Automated Atrial Fibrillation Source Detection Using Shallow Convolutional Neural Networks

Isac do Nascimento Lira1, Pedro Marinho Ramos de Oliveira2, Walter Freitas Jr.1, Vicente Zarzoso2
1Universidade Federal do Ceará, 2Université Côte d'Azur


Atrial Fibrillation (AF) is the most frequent sustained arrhythmia diagnosed in clinical practice, understanding its ectrophysiological mechanisms which requires a precise analysis of the atrial activity (AA) signal in ECG recordings. Over the years, signal processing methods have helped cardiologists in this task by noninvasively extracting the AA from the ECG, which can be carried out using blind source separation (BSS) methods. However, the robust automated selection of the AA source among the other sources is still an open issue. Recently, deep learning architectures like the CNN (Convolutional Neural Network) have gained attention mainly by their power of automatically extract complex features from signals and classifying them. In this scenario, the present work proposes to train a shallow CNN model to automatically detect AA sources without the need of hands-on feature extraction steps. The proposed model requires less parameters than the well established CNN architectures and also less data in the training phase. Consequently the training time is considerably lower. A tensor-based BSS technique is applied to 116 random segments of 58 12-lead ECG recordings from 58 persistent patients, generating 509 sources that are visually labeled as AA, ventricular activity (VA) and Unknown sources (UNK). These segments have around 1 second of duration and a binary classification problem is set to classify AA sources and non-AA sources. In order to increase the amount of training samples, an augmentation technique is applied in these signal sources. The shallow CNN architecture parameters (i.e, kernel size) are selected by a Bayesian optimization algorithm, and the best model is built by a composition of convolutional, max-pooling, batch normalization, dropout and fully connected layers trained over 33 epochs. 96.3% of area under the curve and 93.6% of accuracy is achieved from a 10-fold cross validation, overcoming the performance of all the methods present in literature.