Comparative Analysis of Deep Learning Models for ECG Classification

Abstract

Classification of cardiac time series is an important application area of deep learning. Electrocardiographic signals have great diagnostic value in determining the cardiac health of people. For this reason, they need to be carefully pre-processed and analyzed reliably. Convolutional and recurrent neural architectures show good potential in ECG classification and are currently among the most widely used approaches for modeling cardiac signals. This paper presents a comparative study of a one-dimensional convolutional neural network and a two-way long short-term memory model for ECG classification using the MIT-BIH arrhythmia database. Experiments were conducted on ECG segments important for the heart rhythm and they were assigned to five different arrhythmia classes. Both models were evaluated using conventional classification metrics: accuracy, precision, recall and macro-averaged F1-score. The experimental results obtained show that the CNN model outperforms the BiLSTM model, achieving 95.24% accuracy and a macro F1 score of 0.636, compared to 77.56% accuracy and 0.375 macro F1 score for the BiLSTM model. It can be concluded that the convolutional architecture is more efficient for the considered task and the studied ECG classification database. Both models remain hampered by the classes of rare arrhythmias due to class imbalance. These findings confirm the importance of comparative evaluation of deep learning architectures for ECG classification. The information obtained can be useful for the further development of artificial intelligence-based models, as well as for the analysis of cardiac signals.