Modelagem preditiva utilizando séries temporais para incidência de dengue em londrina-paraná, brasil

Abstract

Introduction: Dengue is a viral disease transmitted by the Aedes aegypti mosquito and represents a significant public health concern in tropical and subtropical regions. Its incidence is strongly influenced by climatic factors such as temperature and precipitation, which contribute to the development of the vector. This study aims to forecast dengue case incidence in the municipality of Londrina, Paraná (Brazil), from 2013 to 2024, using time series-based predictive models. Methods: Monthly dengue case data were obtained from the InfoDengue system and combined with climatic variables such as temperature and precipitation. The methodology involved the application of traditional time series models, including ARIMA and SARIMA, as well as their extension, SARIMAX, which incorporates exogenous climatic variables into the forecasting process. Additionally, Long Short-Term Memory (LSTM) neural networks were used as a comparative model to explore their ability to capture complex temporal patterns. Results: After several inductive adjustments, the SARIMA model (2,1,2)(1,1,1)₁₂ showed the best fit for monthly dengue incidence data, demonstrating strong performance, including the ability to capture the peak oscillations during the year 2020 (COVID-19 pandemic period). For weekly data, the SARIMAX model (2,1,8)(1,1,1)₅₂, which incorporated precipitation from the two preceding weeks and minimum temperature from the three preceding weeks, was identified as the most suitable, showing excellent fit and accurate forecasts. On the other hand, LSTM neural networks were able to identify temporal and seasonal patterns of dengue but showed greater instability and lower ability to anticipate abrupt peaks when compared to the statistical models. Conclusion: This study conclude that SARIMA models—and particularly SARIMAX—are robust and effective tools for forecasting dengue cases, standing out for their stability, accuracy, practical applicability in urban contexts and interpretability of model parameters in the context of arboviruses . These models show great potential to support epidemiological surveillance, public health planning, and preventive strategies. Although LSTM neural networks offer promise for highcomplexity scenarios, statistical models still outperform as the most consistent approach in this study. .