Atrial fibrillation detection using convolutional neural networks on 2-dimensional representation of ECG signal

Abstract

Atrial fibrillation (AFIB), heart condition associated with increased risk of stroke, dementia, heart failure, and mortality, is the most common cardiac arrhythmia. The problem of diagnosis is very often paroxysmal in the first stages of the disease. These challenges are met by solutions using machine learning algorithms that aim to maximize the quality of AFIB detection and minimize the cost. In this study, I compared 3 approaches of using 2D representation of ECG signals as an input to Convolutional Neural Network (CNN), which are known to be the most suitable for image classification. Spectrogram, scalogram, and attractor reconstruction (AR) are used for AFIB detection within 5s windows of raw ECG signal. Such approaches seem to be the perfect way of shortening the time of signal processing, which includes most often steps like filtering, defining detection function, peak finding, and feature computing in most similar systems. Furthermore, this work allows to verify the AR method for AFIB detection, so far successfully used in the analysis of the ECG signal, in terms of gender identification. Sensitivity of 94% (scalogram), 95% (spectrogram), 90% (AR), and the F1 score of 94% (scalogram), 93% (spectrogram) and 89% (AR) were achieved. The comparison of these methods is important in the context of searching for highly effective AFIB detection methods over very short time intervals without the need for signal preprocessing.