Single beat ECG-based Identification System: development and robustness test in different working conditions

Abstract

One-lead electrocardiogram (ECG) tracings have already shown to be a good candidate as a feature for a biometric identification system. Also, the reduced computational burden and the fact that it can ensure that the subject is alive put the ECG ahead of currently used biometric features. Most of the literature provides studies exploiting acquisitions made with clinical instrumentation, preceded by invasive preparation of the subject, in a structured environment and with the subjects at rest. These conditions are not very feasible for an application in a real-world context. Therefore, we are proposing a system that is performant with acquisitions collected with (non-invasive) non-clinical instrumentation and in an unstructured environment, and that is robust to variations of the psychophysical state of the subjects (i.e. at rest or under mental or physical stress). To do so, we developed an acquisition protocol that we followed to collect a new dataset to evaluate our method. The proposed system achieved up to the 97% of single segments (beats) classification accuracy when the test segments come from the same kind of acquisition procedure of the training beats. The same result was obtained by training and testing by combining the three trials. An 88% and 68% of accuracy were achieved by testing the system under mental and physical stress conditions, respectively, while trained at the rest state. Our findings suggest that the proposed method may be put at the base of a future application in a real-world context.