Classificação automática de arritmias cardíacas através de redes neurais convolucionais multimodais com mecanismo de atenção

Abstract

The electrocardiogram (ECG) is a non-invasive technology capable of recording heartbeats and is the most widely used technique for diagnosing heart diseases. Among the conditions that can be diagnosed, cardiac arrhythmia is one of the most common heart disorders, characterized by irregular heartbeats. However, interpreting long ECG signal recordings is a tiring and challenging task when performed visually, which can be time- consuming for medical specialists. Advances in technology and artificial intelligence have enabled progress in the study and development of automatic systems to support medical diagnosis. In this context, this thesis aims to propose a framework for the classification of cardiac arrhythmias based on a multimodal Convolutional Neural Network (CNN) with an attention mechanism. The framework takes as input ECG signal segments transformed into images using the Hilbert Space Filling Curve (HSFC) and Recurrence Plot (RP) techniques. It was developed and evaluated using the public MIT- BIH and PTB databases, following the AAMI (ANSI/AAMI EC57) guidelines and considering both inter-patient and intra-patient paradigms. Due to the high class imbalance in the databases, complementary data augmentation techniques were evaluated during the experimental phase, with two techniques standing out: SMOTE and WGAN-GP. The results achieved, considering variations in the input structure related to the number of ECG leads (MLII lead and V+MLII leads), can be considered competitive with state-of-the-art works. Particularly noteworthy are the results of the structure for two ECG leads, which achieved, for the MIT-BIH database in the intra-patient paradigm, 99.72%, 98.19%, 97.26%, 99.34%, and 97.72% for overall accuracy, precision, sensitivity, specificity, and F1-Score, respectively. In the inter-patient paradigm, the results obtained were 98.48%, 94.15%, 80.23%, 96.34%, and 81.91%, respectively.