Detecção e classificação de patologias cardíacas em eletrocardiogramas utilizando redes neurais profundas

Abstract

Cardiovascular diseases have been, in the last decades, among the main causes of mortality worldwide. Such diseases can be detected/diagnosed by a health professional through a noninvasive test called an electrocardiogram (ECG), which provides a representation of the electrical activity of the heart. Nonetheless, the interpretation of the ECG is not a trivial task, demanding considerable experience from the health professional to properly identify morphological changes. In this context, machine learning (especially, deep learning) techniques have been widely used, in the field of clinical electrophysiology, to construct mathematical models which are able to automatically detect patters associated with the presence or absence of cardiac pathologies. However, despite the good performance achieved, only a few research works published so far have used public datasets for training the introduced architectures, thus making it difficult to validate the results, to obtain fair performance comparisons, as well as to propose improvements. So, the present work focuses on the development of a computational tool for automatic detection and classification of 5 super-classes of cardiac pathologies. Specifically, the architectures introduced in (RAJPURKAR et al., 2017) and (RIBEIRO et al., 2020) are implemented here, using Python language along with Tensorflow and Keras modules, and trained with a public dataset already labeled by cardiologists, named PTB-XL (WAGNER et al., 2020). The performance of the obtained models (after training) is assessed through the multilabel confusion matrix (MCLM), from which some metrics like precision, recall, and F1-score as well as their average values termed micro, macro, and weighted can be computed, according to the methodology described in (HEYDARIAN; DOYLE; SAMAVI, 2022). Still, comparisons involving the computational cost required for implementing and training both architectures are presented, highlighting the number of parameters, number of epochs, and time elapsed until the end of the training. Results showed that the model obtained from the architecture described by (RIBEIRO et al., 2020) presents a similar performance, or even slightly better, for different metrics than the one obtained from the architecture introduced by (RAJPURKAR et al., 2017); especially, when the involved computational cost is taken into account.