Sensitivity Analysis and Forecasting in Network Epidemiology Models

In recent years, several methods have been proposed for real-time modeling and forecasting of the epidemic curve. These methods range from simple compartmental models to complex agent-based models. In this dissertation, we present a model-based reasoning approach to forecasting the epidemic curve and estimating underlying disease model parameters. The method is based on building an epidemic library consisting of past and simulated influenza outbreaks. During an influenza epidemic, we use a combination of statistical, optimization and modeling techniques to either assign the epidemic to one of the cases in the library or propose parameters for modeling the epidemic. The method is presented in four steps. First, we discuss a sensitivity analysis study evaluating how minute changes in the disease model parameters influence the dynamics of simulated epidemics. Next, we present a supervised classification method for predicting the epidemic curve. The epidemic curve is forecasted by matching the partial surveillance curve to at least one of the epidemics in the library. We expand on the classification approach by presenting a method which identifies epidemics similar or different from those in the library. Lastly, we discuss a simulation optimization method for estimating model parameters to forecast the epidemic curve of an epidemic distinct from those in the library.