2026-01-282026-01-282025-09-22RAMOS, Daniel Dantas do Amaral. Classificação de arritmias cardíacas via rede neural convolucional com mecanismo de atenção local. Orientadora: Adriana Rosa Garcez Castro. 2025. 84 f. Dissertação (Mestrado em Engenharia Elétrica) - Instituto de Tecnologia, Universidade Federal do Pará, Belém, 2025. Disponível em: https://repositorio.ufpa.br/handle/2011/17896. Acesso em:.https://repositorio.ufpa.br/handle/2011/17896Cardiac arrhythmias represent alterations in the rhythm or frequency of heartbeats and are associated with significant health risks. The electrocardiogram (ECG) remains the primary non-invasive examination for their diagnosis; however, the manual analysis of long recordings is labor-intensive and prone to errors. Considering the challenges arising from manual ECG signal analysis, automatic arrhythmia classification systems based on artificial intelligence have been proposed in the literature as a promising alternative to support medical diagnosis. Within this context, this work presents the proposal of an automatic cardiac arrhythmia classifier based on Convolutional Neural Networks (CNN – AlexNet), with an integrated local attention mechanism in its architecture, developed under the inter-patient paradigm. The study focused on investigating and defining the most suitable technique, between Gramian Angular Field (GAF) and Hilbert Space-Filling Curve (HSFC), for converting ECG temporal signals into images to be used as input for the CNN, and on evaluating the impact of different internal configurations of the proposed local attention mechanism, specifically regarding the choice of activation function (Hyperbolic Tangent or Sigmoid) and kernel type (fixed or adaptive). For the experiments, the MIT-BIH Arrhythmia Database was used, and the performance of the trained models was evaluated according to standardized metrics, such as accuracy, precision, sensitivity, specificity, and F1-score. The experiments showed that the combination of GAF with the Hyperbolic Tangent activation function and adaptive kernel in the attention module achieved the best result, reaching an accuracy of 97.88% and an F1-score of 0.7678. The model outperformed the baseline CNN (AlexNet) architecture without the attention module and demonstrated competitive performance compared to solutions previously presented in the literature, even without the application of additional data balancing techniques.ptAcesso AbertoEletrocardiograma (ECG)Redes Neurais ConvolucionaiMecanismo de atençãoTransformação de sinais em imagensElectrocardiogram (ECG)Convolutional Neural Networks (CNNs)Attention MechanismSignal-to-Image TransformationDissertaçãoCNPQ::ENGENHARIAS::ENGENHARIA ELETRICAINTELIGÊNCIA COMPUTACIONALCOMPUTAÇÃO APLICADA