Detecção e classificação automática de arritmias cardíacas em sinais de eletrocardiograma utilizando redes convolutivas 2D

Abstract

The ECG is the most widely used diagnostic method for identifying and classifying cardiac arrhythmias, myocardial infarction and coronary artery disease. Most articles presented in the literature for the classification of arrhythmias in ECG use 1D convolutional networks. These networks only explore relationships between left and right neighbors. On the other hand, 2D convolutional networks (CNN) also explore relationships between upstream and downstream neighbors, allowing more complex relationships to be established among the data. This study proposes a method for arrhythmia classification using 2D convolutional networks using CNN. The 1D ECG signal is converted into an intensity image. For CNN training, a new data augmentation technique is proposed using sliding windows. Data augmentation values up to 11 times are achieved. Three CNN architectures and two optimization methods (SGDM, and ADAM) are evaluated. The best result for the classification of 13 arrhythmias, with an overall accuracy of 95.80%, is obtained with the SGDM optimizer and data augmentation.