Enhancing Electrocardiogram Identity Recognition Using Convolutional Neural Networks With a Multisimilarity Loss Model

Identity recognition plays an important role in human civilization. With people’s further cognition of physiological signals, research on electrocardiogram (ECG)-based identity recognition is gradually carried out at home and abroad. At present, the accuracy of the ECG-based identity recognition is not high, and the generalization ability of the model is weak; to solve this problem, we propose a solution to identity recognition based on multisimilarity loss (MS Loss). First, we filter the signal from ECG identity (ECG-ID) database and feed it into the designed convolutional neural networks (CNNs) for training. After that, the extracted embeddings are then trained with K-nearest neighbor (KNN), support vector machines (SVMs), and Siamese neural networks (Siamese NNs) for recognition. To verify the effectiveness of the embeddings obtained by CNN, we also compared the MS Loss method with the direct training method. To verify the generalization ability of the mode, we collected ECG data from 20 members of the laboratory and established self-built database. Results show that the proposed model has advanced generalization ability than the state-of-the-art works and achieves 97.13% accuracy in the 90-person-scale database and 99.86% recognition accuracy in the 20-person-scale self-built database. This article uses a deep CNN to extract features via MS Loss from ECG signals and realizes quickly distinguishing different individuals, which provides a feasible and natural method for small-scale identification and reduces the difficulty of engineering generalization.