Fold Electrocardiogram Into a Fingerprint

Po-Ya Hsu, Po-Han Hsu, Hsin-Li Liu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 828-829

Electrocardiogram (ECG) has become a popular biometric to study since it is highly secured against spoofing attack. In this study, we address the issues of hard-required ECG data and neglected causality in performing ECG identity matching tasks. First, we propose an ECG image generation algorithm that is able to handle any specified number of ECG heartbeats. Such an algorithm uses detected R-peaks as folding points and projects ECG data onto a two-dimensional image, which overcomes the challenge of hardly-required fixed length and truncated ECG. Second, we perform across-session testing. We construct the ECG identification models by using the past ECG data and evaluate their performance on future ECG data. Furthermore, we develop a voting strategy that is able to detect anomaly ECG heartbeats. Our novel ECG image generation approach shows to be a competitive ECG biometric model by leveraging transfer learning method. Such method has been evaluated on MIT-DB and ECG-ID datasets. We observe satisfiable results of the proposed models in both datasets: 100% on the MIT-DB and 94.4% on ECG-ID. More importantly, our method is available to generate satisfying results by using a single ECG beat to conduct identity matching task: 100% on the MIT-DB and 91.7% on ECG-ID. In addition, qualitative analysis presents the perceptual uniqueness of ECG between individuals. We believe that the proposed ECG biometric system is promising to identify humans with short ECG sequence.