Avaliação do impacto do MP2.5 na saúde: comparação entre regressão estatística e rede neural artificial

Abstract

Particulate matter is one of the most harmful pollutants to human health. It is classified according to its size. The particles considered fine, also called PM2.5, have up to 2.5 μm in aerodynamic diameter. High concentrations of particulate matter are related with increased prevalence of adverse health effects, mainly due to respiratory and circulatory problems. Studies in this line of research are more common using Generalized Linear Models (GLM), but studies require a large amount of data and minimal discontinuity. Artificial neural networks (ANNs) are a mathematical modeling of important application in engineering, considered more flexible and adaptable than GLM. In this sense, the objective of the present study was to compare the performance of multilayer perceptron RNA (MLP) with a GLM with Poisson distribution. As a case study, we considered the impact of PM2.5 on the number of hospitalizations due to diseases in the population of the city of Ponta Grossa, Paraná, the 4th most populous city in the state. The database contains 564 samples dated from 01/10/2016 to 17/04/2018. The following input variables were considered: PM2.5 concentration, mean temperature, relative humidity, day of the week and whether or not the day is a holiday, and the output is the expected value of hospitalizations for respiratory problems. For the MLP the data were divided into training (70%), validation (20%) and test (10%), while for the GLM, the data were divided into training (90%) and test (10%). As a result, an error of 37% (MAPE) was obtained for MLP and 42% (MAPE) for GLM, but the other two metrics (MAE and RMSE) were lower for GLM, noting that for this case study, GLM performed better. In many cases, the application of GLM is sufficient, as it is a simpler and faster modeling compared to RNAs. However, RNAs are considered alternatives to the application of GLM, when working with databases with a lot of discontinuity, which is frequent in developing countries, such as Brazil.