Heart beats classification method using a multi-signal ECG spectrogram and convolutional neural network with residual blocks

Abstract

The paper describes a process of formulating a classifier on the basis information contained by MIT-BIH arrhythmia database. This data source contains electrocardiographic signals from two sensors. Both were used, which represent not a typical phenomenon. In the learning process, the classifier uses only information with high certainty. Data are based on expert endorsements and the errors found have been corrected over the years. Specific types of heartbeats were divided into special groups according to the standard "Association for the Advancement of Medical Instrumentation" (AAMI). It recommends splitting the specific types into five separate groups according to physiological origin. Rare heartbeats have a limited number of occurrences. For one group, modifying methods were used which allowed to increase sufficiently the amount of data in training sets. This had a beneficial impact on the results. The solution includes features extraction. The main module of the classifier is a deep neural network. Good result was obtained with tools supporting automatic hyperparameter selection. In ECG signal diagnostics, the most significant task is to properly separate the group of supraventricular and ventricular beats. The study managed to obtain this error at an exceptionally low level and an overall accuracy of 98.37%.