Abstract
The classification of cardiovascular diseases using ECGdata is considered. It is argued that to obtain a satisfactory classification features should be extracted from ECG images in their entirety, instead of translating the image into a 1D time series and only considering a small number of features as is the current common practise. The presented approach used a pre-trained Convolutional Neural Network (CNN) as a features extractor, followed by the application of T-distributed Stochastic Neighbour Embedding (T-SNE) to find the best discriminant features to perform ECG classification. Themotivation using a pre-trained CNNmodel is that available ECG data sets tend to be limited in size; typically insufficient for training a bespoke deep learning model for feature extraction. Using a pre-trained CNN this challenge can be addressed. The features were extracted from the fully connected layers immediately preceding the softmax layer. The use of several pre-trained CNNs is reported on: VGG16, InceptionV3, and ResNet50. The operation of the proposed approach was also compared with recent relevant published approaches. A best AUC value of 0.960 was produced using the proposed approach; while the best alternative approach, out of those considered, produced an AUC of 0.932.