Febre do oropouche: eficácia de modelos matemáticos e de aprendizado de máquina na previsão de casos

Abstract

Abstract: Climate change amplifies viruses such as Orthobunyavirus oropoucheense, which causes Oropouche Fever (OF). In Brazil, it was confined to the Amazon until it expanded beyond its borders, becoming emerging disease. Therefore, new cases of OF in Brazil were estimated using the Autoregressive Integrated Moving Average model with exogenous variables, the Long-Short Term Memory (LSTM) model, and the arithmetic mean between them. Data from the Epidemiological Panel of the Ministry of Health and the Prediction of Worldwide Energy Resources regarding OF cases and average temperatures from 2023 to 2025 were collected. A descriptive analysis of the sample was performed, including sex, age group and region. Stationarity, trend, and autocorrelation of the time series were analyzed using the augmented Dickey-Fuller, Mann-Kendall, and Ljung-Box tests. Furthermore, the curve was decomposed using the additive model. Predictive models were implemented in Python. 26,678 cases were reported between 2023 and 2025, predominantly male (52.36%) and aged 20 to 49 years (56.90%). Cases were concentrated in the North (94.36%) in 2023, shifting to the Southeast in 2025 (85.01%). The series was non-stationary, trendless, and autocorrelated. Curve decomposition showed an increasing trend from June/2023 to April/2025, after which it became decreasing. From July/2025 onwards, the trend disappeared. Regarding seasonality, notifications increased from September to February and decreased from March to October. In model validation, LSTM demonstrated the best performance. Its Mean Absolute Error was three cases predicting 3 months and 515 predicting six. Therefore, LSTM can assist in combating FO. However, it still needs calibration.