Robust ECG signal classification for detection of atrial fibrillation using a novel neural network

Electrocardiograms (ECG) provide a non-invasive approach for clinical diagnosis in patients with cardiac problems, particularly atrial fibrillation (AF). Robust, automatic AF detection in clinics remains challenging. Deep learning has emerged as an effective tool for handling complex data analysis with minimal pre- and post-processing. A 16-layer 1D Convolutional Neural Network (CNN) was designed to classify the ECGs including AF. One of the key advances of the proposed CNN was that skip connections were employed to enhance the rate of information transfer throughout the network by connecting layers earlier in the network with layers later in the network. Skip connections led to a significant increase in the feature learning capabilities of the CNN as well as speeding up the training time. For comparisons, we also have implemented recurrent neural networks (RNN) and spectrogram learning. The CNN was trained on 8,528 ECGs and tested on 3,685 ECGs ranging from 9 to 60 seconds in length. The proposed 16-layer CNN outperformed RNNs and spectrogram learning. The training of the CNN took 2 hours on a Titan XPascal GPU (NVidia) with 3840 cores. The testing accuracy for the CNN was 82% and the runtime was ~0.01 seconds for each signal classification. Particularly, the proposed CNN identified normal rhythm, AF and other rhythms with an accuracy of 90%, 82% and 75% respectively. We have demonstrated a novel CNN with skip connections to perform efficient, automatic ECG signal classification that could potentially aid robust patient diagnosis in real time.