Very deep feature extraction and fusion for arrhythmias detection

The electrocardiogram (ECG) is a picture of heart electrical conduction, which is widely used to diagnose many types of diseases such as abnormal heartbeat rhythm (arrhythmia). However, it is very difficult to detect the abnormal ECG characteristics because of the nonlinearity and the complexity of ECG signals from one side, and the noise effect of these signals from the other side, which make it very difficult to perform direct information extraction. Therefore, in this study we propose a very deep convolutional neural network (VDCNN) by using small filters throughout the whole net to reduce the noise affect and improve the performance. Our approach introduces multi-canonical correlation analysis (MCCA), a method to learn selective adaptive layer's features such that the resulting representations are highly linearly correlated and speed up the training task. Moreover, the Q-Gaussian multi-class support vector machine (QG-MSVM) is introduced for classification, an algorithm which has a better learning performance and generalization ability on ECG signals processing. As a result, we come up with expressively more accurate architecture which is able to differentiate between the normal (NSR) heartbeats and three common types of arrhythmia atrial fibrillation (A-Fib), atrial flutter (AFL), and paroxysmal supraventricular tachycardia (PSVT) without performing any noise filtering or pre-processing techniques. Experimental results show that the proposed algorithm outperforms the state-of-the-art methods.