Abordagens para predição do número de casos semanais de dengue em Regiões Tropicais

Abstract

Arboviruses transmitted by Aedes aegypti and Aedes albopictus mosquitoes represent one of the major public health challenges, with dengue being the most prominent. Managing dengue epidemics requires advanced preparation; thus, predicting the weekly number of cases in a specific region can aid in prevention and control strategies for the epidemic process. In this study, we evaluated the effectiveness of classical statistical techniques and machine learning methods in predicting the number of weekly dengue cases, using geographic data from San Juan, Puerto Rico, and the 27 Brazilian federative units. In San Juan, we selected features based on the cross-correlation matrix with the total number of weekly dengue cases and applied wavelet transformations to the selected feature, where the Linear Regression (LR) model, using precipitation levels and vegetation indices filtered by the Symmlet wavelet (sym20), presented the best results in metrics such as MAE, R-squared , MAPE, RMSE, and BIAS, showing a 67.22% reduction in RMSE compared to the univariate LR without wavelet filtering. For the Brazilian federative units, a similar procedure was carried out, but without using the cross-correlation matrix. The univariate LightGBM model, trained on 26 cities (cross-learning) and validated individually in each federative unit through the univariate leave-one-out technique and one-step predictions, demonstrated superiority both in individual validations and in comparison with the geographic locality of San Juan, where the model was not trained, evidencing better generalization of results compared to other shallow, deep, and founding models such as TimeGPT-1 and MOIRAI.