Modeling and Predicting Incidence: Critical Systems Failures and Flu Infection Cases

Abstract

Given several related critical infrastructure (CI) networks, such as power grid, transportation, and water systems, one crucial question emerges: how to model the propagation of failed facilities and predict their spread over time to the whole system? Given digital surveillance data, can we predict the impact of Influenza-Like Illness (ILI), including the percentage of outpatient doctors visits, the season duration, and peak? These two questions are related to modeling and predicting the incidence of different types of contagions. In the case of CI, the contagions are the failures of facilities. In the case of flu spread, the contagions are the infective ILI. In this thesis, in the case of CI, we give a novel model of failure cascades and use it to identify key facilities in an optimization-based approach, called HotSpots. In the case of flu spread, we develop a deep neural network, EpiDeep, to predict multiple key epidemiology metrics. In both of these applications, we use the dynamics of propagation to develop better approaches. By collaborating with Oak Ridge National Laboratory (ORNL) and working on the real CI networks provided by them, we find that HotSpots helps solve what-if scenarios. By using the digital surveillance data reported by the Centers for Disease Control and Prevention (CDC), we carry on experiments and find that EpiDeep is better than non-trivial baselines and outperforms them by up to 40%. We believe the generality of our approaches, and it can be applied to other propagation-based scenarios in infrastructure and epidemiology.