Comparing methods to find patient zero in infection contact networks
Finding Patient Zero in Temporal Contact Networks: A Comparative Study
Social and Information Networks
Summary
When an outbreak happens, it’s helpful to figure out who started it, called patient zero. The authors compare seven different ways researchers try to find patient zero by looking at how people interacted over time. They tested all methods on several different networks and found that two methods using belief-based approaches worked best overall. Some methods did well only when they had limited information, while others were overconfident in their guesses. The study also points out that making these methods work well on bigger networks is still a challenge.
What this means in practice
- •For public health teams: Improve outbreak response by pinpointing infection origin more accurately using the best source detection methods identified.
- •For epidemiological modelers: Incorporate effective source detection techniques into epidemic simulations to better allocate resources during outbreaks.
Authors
Nathan Cincera Brack, Martin Sterchi, Lorenz Hilfiker, Abraham Bernstein
Abstract
Identifying the source of an epidemic, i.e., the patient zero, in a temporal contact network is important for effective outbreak containment and resource allocation. While several methods have been proposed for this task, they have rarely been evaluated under comparable conditions, and no systematic comparison exists for temporal contact networks. We address this gap with a comparative study of seven major methods, namely Steiner Tree (ST), Individual-Based Approximation (IBA), Dynamic Message Passing (DMP), Monte Carlo Mean-Field (MCMF), Soft Margin Estimator (SME), Backtracking Network (BN, a graph neural network), and Belief Propagation (BP). We unify all methods under a single problem formulation and evaluate them on eight diverse temporal networks. In the main experiment, we compare source detection performance for varying basic reproduction numbers. Three complementary experiments on a subset of networks study (i) learning curves, (ii) partial observability, and (iii) parameter misspecification. We find that BP and BN are the strongest source detectors overall. SME is strong under partial observation but mediocre otherwise. IBA, DMP, and MCMF share a common weakness rooted in their factorized-likelihood assumption, leading to overconfident predictions. We identify scaling these methods to larger networks as the most pressing direction for future work.