Atrial Fibrillation (AF) Detection Using Deep Learning with GAN-based Data Augmentation

Atrial Fibrillation (AF) is the most common cardiac arrhythmia that may lead to stroke and heart failure and because of this, a lot of research has gone into detecting AF from the electrocardiogram (ECG) signal. In this paper, we propose an AF detection pipeline, which first transforms the ECG data into an informative 2-D image by the aid Poincaré recurrence plot, and then addresses the imbalance of the data by augmenting AF samples using a form of generative adversarial network (GAN). The augmented dataset is then used within the training set to train a 5-layer convolutional neural network (CNN) as a classifier. The performance of the proposed classifier is finally evaluated based on a 4-fold cross-validation scheme. The performance metrics suggest that the proposed method provides acceptable sensitivity and specificity.