We detect communities on time-dependent correlation networks to study the geographical spread of disease. Using data on country-wide dengue fever, rubella, and H1N1 influenza occurrences spanning several years, we create multilayer similarity networks, with the provinces of a country as nodes and the correlations between the time series of case numbers giving weights to the edges.
We perform community detection on these temporal networks of disease outbreaks, looking for groups of provinces in which disease patterns change in similar ways. Optimizing multilayer modularity with a Newman-Girvan null model over a wide parameter range, we observe several partitions that corresponding roughly to relevant historical time points, such as large epidemics and introduction of new disease strains, as well as many strongly spatial partitions.
We develop a novel null model for community detection that takes into account spatial information, thereby allows to uncover additional structure that might otherwise be obscured by spatial proximity. The null model is based on a radiation model that was proposed recently for modelling human mobility, and we believe that it might be better at capturing disease spread than existing spatial null models based on gravity models for interaction between nodes.
The radiation null model performs better than the Newman-Girvan null model and similarly to the gravity model on benchmark spatial networks with distance-dependent links and a known community structure (both static and multislice networks), and it strongly outperforms both on flux-based benchmarks. When applied to the disease networks, the radiation null model uncovers novel, clear temporal partitions, that might shed light on disease patterns, the introduction of new strains, and provide epidemic warning signals.