Single-Lead ECG Cross-Session Identification Based on Conditional Domain Adversarial Network

Biometric human identification systems have been mainly implemented based on fingerprint, face, iris, and voice recognition. However, counterfeits generated from deep-learning technologies make such systems more and more vulnerable. On the other hand, the electrocardiogram (ECG) signal, which can only be measured from a living body, provides a secure alternative for identity authentication. For an ECG identification system, the most difficult challenge is to face heart rate variability caused by different physiological states and long-term cardiac states. In other words, the system must have cross-session generalization ability to identify ECG signals recorded in different periods of time. In this article, we propose a robust ECG identification model using a single heartbeat recorded from lead-I by treating the cross-session identification task as a cross-domain task. The proposed model is referred to as the conditional domain adversarial neural network for cross-session ECG signals (CDAN-CS), which combines the temporal convolutional neural network (TCN) and the cross-domain model of conditional domain adversarial network with entropy (CDAN-E). Averaged over experimental results on three databases, the proposed model achieves 100% accuracy and ${F}1$ -score for ECG signals within the same session and 99.76% accuracy and 90.5% ${F}1$ -score for cross-session ECG signals. The averaged ${F}1$ -score of 90.5% is 8.44% higher than the averaged ${F}1$ -score achieved by the baseline TCN model. The robust results from CDAN-CS validate the idea of tackling the cross-session ECG identification task using domain adaptation models.