In an emerging disease epidemic, public health officials must move quickly to contain the spread. Knowledge of key epidemic parameters often informs the development of containment strategies. Inference procedures such as Bayesian Markov chain Monte Carlo allow researchers to estimate these parameters, but are computationally expensive. We explore supervised statistical methods for fast inference, with a focus on deep learning. We bypass the problem of parameter estimation by instead classifying each epidemic curve into one of a set of possible epidemics according to learned features.
We use supervised classification models including random forests, multilayer perceptrons (feed forward neural networks), and three neural networks for time series analysis. We apply our methods to simulated epidemics on two populations of swine farms in Iowa, one dense and one sparse. We use the R package randomForest and Keras, a specialized library in Python, for the neural network models. We also explore the use of deep neural nets for making forecasts of dynamic contact networks for network-based infectious disease systems.