Sistema de apoio ao diagnóstico de arritmias cardíacas

• Ano de defesa: 2018
• Autor(a) principal: Cardozo, Regis Augusto
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Ponta Grossa; Brasil; Programa de Pós-Graduação em Engenharia Elétrica; UTFPR
• Palavras-chave em Português: Coração - Doenças; Arritmia - Diagnóstico; Análise de componentes principais; Heart - Diseases; Arrhythmia - Diagnosis; Principal components analysis; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA; Engenharia Elétrica
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/3297

In 2013, 4.2% of the Brazilian population over the age of 18 had a diagnosis of heart disease, 13.5% of whom had limitations in their usual activities due to the disease. Therefore the use of medical assistance systems that produce a good performance is desirable. Besides helping medical specialists in the diagnosis of diseases, systems for this purpose may be used when the presence of this specialist for the analysis of the results is not always possible. Being useful also in patients monitoring equipment of the mHealth type. Thus, this work proposes and evaluates a diagnosis assistance system of cardiac arrhythmias, to classify 11 types of heart beats that were grouped into 5 supertypes, using electrocardiographic signals. This work evaluates the influence of an adaptive filtering technique with the use of Morphological Filter and a classical technique based on finite impulse response filters, the Discrete Wavelet Transform. The characteristics extracted from the electrocardiogram signal were obtained with the principal component analysis (PCA), varying the number of components between 10, 12 and 14. Some classifiers based on artificial neural networks (ANNs) were also evaluated. Two hybrid radial-based function ANNs (RBFs) with the training algorithm of extreme learning machine (ELM) ANN, one with one hidden layer and the other with two; and two ANNs with only the ELM algorithm, but with two different activation functions (Logistic Function and Gaussian Function). Another point evaluated was the influence of the presence or not of the regularization coefficient in the ELM algorithm. The results were obtained by the k-partitions validation method, with 5 partitions being combined 2 by 2, in order to perform 10 training and tests, using 2 partitions for training and the other 3 for testing. The results obtained demonstrate that the tested activation function do not affect significantly the results on the RNAs with ELM algorithm. It was also observed that the regularization coefficient has only influenced the results when there are more than 1000 neurons in the hidden layer, always presenting better results. It was also concluded that although in most cases the result is not affected by the filtering technique, the Morphological Filter presents slightly better results where there are significantly different results. Finally, the best average accuracy obtained was 96.61 ± 0.51%, with Morphological Filter, RNA with ELM algorithm, 12 principal components and 1500 neurons in the hidden layer.