Segmentação de batimentos cardíacos em eletrocardiograma e identificação de arritmia cardíaca em um único lote de processamento por aprendizado profundo usando uma abordagem de regressão em uma rede neural convolucional

Abstract

Conventional methods for diagnosing cardiovascular disease on electrocardiogram (ECG) depend on obtaining particular characteristics in the signal by a human observer. Due to many patients and the need for continuous observation, several automatic arrhythmia classification techniques have been developed in recent years to solve this problem. Therefore, the thesis proposes a new method for classifying cardiac arrhythmias and wave segmentation in electrocardiogram signals that uses a Deep Learning regression approach. In addition, we evaluate the classic approach to classifying arrhythmias, which independently deals with the stages of segmentation, feature extraction, and classification. The combination with the highest accuracy, 94.3%, in the classical methods was the extractor Estatísticas de Alta Ordem (Higher-Order Statistics, HOS) with the classifier Naive-Bayes using division with five classes. As for the problem with the three classes, the results show that the HOS had a similar performance. This extractor and the Naive-Bayes classifier achieved an accuracy of 94%, sensitivity above 99% for normal heartbeats, and specificity of 99% in the EBV and SVEB classes. Due to the challenges presented in this study group, we propose a new method that uses regression with Deep Learning. Their accuracy using the same evaluation protocol for divisions with 2, 3, and 5 classes was 99.56%, 99.55%, and 99.45%, respectively. The other metrics computed from the confusion matrix were also greater than 99% in all variations. The average classification time of the proposed method using Unidade de Processamento Gráfico (Graphics Processing Unit, GPU) was 5.59E-3s, confirming that it is suitable for real-time systems. As for QRS detection, the proposed method obtained 99.88% of accuracy in the MIT-BIH and QT databases. Regarding QRS segmentation, the mean error of the time difference from the beginning of the QRS, QRSon, was 0.26ms, and the QRSo f f was 2.08ms. In detecting T and P waves, the best sensitivities were 99.85% and 99.52%, while the positive predictive values were 99.92% and 99.69%. The segmentation errors for these wave types were −0.25 ± 13.22 and −2.80 ± 14.79 for T wave and 0.26 ± 9.98 and 0.17 ± 9.68 for P wave. The study presented in this thesis shows that the proposed method proves to be adequate in integrating systems to aid medical diagnosis in Cardiology.