Abstract
Identifying the source of an infectious disease outbreak is a fundamental challenge in epidemiology, often likened to finding a "needle in a haystack." Traditional models frequently rely on assumptions of population homogeneity, failing to capture the complex, multi-layered reality of human interactions. This paper introduces the Heterogeneous Hypergraph Attention Network (HHAN), a framework designed to trace infection origins in complex environments. By combining Agent-Based Modeling (ABM) with a specialized graph neural network that incorporates diverse population types and community structures, HHAN effectively identifies early-stage "patient zero" candidates. Our evaluation across three real-world contact datasets demonstrates that HHAN significantly outperforms existing state-of-the-art tracing methods.
Introduction: The Challenge of Heterogeneity
Tracing the source of an infection is complicated by the messy reality of human interaction. Most traditional epidemic models assume homogeneity—the idea that everyone interacts in the same way and shares the same risk of infection. However, real-world transmission is inherently heterogeneous:
- Diverse Populations: Different groups, such as healthcare workers versus the general public, exhibit distinct contact patterns and risks.
- Multiple Pathways: Diseases spread through direct person-to-person contact (edges) and indirect shared environments, such as crowded offices or contaminated surfaces (hyperedges).
- Community Structures: Families, schools, and workplaces create clusters where infection risks are nonlinear and highly localized.
Existing source-tracing algorithms often struggle with these complexities, either oversimplifying the network structure or failing to account for how different types of nodes (people vs. locations) interact over time.
Methods: The HHAN Framework
The proposed HHAN system consists of two primary modules designed to capture the temporal and structural nuances of disease spread.
Agent-Based Modeling (ABM) Module
The ABM module simulates individual behaviors and state transitions using the Susceptible-Infected-Recovered (SIR) framework. It treats the tracing process as a discrete-step simulation where agents (nodes) update their status based on direct contact or group-based exposure. This module provides essential temporal features by capturing the state of the network at various time steps.
Heterogeneous Hypergraph Attention Network (HHAN) Module
The core of the framework is the HHAN module, which processes the complex interaction data:
- Hypergraph Transformation: To handle group interactions (e.g., multiple people in a single room), the model transforms group-based "hyperedges" into a distinct node type (Type C). This allows the network to explicitly learn the specific infectious properties of different environments.
- Attention Mechanism: Utilizing Graph Attention Networks (GAT), the model assigns varying weights to different connections. This recognizes that a contact with a family member carries more transmission "weight" than a fleeting interaction in a public space.
- Temporal Features: The model incorporates "lagged features"—data from previous time steps—to capture the time-series nature of viral movement through a population.
Experimental Setup and Datasets
The researchers evaluated HHAN on three high-resolution real-world contact datasets:
- ACM Hypertext Conference: Face-to-face contact data among conference attendees over 2.5 days.
- High School: Contact and friendship networks among students in Marseille, France.
- Hospital: A high-resolution contact network between patients and healthcare workers in a Lyon hospital ward.
SIR simulations were used to generate "ground truth" infection paths, starting with 5% of nodes as sources and concluding when 30% of the population was infected.
Key Results and Discussion
HHAN consistently outperformed state-of-the-art baselines, including DDMIX, LPSI, and SLVAE, across all datasets.
- Accuracy: On the Conference dataset, HHAN achieved an impressive 98.1% accuracy, significantly higher than the ~90% achieved by the next best competitor, CULT.
- F1 Score: In the Hospital dataset, HHAN reached an F1 score of 0.694, more than doubling the performance of baselines that struggled to exceed 0.35.
- Robustness: Ablation studies confirmed that both the ABM component and the specific feature engineering are critical; removing either led to a significant degradation in performance.
The model's success lies in its ability to balance precision and recall. While many baseline models tend to flag too many nodes as potential sources (high recall, low precision), HHAN maintains a strong balance, making it a more practical tool for public health interventions where testing resources are limited.
Conclusion and Future Directions
HHAN represents a significant advancement in infectious disease source tracing by explicitly modeling the heterogeneity of human populations and environments. By transforming hyperedges into nodes and utilizing attention mechanisms, it captures the nonlinear dynamics of community spread.
Future work will focus on adapting the model for highly fluid, dynamic real-world networks and exploring advanced automated feature selection techniques. Testing the model against actual clinical outbreak data remains the ultimate goal for real-world deployment.
Summary updated