The objective of biometrics is to identify subjects based on physiological or behavioral characteristics. This paper considers the spatial P-QRS-T loops of the vectorcardiogram (VCG),
IntroductionThe objective of biometrics is to identify subjects based on physiological or behavioural characteristics, such as fingerprint, iris, face, voice, which however could easily be mimicked via fake finger, iris, face photos, playback, do not provide liveness detection [1] and are a topic for discussion on privacy protection [2]. The electrocardiogram (ECG) has been investigated as an advanced signal for human biometrics, presenting vital signs. The human verification or identification solutions employ a single ECG lead [3,4], limb leads [5] or standard 12-lead ECG [6,7], based on temporal and amplitude ECG features [6], cross-correlation analysis [5,7], PQRST pattern matching [3,4].The spatial features of the cardiac vector represented by the vectorcardiogram (VCG) are expected to be useful for biometric applications, considering the inter-subject differences of the VCG loop orientation and shape, and its independence from the heart rate [8]. However, we could find a few studies based on VCG biometrics, all of them solving the human identification task, using:-Support vector machine classifier, applied over QRSand T-loop features derived via inverse Dower transform [9] or pseudo-inverse transform, including only the limb leads [10]; -Neural networks classifier, applied over equal distance descriptor coefficients or Fourier descriptor coefficients of the QRS-loop constructed by plotting the QRS in lead I (x-axis) against lead aVF (y-axis), i.e. the QRS-loop projection in the vertical plane [8]. This paper considers the spatial P-QRS-T loops of the VCG, aiming to identify the most reliable VCG-based features for human verification. Presuming that different techniques for transformation of 12-lead ECG to VCG [11][12][13] and P-QRS-T loops projections in the Frontal, Horizontal and Sagittal planes [14] have specific diagnostic significance, we aim to compare their effect on human verification.
ECG databaseThe study is using a proprietary clinical ECG database (Schiller AG, Switzerland), which contains two 10s-sessions of standard 12-lead resting ECGs from 460 noncardiac patients (235/225 male/female, 18-106 years old), admitted in the emergency department of the University Hospital Basel during the period (2004)(2005)(2006)(2007)(2008)(2009). The ECGs are recorded via the commercial ECG device SCHILLER (500Hz, 2.5µV/LSB, bandwidth 0.05-150Hz) at distant time sessions S1 and S2>S1+1year. The person verification scheme for comparison of subjects between S1 and S2 gives N=460 pairs with equal identity (ID) and N*(N-1)=211140 pairs with different ID. Our approach to handle the imbalance ratio (459:1) of different-to-equal ID pairs considers two independent datasets: -Training dataset: 230/230 ECG pairs of equal/ different IDs, presuming that the verification classifier should be trained on the first half of subjects using...