Toward optimal disease surveillance with graph-based active learning

Tracking the spread of emerging pathogens is critical to the design of timely and effective public health responses. Policymakers face the challenge of allocating finite resources for testing and surveillance across locations, with the goal of maximizing the information obtained about the underlying trends in prevalence and incidence. We model this decision-making process as an iterative node classification problem on an undirected and unweighted graph, in which nodes represent locations and edges represent movement of infectious agents among them. To begin, a single node is randomly selected for testing and determined to be either infected or uninfected. Test feedback is then used to update estimates of the probability of unobserved nodes being infected and to inform the selection of nodes for testing at the next iterations, until certain test budget is exhausted. Using this framework, we evaluate and compare the performance of previously developed active learning policies for node selection, including Node Entropy and Bayesian Active Learning by Disagreement. We explore the performance of these policies under different outbreak scenarios using simulated outbreaks on both synthetic and empirical networks. Further, we propose a policy that considers the distance-weighted average entropy of infection predictions among neighbors of each candidate node. Our proposed policy outperforms existing ones in most outbreak scenarios given small test budgets, highlighting the need to consider an exploration-exploitation trade-off in policy design. Our findings could inform the design of cost-effective surveillance strategies for emerging and endemic pathogens and reduce uncertainties associated with early risk assessments in resource-constrained situations.