Automatic Identification of Arrhythmia from ECG Using AlexNet Convolutional Neural Network

Atrial fibrillation (AF) is one of the most widespread cardiovascular diseases and impacts the overall general population of the world. Most of the current techniques are based on hand-crafted features for automatic AF classification. The primary task of this work is to design a deep learning-based approach that will eliminate the necessity of manual identification of features. We have used a pre-trained convolutional neural network (CNN), namely AlexNet, to train using 5,655 single-lead ECG recordings. Initially, we have extracted a spectrogram for all 30s signals and converted them to RGB images with Continuous Wavelet Transform (CWT); later fed to transferred AlexNet and trained with some changes in specifications. The findings of the study indicate that our technique attains a state-of-the-art accuracy of 97.9% and an F1 score of 98.82% while having higher overall sensitivity (98.9%) and specificity (90.7%) and outperformed all existing methods.