A Deep Learning-Based Algorithm for ECG Arrhythmia Classification

This paper aims to automatically classify five classes of arrhythmia present in Electrocardiograms (ECG) by using two Deep Learning (DL)-based models. One based on Convolutional Neural Network (CNN) and the other based on Residual Networks (ResNet). The main motivation of this research is to enhance the field of medicine and assist doctors in the diagnosis of arrhythmia. The DL-based models were trained using the MIT Arrhythmia database. The evaluation of the DL-based models was performed by separating the data into 70% for training, 20% for testing and 10% for validation. Results with the testing dataset show an accuracy of 96.33% and 95.40%; a F1-Score value of 96.34% and 95.34% for the CNN- and ResNet-based models, respectively. With the validation dataset, CNN-based model achieved an accuracy of 99.32% and a F1-Score of 99.32%; ResNet-based model achieved 98.55% and 98.55% for accuracy and F1-Score, respectively.