Atrial fibrillation detection and ECG classification based on convolutional recurrent neural network

The aim of the 2017 PhysioNet/CinC Challenge [1] is to classify short ECG signals (between 30 seconds and 60 seconds length), as Normal sinus rhythm (N), Atrial Fibrillation (AF), an alternative rhythm (O), or as too noisy to be classified. Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) as classifiers have recently shown improved performances compared to methods established in various sound recognition tasks [2] and interesting result in tasks such as the 2016 Physionet Challenge for the classification of heart sound [3]. Our approach is based on a convolutional recurrent neural network (CRNN), involving two independent CNNs, to extract relevant patterns, one from the ECG and the other from the heart rate, which are then merged into a RNN accounting for the sequence of the extracted patterns. The final decision is then evaluated through a Support Vector Machine (SVM).