Forecasting epidemiological time series with backpropagation neural networks

Abstract

In public health, surveillance is an important issue. To account for the dynamics of diseases in the population, time series methodologies have been used to provide forecasts of future behaviors. Here, we evaluated the use of backpropagation trained multilayer feedforward networks to forecast epidemiological time series. Sixteen different models within this paradigm, differing basically in input layers and training set presentation, were tested and discussed. Six of them produced fair forecasts for the hepatitis B case occurrence in the US time series.