ECG AND RELATED WORK

Electrocardiogram (ECG or EKG), proposed in 1901 by William Einthoven as medical diagnostic tool, is a quasi-periodic, non-stationary signal of approximately 100 Hz, which represents the electrical activity of heart. The backbone of cardiovascular system, ensuring supply of nutrition and oxygen through blood, is human heart. It is a muscle roughly of the size of human fist comprising of four chambers, two atria for collection of blood and two ventricles for pumping out blood. From engineering point of view, it can be considered as a combination of mechanical, hydraulic, and electrical sub-systems.

It is basically a two-stage pump with its pumping action initiated by an electrical stimulus provided by the sinoatrial (SA) node (natural pacemaker of heart). The pulse produced by the SA node results in contraction of atria. The action potential generated propagates through atria, and on reaching atrioventricular (AV) node is delayed before being transmitted to the ventricles. This entire process known as cardiac cycle is composed of relaxation of ventricles for filling of blood (diastole) and contraction of ventricles for pumping the blood out of heart to pulmonary artery and aorta (systole). The contraction of so many cells at one time creates a mass electrical signal that can be detected by electrodes placed on the surface of person’s chest or his extremities. ECG or EKG is a graphic recording or display of these time-varying voltages produced by heart during the cardiac cycle. Depending upon placement of electrodes, 12 lead configurations can be obtained, three limb leads (Lead I, Lead II, and Lead III), three augmented leads (avR, avL, and avF), and six chest leads (V,, V₂, V₃, V₄, V₅, and V₆). A typical ECG waveform is depicted in Figure 6.2.

An ECG waveform consists of P-QRS-T waves; a small U wave may also be sometimes present. Each of these characteristic points in ECG is related to electrical activity in human heart during one cardiac cycle. The P-wave represents the atrial depolarisation, resulting in response to SA node triggering; PR interval indicates the AV node delay; the QRS complex characterizes the depolarisation of the ventricles, and finally the ventricular repolarisation is depicted by the T-wave.

FIGURE 6.2 A typical ECG waveform showing various characteristic points.

6.2.1 Advantages of ECG Biometric

ECG has been extensively used for detection of various abnormalities of the heart. In most of these diagnoses, the time duration and the amplitude of the various waves are used. As stated above, primarily ECG has been employed for detection of cardiovascular disorders but recently there has been an upsurge in activity related to its use for person recognition. The inter individual variability in ECG is attributed to the differences in position, size, and anatomy of the heart, age, sex relative body weight, chest configuration, and various other factors. Figure 6.3 shows that the ECG samples were collected from four different subjects, and the distinction in the ECG signal waveforms in these cases can be inferred easily by even visual inspection [2]. The unique and distinct advantages offered by ECG have made it a popular choice among biometric researchers. The qualities possessed by ECG include (i) inherent liveness property integrated w'ith its universal nature, (ii) the ease of combining it with existing and established biometric modalities, (iii) difficult to copy and

FIGURE 6.3 ECG waveform of four different subjects [40].

manipulate, and (iv) application of ECG for medical domain, e.g., person identification in telemedicine and patient monitoring or automatic drug delivery. Therefore, using ECG together with stable biometrics like fingerprint and iris in a multimodal framework ensures fulfilment of majority of the requirements of an ideal biometric system. The combined system due to complementary features of individual traits will be universal, stable, immune to replay attack, and liveness assured. In order to assess the ability of ECG to distinguish individual under varied conditions, a number of studies have been carried out. Some of these have been discussed in the following sub-section.

6.2.2 Literature Review

ECG-based biometric approaches can be divided into three main categories: (i) Time interval and amplitude-based approaches, (ii) Transform domain approaches, and (iii) model-based approaches. Some researchers consider only two main categories: fiducial-based techniques and other two grouped together under non-fiducial-based techniques.

Studies related to exploring the utility of ECG are not altogether new; in fact around 1977, George Forsen and his fellow researchers had recognised the potential of ECG (modified ECG named “C-trace”) as a personal authentication modality. Then, in 1997, M. Ogawa et al. presented first practical work in this direction, when they used ECG for personal identification system in their home health monitoring system. They used features extracted with Daubechies 4 Wavelet and a three-layer neural network as a classifier for distinguishing between two subjects [3]. However, it was Lena Beil and her colleagues whose work provided the much needed boost and resulted in an upsurge in research activity in this area. They investigated the use of ECG as a biometric measure and tested their algorithm on 20 subjects. A statistical model based on time intervals and amplitudes of characteristic points (PQRST) was constructed for each person. A total of 30 features were extracted which were reduced to 12 after dropping the correlated features which were used to perform the identification tasks [4,5]. T. W. Shen used Lead I and modified Lead I (with electrodes placed on palms) for ECG-based biometric using time domain features with DBNN (Decision-Based Neural Network) for identification [6,7]. R. Palaniappan and S. M. Krishnan used three amplitude and two time interval values extracted from the ECG waveform along with the form factor of the QRS complex as the feature set for their study. Two neural network architectures - multilayer neural network with one hidden layer trained using backpropagation algorithm and simplified Fuzzy ARTMAP were used as classifiers, with better results reported for multilayer perceptron (MLP) [8]. To tackle the problem of varying heart rate, Saechia et al. used a normalisation procedure by assuming heart rate to be 80 times/minute and mapping all ECG waves to it by using scaling factor and reconstruction procedure. Fourier transform of the ECG waveform of one period and also of the three sub-sequences, P-wave, QRS complex, and T-wave was obtained. Out of these significant coefficients of complete heart beat and three sub-sequences were fed to neural network trained with back- propagation algorithms for training. The results obtained depicted that the neural network trained with three sub-sequences performed better than the network trained with Fourier coefficients from one period of ECG [9]. In order to alleviate the need of detection of characteristic points, Konstantinos and his colleagues used a non- fiducial feature extraction scheme for ECG-based authentication. In order to capture the similarity across multiple ECG cycles, the authors proposed the use of normalised autocorrelation for different lags say ‘M’ such that M « N where “N” is the signal length chosen. The choice of the signal length is restricted by the fact that it should contain multiple cycles. Discrete Cosine Transform (DCT) coefficients of normalised autocorrelation were used as features [10-12]. In 2008, Yao and Wanput forward a wavelet-based scheme in which ECG cycles between two consecutive R peaks were selected, and wavelet features were obtained by decomposing the extracted cycles using “biorl .1” wavelet at scale 6. Principal Component Analysis (PCA) was used for classification [13]. In order to improve the classification performance, a three-layer neural network with feedforward architecture (512-64-128-1) was used. Input vector to the neural network comprised wavelet coefficient structures obtained from two different cycles either from same or different subjects. The output of network was +1 when two wavelet coefficient structures were obtained from the same subject and -1 when they were obtained from different subjects [14]. Anthony Kaveh and Wayne Chung extracted 34-dimensional feature set made up of ECG time intervals and morphological information. The features were reduced to 13 by employing PCA for dimension reduction. A multi-class support vector machine (SVM) built on “one versus all” was used for classification [15]. In 2009, S. C. Fang and H.L. Chan proposed a method of identifying individuals in phase space [16]. For this, a three- dimensional (3D) vector was built, with the averaged ECG signal as the first dimension and two other formed using a time delay т each from the averaged signal and second component. For three leads, 3D feature vector was built from the anterior, lateral, and posterior leads after averaging each of the three leads individually as win case of a single lead. The comparison of phase portraits was carried out using spatial correlation and mutual nearest point distance. Authors also compared their work with ECG- based identification using fiducial features. Neural networks MLP, radial Basis function (RBF), and Euclidean distance were used as classifiers. The results obtained showed better performance for three leads rather than single lead ECG. Moreover, the mutual nearest point distance was better than the spatial correlation. Further results were better when only QRS complex was used rather than the complete wave. Can Ye et al. [17] employed wavelet features extracted using Daubechies wavelets clubbed with ICA (Independent Component Analysis) features extracted from two leads of ECG to form the feature vector. SVM was used for classification. Tests conducted on three public dataset showed that good identification accuracy can be achieved when information was combined using decision level fusion for the two lead ECGs. Ming Li and Shrikant Narayan came out with a technique in which two feature sets were extracted from the ECG signal. In one of them, coefficients extracted from Hermite polynomial expansion were modelled using SVM while in the second approach, cepstral features of ECG were used to model subjects using GMM (Gaussian Mixture Models) supervector to get a score. Finally, score level fusion was carried out to get the combined score. Highest accuracy was obtained for the combined scheme [18]. Extended Kalman Filter has also been applied for the ECG biometric task [19]. QRS complex is relatively stable and undergoes very little variation with a change in the heart rate. Recognising this, Tawfik and Kamal carried out normalisation of only the QT segment. DCT features were extracted from normalised QT segment and QRS complex and were fed to separate the feedforward neural network with one hidden layer for classification. Best results were obtained using features from only the QRS complex [20]. Vu Mai and his colleagues also used the QRS complex values as the feature vector and with two neural network architectures; a 2-layer MLP and RBF as classifiers [21]. Taking into note the stability of QRS complex with reference to varying heart rate, L. Hou and his fellow researchers generated spectral coefficients based on Fourier transform and DCT of the QRS complex. The dimension reduction was further carried out using PCA. A two-stage classification was carried out with feature matching as the first check and in case multiple subjects are identified, the neural network was used at the second stage [22]. Loong and his fellow researchers put forward a non-fiducial approach using 40 coefficients of linear prediction coefficients (LPC) spectrum obtained from 5 seconds ECG signal with 50% overlap and multilayer neural network with two hidden layers each having 100 nodes as the classifier. One more scheme using 50 features from wavelet packet decomposition coefficients was also investigated. Experiments conducted by them showed superior performance for the LPC method [23]. In 2010, Nahid Ghofrani and Reza Bostani studied performance of six non-conventional features for ECG biometric application, namely, autoregressive (AR) coefficients, power spectral density (PSD), Lyapunov exponent, approximation entropy, Higuchi fractal dimension and Shannon entropy. For classification, К-nearest neighbours (KNN) and two neural networks multi-layer perceptron and RBF were used. Five-dimensional feature vector formed by combining four coefficients from the fourth order AR model with mean PSD of every window, provided the best results when KNN was used as the classifier [24]. Ikenna Odinaka et al. presented an extensive study of a frequency- based method on a fairly large dataset comprising 5 minute recordings from 269 subjects with multiple session recordings. They computed spectrogram of each ECG pulse by calculating the short-time Fourier transform of the windowed sequence (Flamming window) and finding the logarithm of square of its magnitude. For classification, maximum likelihood-based model using normal distribution for feature vectors was built. Features selection was carried out based on the relative entropy approach. For verification, log likelihood ratio (LLR) was computed, and claim was accepted if its value was greater than a pre-determined threshold. Identification was based on the highest LLR value. The authors reported considerably good performance even when the suggested approach was tested on across sessions recordings [25]. In 2011, Ching-Kun Chen and his fellow workers used non-traditional sixdimensional feature vector made up of correlation dimension, four Lyapunov coefficients, and root-mean-square value and fed it to the multilayer network with two hidden layers (20 and 5 nodes). A recognition rate of 90% was achieved for database of nine subjects [26]. Andre Lourenco, Hugo Silva, Ana Fred, and their co-workers have been trying to develop a finger-based ECG biometric system. They have used fiducial features and template features with Euclidean distance, SVM, KNN, etc. as classifiers. The results reported by authors are quite encouraging [27,28]. A shortterm dataset (65 subjects) and long-term dataset (63 subjects) for testing of ECG biometric methods have also been prepared by the authors [29]. Tantawi et al.

compared the performance of 36-dimensional feature vector comprising majority of fiducial features mentioned in the literature and a subset of 23features derived from the five prominent characteristic points (P, Q, R, S, T) with RBF as the classifier. It was found that the system was able to recognise the individual even with reduced features although with slightly less accuracy [30]. In 2013, Tantawi and his fellow researchers derived level six wavelet features from R-R interval of the ECG waveform by utilising bior 2.6 wavelet. The RBF neural network was employed as the classifier. Experiments were also conducted by feeding RBF with features extracted from the AC/DCT method, QRS complex, and QT interval. Testing of the proposed approach depicted high subject identification and window recognition rate. However, results were not encouraging by using features extracted from the QRS complex, QT interval, and AC/DCT method [31]. In 2014, Eduardo J. S. Luz and his colleagues have studied the viability of the biometric application of ECG signals sampled at a low frequency. For this, they used four set of features, mainly wavelet features and SVM as the classifier and found that signals obtained at low frequencies can be successfully used [32]. In the last couple of years, R. D. Labati and fellow researchers have suggested a scheme named as Deep-ECG which uses convolutional neural network (CNN) for feature extraction and perform identification by score-based matching [33]. Some authors have transformed al-D ECG into a 2-D representation and then used CNN for subject identification [34,35]. In Ref. [36], authors have used a two-dimensional matrix created from QRS segments to obtain the model via CNN for personal identification. The experiments conducted by authors in majority of the works have been conducted on small population datasets generally less than 50. However, in Ref. [37], extensive survey of ECG-based biometric approaches has been presented, and then the comparative study of around 19 non-fiducial algorithms was carried out for authentication applications with experiments conducted on their in- house dataset of 265 individual having multiple session recordings having two weeks to six months gap between recordings. The performance of all algorithms dipped with an increase in duration between testing and training sessions, although some improvement was achieved by using multiple session recordings for training. More importantly, only three of the eighteen algorithms resulted in EER less than 10% for testing across multiple sessions. Overview of ECG biometric with challenges and opportunities related to this biometric trait has also been discussed in Ref. [38,39]. The main issues faced by ECG biometric are related to the population size of datasets used in experiments and variability of ECG with time, stress, and diseased state. Moreover, the user friendliness and convenience of the subjects is another aspect that demands the attention of the biometric researchers. For this, studies need to be carried on a datasets comprising of sufficient number of subjects and reasonable gap between recordings employed for training and testing the developed algorithms. Keeping this in mind, an approach for ECG-based human identification has been explored henceforth in this chapter.