Abstract
When an epidemic spreads into a population, it is often unpractical or impossible to have a continuous monitoring of all subjects involved. As an alternative, algorithmic solutions can be used to infer the state of the whole population from a limited amount of measures. We analyze the capability of deep neural networks to solve this challenging task. Our proposed architecture is based on Graph Convolutional Neural Networks. As such it can reason on the effect of the underlying social network structure, which is recognized as the main component in the spreading of an epidemic. We test the proposed architecture with two scenarios modeled on the COVID-19 pandemic: a generic homogeneous population, and a toy model of the Boston metropolitan area.
Introduction
Most epidemiological modeling focuses on "macro" statistics—predicting the total number of infected or recovered people in a city or country. While useful for general planning, these models do not identify who is sick. To implement effective "micro" interventions, such as precise contact tracing or localized quarantines, health officials need to know the state of every individual.
However, continuous monitoring of an entire population is impossible. This paper introduces a Graph Neural Network (GNN) framework to solve this difficult inverse problem: reconstructing the health status of an entire population from a tiny fraction (as low as 5%) of monitored individuals. The challenge is to "reason backward": if we know Person A is sick, and we know the social graph, can we infer the status of their neighbors, and their neighbors' neighbors?
Methodology
The proposed architecture leverages Graph Convolutional Neural Networks (GCNs) to reason on the effect of the underlying social network structure. The model consists of three primary stages:
- Time and Uncertainty Encoding: Instead of complex Recurrent Neural Networks (RNNs), the authors use a "bag-of-words" approach. For each monitored node, they count the occurrences of health states (Susceptible, Exposed/Infected, Recovered/Deceased) over the past $k$ days. This creates a feature vector capturing temporal trends simply and effectively.
- Graph Convolutional Layers: The core consists of three GCN layers. These layers aggregate information from a node’s neighborhood, effectively "spreading" known information across the social graph to unlabeled nodes.
- Output Layer: A fully connected layer with a Softmax activation classifies each node into one of three categories: Susceptible (S), Exposed/Infected (E+I), or Recovered/Deceased (R+D).
The model is trained using a weighted cross-entropy loss to handle the inherent class imbalance, as most individuals remain healthy during early stages of an outbreak.
Experimental Setup
The study utilizes synthetic datasets generated to mimic COVID-19 dynamics across two scenarios:
- Scenario 1 (Random Network): Erdős-Rényi graphs of 500 nodes, serving as a baseline for fundamental inference capabilities.
- Scenario 2 (Boston Metropolitan Area): A sophisticated toy model of Boston and Cambridge, MA. This multi-layer graph integrated geography (neighborhood-level connectivity) and demographics (age-based contact patterns). Testing was scaled up to $10^5$ nodes to evaluate performance in a city-sized environment.
Key Results and Discussion
The GNN approach demonstrated significant robustness and accuracy:
- High Accuracy with Sparse Data: Even with only 5% of the population monitored, the model achieved an overall accuracy of 75% on random networks and 72% on the complex Boston model.
- Zero-Shot Scaling: A model trained on a small "village" (500 nodes) could be applied to a "medium city" ($10^5$ nodes) without retraining, maintaining 83% accuracy. This suggests the GCN learns local spreading rules independent of total population size.
- Precision for Healthy Individuals: In the Boston scenario, precision for the "Susceptible" class exceeded 97%. This high precision is critical for public health to avoid mistakenly labeling healthy individuals as infected, thereby maintaining public trust.
Conclusion
While the model assumes a known social network structure, it provides a powerful proof-of-concept for targeted isolation policies. Future work involves integrating dynamic mobility data to infer the graph itself and closing the loop by using these estimates to trigger real-time interventions.
Summary updated