Comparative Analysis of Deep Neural Network Architectures for Heart Disease Classification in Electrocardiography Signals

Andrés Munõz, Juan Torres-Santamaria, Marcela Iregui, Eduardo Romero, Angel Cruz-Roa

Abstract: Electrocardiography (ECG) is one of the most used non-invasive techniques to evaluate the cardiac condition. This technique provides a graphic representation of the heart's electrical activity, which allows the identification of characteristic patterns associated with the pumping of the heart's blood. There-fore, improving the accuracy and efficiency of ECG interpretation could significantly impact the early detection and treatment of heart diseases. This work presents a comparative analysis of three deep learning architectures Convolutional neural network (CNN), a CNN that integrates an LSTM module (CNN-LSTM) and a neural network that adds an attention module to the CNN-LSTM architecture (CNN-LSTM-Attention) for the task of classifying heart diseases into classification tasks: 2-class (binary: Normal or abnormal) and 5-classes (multiclass: Normal, conduction disturbance, myocardial infarction, hypertrophy, ST/T changes). These methods and tasks were evaluated over the publicly available and large PTB-XL dataset of 12-lead ECG signals with their corresponding diagnosis per task. The CNN model obtained the best classification results for both binary and multiclass tasks. For binary task, the model achieved an accuracy of 86.58%, a precision of 87.15%, a recall of 86.58%, and an F1 score of 86.62%. Nevertheless, for multiclass task, the model achieved an accuracy of 78.12%, a precision of 79.28%, a recall of 78.12%, and an F1-score of 78.45%. However, the performance measures of the other methods are close, with an accuracy of 85.71% for binary and 74.84% for multiclass tasks using CNN-LSTM model; and 83.86% for binary and 72.52% for multiclass tasks using CNN-LSTM-Attention model.