ECG heartbeat arrhythmias classification: a comparison study between different types of spectrum representation and convolutional neural networks architectures

This research presents a comparison study between different representations of spectrograms and then feeding them to different convolutional neural network (CNN) architectures. The study uses two short-time Fourier transform (STFT) representations, namely, Log-scale and Mel-Scale in addition to Bi-Spectrum and the third-order cumulant. Meanwhile, four different CNN architectures have been utilized in the present study, namely, AOCT-NET, Mobile-Net, Squeeze-Net, and Shuffle-Net. The study has exploited 10,502 beats extracted from the standard MIT-BIH arrhythmia database and represent six different classes: normal beat (N), left bundle branch block beat (LBBB), right bundle branch block beat (RBBB), premature ventricular contraction (PVC), atrial premature beat (APB), and aberrated atrial premature (aAP). The study compares the accuracy, sensitivity, precision, and specificity rates of the spectrogram-based and CNN architecture models under study. This paper hypothesizes that ECG features can be extracted from different spectral representations and can lead to improving the understanding and detection of the human heart's different arrhythmias by feeding these features to different CNN models. The suggested models’ performance was evaluated by dividing the dataset into three subsets (Training 70%, Validation 15%, and Testing 15%) and the best overall performance among all used CNN architectures was MobileNet with an overall accuracy of 93.8%, while the best spectrum representation among all used was the bispectrum with an overall accuracy of 93.7%. It has been shown that the spectrum representations of ECG beat have provided significant information about heart performance and can be used significantly in arrhythmia classification using deep learning techniques.