TL;DR
Existing epidemic forecasting models rely almost exclusively on geographical adjacency, overlooking non-spatial transmission pathways. The authors propose BLUE (Bi-Layer heterogeneous graph fUsion pipEline), a framework that integrates case-level viral genomics, county-level bird abundance, and geographic connectivity to predict Avian Influenza Virus (AIV) outbreaks. Supported by a new multimodal dataset (Avian-US), BLUE uses spectral-regularized graph fusion to compress multi-layer graphs without discarding crucial transmission geometry, achieving state-of-the-art forecasting performance.
What Problem Does This Paper Address?
Standard spatio-temporal Graph Neural Networks (GNNs) for diseases like COVID-19 or human influenza assume that transmission is governed primarily by inter-region mobility and geographic proximity. Nodes represent geographic locations connected by static or learned spatial edges.
Avian influenza spreads very differently. Wild migratory birds carry viral lineages across vast continental distances, seeding outbreaks in non-neighboring regions while viruses undergo rapid reassortment and mutation. Conventional homogeneous models fail here because they treat all reported cases identically and assume spatial proximity is the sole transmission vector. They cannot capture case-level genetic linkages that explain why an outbreak in one county may be genealogically linked to a strain hundreds of miles away.
Modeling this dynamic requires multi-layer heterogeneous graphs (cases and locations) with time-varying topologies, extreme data sparsity (over 98% of county-weeks report zero cases), and the challenge of compressing these complex graphs into unified representations without losing essential epidemiological structure.
Methods
BLUE addresses these challenges through a modular pipeline:
- Bi-Layer Heterogeneous Graph Construction: Constructs two node tiers: location nodes (counties attributed with reported cases and eBird abundance data) and case nodes (individual infected samples attributed with genetic features derived via the Kimura 2-parameter [K80] evolutionary distance model on hemagglutinin [HA] segments). Edges include spatial distance kernels, genetic similarity, and case-to-location assignments.
- Cross-Layer Smoothing Block: Uses Markov Random Field (MRF)-inspired mean-field message passing across distinct relation types to harmonize representations between cases and locations before fusion.
- Information-Preserving Fusion Graphs: Fuses case and location features into unified county-level nodes. To scale edge formation without an $O(N^2)$ brute-force search, it employs Locality-Sensitive Hashing (LSH) to restrict candidate pairs to $O(N + M_{\max})$, followed by an attention-based gate that adaptively balances spatial and genetic edge weights.
- Spectral Regularization ($\mathcal{L}_{spec}$): Compressing bi-layer graphs risks acting as a low-pass filter that drops high-frequency case signals. BLUE adds a regularizer penalizing the difference between the graph Laplacians ($\|L_{hetero} - \tilde{L}_f\|_F^2$). The authors theoretically prove (Theorem 3.1) that GraphSAGE bounds the cumulative spectral distortion to $O(L\epsilon)$.
- Autoregressive Sequence-to-Sequence Forecasting: Encodes historical fused graphs with GraphSAGE and multi-head temporal attention, predicting future case counts through an autoregressive decoder trained with a stratified, severity-weighted loss to prioritize rare, high-consequence outbreak spikes.
Datasets
- Avian-US (Newly introduced): A multi-modal benchmark spanning 3,227 contiguous U.S. counties over 104 weeks (2021–2024). Integrates >12,000 H5-positive wild bird cases, ~8,000 matched HA viral genomic sequences from NCBI, and weekly species-level bird abundance from eBird Status and Trends. It reflects real-world surveillance challenges, notably extreme sparsity (98.4% zeros).
- Flu-Japan (Benchmark): A standard homogeneous dataset covering 47 prefectures over 348 weeks (2012–2019) with 15.8% zeros, used to validate scalability and baseline temporal forecasting under continuous outbreak conditions.
Key Insights and Ideas
- Genomics reveals invisible spatial pathways: Ablation tests show removing genetic edges (
w/o gen) degrades MAE from 0.0848 to 0.1667, proving that spatial distance alone cannot explain migratory seedings. - Spectral alignment prevents information loss: Enforcing spectral consistency between the raw heterogeneous graph and the fused homogeneous graph ensures that global diffusion modes are retained despite aggressive node compression.
- LSH as implicit regularization: Replacing dense all-to-all attention with LSH-based candidate selection not only reduced runtime from quadratic to linear, but also improved outbreak detection F1 score (0.1001 vs. 0.0901) by pruning spurious long-range noise in sparse data.
- GraphSAGE over GCN/GAT: The authors demonstrate that GraphSAGE's Lipschitz-bounded aggregation mathematically aligns with their spectral objective, whereas GCN accumulates depth-wise error linearly and GAT dynamically alters adjacency outside the Laplacian constraint.
Key Results
- Avian-US Outbreak Forecasting ($T=4, H=4$): BLUE outperformed all baselines (including EpiGNN, Cola-GNN, and HGT), obtaining the lowest RMSE (0.611) and MAE (0.085), and the highest PCC (0.087), SCC (0.122), and F1 score (0.100). It reduced RMSE by 0.041 over the runner-up heterogeneous baseline (HGT).
- Flu-Japan: Even on homogeneous data without genomic sequences, BLUE achieved state-of-the-art results (RMSE: 1511, MAE: 553, F1: 0.7234, PCC: 0.7954).
Limitations and Future Work
- Missing Environmental Drivers: Meteorological variables (temperature, humidity, precipitation) known to influence viral persistence in water bodies are not yet integrated.
- Single-Host Focus: The dataset and evaluation are restricted to wild birds, excluding transmission interfaces with domestic poultry, dairy cattle, or mammalian/human spillover risks.
- Long-Horizon Degradation: Performance declines significantly at longer horizons ($H=8$), primarily due to autoregressive error accumulation on sparse count series.
Summary updated