Ensemble method based on Artificial Neural Networks to estimate air pollution health risks

Abstract

Estimating of daily hospital admissions due to air pollution is a leading issue in environmental science. To better understand this problem, it is essential to improve the applied methodologies. The use of Generalized Linear Models (GLM) is well known. However, they may be improved using different methods to coefficients estimation and to consider seasonality. Alternative methodologies, rarely applied in such topic, are Artificial Neural Networks (ANN), efficient to solve non-linear problems and; ensembles, which combine various models outputs. This research aims to apply 10 distinct ANN and 4 ensemble to estimate hospital admissions for respiratory diseases caused by particulate matter and meteorological variables of Campinas and São Paulo cities, Brazil. In addition, a new proposal of GLM was introduced, considering coefficients calculation via particle swarm optimization and seasonality via normalization procedure. ANN and ensembles use showed significant improvements and may allow studies into areas with flawing database, developing countries reality.