Cardiac arrhythmia classification by time-frequency features inputted to the designed convolutional neural networks

The electrocardiogram (ECG) plays a vital auxiliary role in medical diagnosis, but due to the very low amplitude of the ECG signals, it is challenging and time-consuming to conduct artificial visual evaluation of the ECG signals. In recent years, medical aid research methods through ECG have emerged one after another. However, most of them have defects such as poor model generalization ability and obvious individual differences. This paper constructs two-way multiplex convolutional neural networks (CNNs) based on time-frequency features to classify normal cardiac rhythm (NOR) and seven cardiac arrhythmias including atrial premature contraction (APC), ventricular premature beat (PVC), left bundle branch block (LBBB), right bundle branch block (RBBB), signal quality change (similar to), ventricular fused heart beat (FVN), and pacing heart beat (/). Firstly, the preprocessing steps of the original rough ECG signal are arranged in a unique order, including wavelet transform, threshold denoising, normalization, chopping, mel-frequency cepstral coefficients (MFCC). Secondly, a 12-layer one-dimensional CNN model with block representation and a 11-layer auxiliary-two-dimensional CNN architecture are designed for the time-domain feature and the frequency-domain feature, respectively, where the focal loss function is defined to solve the problem of data categories imbalance. Finally, the experimental results show that the proposed algorithm presents excellent performances in processing variable length ECGs, the average accuracy of time-domain model is 99.1 %, and the classification accuracy of APC in frequency-domain model is 96.3 %.