A Comparative Analysis of 2D Deep CNN Models for Arrhythmia Detection Using STFT-Based Long Duration ECG Spectrogram

Cardiac arrhythmia detection exploiting electrocardiogram (ECG) signal has gradually become mature in the recent decade. Recent advancement in the deep learning (DL) area has accelerated its extensive application in this area of healthcare research. Especially, convolutional neural network (CNN) based architecture has become more popular and has drawn a lot of research attention for the precise detection of cardiac arrhythmia utilizing single or few QRS complexes or beats of ECG signals. In this article, we have investigated the performance of six CNN based classifiers for cardiac arrhythmia (15 classes) detection based on the spectrogram of the long duration ECG signals. Six pre-trained state-of-the-art CNN based models, i.e. VGG-16, ResNet-50, Inception, MobileNet, DenseNet, and EfficientNet have been used as feature extractors. At the beginning of this investigation, short term fourier transform has been used to extract time-frequency domain information’s from the long duration ECG signals as well as generate two dimensional spectrogram images to be fed into these CNN based architectures as input. However, the overall performances of the six ECG rhythm classifiers in terms of accuracy, precision, recall, and F1 score, are also evaluated and investigated explicitly to demonstrate the comparative analysis. In addition, the experiment has been conducted varying the learning rate during training of the classifiers to study the impact of the learning rates on discriminative feature learning.