Identification of Atrial Fibrillation from Electrocardiogram Signals Based on Deep Neural Network

As the most common continuous arrhythmia, atrial fibrillation is related to significant morbidity and death rate. Furthermore, the automatic detection of atrial fibrillation remains a challenging work all the time. This paper develops a deep neural network for automatic atrial fibrillation identification based on stacked sparse autoencoder and softmax layer. Nineteen features are extracted from each electrocardiogram sample through Hilbert-Huang transform, wavelet decomposition and statistical measures. Stacked sparse autoencoder including two layer sparse autoencoder is employed to learn advanced features. The softmax classification layer is connected to the top of the stacked sparse autoencoder, in order to map the advanced features into the classes of the electrocardiogram samples. Experimental results illustrate that compared with extreme learning machine and support vector machine, the identification performance of the deep neural network is better and its value of accuracy, sensitivity, specificity, precision of atrial fibrillation reach 96.00%, 90.00%, 98.00%, 93.75%, respectively. Compared with some super existing methods, the deep neural network has the better performance. Therefore, the deep neural network proposed in this task could be effective in the automatic detection of atrial fibrillation.