Deep learning-assisted arrhythmia classification using 2-D ECG spectrograms

Abstract This article studies modern classification techniques in ECG signals through the transfer learning approach with CNN (Convolutional Neural Network). The proposed pre-trained network combines an Imagenet with huge labeled image datasets and a separate network composed of fully connected layers. This method uses the CWT (Continuous Wavelet Transform) to construct a time-frequency visualization of ECG signals, which are subsequently transformed into RGB images. The developed images are plugged into a pre-trained CNN to retrieve the desired features. We next employ supervised learning to train the neural network on the ECG labeled data using CNN features. To train a Deep Neural Network, three sets of PhysioNet databases are used: MIT-BIH (ARR) Arrhythmia, NSR (Normal Sinus Rhythm), and BIDMC CHF (Congestive Heart Failure). The classification Accuracy, Sensitivity, Specificity, F1-score, Precision, and Detection Error Rate of the CNN classifier are compared to AlexNet, GoogleNet, Vgg16, and SqueezeNet pre-trained networks. Among all these networks, SqueezeNet provides an Acc of 98.7%, Se of 99.1%, Sp of 99.20%, F1-score of 98.33%, Precision of 98.67%, and DER of 0.89%. For further investigation, the technique suggested can be implemented in addition to Bi-LSTM on some real ECG data.