Speaker
Dr
Sebastian Funk
(Institute of Zoology)
Description
Community structure is a ubiquitous feature of complex
networks, and methods for its detection has gained much
attention in recent years. Beyond the study social networks
with well defined links, these methods can be generalised to
operate on any dataset in which different entities are
similar in one or more traits, and be used to identify
meaningful groupings. Here, we describe the application of
network theory and methods for finding community structure
to identify undiagnosed disease outbreaks reported in online
surveillance systems. The efficacy of these programs is
often inhibited by the anecdotal nature of informal or
rumour-based reporting, and uncertainty of pathogen
identity. We create associations between disease outbreaks
and and their symptoms, case fatality ratio, and
seasonality, and represent them in an abstract network. We
train the model with a set of outbreaks reports of 10 known
infectious diseases causing encephalitis and combine methods
for community detection with an optimisation procedure for
symptom weights to generate networks of maximal modularity.
We then use these to determine a most probable
identification for 97 outbreaks of encephalitis reported in
an online surveillance system as undiagnosed or ‘mystery
illness’, by determining the best association with
communities in the reference networks. This illustrates the
general use of methods from network analysis for the study
of datasets even where links are not obvious physical
entities but mere measures of similarity.