Previsão de internações hospitalares de dengue por meio de séries temporais

Abstract

Aedes aegypti is the vector of comorbidities yellow fever, dengue fever, chikungunya and zika. Among them, dengue is the most common disease and the one that causes the most deaths, becoming a public health concern in Brazil and worldwide. Brazil presents ideal conditions for mosquito proliferation and concentrates most dengue cases in the Americas. Controlling the vector is challenging and needs strategic action involving government and civil society. Infection by dengue virus can be asymptomatic, mild or cause a serious disease that puts the patient's life at risk, who will need to be admitted to a hospital for treatment. It is known that hospital beds are limited, the demand for dengue is representative and hospitalization generates costs. The incidence of dengue has tendency and seasonality, the main components of a time series. To contribute to state and municipal health management in combating Aedes aegypti, as well as hospital management of dengue hospitalizations, this research created an algorithm for the prediction of dengue hospitalizations using computational statistical models for time series analysis focusing on planning and management of disease control. The main statistical techniques used (Exponential Smoothing, Autoregressive Integrated Moving Average Model (ARIMA), Autoregressive Artificial Neural Networks, Model combination, Linear Regression and Naive Methods) to forecast 8-week time series and a real estimate of 4-weeks, because DATASUS releases data is 4-weeks late. The adjustment of the parameters of each model was performed automatically in the Integrated Development Environment (IDE) RStudio, by functions of the forecast package that contains implementations of the statistical models of time series for the programming language R. Using multiple forecasting methods applied concomitantly over the same time series has improved forecast accuracy. The algorithm made about 8 predictions every 10 with Mean Absolute Percentage Error (MAPE) below 26%. This is another strategy with potential to be used in dengue control in Brazil, because the algorithm created may be the basis for the development, in the future works, of a web service that provides health managers (Ministers, Secretaries, Directors and Coordinators) the possibility of predicting hospitalizations for dengue according to reality of each one. From this study, research opportunities arose to explore statistical methods that deal with count time series for predicting dengue hospitalizations with a focus on the health establishment where the data occur, as well as conducting a Systematic Literature Review, considering the classical methods and machine learning techniques, to answer the questions: Which statistical methods for forecasting time series (discrete/count or continuous) are most used? Are classical methods more efficient than machine learning approaches? Which error metrics are safer for choosing the most accurate model among candidates for a given time series forecast?