Utilização de redes neurais artificiais para previsão de internações causadas pela poluição atmosférica

Abstract

Atmospheric pollution is considered one of the main causes of respiratory problems for the urban population, due to the high volume of polluting sources, such as automobiles and industries. Studies in the literature often assess the impact of pollutants on health through statistical regression models, but recently, the use of artificial neural networks has been highlighted, due to its efficiency in mapping problems. The objective of this work was to compare two neural network architectures, the MLP (Multilayer Perceptron) and ELM (Extreme Learning Machine), to predict the number of hospitalizations due to respiratory diseases (J00-J99 of International Classification of Diseases - CID-10) caused by particulate matter with an aerodynamic diameter less than or equal to 10 μm. As input, meteorological variables (temperature and humidity) were also considered. As a case study, the city of São Paulo was considered. Both neural networks were implemented using the programming language Python version 3.7 with the help of the Scikit-learn and Numpy libraries. The health effects of pollution may occur sometime after exposure. Therefore, it is essential to model health effects on the day of pollution exposure (lag0) to seven days after exposure (lag7). The metrics used to evaluate the networks were Mean Square Error (MSE) and Mean Absolute Error (MAE). The best results presented by MAE for the tested MLP were for lag 0, presenting a value of 34.38, while for ELM the best result was for lag 7, presenting a value of 40.35. For the case of the MSE, the MLP presented the best result for lag 0, with a value of 1,820.18, while for the ELM, the best result was for lag 2, with a value of 2,585.73. The neural networks proved to be a robust forecasting tool, and the MLP had a better ability to perform forecasts. It was possible to observe a limitation of the generalization capacity of the networks, which is justifiable by the complexity of the problem and because it depends on other variables (age, gender, social status, among others) that were not considered in the work.