ArithNet: A Promising Multi-Scale Feature Fused Convolutional Network for Arrhythmia Identification from Electrocardiogram Signals

Abstract

Arrhythmia is a highly prevalent chronic cardiac disorder in senior citizens and is related to the high severity including cardiovascular accidents, heart failure and myocardial ischemia. It is essential to instantly identify and categorize arrhythmia rhythms from Electrocardiogram (ECG) signals. From this viewpoint, a Multi-Scale Fusion-Convolutional Neural Network (MSF-CNN) was developed, which uses multi-scale features from the ECG signal for identifying arrhythmia classes. But, it needs a vast number of ECG signals and takes more time to train the model because of using cross-validation. As a result, this article designs a new lightweight end-to-end MSF-CNN with Long Short-Term Memory-Gated Recurrent Unit (LSTM-GRU) structure called an ArithNet model for recognizing arrhythmia automatically. In this model, two different training schemes are applied such as representation training and sequence residual training. At first, the ECG signal database is collected and preprocessed to remove the noisy signals. Then, the noiseless ECG waves are partitioned into regular (R), supraventricular ectopic beat (SV), ventricular ectopic beat (V), merging beat (M) and unfamiliar beat (U) based on the labeling from heart specialists. Such waves are given to the representation training, which extracts time-variant salient characteristics from the ECG signals. Moreover, the sequence residual training is performed, which extracts the temporal characteristics using bidirectional links. Further, the obtained salient and temporal characteristics are fused and categorized by the softmax layer to identify arrhythmia. Finally, the experimental results illustrate that the ArithNet on MIT-BIH and Arrhythmia Data Set achieves an accuracy of 93.09% and 92.84%, respectively than the other classical deep learning models for arrhythmia identification.