Network-Based Multi-Layer Model Using Machine Learning for Optimal Vaccine Prioritization in Heterogeneous Populations

Technical Summary: Network-Based Multi-Layer Model Using Machine Learning for Optimal Vaccine Prioritization in Heterogeneous Populations

Problem Statement
Traditional epidemic control models, such as SIR and SIRV, often rely on the assumption of homogeneous mixing, which fails to capture the heterogeneous contact patterns inherent in real-world populations. Individuals vary significantly in their number of contacts, community roles, and transmission potential. While network-based models address this by representing individuals as nodes and interactions as edges, existing strategies for vaccine prioritization often rely on static, first-order heuristics (e.g., degree centrality, betweenness centrality) or single-layer network assumptions. These approaches may overlook higher-order relational patterns and cross-layer dependencies essential for effective intervention in complex, multi-layer social structures. Furthermore, there is a gap in integrating machine learning to dynamically identify structurally critical nodes that classical metrics might miss.

Methodology
The authors propose a multi-layer network SIRV (Susceptible-Infected-Recovered-Vaccinated) model that incorporates heterogeneous transmission rates across distinct social layers (e.g., households, workplaces). The study employs a comparative framework evaluating three classical vaccination strategies against machine learning-based approaches:

1. Classical Heuristics:

   • High-Degree Prioritization: Targets nodes with the highest total number of contacts across all layers.
   • Betweenness Centrality Prioritization: Targets nodes acting as structural bottlenecks on shortest paths between other nodes.
   • Layer-Specific Prioritization: Assigns weights to layers based on community structure and prioritizes nodes with high weighted degrees.

2. Mathematical Analysis:

   • The study derives the basic reproduction number ($R_0$) and the effective reproduction number ($R_t$) for the multi-layer system.
   • A stability analysis of the disease-free equilibrium (DFE) is conducted using spectral radius theory, establishing that $R_0 < 1$ is required for stability.
   • Sensitivity analysis demonstrates how structural perturbations (edge addition/removal) affect the spectral radius, suggesting that removing high-betweenness nodes fragments the network more effectively than removing high-degree nodes.
   • The problem is formulated as a multi-objective optimization task balancing infection minimization, cost, and fairness (using a Gini coefficient), solved via weighted-sum scalarization.
   • An optimal control approach is introduced using time-dependent vaccination rates ($v_i(t)$), solved via a Hamiltonian formulation to identify the marginal benefit of vaccinating specific susceptible nodes over time.

3. Machine Learning Framework:

   • Graph Neural Networks (GNN): A static prioritization strategy where a GNN learns structural risk patterns directly from the multi-layer topology. Node features include structural metrics, layer membership, and epidemic states. The GNN is trained to minimize a surrogate loss function combining epidemic severity, cost, and fairness.
   • Reinforcement Learning (RL): A dynamic policy approach modeled as a Markov Decision Process (MDP). An agent uses a GNN-based policy (optimized via Proximal Policy Optimization) to select vaccination actions based on the current epidemic state and network structure.

Experimental Setup
The models were evaluated using the Email-Eu-Core contact network (Yin et al., 2017), a real-world dataset representing email interactions. The study conducted 30 independent stochastic SIRV simulations for each strategy. Performance was measured using three epidemiological metrics:

• Peak infection level.
• Final epidemic size.
• Time to peak infection.

Key Results

• Classical Heuristics: Degree-based, betweenness-based, and layer-based strategies exhibited broadly similar performance across all metrics. This clustering is attributed to the Email-Eu-Core network's dense connectivity and modest community structure, which limits the discriminative power of traditional centrality measures.
• GNN-Based Strategy: The GNN-based approach substantially outperformed all classical heuristics. It achieved:
  • The lowest peak infection (381.4 vs. ~480–500 for classical methods).
  • The smallest final epidemic size (574.0 vs. ~713–733).
  • The earliest time to peak (9.2 vs. 10.5–13.7).
  • These results suggest the GNN successfully identified structurally critical nodes, such as those in overlapping communities or acting as multi-layer brokers, which classical metrics overlooked.
• RL-Based Baseline: An untrained RL baseline performed the worst, yielding the highest peak infection and final epidemic size, serving as a lower bound that highlights the necessity of training for dynamic policies.
• Optimal Control: The time-dependent control analysis revealed that intervention effectiveness is highest during the early epidemic phase, targeting highly connected hubs. As the epidemic progresses, the optimal targets shift toward peripheral nodes, confirming that optimal targeting is state-dependent rather than static.

Significance and Claims
The paper claims that machine learning-based vaccination policies can significantly outperform traditional heuristics by exploiting higher-order relational patterns in real-world networks. Specifically, the study demonstrates that:

1. Limitations of Classical Metrics: In dense, real-world networks, traditional centrality measures often yield similar, suboptimal outcomes because they fail to capture complex, multi-hop dependencies.
2. Superiority of Learning-Based Approaches: Graph Neural Networks can learn structural risk patterns directly from topology, identifying non-intuitive vulnerabilities that lead to superior epidemic suppression.
3. Framework for Targeted Intervention: The proposed framework offers a mathematically rigorous and data-driven method for dynamic vaccine prioritization, bridging the gap between classical network science and modern machine learning.

The authors conclude that while classical strategies provide a baseline, learning-based policies offer a powerful framework for targeted epidemic intervention, particularly in heterogeneous and multi-layered populations. The study does not claim that RL is currently superior to GNNs in this specific implementation (as the RL agent was untrained) but posits that future work with fully trained RL agents could further enhance dynamic, temporally adaptive strategies.