Abstract
In this paper, we propose a transfer learning approach for Arrhythmia Detection and Classification in Cross ECG Databases. This approach relies on a deep convolutional neural network (CNN) pretrained on an auxiliary domain (called ImageNet) with very large labelled images coupled with an additional network composed of fully connected layers. As the pretrained CNN accepts only RGB images as the input, we apply continuous wavelet transform (CWT) to the ECG signals under analysis to generate an over-complete time-frequency representation. Then, we feed the resulting image-like representations as inputs into the pretrained CNN to generate the CNN features. Next, we train the additional fully connected network on the ECG labeled data represented by the CNN features in a supervised way by minimizing cross-entropy error with dropout regularization. The experiments reported in the MIT-BIH arrhythmia, the INCART and the SVDB databases show that the proposed method can achieve better results for the detection of ventricular ectopic beats (VEB) and supraventricular ectopic beats (SVEB) compared to state-of-the-art methods.