Detecção do complexo QRS em eletrocardiogramas com 12 derivações utilizando redes neurais convolucionais

Abstract

Cardiovascular diseases (CVDs) are the leading cause of death in the world. According to the World Health Organization (WHO) about 17.9 million people died from CVDs in 2019, with the highest occurrence in low- and middle-income countries. CVDs have significant economic consequences that affect individuals, healthcare systems, and society. Early detection is the main form of prevention. The Electrocardiogram (ECG) is the widely used technique for detecting CVDs. It represents the heart’s electrical activity, and has several waves, the most prominent of which is the QRS complex. The ECG signals are captured in various regions of the body, and the 12-lead ECG is commonly used in clinical practice. The use of machine learning techniques, and in recent times deep machine learning, has been the focus of much research in order to detect the QRS complex. However, a literature database search evidenced the lack of such studies applied individually to the 12 leads of the ECG. In this study, we propose to advance the detection of the QRS complex in all 12 ECG leads. To this end, we use the public dataset, INCART, in which we label the QRS intervals of all the signals available in this dataset. The methodology investigates three convolutional neural network architectures (CNN1, 2 and 3) with different sizes of receptive fields. The architecture that showed the best average performance for the 12 leads-ECG was CNN2, with an accuracy of 76.44%, recall of 99.90%, and F1-score of 86.50%. The accuracy and recall values were affected by the model’s false positives. These were minimized by an additional post-processing step. With this post-processing, the performance metrics became: accuracy 99,98%, recall de 99,90% and F1-score, outperforming the CNN model presented by Xiang et. al. (2018) on the same dataset.