USE OF OPTIMIZATION AND COMPUTER VISION FOR DENGUE CLASSIFICATION WITH TABULAR DATA FROM THE SUS

 This work presents a hybrid approach that combines optimization techniques and computer vision for the automatic classification of dengue cases based on tabular data from public records of the Brazilian Unified Health System (SUS). Although convolutional neural networks (CNNs) are highly effective in image classification tasks, their direct application to tabular data—common in public health contexts—is challenging due to the lack of explicit spatial structure in such data. To overcome this limitation and leverage the potential of CNNs, recent studies have proposed transforming feature vectors into two-dimensional images while preserving local relationships among variables. The transformation used in this study is based on the t-Distributed Stochastic Neighbor Embedding (t-SNE) technique, which employs an iterative optimization process to minimize the Kullback–Leibler divergence between the data distributions in the original and projected spaces. The resulting images are then used as input to CNNs trained to classify notifications as positive or negative dengue cases. Experiments conducted with real data from Open DataSUS show that the proposed approach can outperform classical supervised learning models in diagnostic prediction tasks. These results reinforce the potential of combining optimization techniques and deep learning in the development of tools to support epidemiological surveillance and public health decision-making.