BLUE: Bi-layer Heterogeneous Graph Fusion Network for Avian Influenza Forecasting

Jing Du, Haley Stone, Yang Yang, Ashna Desai, Hao Xue, Andreas Züfle, C. Raina MacIntyre, Flora D. Salim

Abstract

Accurate forecasting of avian influenza outbreaks within wild bird populations requires models that account for complex, multi-scale transmission patterns driven by various factors. Spatio-temporal GNN-based models have recently gained traction for infection forecasting due to their ability to capture relations and flow between spatial regions, but most existing frameworks rely solely on spatial connections. This overlooks valuable genetic information at the case level, such as cases in one region being genetically descended from strains in another, which is essential for understanding how infectious diseases spread through epidemiological linkages beyond geography. We address this gap with BLUE, a Bi-Layer heterogeneous graph fUsion nEtwork designed to integrate genetic, spatial, and ecological data for accurate outbreak forecasting. The framework builds heterogeneous graphs from multiple information sources, smooths across relation types, performs fusion while retaining structural patterns, and predicts future outbreaks via an autoregressive graph sequence model. To facilitate further research, we introduce the Avian-US dataset, incorporating genetic, spatial, and ecological data across the United States. BLUE achieves superior performance over existing baselines, highlighting the value of incorporating multi-layer information into infectious disease forecasting.

Introduction

Predicting the spread of Avian Influenza Virus (AIV) in wild birds is a significant challenge for epidemiologists. Unlike human-centric diseases which largely spread through local contact and mobility networks, AIV is multi-source and multi-level. Traditional models often rely on spatial homophily—the assumption that geographic proximity is the primary driver of transmission. However, AIV can jump thousands of miles via migratory birds or evolve through genetic reassortment. Consequently, two outbreaks on opposite sides of a continent might be genetically linked, while neighboring cases might be unrelated. Existing models fail to capture these "invisible linkages" found in the virus's genetic code, treating cases as simple count data.

Methodology

The proposed BLUE framework moves beyond simple spatial grids to a sophisticated graph-based approach that integrates multi-modal data streams.

Overview of the BLUE framework architecture, showing the bi-layer graph construction, cross-layer smoothing, and autoregressive forecasting components.

Bi-layer Heterogeneous Graph

BLUE utilizes two distinct layers: a Location Layer representing static geographic counties and a Case Layer representing individual infected birds. Genetic similarities in the Case Layer are calculated using the K80 evolutionary distance model, allowing the network to model transmission pathways that bypass geographic constraints.

Cross-layer Smoothing and Fusion

To bridge these layers, the model employs a Markov Random Field (MRF)-inspired block for cross-layer smoothing, ensuring coherent information flow between individual cases and their broader locations. To manage computational complexity, BLUE fuses these layers into a simplified graph using a spectral regularizer. This mathematical constraint ensures that the simplified graph preserves the essential "diffusion geometry" and transmission structure of the original complex data.

Autoregressive Forecasting

The final stage uses an encoder-decoder architecture with a GraphSAGE backbone. This processes the fused graphs over time to predict infection counts for future weeks, capturing the temporal dynamics of the epidemic.

The Avian-US Dataset

A key contribution of this work is the Avian-US dataset, a benchmark for AIV forecasting covering 3,227 U.S. counties from 2021 to 2024. It integrates three distinct streams:

  • Epidemiological: Over 12,000 reported H5-positive wild bird cases.
  • Genomic: 8,000 hemagglutinin (HA) segment sequences from GenBank.
  • Ecological: Weekly bird abundance estimates for over 630 species sourced from eBird.

Results and Discussion

BLUE was evaluated against several baselines, including standard spatiotemporal models (STGCN, DCRNN) and advanced heterogeneous graphs (HGT). BLUE achieved the lowest error metrics with an RMSE of 0.6106 and MAE of 0.0848. In outbreak detection, BLUE showed a 17.1% improvement in F1 Score over baselines when the spectral regularizer was fully tuned.

Per-step evaluation metrics heatmap showing the performance of BLUE across 8 prediction steps.

Ablation studies confirmed that removing genomic data ("w/o gen") caused a sharp increase in error, proving that genetic information is a primary driver of forecasting accuracy. However, performance naturally decays as the forecasting horizon increases, a common challenge in autoregressive modeling.

Conclusion

BLUE represents a major step forward in epidemic forecasting by treating genetic similarity as a functional connection. Future work will focus on addressing data sparsity in wild bird surveillance and extending the model to include cross-species transmission (e.g., from wild birds to poultry), which is critical for pandemic preparedness.

Summary updated

Abstract

Accurate forecasting of avian influenza outbreaks within wild bird populations requires models that account for complex, multi-scale transmission patterns driven by various factors. Spatio-temporal GNN-based models have recently gained traction for infection forecasting due to their ability to capture relations and flow between spatial regions, but most existing frameworks rely solely on spatial connections and their connections. This overlooks valuable genetic information at the case level, such as cases in one region being genetically descended from strains in another, which is essential for understanding how infectious diseases spread through epidemiological linkages beyond geography. We address this gap with BLUE, a B}i-Layer heterogeneous graph fUsion nEtwork designed to integrate genetic, spatial, and ecological data for accurate outbreak forecasting. The framework 1) builds heterogeneous graphs from multiple information sources and multiple layers, 2) smooths across relation types, 3) performs fusion while retaining structural patterns, and 4) predicts future outbreaks via an autoregressive graph sequence model that captures transmission dynamics over time. To facilitate further research, we introduce \textbf{Avian-US} dataset, the dataset for avian influenza outbreak forecasting in the United States, incorporating genetic, spatial, and ecological data across locations. BLUE achieves superior performance over existing baselines, highlighting the value of incorporating multi-layer information into infectious disease forecasting.
BibTeX
@misc{du2025blue,
  author = {Jing Du and Haley Stone and Yang Yang and Ashna Desai and Hao Xue and Andreas Züfle and C. Raina MacIntyre and Flora D. Salim},
  title = {{BLUE: Bi-layer Heterogeneous Graph Fusion Network for Avian Influenza Forecasting}},
  year = {2025},
  eprint = {2505.22692},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.22692},
}
Figure