Classification of heart diseases using time-frequency representations of electrocardiogram signals by transfer learning networks

Abstract

It is crucial to monitor and diagnose cardiac function early to prevent the development of future, more severe issues. This study categorized 193 male and female subjects into three groups based on their ECG signals obtained during an exercise test: healthy, myocardial infarction, and left bundle branch block. The data were then processed and converted into images representing three time-frequency representations: a spectrogram, a scalogram, and a spectrum. These images were used as input for two pre-trained networks through transfer learning. The ResNet-18 and GoogLeNet networks were utilized in this study. The ResNet-18 network achieved an accuracy of 88.64% for the spectrogram, 98.41% for the scalogram, and 83.33% of the spectrum. The results for the GoogLeNet network were as follows: 77.27% for the spectrogram, 97.62% for the scalogram, and 78.57% of the spectrum.