ECG Arrhythmia Classification for Comparing Pre-Trained Deep Learning Models

In order to perform the diverse tasks of deep learning, different models of neural networks have been proposed, and for several years now many deep neural networks have been developed for a variety of uses. In general, the labor of training these networks is the most challenging part of this methodology, and is by far the most time consuming, both in terms of effort required to configure the process and computational complexity required to perform it. To relieve this process, pre-trained models have been made available that already learned to extract powerful and informative features from big sets of images, thus reducing training time for particular problems, but, given the large number of options, it would consume a lot of time testing each of them. In this study, we tested the performance of nineteen pre-trained networks that are available for MATLAB and compared their performance for classification of ECG signals from MIT-BIH arrhythmia dataset. We report the main characteristics of these networks, training times, and performance for this classification job to serve as reference for their use by other researchers.