Advanced Time-Frequency Methods for ECG Waves Recognition

ECG wave recognition is one of the new topics where only one of the ECG beat waves (P-QRS-T) was used to detect heart diseases. Normal, tachycardia, and bradycardia heart rhythm are hard to detect using either time-domain or frequency-domain features solely, and a time-frequency analysis is required to extract representative features. This paper studies the performance of two different spectrum representations, iris-spectrogram and scalogram, for different ECG beat waves in terms of recognition of normal, tachycardia, and bradycardia classes. These two different spectra are then sent to two different deep convolutional neural networks (CNN), i.e., Resnet101 and ShuffleNet, for deep feature extraction and classification. The results show that the best accuracy for detection of beats rhythm was using ResNet101 and scalogram of T-wave with an accuracy of 98.3%, while accuracy was 94.4% for detection using iris-spectrogram using also ResNet101 and QRS-Wave. Finally, based on these results we note that using deep features from time-frequency representation using one wave of ECG beat we can accurately detect basic rhythms such as normal, tachycardia, and bradycardia.

Automated summary

This study used two different spectrum representations, iris-spectrogram and scalogram, to extract deep features and classify different ECG beat waves using two deep convolutional neural networks (CNN), ResNet101 and ShuffleNet. The results showed that using ResNet101 and scalogram of T-wave achieved the highest accuracy of 98.3% for beat rhythm detection, while using iris-spectrogram and ResNet101 for QRS-wave achieved an accuracy of 94.4%. In conclusion, deep features from time-frequency representation of ECG beat waves can accurately detect basic rhythms such as normal, tachycardia, and bradycardia.