Método ensemble baseado em redes neurais artificiais para estimação de internações por doenças respiratórias

Abstract

Artificial neural networks are universal approximations and efficient tools for solving mapping, interpolation and prediction problems. Another approach that has been explored in the last decade is the ensemble method, which combines the outputs of various predictors. These tools are efficient in solving a wide variety of problems. However, it is not common to use them in forecasting health problems caused by atmospheric pollution. Air Pollution is a theme that for years has attracted the attention of researchers in many areas, mainly because of the damage it causes to human health. There are several studies that relate some pollutants found in the air to the number of hospitalizations for respiratory diseases, and these studies generally apply methods of statistical regression. Therefore, this paper analyzes the application of 10 neural networks models and 4 possibilities of forecasting the number of hospitalizations for respiratory diseases caused by air pollutants. These new approaches were compared with the results of the generalized linear model (GLM) with Poisson regression, which is commonly used in this kind of problem. In addition, in the present study, a new proposal for the use of the GLM is introduced, with the simplification of the pre-processing through the use of a normalization, in which the spline used in the MLG is replaced by the seasonality of the data together with optimal coefficients calculation by means of swarm of particles. The case study involves pollution data, particulate matter with aerodynamic diameter less than or equal to 10𝜇m (MP10) and meteological data, such as mean temperature and relative humidity, of Campinas and São Paulo cities. The desired output of the models is the number of hospitalizations for respiratory diseases. The results showed that, GLM/PSO was better than the classical proposal to Sao Paulo. However, we highlight that the general results strongly indicates that the linear models are not able to reach superior performances thant the non-linear ones. In this same scenario, the unorganized Ozturk ESN model surpassed all the recurrent classic methodologies. To the effect seven days after exposure, it was found that ensemble using multiple layers percetron as combinator reached the best overall results in terms of mean square error, in both scenarios. According to the results, we may conclude that the use of artificial neural networks brought benefits to the addressed topic and had a better adjustment than the classical models of statistical regression.