Daily mortality analysis in relation to meteorological and air pollutant data using compound backpropagation multi-layer perceptron neural network models

Abstract

A backpropagation algorithm with multi-layer perceptrons is implemented for a mortality analysis study. It is a new approach to explore the health impact of air pollution and other environmental factors on daily mortality in Christchurch of New Zealand for the period 1988-1997. The conventional multi-layer neural network was architecturally extended to cope with data which has non-linearities and interdependencies. Despite extensive research on the meteorological effect of air pollution trends and pollutant effects on public health, there have been no attempts to combine both meteorological factors and pollutant factors to estimate the levels of risk to human health. Due to the complexity of uncertainty in the environmental problems, it is often difficult to apply full numerical models. Therefore statistical approaches have been introduced and applied. However Statistical approaches are not fully justified due to the assumptions made in the most statistical models regarding the data distribution and inter-attribute dependency. In this research, the compound neural network, architectural extension of the original multi-layer perceptron neural network, is employed and various experiments on both conventional and compound models are presented. The results of these experiments and their implications on health science are discussed.