Deep learning feature representation for electrocardiogram identification

Abstract

This paper presents an efficient and powerful method of electrocardiogram (ECG) identification. Specially, a set of discriminative feature representations can be learned from one-dimensional ECG signals with an arbitrary origin through deep learning, referred to as deep fusion features. Non-linear classifier is adopted to classify test ECG signals. A final simple voting step can further enhance performance. Based on the above steps, our method can reduce the dependence of algorithm accuracy on the origin and length of the ECG signal. Unlike traditional methods, detecting fiducial points and combining features are not required. Moreover, test process can use parallel processing to improve efficiency. The method achieves 99.33% accuracy for a publicly available database. The experiments demonstrate that our method is efficient and powerful in real applications.