Toward optimal disease surveillance with graph-based active learning.

Area of Science:

• Epidemiology
• Network Science
• Public Health

Background:

• Effective public health responses rely on tracking emerging pathogens.
• Resource allocation for testing and surveillance is a key challenge for policymakers.
• Understanding disease spread requires modeling pathogen movement between locations.

Purpose of the Study:

• To model pathogen spread as a node classification problem on a graph.
• To evaluate and compare active learning policies for optimal node selection in surveillance.
• To propose a novel policy considering neighbor infection probabilities for improved testing strategies.

Main Methods:

• Modeled disease spread using an iterative node classification approach on a graph.
• Compared existing active learning policies (Node Entropy, Bayesian Active Learning by Disagreement) with a proposed policy.
• Simulated outbreaks on synthetic and empirical networks to assess policy performance under various scenarios.

Main Results:

• The proposed policy, incorporating distance-weighted average entropy of neighbor predictions, outperformed existing methods in most scenarios with small test budgets.
• Demonstrated the importance of balancing exploration and exploitation in active learning policy design for surveillance.
• Identified the proposed policy's effectiveness in resource-constrained situations.

Conclusions:

• The developed policy offers a cost-effective approach for surveillance of emerging and endemic pathogens.
• Findings can reduce uncertainties in early risk assessments for public health planning.
• Highlights the need for advanced strategies in optimizing limited surveillance resources.