases as many nodes lie on shortest paths between symptomatic individuals, making it harder to distinguish susceptible nodes from asymptomatic ones.
Conclusion
This work demonstrates that network topology can serve as a powerful proxy for identifying asymptomatic carriers. The proposed modified betweenness centrality offers a strategic alternative to mass testing by prioritizing individuals who are structurally likely to be infected. Future work could extend this to SIR models, where recovered nodes provide additional historical data, or explore the dynamics of the epidemic over time rather than relying on a static snapshot.
Summary updated