About the Choice of Data Balance Method for Neural Network Classification of Electrocardiogram Signals

Abstract

Diseases of the cardiovascular system are the main cause of death in the world population. Classification of electrocardiogram (ECG) signals is a reliable method for diagnosing cardiac pathologies. The available ECG databases consist of an unequal number of signals from various pathologies. This article analyzes the impact of using class alignment methods on the result of neural network classification of ECG signals. The results demonstrate that the SMOTE GRU algorithm provides high performance in classifying ECG segments, while the BiLSTM ROS algorithm provides high performance in classifying full ECG signals. The Accuracy, Loss, Recall, Precision, F-score values are respectively 70.31% and 77.73%, 0.29 and 0.41, 90.1% and 96.0%, 78.8% and 83.4%, 88.5% and 95.3%.