A Comparative Study of Deep Learning Techniques Aimed at Detection of Arrhythmias from ECG Signals

This article presents a comparative study of the performance of Deep Learning techniques (LSTM, Dense, and CNN) applied to the detection and classification of arrhythmias in an ECG signal. The objective is to obtain the best prediction model for cardiac arrhythmias with each Deep Learning technique and compare their performances. The training was done with the Physionet MIT-BIH database, from which 108854 samples were taken from a total of 48 patients with and without proven arrhythmia. With this information, the three indicated neural networks were trained, and the performance comparison was carried out through the ROC curve. The one-dimensional convolutional neural network (1D-CNN) achieved the best performance. It was used to build an arrhythmia detection algorithm using one hidden convolutional layer and 3600 input samples at a time (10-s signal fragments). The obtained results were very satisfactory, achieving a precision of 92.07% for the 1D-CNN technique, 82.10% for LSTM, and 75% for Dense.