Utilização de técnicas de machine learning e de deep learning para a predição de casos de dengue nos municípios da Paraíba

Abstract

Dengue is a disease caused by the DENV virus and transmitted to humans through the Aedes aegypti mosquito. Although it is not a new disease, there is still no regulated vaccine in Brazil that can be used without restriction in the population. Therefore, the fight against the disease is done through actions to eliminate the transmitting mosquito. Dengue numbers returned to grow in Brazil and Paraíba. According to the seventh epidemiological bulletin of arbovirus in Paraíba, there was an increase of 53% of dengue cases in relation to the cases of the previous year. The objective of this work was to create a system capable of forecasting notifications and hospitalizations caused by dengue in the municipalities of Paraíba. Through Machine Learning (Random Forest and Support Vector Regression) and Deep Learning (Multilayer Perceptron, Long Short-Term Memory and Convolutional Neural Network) techniques and using epidemiological, climatic and sanitary data, between 2010 and 2019, the system was able to find the best combination of predictive attributes, the best parameters for the techniques, make predictions of cases of hospitalizations and notifications caused by dengue for the municipalities of Paraíba Bayeux, Cabedelo, Cajazeiras, Campina Grande, Catolé do Rocha, João Pessoa, Monteiro, Patos and Santa Rita, determine which techniques produce better results per city and, finally, the statistical difference between the approaches was demonstrated. The results produced demonstrate the superiority of Deep Learning techniques in comparison to Machine learning techniques. During notification case forecasting, the Long Short-Term Memory (LSTM) technique obtained better results in 66.67% of cities, Convolutional Neural Network (CNN) in 22.22% and Multilayer Perceptron (MLP) in 11.11 %. Regarding hospitalizations, LSTM had the lowest error rate in 33.34% of the municipalities, CNN, MLP and Random Forest (RF) each obtained better results in 22.22% of the cities.