by saeid wahabi - university of toronto t-space · abstract variability in ecg biometrics: state of...

variability in ecg biometrics: state of the art and subspace methods

by

Saeid Wahabi

A thesis submitted in conformity with the requirementsfor the degree of Master of Applied Science

Graduate Department of Electrical and Computer EngineeringUniversity of Toronto

c© Copyright 2015 by Saeid Wahabi

Abstract

variability in ecg biometrics: state of the art and subspace methods

Saeid Wahabi

Master of Applied Science

Graduate Department of Electrical and Computer Engineering

University of Toronto

2015

This work addresses the challenges of evaluating electrocardiogram (ECG) biometric recognition

systems. While ECG has attracted significant interest, most of the prior art have approached it from

the perspective of a typical biometric modality and have neglected the physiological factors that affect the

respective systems. This work presents the UofT ECG Database (UofTDB) and offers a comprehensive

analysis of the underlying inter-individual variability under a number of conditions such as body posture,

physical activity and time-lapse. The performance of various methodologies is reported under the above

mentioned conditions and a method based on template fusion is proposed to address these shortcomings.

Furthermore this work proposes a ECG-based posture-detection algorithm and state-detection veri-

fication method to overcome the problem of verification under uncontrolled posture settings.

Finally we evaluate different subspace methodologies for ECG recognition under different ECG vari-

ability to establish a baseline performance for future works.

ii

Dedication

To my wonderful wife, Shahrzad.

iii

Acknowledgements

First and foremost, I would like to express my sincere gratitude to my advisor Professor Hatzinakos

for giving me guidance and his continuous support of my master’s study and related research. I could

not have imagined having a better advisor for my master’s studies.

My sincere thanks also goes to Foteini Agrafioti and Siddarth Hari for the stimulating discussions,

their insightful comments and encouragements which helped me to widen my research in many perspec-

tives.

I also would like to thank my thesis committee: Prof. Valaee, Prof. Draper, and Prof. Sarris, for

the challenging questions and their helpful feedback.

Last but not least, I would like to thank my family and friends for their constant encouragement

during my studies.

iv

Contents

1 Introduction 1

1.1 Research Goals and Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 Taxonomy of Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.3 Related Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.4 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Benchmarking ECG biometric testing techniques 7

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2 ECG as a Biometric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2.1 ECG Wave . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.2.2 ECG Recording Standards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.3 Factors affecting the ECG signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.4 Existing Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.5 Literature Review of Testing Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.6 Benchmarking an ECG database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.6.1 The Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.6.2 Acquisition Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 Evaluation of the state of the art ECG biometric methods 20

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3.3 ECG Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.4 Review of the Implemented Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.5 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

v

3.6 Enhancing Biometric Templates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.6.1 State-agnostic verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.6.2 State-detection verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

4 Posture Detection for ECG Biometric Systems 36

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

4.2 Effect of Posture on ECG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

4.3 Literature Review of Posture Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4.3.1 Method of Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4.3.1.1 Signal Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4.3.1.2 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

4.3.1.3 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4.3.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4.3.3 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4.3.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4.4 Posture Detection For ECG Biometrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

5 ECG Recognition in Subspaces 46

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

5.2 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

5.3 Linear Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

5.3.1 Principal Component Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

5.3.1.1 Optimum Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

5.3.1.2 Summary of Principal Component Analysis . . . . . . . . . . . . . . . . . 49

5.3.2 Linear Discriminant Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

5.3.2.1 Small Sample Size Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 52

5.3.3 Independent Component Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

5.3.3.1 ICA Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

5.4 Nonlinear Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

5.4.1 Kernel methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

5.4.1.1 Kernel PCA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

5.4.1.2 Kernel LDA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

vi

5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

6 Empirical Comparison of Subspace Methods 61

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

6.1.1 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

6.1.2 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

6.2 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

6.2.1 Experiment Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

6.2.2 Robust ECG Features for Across Sessions Variability . . . . . . . . . . . . . . . . . 64

6.2.3 PCA-based Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

6.2.4 ICA-based Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

6.2.5 LDA-based Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

6.2.6 Kernel-based Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

6.3 Effectiveness of Template Updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

6.4 Effect of Posture and Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

7 Conclusion and Future Directions 71

7.1 Future Directions and Improvements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

Bibliography 72

vii

Acronyms

AC Autocorrelation

AC/LDA Autocorrelation/Linear Discriminant Analysis

ANS Autonomic Nervous System

DBNN Decision-based Neural Network

DLDA Direct Linear Discriminant Analysis

DWT Discrete Wavelet Transform

ECG Electrocardiogram

EEMD Ensemble Empirical Mode Decomposition

EER Equal Error Rate

FAR False Accept Rate

FRR False Reject Rate

FT Fourier Transform

ICA Independent Component Analysis

IMF Intrinsic Mode Function

IR Identification Rate

K-NN K-Nearest Neighbours

KDLDA Kernel Direct Linear Discriminant Analysis

KLT Karhunen-Loeve Transform

KPCA Kernel Principal Component Analysis

LDA Linear Discriminant Analysis

LLRT Log Likelihood Ratio Test

MLE Maximum Likelihood Estimation

PCA Principal Component Analysis

viii

PTBDB Physikalisch-Technische Bundesanstalt Database

QRS Complex made of Q R and S fiducial points on the ECG heartbeat

ROC Receiver Operating Characteristic

SVD Singular Mode Decomposition

SVM Support Vector Machine

UofTDB University of Toronto ECG Database

WDIST Wavelet Distance Measure

ix

List of Tables

2.1 UofT ECG Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.2 Vernier EKG sensor and USB interface specifications . . . . . . . . . . . . . . . . . . . . . 18

3.1 Equal Error Rate for the four methodologies tested in different body postures and exercise

condition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3.2 State detection rate of the methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4.1 Personalized testing case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4.2 Generic testing case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

x

List of Figures

2.1 Sample Electrocardiogram (ECG) recording of a subject. . . . . . . . . . . . . . . . . . . . 8

2.2 Typical components of an ECG heartbeat. . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.3 The limb leads. Photo by Nicholas Patchett,http : //en.wikipedia.org/wiki/F ile :

EKGlimbleads.png . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.4 Heartbeats of a subject in rest and exercise condition from UofTDB. The T-wave moves

closer to the QRS complex as the heart rate increases. . . . . . . . . . . . . . . . . . . . . 11

2.5 (a) Standing (b) Sitting (c) Tripod and (d) Supine . . . . . . . . . . . . . . . . . . . . . . 17

2.6 Placement of electrodes according to lead I configuration . . . . . . . . . . . . . . . . . . . 18

3.1 (a) Raw, (b) filtered ECG signal and (c) shows extracted and aligned heartbeats of a

subject. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.2 Power Spectral Density of raw ECG signal and the magnitude impulse response of the

fourth order Butterworth bandpass filter with cut-off frequencies at 0.5Hz and 40Hz. . . . 22

3.3 Feature extraction block diagram for discrete wavelet transform approach [17]. . . . . . . 22

3.4 Feature vector γ formed from level 1-5 detail coefficients (D1, D2, D3, D4 and D5) of

the 5-level wavelet decomposition of the input heartbeat. The Daubechies scalar wavelet

(Db3) was used as the wavelet function. DX represents the detail coefficients and AX is

the approximation coefficients at level X. . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.5 Feature extraction block diagram for Time-Frequency content approach [64]. . . . . . . . 24

3.6 Feature extraction block diagram for eigenPulse [43] method. . . . . . . . . . . . . . . . . 25

3.7 The estimated normalized autocorrelation of five different subjects. . . . . . . . . . . . . 26

3.8 Feature extraction block diagram for AC/LDA [3] method. . . . . . . . . . . . . . . . . . 26

3.9 Receiver operating characteristic (ROC) curve for the four methodologies. . . . . . . . . . 27

3.10 Performance evaluation of the four methodologies across different sessions with enrol data

from session 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

xi

3.11 Performance evaluation of the four methodologies across different sessions with enrol data

from session 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.12 Flow chart diagram for state-detection verification for the AC/LDA algorithm. . . . . . . 32

3.13 Equal error rates for the two verification methods for Chan et al. [17] and Odinaka et

al. [64] methods. Enrolment with sit, stand, supine, tripod and exercise state ECG signals. 33

3.14 Equal error rates for the two verification methods for Irvine et al. [43] and Agrafioti et

al. [3] methods. Enrolment with sit, stand, supine, tripod and exercise state ECG signals. 34

4.1 System block diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4.2 Heartbeats of a subject under different postures. . . . . . . . . . . . . . . . . . . . . . . . 39

4.3 Normalized autocorrelations of a subject under different postures. . . . . . . . . . . . . . . 40

4.4 Fiducial features of a subject extracted from its heartbeat. . . . . . . . . . . . . . . . . . . 41

4.5 Fourier transform of a subject under different postures. . . . . . . . . . . . . . . . . . . . 42

4.6 Level three approximation coefficients of discrete wavelet transform of a subject’s heart-

beat under different postures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

4.7 System block diagram of state-detection verification system. . . . . . . . . . . . . . . . . 44

4.8 ROC curves of state-detection verification system under different test postures. . . . . . . 45

5.1 PCA of a multivariate Gaussian distribution centred at (1.01,1.01) with variance of 3.95

in direction of (0.705,0.709) and variance of 0.1 in the orthogonal direction. . . . . . . . . 50

5.2 Projection vectors of PCA and LDA for the case of samples from two classes. PCA vector

is in the direction of maximum overall variance in the data whereas LDA finds a direction

so that the class separability is maximized. . . . . . . . . . . . . . . . . . . . . . . . . . . 51

5.3 A principal curve represents the underlying structure of data. . . . . . . . . . . . . . . . . 56

5.4 An auto-associative neural network for computing principal manifolds. Number of ele-

ments in the bottleneck represent the dimensionality of the principal curve. . . . . . . . . 57

6.1 The ECG verification accuracy under across-session analysis when different waves are used. 65

6.2 The accuracy of ECG biometrics when different number of principal components are

used. The top 20 principal components capture approximately 72% of the variance from

the original space. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

6.3 The ROC of different subspace methods under across-session analysis. . . . . . . . . . . 68

6.4 The verification performance of subspace methods with template updating. . . . . . . . . 69

6.5 The identification performance of subspace methods with template updating. . . . . . . . 69

xii

6.6 Performance summary of DLDA and PCA subspace methods. . . . . . . . . . . . . . . . . 70

xiii

Chapter 1

Introduction

Biometrics refer to physiological and behavioural characteristics that are unique to an individual which

can be used to establish their identity. Physiological characteristics relate to physical properties of the

human body while behavioural characteristics relate to adopted behaviours and idiosyncrasies that are

not attributed to genetic factors. Some examples of physiological biometrics include the fingerprints, the

facial structure, hand geometry, DNA, retina, and the iris while examples of behavioural characteristics

are the voice, gait, signature, and keystroke dynamics. Unlike the traditional means of authentication,

such as passwords or ID cards, biometric systems offer a strong connection between an individual and

his/her identity. This link is the primary reason why biometric systems have drawn significant attention

in physical and logical access control applications.

A biometric system is essentially a pattern recognition system that collects biometric data from an

individual, designs a biometric template, and compares this signature against other templates in the

database.

There are three modes of operations for a biometric system: 1) enrolment, 2) verification and 3)

identification. During the enrolment mode of operation the system extracts a biometric template from

users and saves the template into a database. Depending on the use case of the system, it can be

operating in verification or identification mode. In verification mode of operation the users claim an

identity and the system should either accept or reject the claim. This means that the system needs to

decide between two hypothesis: 1) it is a genuine claim and the person is who she claims to be or 2) it is

an impostor and the person is not who she claims to be. During the identification mode of operation, the

system searches the database in order to find an identity match for the subject. In this case the system

should find the cross match with all the templates in the database in order to find the best match.

1

Chapter 1. Introduction 2

Not all attributes of the human body can be used for biometric recognition. For a signal to qualify

for biometric use the following criteria must be met:

• Universality - the signal must be collectable from all human subjects.

• Uniqueness - the signal must be in position to distinguish different individuals.

• Permanence - the signal should not be altered extensively over time.

• Robustness to attacks - one must not be able to easily imitate or steal this signal.

While, most biometric signals satisfy the above criteria, it is impractical to define the golden bio-

metric characteristic i.e., the best biometric modality in existence. This is because in the spectrum of

physiological and behavioural biometrics certain modalities may better meet the needs of an application

environment than other. Furthermore, the ease-of-acquisition and cost of deployment are additional

factors that are typically considering when evaluating biometric modalities.

The electrocardiogram (ECG) signal is among the younger additions to the biometrics family. The

ECG reflects the electrical activity of the heart over time. Relative to the above criteria, the ECG is

a vital signal, and as such it can be recorded from every living individual (universality). Furthermore,

for healthy individuals, the ECG waveform does not change significantly over time and it maintains

characteristics that have been shown to be unique to every individual [60] [92] (permanence). From

a security point of view, ECG guarantees some level of robustness since it is difficult for an impostor

to acquire someones ECG without their consent (robustness to attacks). The presence of the signal

automatically guarantees liveness of the sensor reading and it cannot be mimicked or obfuscated like

other modalities such as gait and keystroke.

While the use of the ECG in biometric recognition is relatively new, the same signal has been

extensively studied for medical diagnostics. The ECG is directly linked to the geometrical and elec-

trophysiological properties of the myocardium [34]. Unique ECG morphologies are the result of unique

muscular structures in the myocardium. In fact, the medical community has been aware of the inter-

individual differences of the ECG and concentrated efforts have been made to eliminate this variability

in order to establish universal diagnostic standards [25].


1.1 Research Goals and Contributions

Unlike most biometric characteristics, ECG is time-dependent and naturally affected by the physical

and psychological activity of the human body. This indeed presents a challenge for ECG biometric

deployment and measures have to be taken to ensure that ECG-enabled biometric systems are robust

to such variations. However these variations have not been sufficiently addressed by the ECG biometric

community.

In this work we first look into several factors that can affect the ECG signal and consequently the

performance of the ECG biometric systems. Then we introduce a new ECG Biometric database that

has recordings under those factors. With the new database we systematically evaluate the performance

of different existing ECG biometric methodologies.

Based on the reported results, we propose propose a new template updating method to improve the

performance of ECG biometric systems in various settings. Further we explore and propose a posture-

detection algorithm based on the ECG signal for ECG biometric systems in order to improve the systems’

accuracy.

Finally we explored different subspace methods with the goal of finding the best subspace for repre-

senting the ECG signal. The performance of different methods (PCA, KPCA, ICA, LDA, DLDA and

KDLDA ) were evaluated in the context of ECG recognition.

The following is the summary of contributions of this work:

1. Benchmarked the performance evaluation of ECG biometric methods and collected UofT ECG

Database (UofTDB) which compared to existing databases, it has four important characteristics

namely, large population size (1020 individuals), varying body postures, physical exercise, acquisition

over a long period, acquisition from fingertips (chapter 2).

2. Implemented and systematically analyzed the performance of the state of the art ECG biometric

methods namely the AC/LDA [3], EigenPulse [43], Odinaka et al. [64] and Chan et al. [17] methods

under various ECG variability conditions (chapter 3).

3. proposed a new approach to address the problem of system deployment in uncontrolled settings

such as body posture changes and heart rate fluctuations through a way of enhancing the biometric

template appropriately (chapter 3).

4. Evaluated and compared the template updating approach performance on the state of the art ECG

methods (chapter 3).


5. Explored the feasibility of posture detection based on the ECG signal by exploring different feature

spaces. Proposed and evaluated a new posture detection algorithm based on the ECG signal for

improving the performance of ECG biometric systems (chapter 4).

6. Evaluated the performance of different subspace methods (PCA, DLDA, ICA, KPCA and KDLDA)

for ECG recognition under different ECG variability conditions (chapter 5 and 6).

1.2 Taxonomy of Errors

Matching two biometric feature vectors does not have a single positive or negative answer. In biometrics,

even though features originate from the same subject, significant intra-class variability is usually observed

that renders classification very difficult. The error of such systems is directly linked to the mode under

which they operate. For this reason, the biometric modes of operation are herein presented first.

1. Enrolment: During this mode of operation the biometric system collects the recognizing modality

(ex. ECG, face, iris), performs some quality check, processes to extract discriminative features,

and stores the result in the gallery set.

2. Identification: During this mode of operation, the system uses an input biometric reading to

perform one-to-many matches with the gallery set. The purpose of this operation is to answer the

question: What is the identity of this user?

3. Verification (or Authentication): During this mode of operation, the user submits to the system

a biometric sample along with an identity claim. The system compares the input sample with

the corresponding record from the gallery set, and a Good/Poor match decision is returned. The

purpose of this operation is to answer the question: Is the user who he/she claims to be?

Instead of the credential information, the output of a biometric system is often a match score,

revealing the degree of resemblance for a given pair of biometric samples. In essence, the match score

expresses the degree of certainty (or uncertainty) about a users identity. Match scores can take the form

of a probability, similarity, or distance, and authentication is then carried out by setting a threshold

empirically.

In order to distinguish the types of errors that a biometric system can make, it is important to outline

the following states as the systems possible conditions:

1. Identify an individual correctly, which is measured in identification rates.

2. Misidentify an enrolled individual, which is measured in mis-identification rates.


For more complex systems, authentication of legitimate subjects is referred to as sensitivity and

measured in authentication (or verification) rates.

Deny identity authentication to a legitimate subject, measured in false rejection rates (FRR). Deny

identity authentication to intruders, referred to as specificity of the system. Authenticate intruders,

which is measured in false acceptance rates (FAR). Specifically, the false acceptance and rejection statis-

tics are computed as:

FAR =Number of falsely authenticated subjects

Total number of intrudersFRR =

Number of rejected legitimate subjects

Total number of ID attempts

(1.1)

The equal error rate (EER) is also defined as the point in the FAR and FRR curves, where false

acceptance is equal to false rejection i.e., EER = FAR = FRR. The lower the equal error rate, the better

the authentication performance of the system.

Depending on the employed similarity measure, the appearance of the FAR and FRR distribution

can vary. When distance is used to associate two records, suggesting that the higher the score the less

the resemblance, FAR is expected to increase as the distance threshold increases. This way, for a higher

selection of the threshold, intruder authentication is rendered more likely. Similarly, the false rejection

percentage is expected to fall as the distance threshold increases, because more legitimate subjects will

be rejected. Obviously, there is a trade-off between false acceptance and rejection cases, and even though

ideally a biometric system would demand both of them to be low, it is usually left up to the designer to

decide on the specifics of the application.

1.3 Related Publications

The research findings of this work resulted in the following publications:

1. [88] S. Wahabi, S. Pouryayevali, S. Hari, and D. Hatzinakos. On evaluating ECG biometric

systems: Session-dependence and body posture. IEEE Transactions on Information Forensics and

Security, 9(11):2002-2013, 2014.

2. [71] S. Pouryayevali, S. Wahabi, S. Hari, and D. Hatzinakos. On establishing evaluation standards

for ECG biometrics. In 2014 IEEE International Conference on Acoustics, Speech and Signal

Processing (ICASSP), pages 3774-3778, 2014.

3. [89] S. Wahabi, S. Pouryayevali and D. Hatzinakos. Posture-invariant ECG recognition with pos-

ture detection. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing


(ICASSP), 2015.

1.4 Thesis Outline

The structure of the thesis is described here. Chapter 2 discusses the ECG signal, its recording standards

and the factors that affect the shape of ECG signal such as exercise, body posture and etc. Furthermore

we give a literature review of ECG methodologies and the existing databases. Finally we introduce the

UofT ECG Database (UofTDB) which has recordings from a large population size as well as recordings

under different postures and conditions.

In chapter 3 we systematically evaluate the performance of different existing ECG biometric methods

and at the end we propose a state-detection verification method to address the problem of system

deployment in uncontrolled settings through a way of enhancing the biometric template.

Chapter 4 introduces a new posture-detection technique based on the ECG signal for the proposed

state-detection method. Different feature spaces were investigated and the method was tested in con-

junction with the AC/LDA [3] ECG biometric method.

Chapter 5 and 6 introduce and evaluate various subspace dimensionality reduction techniques in the

context of ECG biometrics. While chapter 5 provides the details of each method, chapter 6 evaluate their

accuracy when deployed for ECG recognition. The methods are compared based on their performance

under different settings and also based on their subspace compactness.

Finally chapter 7 concludes this work by summarizing the main achievements and gives suggestions

for future improvements and directions.

Chapter 2

Benchmarking ECG biometric

testing techniques

2.1 Introduction

This chapter presents the ECG signal, describes its individual waves and recording standards. Then

we discuss the factors that affect the ECG signal such as posture changes, exercise, stress, day to day

ECG changes and etc. We further review the ECG biometric literature in terms of methodologies and

the databases used for evaluations and show that most of the ECG variability factors are absent in the

evaluations of ECG biometrics. Then we introduce a new ECG Biometric database that has recordings

under different conditions such as different recording sessions, multiple postures and exercise condition.

Therefore with the new database we can reliably evaluate the performance of different existing ECG

biometric methodologies (Section 4).

2.2 ECG as a Biometric

Biometrics provide a means of identifying individuals based on anatomical and behavioural character-

istics which are unique to each individual. Currently, there are many biometric modalities being used,

such as face, voice, fingerprint, signature, iris, etc. All of the mentioned biometrics have been success-

fully used in many applications but a major weakness of many of them is the lack of liveliness detection

which can lead to spoof attacks. The biometric template of a genuine user can be acquired by synthetic

reproduction to produce an artifact. The artifact then can be presented to a biometric system to get

7

Chapter 2. Benchmarking ECG biometric testing techniques 8

access as a genuine user and consequently delude the system. Therefore the presence of vitality detection

can protect the system from spoof attacks. The Electrocardiogram (ECG) biometric systems inherently

have liveness detection thus can ensure that the individual is present at the time of enrolment. Also the

ECG signal is very difficult to regenerate or mimic. Therefore the ECG signal has strong characteristics

that can address the issues of previous biometrics. However unlike most biometrics, ECG is naturally

affected by physical and psychological activity of the human body. The unique characteristics of the

ECG signal presents a challenge for biometric deployment and measures have to be taken to ensure that

ECG biometric systems are robust to such changes.

2.2.1 ECG Wave

Electrocardiogram (ECG) reflects the electrical activity of heart over time which is recorded by placing

electrode sensors on the body surface. The electrodes sense the small electrical changes (in the order of 1

mV) from the subject skins which are due to depolarization of the heart muscles during each heartbeat.

As shown in Figure 2.1, ECG is a quasiperiodic signal which is composed of repeating heartbeats.

Figure 2.1: Sample Electrocardiogram (ECG) recording of a subject.

The different waves that comprise a heartbeat represent the sequence of depolarization and repolar-

ization of the atria and ventricles heart muscles (Figure 2.2):

• P-wave represents the depolarization (contraction) of the atria with a duration of less than 80

msec. Its frequency spectral components is in the range of 10 Hz- 15 Hz.

• QRS complex which represents the rapid depolarization of the right and left ventricles with

higher frequency components, 10 Hz - 40 Hz. The ventricles have a large muscle mass compared

to the atria, therefore the QRS complex has a much larger amplitude than the P-wave and its

duration ranges from 80-100 msec.


• T-wave is the result of repolarization of the ventricles with an average duration of 160 msec. Its

position relative to the QRS is dependent on the heart rate of the subject; closer to the QRS

complex when the heart rate is higher.

Figure 2.2: Typical components of an ECG heartbeat.

In 1901, a Dutch physician named Willem Einthoven invented the first practical string galvanometer

that was capable of recording small electrical currents produced by human hearts. He then named each

deflection in the heartbeat with letters P, Q, R, S and T. He also associated electrocardiographic features

with various heart disorders. Still in modern medicine the ECG signal is the primary diagnostic tool for

patients with cardiac diseases.

2.2.2 ECG Recording Standards

As an electrical signal, the ECG can be recorded at the surface of the body using electrodes that

are positioned at various locations across the heart. The most widely used acquisition method is the

standard 12-lead ECG system which includes three main sets of lead orientations. The bipolar limb

leads are usually denoted as I, II and III and they track the electrical potential of the heart when three

electrodes are attached at the right and left wrist and left ankle [83]. By convention, Lead I measures the

potential difference between the two arms. In Lead II, one electrode is attached on the left leg and the

other on the right hand. In Lead III, the measured potential is between the left leg and right hand (Figure

2.3). Each Lead offers a signal of unique waveform characteristics. Healthcare practitioners typically

extract and combine information from all ECG Leads, which are recorded simultaneously. However, the

standard 12-Lead ECG system is not practical for real-world biometric systems unless those that are


Figure 2.3: The limb leads. Photo by Nicholas Patchett,http : //en.wikipedia.org/wiki/F ile :EKGlimbleads.png

deployed in healthcare environments. Nevertheless, the ECG can be easily captured with sensor contact

of any two points across the heart. As such, the hands and the fingers present a practical solution for

real-world physical and logical access control systems [17,18,24,57,78].

2.3 Factors affecting the ECG signal

As discussed earlier, ECG offers a unique combination of attributes that differentiate it from conventional

biometric modalities. To that end, evaluation of ECG biometric systems should take into account the

idiosyncrasy of this modality and assess its performance under the various conditions that affect it, such

as:

• Exercise: Exercise affects the heart rate and alters the frequency content of the ECG signal. The

accuracy of an ECG biometric system can potentially be deteriorated by this effect, if not taken into

consideration within the biometric feature space. Variations in the heart rate are more noticeable

on the relative position of the T-wave rather than the P-wave or the QRS-complex which are more

stable. When comparing a resting and an active state it is observed that the T-wave moves closer

to the QRS-complex [80] which also can be seen from Figure 2.4. Prior work has noted that heart

rate effects can be eliminated via normalization of time-domain ECG features [49].

• Cardiac Disorders: Another factor that can affect the ECG signal are cardiac abnormalities

(arrhythimias). Examples of arrhythmias that affect the ECG waveform are premature heartbeats

and atrial or ventricular arrhythmias. Depending on the origin of the disorder, arrhythmias can

alter the geometrical characteristics of the ECG signal to a varying extend. An approach has

been proposed to biometrically analyze ECG signals with cardiac disorders similarly to healthy


0.2 0.4 0.6

0

0.5

1

Time (s)

Am

pli

tud

e (m

V)

Rest Condition

Exercise Condition

Figure 2.4: Heartbeats of a subject in rest and exercise condition from UofTDB. The T-wave moves closer tothe QRS complex as the heart rate increases.

ones [15].

• Body posture: A factor that has been marginally examined in the prior art is the effect of body

posture. The electrical heart vector changes in relation to body position [26] . Using limb leads,

this work demonstrated that body position influences the cardiac response. Based on the findings

of Sutherland et al. [84], when transitioning from the supine to upright position, there is an interior

torso shift of the P-wave but no change in the amplitude of the local extrema.

• Time evolution: The ECG signal can change over time. Examples of factors that may impact

the ECG waveform over time include daily activities, dietary choices, substance consumption, drug

use and other. In order for the ECG recognition system to reliably authenticate users, it should

be designed to compensate for such variations. Odinaka et al. [65] compared the authentication

performance of different methods for across-session and within-session recordings,and reported that

higher accuracy can be achieved when enrolment and testing are done within the same session.

Only a few methods in the prior art have considered this property [6, 12,17,30,62,64,85,92,97].

• Psychology: A person’s emotions are continuously changing. These changes are reflected on

the ECG signal as variations in the rhythm at which the muscle pumps blood. The impact of

psychological changes on the ECG waveform has been documented in [7].

• Database size: Typical ECG evaluation databases are limited to about 100 individuals. Evalua-

tion over larger databases can provide a better insight in the uniqueness of this biometric. In [64],

it was demonstrated that the error rate of a biometric system can potentially increase with an

increase of the size of the testing database.


2.4 Existing Databases

The largest private database is that of Zhang and Wei [102], which consists of 520 subjects. The largest

public database is the PTBDB [31] which offers ECG readings from 290 subjects. While, the database

by Odinaka et al. [64] is smaller (269 subjects), it is unique in that it includes recordings from multiple

recording sessions spaced up to six months apart. The database by Irvine et al. [43] (43 subjects)

contains recordings acquired after each subject performed tasks specifically designed to elicit various

levels of anxiety. The database of Kim et al. [49] (10 subjects) includes recording under both normal

and elevated heart rate conditions. The PTBDB [31] and Odinaka et al. [64] databases also include some

subjects with heart-related disorders.

As previously discussed, the ECG signal reflects the electrophysiological properties of the heart but

it is also affected by the various sympathetic and para-sympathetic processes of the autonomic nervous

system (ANS) [7]. The latter contribute to the time-varying nature of the ECG signal. Furthermore,

the ECG is affected by body posture [1] and heart rate changes. To date, the investigation of the above-

mentioned factors for the ECG, in a biometric context, is very limited. This is primarily due to the lack

of availability of a signal database that encompasses all this information simultaneously. It is important

to note that the majority of the existing methods was evaluated on readings that were acquired within

the same experimental session or within the same day. In [65], the authors demonstrated the degradation

of the ECG biometric accuracy using data that are recorded over multiple days.

Additionally, most of the early ECG databases have a small population size (couple of hundreds

of subjects). Therefore, while these works can demonstrate a proof of concept for ECG biometric

recognition, feasibility studies for large-scale deployments has, so far, been limited.

2.5 Literature Review of Testing Techniques

A biometric system is essentially a pattern recognition system that acquires biometric data from an

individual, designs a biometric signature, and compares this signature against other templates in the

database. In the ECG biometrics literature, data acquisition typically refers to single channel recordings.

Although the use of the complete medical (12-lead ECG system) setup has been shown to increase the

overall biometric accuracy [4,12,27,92,99], Biel et al. [12] showed that a single channel ECG has sufficient

information for biometric purposes and is far more convenient for the user than the 12-lead configuration.

Similar to Biel et al.’s work are Chan et al.’s [17] proposal for ECG recording from the thumbs, Shen et

al.’s [78] proposal from the palms and Odinaka et al.’s [64] from the lower rib cage.


The process of designing the biometric signature from the ECG signal typically encompasses three

steps: 1) Pre-processing, 2) Feature Extraction and 3) Feature Selection. The preprocessing step gener-

ally consists of filtering the raw ECG signal to eliminate noise caused by body movement, misplacement

of electrodes and any interference by nearby devices. Thanks to the advancements in ECG signal pro-

cessing for medical applications a powerful set of algorithms and processes are available to clean and

enhance the quality of the ECG signals.

Feature extraction techniques aim to detect attributes of the ECG signal that are unique and

permanent for each subject. The early works in this field (namely, fiducial dependent approaches)

examined local characteristics of the heartbeats such as the amplitude and angles of certain waves

[12,28,32,40–42,44,49,51,52,66,81,82,87,102]. Due to the limited number of features proposed by these

methods, and due to the limitations in accurately locating these waves’ boundaries [59] efforts shifted to

the analysis of the overall shape of the ECG. In various methods, the QRS complex is detected and the

ECG signal is segmented into individual heartbeats [14,17,19,22,27,35,43,45,55,62,64,73,90,92,97–99].

Features are typically extracted by modelling or transforming the heartbeat in another domain.

In another set of non-fiducial approaches to ECG biometric recognition, the signal is windowed using

either overlapping or non-overlapping segments, regardless of the location of the heartbeats. Some of

these methods suggested that the autocorrelation is used to blend in all the samples of the window

[2–6,70]. In [56] the ECG signal was windowed with overlapping windows and Linear Predictive Coding

and Wavelet Packet Decomposition were then applied. Ghofrani et al. [30] also windowed the signal

using overlapping windows and used the coefficients of the autoregressive model as features.

After feature extraction, feature selection aims to choose the most robust features or to transform

the whole set of features into a more robust form. In order to effectively select or transform features,

a rule of high inter-variability is established to increase the distinctiveness of the signatures, and a rule

of low intra-variability is also used to decrease the variance among recordings of the same individual.

Various feature selection methods have been evaluated, some of the most common techniques are LDA

[28, 40, 41, 44, 49, 51, 52, 91], Generative models [42, 55, 64, 70] and Support Vector Machines [55, 99]. A

comprehensive analysis of the above-mentioned methods is provided in [65].

The following is a summary of the performances reported by researchers.

• Biel et al. [12] proposed a fiducial based method using ECG recordings from 20 subjects. The

extracted features are the local characteristics of the QRS complex, T wave and P wave. The

best performance that was achieved by combining 10 different features was an identification rate

of 100%.


• Kyoso et al. [52] proposed a fiducial method using four features i.e. the P wave duration, PQ

interval, QRS and QT durations. The highest performance reported was an identification rate of

94% using the QRS and QT intervals of nine subjects.

• Shen et al. [77] proposed an ECG based recognition method with seven fiducial based features

that relate to the QRS complex. This research applied two techniques, template matching and

a decision-based neural network (DBNN), to implement the identity verification. Using each of

the two methods separate on a predetermined group of 20 subjects. The experimental results

showed that the rate of correct identity verification was 95% for template matching and 80% for

the DBNN. Combining the two methods produced a 100% recognition rate.

• Israel et al. [44] presented the three clear stages of ECG biometric recognition i.e., preprocessing,

feature extraction and classification.The proposed system employed 15 temporal features. The

highest achieved performance was reported to be an identification rate of 100% tested on 29

individuals.

• Plataniotis et al. [70] are one of the first to propose a non fiducial based method. They take the

autocorrelation of the windowed ECG signals and then performed the Discrete Cosine Transform

for dimensionality reduction. With the objective of capturing the repetitive pattern of ECG, the

authors suggested the AC of an ECG segment as a way to avoid fiducial points detection. The

method was tested on 14 subjects and achieved an identification rate of 100%.

• Wubbeler et al. [92] also proposed a non fiducial method by analyzing the distance between two

ECGs which was determined utilizing one single heart beat of the two-dimensional heart vector

which is known as a characteristic of the ECG. The method was tested on 74 subjects and achieved

a 2.8% EER.

• Irvine et al. [43] proposed a non fiducial method by using the extracted heartbeats of the ECG

signal and normalizing them to have a dynamic range between zero and one. Furthermore apply

Principle Component Analysis for feature extraction. Then an euclidean distance measure along

with majority voting is used to make the final decision. They achieve an identification rate of more

than 80% based on 43 subjects.

• Chan et al. [17] also proposed a non fiducial method which utilizes the detail coefficients of the

discrete wavelet transform of PQRST complexes for the feature space and a proposed wavelet

distance measure for classification. The ECG data were collected from 50 subjects during three


data-recording sessions on different days. Data from session 1 were used to establish an enrolled

database, and data from the remaining two sessions were used as test cases. Classification was

performed using three different quantitative measures: percent residual difference, correlation co-

efficient, and a novel distance measure based on wavelet transform. The highest performance

reported was an identification rate of 89% using the wavelet distance measure.

• Agrafioti et al. [3] method used the autocorrelation as a source of discriminative information in a

population, to eliminate the need for fiducial points detection. On top of the autocorrelated ECG

signals, discrete cosine transform or linear discriminant analysis are applied for dimensionality

reduction. The best performance reported is an EER of less than 1% tested on 27 subjects.

• Odinaka et al. [64] proposed the time-frequency content of the heartbeats as the feature space and

proposed a feature selection method for dimensionality reduction and performance enhancement.

They test their method on a database containing recordings from 269 subjects and the best reported

EER was less than 1%.

• Zhao et al. [104] proposed a non fiducial method based on ensemble empirical mode decomposition

(EEMD). They first eliminate the noise of the ECG signal using wavelet decomposition and then

they extract the individual heart beats. Furthermore heartbeat normalization and quality mea-

surement is performed to eliminate the effects heart rate variability. Then the ECG heartbeats are

decomposed into a number of intrinsic mode functions (IMFs). Principal component analysis is

employed to reduce the dimensionality of the feature space, and the K-nearest neighbors (K-NN)

method is applied as the classifier tool. The system achieved an identification accuracy of 95% for

90 subjects.

The majority of the prior work did not examine the evolution of ECG signal with time, posture changes

and exercise condition. To some extent, the sources of intra-subject variability of the ECG signal

have been ignored. Furthermore the methods were tested on relatively small datasets thus leading to

optimistic and often unreliable results.

2.6 Benchmarking an ECG database

2.6.1 The Database

To address the above-mentioned shortcomings, this work presents a new biometric database for evaluation

of ECG biometric systems. The UofT database(UofTDB), is an ECG database captured from subjects’


fingers, thereby mimicking real world application settings. Emphasis has been given to the collection of

ECGs from a large number of individuals under several conditions and in uncontrolled environments.

The database can be divided into two portions, one is the single session and the other is the follow-up

sessions.

The single session recordings took place at multiple public locations in the University of Toronto;

Sandford Fleming Building, Bahen Centre for Information Technology, Gerstein Science Information

Centre in the Sigmund Samuel Library Building, J. M. Kelly Library and the Athletic Centre. In order

to explain our intention for this experiment and increase participation, a set of promotional banners was

created and placed close to the workbench and a member of our signal collecting team was in charge

of motivating potential candidates. The principal members of our collecting team were Saeid Wahabi,

Shahrzad Pouryayevali and Foteini Agrafioti. The experimental procedure was disclosed to the volunteers

in the beginning of the experiment, who also signed consent forms. None of the participants were asked

about any health problems therefore the database can have a mixture of healthy and non healthy subjects.

As a reward for participating in the experiments, a gift card was given to the participants. The single

session data collection process was spanned over three months, and in the end data had been collected

for an overall total of 1020 participants, the majority of which were students.

The follow-up sessions recording took place at the Biometrics Security Laboratory of the University

of Toronto. Data recollection from the set of volunteers previously enrolled in the experiment was

done with the purpose of studying the changes in the ECG morphology over time and the effect of

posture change and heart rate variability. Overall 100 volunteers participated in these experiments.

Five recording sessions took place several weeks apart, however not all the 100 subjects participated in

the five follow-up sessions only 43 subjects were recorded at all five sessions. Similarly the experimental

procedure was disclosed to the volunteers in the beginning of the experiment, who also signed consent

forms.

Overall, the age range of the participants was between 18 and 52 and 61% were female and 39% were

male. The whole signal collection process took 6 months and ended in the Fall of 2013. More specific,

the characteristics of the UofTDB are as follows (See Table 2.1 ):


Table 2.1: UofT ECG Database

Number of Recording Sessions Conditions

single two three four five six supine tripod exercise sit stand

Number of

subjects1020 72 65 54 47 43 63 63 71 1020 81

Age Range: 18 - 52, Gender: 61% Female, 39% Male, Total number of subjects = 1020

• Database size: The UofTDB has ECG records from 1020 subjects. The length of the recordings

varies between 2-5 minutes.

• Number of sessions: 100 subjects, randomly selected , participated in follow-up recordings. The

follow-up study consists of five sessions of recordings spanned over a period of six months.

• Testing conditions: During each follow-up session, the subjects were recorded under five different

conditions (Refer to Figure 2.5):

(a) (b) (c) (d)

Figure 2.5: (a) Standing (b) Sitting (c) Tripod and (d) Supine

1. Sitting: The subject’s ECG was recorded while sitting comfortable on a chair.

2. Standing: The subject was asked to stand still during the recording.

3. Supine: In this position, the subject was asked to lay back and relax

4. Tripod: The subject completely leaned forward while sitting on a chair.

5. Exercise: The subject performed basic structural workouts such as jumping jacks and push-

ups for 3-5 minutes depending on their physical status. After exercise, the heart rate raised

on average to 132 beats per minute.


Figure 2.6: Placement of electrodes according to lead I configuration

2.6.2 Acquisition Protocol

The ECG signals recorded from the subjects were anonymized in order to protect their personal infor-

mation in accordance with the University of Toronto Research ethics policy. At the beginning of the

session each subject signed a consent form in order to qualify their participation in the study.

The acquisition device: The Vernier EKG Sensor and the Vernier Go!Link interface1 were used

for recording. The device resolution is 12 bit per sample and the sampling frequency was set at 200 Hz.

Table 2.2 summarizes the specifications of the acquisition device.

Interface Sampling Rate 200Hz

Interface Resolution 12 bits / sample

Sensor Offset 1.00 V(±0.3 V)

Sensor Gain 1 mV body potential /

1 V sensor output

Sensor Roll-off Frequency 250Hz

Table 2.2: Vernier EKG sensor and USB interface specifications

Placement of electrodes: three dry AgCl electrodes were used to capture the ECG from the

subjects’ fingertips similar to a lead I configuration. The left thumb was positioned on the positive

electrode, the right thumb on the negative and a reference electrode was was in contact with the right

index finger. All the three electrodes were attached to a pad so that the subject could rest their fingers

on the electrodes by holding the pad as shown in Figure 2.6.

Signal quality: during each recording session, an operator verified that the acquired signal is a

valid ECG. The operator instructed the volunteer to position their fingers according to the experimental

1www.vernier.com


protocol. High frequency noise was observed in the recordings of certain volunteers, however the current

analysis did not discard any such records. Low quality signals are a direct consequence of the uncontrolled

environment and the subjects’ natural movements. The inclusion of such signals in the database simulates

the challenge that all ECG biometric systems encounter in real-world scenarios.

2.7 Conclusion

In summary, this chapter brings to the table the issue of establishing testing standards for ECG biometric

systems. Factors that directly affect the ECG waveform and impact the biometric accuracy are identified

and a new testing database is presented as a benchmark for algorithmic evaluation. Therefore in the

evaluation of ECG biometric systems, performance under such factors must be assessed. The UofTDB

database covers conditions such as exercise and day-to-day variability as well as morphology evolution

over a period of six months. Furthermore the signals in this database have been acquired from the

volunteers’ fingertips thereby mimicking real world application settings.

Chapter 3

Evaluation of the state of the art

ECG biometric methods

3.1 Introduction

In this chapter, with the new database we presented in the previous chapter we systematically evaluate

the performance of different existing ECG biometric methodologies under novel experimental conditions

such as different sessions, different postures and in exercise condition. We review in detail the methods

chosen to be implemented and compare their performances in various settings. At the end we propose

a new method to address the problem of system deployment in uncontrolled settings through a way of

enhancing the biometric template appropriately.

3.2 Background

Among the ECG biometric methods reviewed in the previous section, four of the non-fiducial based

algorithms are herein analyzed with the UofTDB database: the discrete wavelet transform method by

Chan et al. [17], the time-frequency content approach by Odinaka et al. [64], the eigenPulse method

proposed by Irvine et al. [43] and the AC/LDA method by Agrafioti et al. [3]. The main reason that

fiducial based methodologies were excluded from this analysis is that for a large number of the subjects

enrolment may not be feasible due to the presence of noise in the data that renders the detection of

fiducial points problematic. Similar observations have been reported by Irvine et al. [43] where the

fiducial-based feature extraction methods failed to enroll approximately 30% of the testing population.

20

Chapter 3. Evaluation of the state of the art ECG biometric methods 21

1 1.5 2

0.8

1

1.2

1.4

1.6

Time (s)

Am

pli

tud

e (m

V)

(a)

1 1.5 2−0.2

0

0.2

0.4

0.6

Time (s)

Am

pli

tud

e (m

V)

(b)

0.1 0.2 0.3 0.4 0.5 0.6 0.7−0.2

0

0.2

0.4

0.6

0.8

Time (s)

Am

pli

tud

e (m

V)

(c)

Figure 3.1: (a) Raw, (b) filtered ECG signal and (c) shows extracted and aligned heartbeats of a subject.

3.3 ECG Preprocessing

The raw ECG signals were first filtered using a 4th order bandpass Butterworth filter with cut-off

frequencies between 0.5Hz and 40Hz. Figure 3.1 shows an example of noise removal using this filter

and Figure 3.2 shows the normalized power spectral density of the raw ECG signal and the magnitude

impulse response of the filter. Under 0.5Hz the signal is corrupted by baseline wander, and over 40 Hz

there is distortion due to muscle movement, power-line interference etc. [83]. Odinaka et al. [64] used a

0.5 Hz high pass filer and a 500Hz low pass filter while Irvine et al. [43] used a Fourier bandpass filter

described in [86]. Chan et al. [17] used a high-gain AC amplifier with bandwidth set at 1Hz -100Hz and

a notch filter was employed to reduce the power-line interference noise.

Except for the AC/LDA, all the methods require the filtered ECG signal to be segmented into PQRST

segments (heartbeats), and to be aligned at the R peaks. The QRS detection algorithm described in [67]

was employed to segment and align the heartbeats. Since the heartbeat durations may vary from


Figure 3.2: Power Spectral Density of raw ECG signal and the magnitude impulse response of the fourth orderButterworth bandpass filter with cut-off frequencies at 0.5Hz and 40Hz.

heartbeat to heartbeat and from one subject to the other, all heartbeats were fixed to a length of 700

msec with a 200 msec segment fixed before the R peak (see Figure 3.1). Furthermore, outlier heartbeats

were discarded by estimating their Euclidean distance from the subject’s median heartbeat. This ensures

the exclusion of the heartbeats that have been corrupted by a large amount of noise and artifacts.

3.4 Review of the Implemented Methods

In this section an overview of the analyzed methods along with any deviations from the original imple-

mentations are discussed.

• Discrete wavelet transform approach [17]: The average heartbeat of each subject is estimated

and further analyzed for feature extraction. The feature extraction process is displayed in Figure

3.3.

Figure 3.3: Feature extraction block diagram for discrete wavelet transform approach [17].

A five level discrete wavelet transform of the average heartbeat is computed for each subject using

the Daubechies scalar wavelet (Db3) as the wavelet function. Figure 3.4 shows how the feature

vector is composed of the five level detail coefficients.


Figure 3.4: Feature vector γ formed from level 1-5 detail coefficients (D1, D2, D3, D4 and D5) of the 5-levelwavelet decomposition of the input heartbeat. The Daubechies scalar wavelet (Db3) was used as the waveletfunction. DX represents the detail coefficients and AX is the approximation coefficients at level X.

Let γp,qclaim denote the detail coefficient q from the pth level of decomposition. For verification, the

test average heartbeat undergoes the same steps to calculate the γp,qinput. Using these coefficients,

a distance can be estimated as:

WDIST =

P∑p=1

Q∑q=1

| γp,qinput − γp,qclaim |

max(γp,qh , τ)

where τ is a threshold value used to avoid small coefficients from magnifying variations, and γp,qclaim

is the detail coefficients of the claimed identity. Chan et al. [17] evaluated the proposed method

in the identification mode of operation whereby the algorithm associates the best match with the

enrolled subject that presents the smallest WDIST. Since the present analysis is performed in

the verification mode of operation, a threshold η is introduced to control the biometric decision

making.

• Time-frequency content method [64]: For this method, each heartbeat was first normalized

by subtracting its mean and dividing by its standard deviation:

hbnormalized =hb−mean(hb)

std(hb),

where hb represents a heartbeat. The spectrogram of each heartbeat was then estimated using a

Hamming window of length 64msec with 54msec overlap. The spectrogram is the square of the

magnitude of the short-time Fourier transform of a signal.


Let Y (l) denote the time-frequency components of the spectrogram where l is the index of each

point, also known as time-frequency bins [64]. Independent Gaussian distributions were used to

model each time-frequency bin. In order to enrol a subject, the means and variances are calculated

using the maximum likelihood estimation. For each time-frequency bin l of subject i the feature

vector is denoted as θi(l) = (µil, σ2il).

For dimensionality reduction, the time-frequency bins (N (µil, σ2il)) whose symmetric relative en-

tropy with the nominal distribution (N (µ0l, σ20l)) is larger than some threshold κ > 0 are selected as

the feature vector. The symmetric relative entropy between N (µil, σ2il) and N (µ0l, σ

20l) is estimated

as:

d(θi(l), θ0(l)) =σ2il + (µil − µ0l)

2

2σ20l

+σ20l + (µil − µ0l)

2

2σ2il

− 1

where the nominal distribution is found by using the spectrogram of all subjects in the database.

The process of feature extraction is shown in Figure 3.5.

Figure 3.5: Feature extraction block diagram for Time-Frequency content approach [64].

In the verification mode of operation, the Y (l) for a heartbeat claiming to be the ith subject is

computed. By using the log-likelihood ratio (LLR), a score is assigned to such a claim:

Λi =

L∑l=1

log

[pi(Y (l)|θi(l))p0(Y (l)|θ0(l))

]Id(θi(l),θ0(l))>κ

where l is the index of bins and I{.} is the truth function indicating which time-frequency bins are

selected for subject i. The test heartbeat with the claimed identity is accepted if the estimated

LLR is greater than a threshold τ , i.e. Λi > τ , otherwise the heartbeat is rejected. Since the

authors did not mention how the method should make the final decision of the claim the average

heartbeat is used in the present analysis to make the final decision.

• EigenPulse method [43]: The extracted heartbeats are normalized to have a dynamic range

between zero and one:

Zh,i =Zh,i −min(Zh,i)

max(Zh,i)−min(Zh,i)


where Zth,i is the ith heartbeat of the hth subject. The covariance matrix Σ, for all the normalized

heartbeats in the database is estimated and the heartbeats are projected onto the eigenvectors of

Σ that corresponds to the K largest eigenvalues. In this analysis, a K=58 is empirically chosen

since it yields robust performance in terms of EER. Let V = [V1, . . . , VK ]T denote the matrix of

eigenvectors and Z represents the mean heartbeat. Then

Ci,h = [Zi,h − Z]V

Ci,h is the projection of the ith heartbeat of the hth subject onto the subspace spanned by the

eigenvectors. The mean of Ci,h over all the heartbeats of a subject represents the feature vector

Ch, for that subject.

Figure 3.6: Feature extraction block diagram for eigenPulse [43] method.

A test signal undergoes the same steps as described above to estimate Ci,test as shown in Figure 3.6.

If the Euclidean distance between Ch and Ci,test is less than a threshold τ , the pair is considered

an acceptable match otherwise it is rejected. Majority voting is used for the final decision on the

claim.

• AC/LDA method [3]: Unlike the previous methods, the AC/LDA does not require heartbeat

segmentation. The filtered ECG signal is blindly segmented into overlapping windows of 6 seconds

length with 50% overlap. The normalized autocorrelation coefficients, AC, of each window can be

estimated using:

Rxx[m] =

N−|m|−1∑i=0

x[i]x[i+m]

Rxx[0]

where x[i] represents the windowed ECG and x[i+m] is the delayed version of ECG window with

a time lag of m = 0, 1, . . .M − 1 and M << N . In this work we empirically chose 50 lags which

is much smaller than N = 6 seconds × 200 samplesecond = 1200 samples in each window. The Linear

Discriminant Analysis (LDA) method is applied as a second step, to project the AC windows into

a lower dimensional subspace while improving the separability for different subjects. Since the


Figure 3.7: The estimated normalized autocorrelation of five different subjects.

number of training AC windows available for each subject is smaller than the dimensionality of

the sample space, the LDA suffers from a so-called small sample sized problem [72]. To overcome

this problem we used the Principal Component Analysis (PCA) as a preprocessing step for LDA

in order to reduce the dimension of the AC windows. Henceforth, we refer to this method as

AC/LDA. The feature extraction process is shown in Figure 3.8.

Figure 3.8: Feature extraction block diagram for AC/LDA [3] method.

The projected AC windows, ACproj , serve as templates for each subject and the main task of

verification is to match the templates of the unknown subject to the ones of the claimed subject.

The template matching is done by estimating the Euclidean distance and regulating it using a

threshold τ .


Figure 3.9: Receiver operating characteristic (ROC) curve for the four methodologies.

3.5 Results and Discussion

In the first experiment the methods were tested on the single session portion of the UofT database which

includes recordings from 1020 subjects in sitting posture. For each subject, the ECG recording was

divided into two halves (same-session analysis). The first half of the recording was used for training the

algorithms and enrolment and the second half for testing. A summary of the experimental performance is

shown in Figure 3.9. These results indicate a significant drop from the previously reported performance

by Agrafioti et al. [3] and Odinaka et al. [64] which were both reported to offer less than 1% equal error

rate. This can be attributed to the increased population size as well as the noise that is present in the

UofTDB database. The performance of the eigenPulse and the discrete wavelet transform methods were

reported only in identification mode and a comparison with the previously published results cannot be

made.

The robustness of the algorithms was further evaluated in across-session testing. For this evaluation

data from 47 subjects spread over a six month period in sitting posture were used. Session 5 recordings

are not in sitting posture therefore this session was excluded from this analysis. In this experiment, the


Figure 3.10: Performance evaluation of the four methodologies across different sessions with enrol data fromsession 1.

subjects were enrolled using from the first session and then tested using ECG recordings from subsequent

sessions. The enrolment set was also used for training the algorithms. Same experiment was performed

by training and enrolling with data from session two. The results are shown in Figures 3.10 and 3.11.

From these figures, it is clear that there are factors that affect the biometric accuracy over time.

Furthermore the performance of the methods proposed by Agrafioti et al. [3], Odinaka et al. [64] and

Chan et al. [17] significantly improve when the training data are from session two. This can be due to

the fact that session one recordings were collected in crowded hallways and as a result are more noisy

whereas the follow-up sessions were recorded in a more quiet lab environment. Another reason could be

that the interval between session one and the follow-up sessions is larger than the interval between the

follow-up sessions.

The effect of body posture on ECG biometric accuracy was further investigated. In each experiment

the enrolment and testing sets are from different body postures and conditions but from the same record-

ing sessions. Similar to previous experiments the enrolment set was used for training the algorithms.

The results of the four methods are summarized in Table 3.1. The average population size for each

experiment was 69 subjects. The performance of all the methods degrades when the train and testing


Figure 3.11: Performance evaluation of the four methodologies across different sessions with enrol data fromsession 2.

data are not from the same body position. However, this degradation is not uniform across all the

postures. For example, the EER of the four methods is lower when enrolling in tripod and testing in

supine position compared to enrolling in supine and testing in tripod position.

Another factor that impacts the performance of the examined methods is the effect of heart rate

fluctuation. The algorithms were trained using sitting ECG signals, then tested using exercise signals

and vice-versa. From Table 3.1 it can be seen that heart rate fluctuations affect the performance of most

methods. However, the AC/LDA method outperforms the other methods by around 10%.

3.6 Enhancing Biometric Templates

The experimental results of the previous section show that the performance of the biometric system is

lower when the ECG signals used for training and testing are from different body postures or condi-

tions. We refer to the body posture (sit, stand, supine and tripod) and condition (post-exercise) as the

underlying state of the acquired ECG signal.

The feature space of the implemented methods are considered for the current analysis. Consider


(a) Chan et al. [17]XXXXXXXXXXEnrol

TestSit Stand Supine Tripod Exercise

Sit 2.62% 13.69% - - 32.42%Stand 14.08% 3.01% - - -Supine - - 3.17% 14.52% -Tripod - - 12.92% 3.21% -

Exercise 32.87% - - - 19.07%

(b) Odinaka et al. [64]XXXXXXXXXXEnrol


Sit 4.68% 8.43% - - 30.88%Stand 8.71% 4.53% - -Supine - - 3.18% 8.07%Tripod - - 7.41% 1.58%

Exercise 33.82% - - - 14.51%

(c) Irvine et al. [43]XXXXXXXXXXEnrol


Sit 15.32% 23.87% - - 39.74%Stand 21.62% 13.76% - -Supine - - 11.06% 17.42%Tripod - - 14.76% 12.61%

Exercise 39.39% - - - 25.63%

(d) Agrafioti et al. [3]XXXXXXXXXXEnrol


Sit 2.57% 8.16% - - 24.10%Stand 8.16% 4.09% - - -Supine - - 1.44% 6.24% -Tripod - - 5.61% 1.73% -

Exercise 20.08% - - - 6.73%

Table 3.1: Equal Error Rate for the four methodologies tested in different body postures and exercise condition.


an enrolment protocol in which the user’s ECG signal is collected under each of the above-mentioned

postures and conditions. In real-world settings when the system is in normal verification operation the

underlying state of the user may be random and unknown to the system. In this analysis, we consider

two cases: state-agnostic and state-detection verification.

3.6.1 State-agnostic verification

In the training-phase, all ECG signals of a user are grouped together regardless of the underlying state.

A class encompasses AC/LDA feature vectors of an individual under all possible conditions. For the

discrete wavelet transform method and the time-frequency content approach, an individual’s template

is modelled by using the ECG signal under all the states. For the EigenPulse method the covariance

matrix Σ is obtained from the heartbeats of all the individuals recorded under different states. During

verification, an input of unknown state is matched against the entire template. Since the template

is a composition of various user states it is expected that it can authenticate inputs under unknown

conditions and postures.

3.6.2 State-detection verification

In this case the biometric template is composed of ECG signals under all postures and conditions but

the algorithms are trained individually for each posture and condition. Using the UofTDB database,

for example, five different AC/LDAs are trained, each corresponding to a different underlying-state. In

addition, the methods are trained per subject for state detection by extracting the state’s features (Chan

et al. [17] method), statistically modelling the states (Odinaka et al. [64]) and obtaining a projection

matrix (AC/LDA [3] and Irvine et al. [43] methods). For instance, in the AC/LDA method [3] the AC

windows of a state is assigned to one class of LDA. The projection matrices and the projected feature

vectors collectively form a subject’s biometric template. During verification, as shown in Figure 3.12,

the state of the user is first classified prior to biometric matching.


Claim ID

User AC

windows

Select AC

windows of

claimed ID

Select pro-

jection

matrix

for state

detection

Project

Project

Iden

tifystate

Select

projection

matrix for

identity

verification

Project

Project

Verify

claimed

iden

tityDecision

result

Figure 3.12: Flow chart diagram for state-detection verification for the AC/LDA algorithm.

For Odinaka et al. [64] and Chan et al. [17] methods we model each state and using the time-averaged

heartbeat of the test signal the detected posture is the one that produces the smallest proposed wavelet

distance(WDIST) for Chan et al. [17] method and largest score for Odinaka et al. [64]. For the Eigenpulse

method [43], each test heartbeat is classified into a posture by using an Euclidean distance classifier and

a majority voting scheme was used for the final decision on the subject’s posture. For the AC/LDA

method [3], a k-NN classifier with k=5 was used to classify each projected AC window and a majority

voting rule was employed to detect the posture of the subject.

UofTDB has 50 subjects for which recordings from all states are available. However these states are

not necessarily recorded in the same session. For example sit and stand posture were recorded in one

session and supine and exercise in another session. In this analysis we ignored the effect of the time of

recordings.

PPPPPPPPPPPPPMethod

PostureSit Stand Supine Tripod Exercise

Chan et al. [17] 88% 94% 94% 96% 84%

Odinaka et al. [64] 66% 58% 84% 78% 90%

Irvine et al. [43] 78% 88% 84% 82% 92%

Agrafioti et al. [3] 77% 75% 90% 94% 52%

Table 3.2: State detection rate of the methods

Table 3.2 shows the state detection rates for each method and Figures 3.13 and 3.14 illustrate the


performance of the above two treatments for each method.

Sit Stand Supine Tripod Exercise0

10

20

30

40

Test Condition

Eq

ua

l E

rro

r R

ate

(%

)

State−agnostic verification

State−detection verification

(a) Chan et al. [17]


5

10

15

20

25

30

Test Condition

Eq

ua

l E

rro

r R

ate

(%

)



(b) Odinaka et al. [64]

Figure 3.13: Equal error rates for the two verification methods for Chan et al. [17] and Odinaka et al. [64]methods. Enrolment with sit, stand, supine, tripod and exercise state ECG signals.

The enrolment template consists of a user’s ECG signals from each of the five underlying states,

namely sit, stand, supine, tripod and exercise. The equal error rates for different test signals is shown,

and we can observe that in most cases the state-detection verification approach outperforms the state-

agnostic approach especially for AC/LDA and discrete-wavelet transform methods. This difference in

performance is especially significant when the test signal is from the exercise state.



10

20

30

40

50

Test Condition

Eq

ua

l E

rro

r R

ate

(%

)



(a) Irvine et al. [43]


5

10

15

20

Test Condition

Eq

ua

l E

rro

r R

ate

(%

)



(b) Agrafioti et al. [3]

Figure 3.14: Equal error rates for the two verification methods for Irvine et al. [43] and Agrafioti et al. [3]methods. Enrolment with sit, stand, supine, tripod and exercise state ECG signals.

3.7 Conclusion

In this section it was shown that the practise of evaluating an ECG biometric system using signals that

are collected within a single recording session may not be indicative of the system’s accuracy. In addition,

this section illustrates a dependency of the accuracy to the body posture and condition under which the

ECG signals are captured. The analysis includes ECG samples under sitting, standing, supine, tripod

and exercise conditions and demonstrates that while the biometric performance may be compromised


when tested in such uncontrolled settings, there are treatments that can be considered to potentially

address this problem.

Chapter 4

Posture Detection for ECG

Biometric Systems

4.1 Introduction

As shown in the previous chapter, posture changes can severely degrade the performance of the ECG

biometric systems. Under the assumption that the ECG signal of subjects under different postures are

available at enrolment, we showed that state-detection verification enhances the biometric templates and

consequently improves the overall authentication accuracy. In this chapter we first explore the idea of

using the ECG signal for posture recognition and propose a method based on wavelet detail coefficients

of heartbeats for posture-detection. We further evaluate our system in conjunction with the AC/LDA [3]

ECG biometric method.

4.2 Effect of Posture on ECG

Among the factors that can affect the ECG signal is body posture changes. In the medical literature it has

been extensively shown that alterations of posture can change the ECG signal. They have demonstrated

that changes of body posture cause the heart to move in thoracic cavity (or chest cavity), as a result of

these movements the ECG signal also changes [46,63].

Sigler [79] is among the pioneers who studied the changes in ECG due to posture changes from

standing to recumbent positions based on 31 subjects. In many cases large changes in T and QRS waves

was noted, however these changes were not uniform across all the subjects. For instance in some subjects

36

Chapter 4. Posture Detection for ECG Biometric Systems 37

the change occurred only in the T-wave.

Jones et al. [48] also studied the changes in heart rate and R-wave amplitude with posture. Based

on 10 healthy male subjects, they found a significant increase in heart rate when the subject changes

position from horizontal to vertical and reaches the highest in standing position. The average resting

heart rate in supine, sitting and standing was 66.1± 12.1,72.1± 11.2 and 79.3± 10.6 respectively. Also

the R-wave amplitude is lowest in the standing posture which can be due to the increase of heart rate.

Jernberg et al. [46] studied the effect of QRS-waveforms and ST-T-segment by changes in body posi-

tion. Based on 19 subjects (10 healthy subjects and 9 patients with ECG abnormalities), they concluded

that ST-segment changes due to posture changes are usually small, however it may reach significant lev-

els in some patients. Their findings also suggest that the QRS complex is the most influenced waveform

by body posture changes and therefore it is not useful for routine patient monitoring.

Although, to our knowledge, there has not been studies on the effect of posture change in biometric

research, recently Pathoumvanh et al. [68] studied the robustness of ECG biometric systems in heart

rate variability conditions caused by physical activity. They show that the performance of the system

decreases by 17% in such conditions.

4.3 Literature Review of Posture Detection

In many health-care applications, body posture recognition is now becoming a very important task.

Although there are many approaches to posture detection such as accelerometer based, pressure sensor

arrays based and vision-based, recently methods based on the electrocardiogram (ECG) signal have been

proposed. Posture change or posture recognition methods based on the ECG signal has been proposed

by researchers for variety purposes such as vehicle drivers sitting posture monitoring [94], patient posture

monitoring on a bed [93], fall detection for aged people and detection of posture changes during Ischemia

monitoring [29] are a few examples to name. In addition because ECG is a vital signal, in many health

monitoring systems ECG recorders already exist and methods that detect the posture based on the ECG

signal do not require extra sensors and hardware. The related work in the literature can be divided in

two categories:

1. Detecting the change of posture: Minchole et al. [61] method detects changes of postures between

right-lateral and left-lateral whereas Garcia et al. [29] and Astrom et al. [8] methods detect the pos-

ture changes between supine to left-lateral and right-lateral and vice versa. The method proposed

by Pawar et al. [69] detects body movement activity including change of posture. The subjects

changed postures between stand and sit, supine and left-lateral and supine and right-lateral.


Preprocessing Feature Extraction ClassificationECG Signal

Figure 4.1: System block diagram.

2. Recognizing the posture: The methods in this category have focused on detecting the posture

of the subjects based on the ECG signal. Lee et al. [53] method achieved a posture detection

accuracy of 98.4% with four postures: supine, left-lateral, right-lateral and prone. The feature

space includes Q, R and S wave amplitudes. Shen et al. [76] extracted features from the heart

axis (magnitude and angle) and achieved a detection rate of 66.67% for classification of lying from

standing posture.

4.3.1 Method of Analysis

In this section we present a new method for the body posture detection problem. We further investigate

different feature spaces extracted from time and frequency domains. Finally the performance of the

system is evaluated under two testing scenarios:1) the generic case where the method does not require

to have a prior knowledge on the ECG signal of the user and 2) the personalized case where the system

is trained and optimized for each user.

The proposed body posture detection method, as shown in Figure 4.1, consists of three stages.

The preprocessing step is for signal quality enhancement, outlier removal, heartbeat extraction and

normalization. Then features are extracted from the preprocessed signal and used to train a Support

Vector Machine (SVM) classifier. In this work, a number of different feature spaces are considered for

body posture detection. The rest of the section gives details of each step.

4.3.1.1 Signal Preprocessing

The ECG recordings were first filtered using a fourth order bandpass Butterworth filter with cut off

frequencies at 0.5 and 40Hz. Below 0.5Hz the signal is corrupted by baseline wander and the frequency

contents beyond 40 Hz mainly correspond to noise created by muscle movements, 60Hz powerline noise

and etc. [83].

Features are extracted from the PQRST complexes (heartbeats). Therefore after filtering, the signals

were segmented into heartbeats and were aligned from their respective R peaks. The QRS detection

method described in [67] was employed to extract the heartbeats. Since the duration of the heartbeats

varies among subjects, all the heartbeats were fixed to have the same length. Moreover, the outlier


Figure 4.2: Heartbeats of a subject under different postures.

heartbeats were detected and discarded by measuring their Euclidean distance from the median heart-

beat.

The heartbeats were then normalized to have a dynamic range of one using the following formula:

hbnorm =hb−min(hb)

max(hb)−min(hb)

Furthermore, every ten heartbeats were time averaged to further reduce the amount of high frequency

noise in the heartbeats. Also in real life scenarios it might not be very useful to detect the posture of a

person every second.

4.3.1.2 Feature Extraction

A set of different feature spaces are extracted from time and frequency domains of the preprocessed

signal:

• Heartbeat: The heartbeats extracted and normalized in the preprocessing step are used as a

feature space (Figure 4.2).

• Normalized Autocorrelation: The filtered ECG signal was windowed with an overlapping

rectangular window of 10 seconds (with 50% overlap). Then the normalized autocorrelation (AC)


Figure 4.3: Normalized autocorrelations of a subject under different postures.

of each window was estimated using:

Rxx[m] =

N−|m|−1∑i=0

x[i]x[i+m]

Rxx[0]

where x[i] represents the windowed ECG, x[i + m] is the delayed version of ECG window with a

time lag of m = 0, 1, . . .M − 1 and M = 70 (Figure4.3).

• Fiducial: This feature space corresponds to wave amplitudes and durations extracted from a

heartbeat. P, Q, R, S and T wave amplitudes and also S-R, T-P, T-Q, T-S and T-R durations

were used as the feature space. Figure 4.4 shows the features extracted from a sample heartbeat.

The R peak was detected similar to the heartbeat extraction step described in the preprocessing

step by using QRS detection algorithm in [67]. The Q and S amplitudes were detected by finding

the minimum value between the R peak and 30 msec before and after the R peak time index

respectively. For T and P peaks, a time window that approximately contains the respective waves

was set heuristically and the maximum value in those windows were detected as the P and T wave

amplitudes. The time interval between peaks are normalized by the sampling frequency.

• Frequency content: Fourier Transform (FT) and multilevel discrete wavelet transform (DWT)

of the heartbeats were also considered for feature extraction. For DWT, the level-3 approximation

coefficients were chosen as a feature space which was empirically found to yield optimal results in


Figure 4.4: Fiducial features of a subject extracted from its heartbeat.

terms of detection rate. Daubechies scalar wavelet (Db3) was used as the wavelet function. For

FT, the magnitude of the discrete Fourier transform of the heartbeat was used as a feature space.

(Figure 4.5 and 4.6)

4.3.1.3 Classification

Support Vector Machine (SVM) was used for classifying each posture’s feature vectors. SVM is a

supervised learning model that constructs hyperplanes with the largest margins between two classes.

SVM is used for classes that can not be divided linearly. It maps the data to a higher dimension, using

a kernel, where they can be separated linearly. SVMs were originally designed for binary classifications,

therefore we used an extended multiclass SVM using one-against-all method [36] with the linear kernel

function.

4.3.2 Evaluation

4.3.3 Dataset

In this work we used the UofT ECG Database (UofTDB) recordings. There are 56 subjects in UofTDB

that have recordings in three different body postures namely sit, stand and supine. For the sit position

the subjects were asked to sit on a chair and rest for 2 minutes before the recording procedure starts. In

the standing position, the subjects were standing still during the recording and in the supine position

the subjects were asked to lay back and relax. Under supine posture the subjects completely leaned

forward while sitting on a chair. The recordings were captured from subjects fingertips similar to Lead-I


Figure 4.5: Fourier transform of a subject under different postures.

Figure 4.6: Level three approximation coefficients of discrete wavelet transform of a subject’s heartbeat underdifferent postures.


hhhhhhhhhhhhhhhPostureFeature Space

Wavelet AC FT Heartbeat Fiducial

Stand & Supine 99.5% 98.5% 97% 99.50% 88.89%Stand & Sit 97.5% 96.5% 96.5% 100% 90.83%Sit & Supine 98% 96.5% 96% 99.5% 91.96%Sit, Stand & Supine 97% 94.3% 96.67%Sit, Stand, Supine & Tripod 97%

Table 4.1: Personalized testing case.

configuration. The length of the recordings varies from 2-5 minutes and the device sampling rate was

set to 200 Hz.

4.3.4 Results

For each subject 80 heartbeats were selected for this analysis: 40 heartbeats for training the SVM

classifier and 40 heartbeats for testing. The proposed method is tested under two different scenarios.

The first scenario is the personalized case where the ECG signal of the user under different body postures

is known to the system. This assumption mainly applies to ECG biometric systems or systems that need

to learn the specific ECG morphologies of the user. The SVM classifier is trained and tested for each

subject in the database separately.

The other testing scenario is the generic case where the system does not have a prior knowledge of

the users’ ECG under different postures. The system is trained using a generic database to learn how to

differentiate the body postures for all subjects. Leave-one-out cross validation was employed to form the

testing and the training sets: out of the 56 subjects in the database 55 subjects were chosen for training

and one subject for testing. This process was repeated 56 times in order to test the method with all the

subjects in the database. For both testing scenarios, the average detection rates are reported in Table

4.1 and Table 4.2 where the empty cells are due to the fact that the SVM classifier did not converge in

those cases. Clearly the personalized testing case produces much higher detection rates than the generic

testing case which is due to the fact that ECG variability under posture changes is not uniform across

different subjects. In the generic testing case, differentiating sit and supine from standing posture is

easier than differentiating sit from supine. In the personalized testing case, features from the wavelet

transform of the heartbeats produced a detection rate of 97% which will be used in the following section

for ECG state-detection verification model.


hhhhhhhhhhhhhhhPostureFeature Space

Wavelet AC FT Heartbeat Fiducial

Stand & Supine 78.29% 66.95% 63.18% 71.32% 66.67%Stand & Sit 71.38% 68.62% 68.12% 69.57% 65.83%Sit & Supine 56.38% 53.44% 48.52% 54.26% 56.25%Sit, Stand & Supine 48.51%

Table 4.2: Generic testing case.

Figure 4.7: System block diagram of state-detection verification system.

4.4 Posture Detection For ECG Biometrics

The experimental results of the previous chapter show that the performance of the biometric system is

decreased when the ECG signals used for training and testing are from different body postures. One

solution is to primarily detect the posture and then identify the individual.

Consider an enrolment protocol in which the user’s ECG signal is collected under sit, stand, supine

and tripod postures. In real-world settings when the system is in normal verification operation the

underlying posture of the user may be random and unknown to the system.

In the state-detection verification method described in chapter 3, the biometric template is composed

of ECG signals under all postures but the biometric algorithm is trained individually for each posture.

Using the UofTDB database, for instance, four different AC/LDAs are trained, each corresponding to

a different underlying posture. In addition, the system is trained per subject for posture detection by

using the proposed method.

The projection matrices and the projected feature vectors collectively form a subject’s biometric

template. During verification, as shown in Figure 4.7, the posture of the user is first classified prior

to biometric matching. Then the system compares the test biometric signal with the detected posture

template of the claimed identity. Using the 56 subjects for whom recording under all postures (sit, stand,

supine and tripod) are available, the system was trained and tested as follows: for each subject, the first

half of the signal for each posture was used for training and the second half was used for testing.

For posture detection, the time-averaged heartbeat was used for testing. The posture detection rates

for sit, stand and supine were 98.04% and 94.12% for tripod. Figure 4.8 shows the ROC curves for


0 1 2 3 4 50

2

4

6

8

10

12

14

16

18

20

False Accept Rate (%)

Fa

lse R

eje

ct

Ra

te (

%)

Test: sit

Test: stand

Test: supine

Test: tripod

Average ROC

Figure 4.8: ROC curves of state-detection verification system under different test postures.

state-detection verification. The average EER of the system reduced to 1.86% which is very close to

the case where the test and enrol ECG are from same posture (1.50%). This great improvement can be

explained by the high classification rates of the posture detection method.

4.5 Conclusion

One of the main challenges of working with the ECG signal as a biometric is susceptibility to physical

activity. To our knowledge there is no article in the ECG biometric research devoted to studying the

impact of body posture changes, which may occur frequently in practical use. State-detection verifica-

tion model explained in chapter 3 can ultimately enhance ECG biometric templates and increase the

performance of the system. In this chapter we evaluated the possibility of detecting subject postures

based on different features extracted from ECG signal. Features from the discrete-wavelet transform of

heartbeats yielded an average detection rate of 97% with postures from sitting, stand, supine and tri-

pod. Furthermore we showed that by using the proposed posture-detection method in the state-detection

algorithm the performance of the overall system improves to 1.86% EER.

Chapter 5

ECG Recognition in Subspaces

5.1 Introduction

In the context of ECG biometrics, the feature vectors represented as high dimensional vectors, often

belong to a subspace of intrinsically low dimension. The value of a time sample in a heartbeat is highly

correlated with the value of the surrounding time samples. Therefore not all the samples in a heartbeat

are needed to represent the actual identity of a subject. While the actual heartbeat belong to the ECG

space, the identity information lie on a subspace of possibly lower dimension called identity-space. The

goal of dimensionality reduction techniques are to find the dimensionality of the identity-space and to

find a projection matrix to project the ECG data to the identity-space.

Researchers have employed different dimensionality reduction techniques to the extracted features

from the ECG signal for a variety of reasons such as reducing the computational cost of matching, re-

ducing the sample complexity for non-parametric methods (number of examples required to efficiently

estimate the underlying distribution of the data) and to avoid the ill-posed problem in practical appli-

cations of parametric methods.

While Chapter 3 provides a systematic comparison of the state of the art ECG biometric methods,

chapter 5 and 6 aim to empirically compare different subspace methods in ECG biometrics. The objective

is not to improve the state of the art performance by exploring different feature spaces and optimizing

classification methods, but to investigate which subspace provides a better representation of ECG signal

in terms of performance.

46

Chapter 5. ECG Recognition in Subspaces 47

5.2 Notations

Unless otherwise stated, this chapter assumes the following notations. Upper letter case letters denote

matrices (X) and the identity matrix is represented by I. Bold lower-case letters (x) denote vectors while

lower-case letters that are not bold (x) are scalar values. Finally E{.} and var(.) are the expectation

and the variance operators respectively.

5.3 Linear Subspaces

Linear subspaces arise from the assumption that the identity-space is a linear subspace of the ECG space.

In this section we review the three popular linear subspace methods namely the Principal Component

Analysis (PCA), Independent Component Analysis (ICA) and Linear Discriminant Analysis (LDA).

5.3.1 Principal Component Analysis

Principal Component Analysis (PCA) [47] is an orthonormal dimensionality reduction techniques that

aims to project data to a lower dimensional subspace, such that most of the energy in the signal (variance)

is captured. PCA is an orthogonal type of transformation which means that the set of possibly correlated

variables are linearly uncorrelated in the projected subspace. As shown in Figure 5.1, PCA finds the

first principal axis φ1 that corresponds to the direction with maximum variance and the second principal

component,φ2 , is found by the orthogonality constraint of PCA. The n-th principal component is the

linear combination with the maximum variance constrained to be orthogonal (uncorrelated) to the n-1

first principal components. We assume the input vector is zero mean:

E{x} = 0.

Therefore for the first principal component one needs to find the projection vector α1 such that the

projected data has the maximum possible variance:

maximizeα1

var(α1Tx) = α1

TΣα1

subject to α1Tα1 = 1.

(5.1)


where Σ represents the covariance matrix of the input data. The n-th principal component, αn, should

also maximize αnT∑

αn and be orthogonal to the n-1 first principal components:

maximizeαn

αnTΣαn

subject to αnTαn = 1.

α1Tαn = 0

...

αn−1Tαn = 0

(5.2)

One can show that the solution to equations 5.1 and 5.2 are the eigenvectors of the input covariance

matrix Σ. In other words, the n-th principal component is the eigenvector of Σ that corresponds to

the n-th largest eigenvalue. It can also be shown that the variance of the data projected onto the n-th

principal component is the same as the eigenvalue λn.

In order to perform PCA and project data onto its k principal components, one needs to project

the data onto A which is formed by k eigenvectors of the data’s covariance matrix that correspond to

the first k largest eigenvalues. Let X be the P ×M data matrix where its columns are the observation

vectors x1, . . . ,xM. The linear transformation from a p-dimensional space X to a k-dimensional space

Y with k < p is described as

Y = ATX, (5.3)

PCA can also be implemented by Singular Value Decomposition (SVD) of the observation matrix X.

SVD of a M × P (M ≥ P ) matrix X is:

X = UDVT , (5.4)

where U is a M × P ,D is P × P and V is P × P matrix.U and V have orthonormal columns, D is a

diagonal matrix which contains the square root of the eigenvalues of XXT on its main diagonal:

XXT = UDVT (UDVT )T = UDVTVDUT = UD2UT (5.5)

From the Equation 5.5 we can observe that the columns of U are the eigenvectors for the sample

covariance matrix XXT , i.e. U = A. Therefore there is no need to compute the covariance matrix which

can result in loss of precision specifically when the number of observations M is much smaller than the

dimension p.


PCA is closely related to Karhunen-Loeve Transform (KLT) [54] which was derived in the context

of signal processing. KLT is also an orthogonal transform which aims to minimize the average L2

reconstruction error.

ε(x) =

∥∥∥∥∥x−k∑i=1

(αTi x)αi

∥∥∥∥∥2

(5.6)

Under the zero mean assumption of the input data, the formulations of PCA and KLT are identical.

Therefore if the most of the energy (variance) of the signal is concentrated in the first k-principal

components, then the error of the approximate reconstruction of the data is small.

5.3.1.1 Optimum Dimension

One question that remains unsolved is the optimum choice of k- the number of principal components.

This is rather an ”engineering” problem and there is no analytical solution to find the optimal choice of

k. However eigenspectrum can be used as a tool to find k. Eigenspectrum is a plot of the eigenvalues of

the data covariance matrix Σ in an descending order, i.e. λ1 > λ2 > . . . > λP . Since the n-th eigenvalue

represents the variance along the n-principal component, a good choice of k can be where the eigenvalues

drop significantly.

Also it should be noted that in practice, the choice of k can be also constrained by computational

complexity such as the cost of matching the projected principal components.

5.3.1.2 Summary of Principal Component Analysis

PCA is an orthonormal dimensionality reduction technique that projects the input data x from RP → RK

where K << P . The main goal of PCA is to project data to subspace RK such that the maximum

energy of the signal is retained. The three main properties of PCA are the approximate reconstruction,

decorrelation of the data and orthogonality of the basis:

x ≈ Ay

E{yiyj} = 0 ∀i, j, i 6= j

ATA = I

(5.7)

5.3.2 Linear Discriminant Analysis

Linear Discriminant Analysis (LDA) [13] is a supervised dimensionality reduction technique that aims

to find a linear subspace that maximizes the class separability. Given that the observation vectors x


Figure 5.1: PCA of a multivariate Gaussian distribution centred at (1.01,1.01) with variance of 3.95 in directionof (0.705,0.709) and variance of 0.1 in the orthogonal direction.

belong to one of the classes c = 1, . . . , C, LDA finds an orientation for which the projected data are ”well

separated”. Consider a two class example in Figure 5.2 where we want to project the two dimensional

data to one dimension without loss of discriminative features. PCA tends to project in a direction that

the overall variance of the observation is maximized however LDA’s projection direction is in favour of

improved class separability. In order to formulate a class separability criterion for LDA, one needs to

consider the distance between class centres as well as the sample covariance matrix of each individual

class. The Fisher Criterion for class separability is described by

J = tr(Sw−1Sb) (5.8)

where

Sw =

C∑i=1

∑x∈Xi

(x− µi)(x− µi)T

is the within-class scatter matrix, and

Sb =

C∑i=1

Ni(µi − µ)(µi − µ)T


Figure 5.2: Projection vectors of PCA and LDA for the case of samples from two classes. PCA vector is in thedirection of maximum overall variance in the data whereas LDA finds a direction so that the class separabilityis maximized.

is the between-class scatter matrix; µi is the mean of class i and µ is mean of all classes and Ni is the

number of samples per class i. Using the linear transformation (Y = ATX), the scatter matrices in the

projected spaces will be:

Sw = ATSwA (5.9)

Sb = ATSbA (5.10)

Therefore the problem is to find the projection matrix A such that the Fisher Criterion is maximized in

the projected space:

max J(Y) = tr(Sw−1

Sb) = tr(ATSbAATSwA

) (5.11)

It can be shown that the columns of the projection matrix A solution are the eigenvectors of S−1w Sb.

Rank of Sb is at most C-1 while Sw is full rank. Therefore there are at most C-1 none-zero eigenvalues

for S−1w Sb that results in a C-1 dimensional subspace at most. Also S−1w Sb is not necessary symmetric,

therefore, unlike PCA, the projection matrix A need not be orthogonal. LDA solution is statistically

optimal when the distributions of all classes are Gaussian with the same covariance matrices but different

means.


5.3.2.1 Small Sample Size Problem

One of the main drawbacks of LDA is the Small Sample Size Problem (SSS) which arises from the fact

that in most applications Sw is singular. The reason of the singularity is because the dimension of

the feature vectors x is typically larger than the number of observations in the training set hence the

eigenvalue problem of S−1w Sb cannot be solved directly. This problem is mainly studied in the context of

face recognition where the dimension of input image is much larger than the number of training samples

and each class is represented by an individual. In this section we give an overview of the techniques used

in literature for overcoming the SSS problem:

• Regulized LDA: Zhao et al. [103] proposed to add a small perturbation to the within-class scatter

matrix. The diagonal matrix κI with κ > 0 and I an identity matrix, is added to the between-

scatter matrix. Since Sw is positive semidefinite, Sw + κI is non-singular and they show that the

resulting matrix has the same eigenvectors as Sw hence the projection matrix does not change.

• Fisherfaces: Belhumeur et al. [10] proposed to solve the problem by first reducing dimension of

the input images by using PCA so that Sw is no longer singular. Then LDA can be applied to

further reduce the dimensionality to C-1 and obtain the discriminative features.

• Complete LDA + PCA: Yang and Yang [96] argue that in the step where PCA is applied the

small principal components may include important discriminative information. They prove that

without loss of discriminative information, the optimal projection matrix can be derived from the

null space of Sw. In fact, they first project the Sb onto the null space of Sw. The columns of

the final transformation matrix are the eigenvectors of the projected between-class scatter matrix

corresponding to the largest eigenvalues.

• Direct LDA (DLDA): Proposed by Yu and Yang [101], the authors argue that the null space

of Sb contains no useful information. They purpose to discard the null space of Sb and at the

same time keep the important null space of Sw. To perform DLDA [101], first Sb is diagonalized

and then its null space is discarded. Second Sw is projected onto this new space. Finally, the

eigenvectors corresponding to the largest eigenvalues of the projected within-class scatter matrix

are selected as the final transformation matrix.

5.3.3 Independent Component Analysis

Independent Component Analysis (ICA) [20] is closely related to the problem of blind source separa-

tion. Consider a p-dimensional zero-mean observation random variable x generated from p different


sources,s = [s1 . . . sp]T , via the mixing matrix A:

x = As (5.12)

ICA is a generative model since the model describes how the observations are generated. Essentially the

goal of ICA is to estimate A and s from x. W = A−1 is called the separating matrix which is required

to find the sources from the observation data. In order to estimate A, ICA makes four assumptions:

1. si are statistically independent: in many applications this is a reasonable assumption. For example

in the context of speech where there are many sources talking simultaneously in a room, it is usually

assumed that the speech waveforms are independent.

2. The independent components, si, are non-Gaussian.

3. There is no prior knowledge on the distribution of si.

4. A is square.

The key to estimating the ICA model is non-Gaussianity of independent Components (ICs). From the

Central Limit Theorem (CLT), under certain conditions the distribution of the sum of independent

random variables tend toward a Gaussian distribution. Loosely speaking the sum of two independent

random variables has a distribution that is closer to a Gaussian distribution than the original random

variable distributions.

For simplicity lets assume ICs have identical distributions. To estimate one of the ICs:

y = wTx =∑i

wixi (5.13)

where w is a vector to be determined. Let z = ATw, then

y = wTx = wTAs = zT s. (5.14)

Therefore y is a linear combination of si with weights given by zi. By CLT y is more Gaussian than any

si and is least Gaussian when only one of the components of z is non-zero. So if we take w as a vector

to maximize the non-Gaussianity of wTx, such a vector would necessarily correspond (in a transformed

coordinate) to a z which has only one non-zero component. That is y = wTx = zT s equals one of the

independent components. There are two measures of non-Gaussianity:


1. Kurtosis: is the fourth order cumulant of a distribution and measures the degree of its peakedness:

kurt(y) = E{y4} − 3(E{y2})2 (5.15)

The kurtosis for Gaussian distributions is zero and for most of the non-Gaussian random variables

kurtosis is non-zero. Distributions with positive kurtosis are spikier than of a Gaussian distribution

and distributions with negative kurtosis are flatter than a Gaussian distribution. Therefore the

absolute value of kurtosis can be used as a measure of non-Gaussianity, however since kurtosis is

very sensitive to outliers [37] it is not a very robust measure of non-Gaussianity.

2. Differential Entropy and Negentropy: A fundamental result of information theory is that

over the entire real axis (−∞,+∞), a Gaussian random variable has the largest entropy among

all the random variables of equal variance [23]. Therefore entropy could be used as a measure of

non-Gaussianity. The differential entropy H of a random vector with probability density f(y) is

defined as [23]:

H(y) = −∫f(y) log f(y)dy (5.16)

However in practice a modified version of entropy is used as a measure of non-Gausianity called

negentropy:

J(y) = H(ygaussian)−H(y) (5.17)

where ygaussian is a Gaussian random vector of the same covariance matrix of y. Negentropy is

always non-negative and is zero when y is Gaussian. Estimating negentropy requires the estimation

of the pdf which can be computationally difficult. However there are simpler approximations of

negentropy that can be used in practice [38]:

J(y) ≈p∑i=1

ki[E{Gi(y)− E{Gi(g)}}]2 (5.18)

where ki are some positive constants, g is a Gaussian random variable with zero mean and unit

variance and y is assumed to also have a zero mean and unit variance, and Gi are some non-

quadratic functions such as:

G1(u) =1

alog cosh au, G2(u) = − exp(−u2/2) (5.19)

where 1 ≤ a ≤ 2 is some constant.


A very efficient implementation of ICA called FastICA is based on the equation 5.18 [33] [39]. The

algorithm is based on a fixed-point iteration for finding the maximum non-Gaussianity of y = wTx

based on the negentropy approximation in (5.18). There two other notable ICA algorithms called

Jade [16] and InfoMax [11], however we use the FastICA algorithm in this work.

5.3.3.1 ICA Properties

As discussed in the previous section ICA yields a linear projection RP → RK that provides the indepen-

dent components. Like PCA, ICA aims to minimize second order dependency of the data as well as the

higher order dependencies. The main properties of ICA are:

x ≈ Ay

P (y) =

K∏i=1

p(yi)

ATA 6= I

(5.20)

approximate reconstruction, near factorization of the joint distribution P (y) into marginal distribu-

tion of non-Gaussian ICs and non-orthogonal projection.

5.4 Nonlinear Subspaces

In many instances a linear projection cannot preserve the shape of the data. As shown is Figure 5.3, the

data is best summarized by a principal curve. The inverse image of principal manifolds in the original

space RP is a non-linear lower dimensional surface that minimizes the sum of total distance between

data samples and their projections on the surface.

Autoencoders can be employed for computing the non-linear principal manifolds [50]. Autoencoders

are feed-forward multi-layer neural networks that are trained to reproduce the input as accurately as

possible. For dimensionality reduction purposes, one of the hidden layers so called the ”bottleneck” layer

contains lower number of units than the input (Figure 5.4). The first half of the network, represented

by h(x), finds a lower dimensional principal manifold whereas the the second half, g(y), is an inverse

mapping of the projected data to the original space that minimizes the reconstruction error. The

principal curve in Figure 5.3 was generated with a 2-5-1-5-2 layer neural network of type in Figure 5.4.

One of the problems with the autoencoder implementation of the non-linear PCA is that it requires

prior knowledge of the network architecture and is vulnerable to over-fitting.


Figure 5.3: A principal curve represents the underlying structure of data.

5.4.1 Kernel methods

It is possible to make linear methods, such as LDA and PCA, nonlinear via a nonlinear mapping of the

input vectors into a high-dimensional feature space ∈ RL. Applying standard linear methods in the new

space is equivalent to applying nonlinear methods in the original space. By utilizing the so-called kernel

trick one can avoid the computation of direct nonlinear mapping. The dot product in the feature space

is expressed as kernel function evaluation in the input space and therefore all the computations can be

performed in the lower dimensional input space.

5.4.1.1 Kernel PCA

The idea of Kernel PCA is originated from the ”kernel eigenvalue” method of Scholkopf et al. [75] that

is to map the input data to a higher dimensional space (possibly infinite )through a nonlinear mapping

function and perform PCA in the new dimension. Let φ be the nonlinear mapping function from the

input space RP to the new space RL where L >> P and is possibly infinite. Let the upper-case characters

represent the elements of RL and lower-case characters are used for elements of RP :

φ(xi) = Xi


Figure 5.4: An auto-associative neural network for computing principal manifolds. Number of elements in thebottleneck represent the dimensionality of the principal curve.

Assuming the data is zero mean in RL, the covariance matrix for M observations xi ∈ RP is then:

C =1

M

M∑i=1

φ(xi)φ(xi)T (5.21)

The principal components are then eigenvectors V of the covariance matrix C that corresponds to the

largest eigenvalues λ

λV = CV (5.22)

Since CV = 1M

∑Mi=1(φ(xi).V)φ(xi), all the solutions V must lie in the span of the training data

φ(x1), . . . ,φ(xM). Hence there exist coefficients αi such that

V =

M∑j=1

αjφ(xj) for all j = 1, . . . ,M (5.23)


Also ((5.22)) must be true for each training sample:

λ(φ(xi).V) = (φ(xi).CV) for all i = 1, . . . ,M. (5.24)

By defining the M ×M matrix K:

Kij = (φ(xi).φ(xj)) (5.25)

and combining (5.23) and (5.24) results in:

MλKα = K2α (5.26)

where α is a column vector with entries α1, . . . , αM .Note that K is symmetric positive semidefinite and

its eigenvectors span the whole space, hence

Mλα = Kα (5.27)

gives the solutions α. Therefore one needs to diagonalize K to obtain the complete set of eigenvalues

λ1 > . . . > λM and the corresponding eigenvectors α1, . . . ,αM . Orthonormality requirements of the

eigenvectors (Vk.Vk = 1) leads to the normalization condition for α:

λk(αk.αk) = 1. (5.28)

In order to compute the dot product in (5.25) (φ(xi).φ(xj)), one can use the kernel representation

of the form

k(xi,xj) = (φ(xi).φ(xj)), (5.29)

where kernel evaluation k(xi,xj) in the input space RK corresponds to dot product in higher dimensional

feature space RL. This trick , also known as kernel trick, avoids the direct computation of φ(x) which can

be prohibitively expensive. The kernel function needs to satisfy the Mercer’s theorem [21]. Popular kernel

functions include Gaussians exp (−(xi − xj)T (xi − xj)), sigmoids tanh (a(xi.xj) + b) and polynomials

(xi.xj)d.

The principal components of any input vector can be efficiently computed by kernel evaluation of

the dataset. The n-th principal component yn of input x is given by

yn = Vn.φ(x) =

M∑i=1

αni k(x,xi) (5.30)


Note that the number of principal components (eigenvectors) is limited to M which is typically more

than the input dimensionality of the data.

In summary, to perform the kernel PCA method for input observations xi for i = 1, . . . ,M , the

following steps need to be carried out:

1. Select an appropriate kernel function k(.) and compute the kernel matrix K using (5.25) and

(5.29):

Kij = k(xi,xj) (5.31)

2. Obtain the eigenvalues λ1 > . . . > λM and the corresponding eigenvectors α1, . . . ,αM of K.

3. Normalize the eigenvectors αn required by Equation (5.28)

λk(αk.αk) = 1.

4. The n-th principal component yn of a test point x is computed by projecting onto the n-th eigen-

vector by Equation 5.30

yn = Vn.φ(x) =

M∑i=1

αni k(x,xi)

5.4.1.2 Kernel LDA

One can extend the idea of LDA to kernel LDA by applying LDA in the feature space. Mika et al. [74]

formulated the kernel Fisher discriminant for two-class cases and Baudat and Anouar [9] proposed the

generalized kernel discriminant analysis for multi-class problems. As discussed in Section 5.3.2.1 SSS

problems are very common in the input space due to small number of observations compared to the

dimensionality of the data . Also due to the implicit high dimensional mapping in the kernel methods,

many large sample size problems in the input space turn into SSS in the feature space.

Lu et al. [58] have taken this issue into account and developed kernel direct discriminant analysis

(KDDA) by generalization of the direct-LDA [100] discussed before. Similarly Yang et al. [95] adapted

the idea of using KPCA+LDA from PCA+LDA. The KPCA+LDA is performed in two steps, in the

first step via KPCA the observations are transformed from RP to RT using the normalized eigenvectors

of the kernel matrix K that correspond to its T largest eigenvalues. The second step involves performing

LDA in the KPCA transformed space RT .


5.5 Conclusion

This chapter provides a review of popular subspace dimensionality reduction methods namely the PCA,

LDA, ICA and nonlinear PCA and nonlinear LDA. The linear methods aim to find a linear subspace

from the input space that optimizes some criteria. The goal of PCA is to find an orthonormal projection

matrix where most of the variance is captured in the top principal components. While the vectors in the

PCA subspace are uncorrelated, in the ICA subspace the vectors are statistically independent. In other

words, ICA minimizes the second-order and higher order dependencies of input data. LDA is a supervised

form of dimensionality reduction technique where a class label is associated with each observation. Based

on Fisher discriminant criterion, LDA finds a linear subspace such that the ”separability” of classes are

optimized. Finally this chapter reviewed the nonlinear variations of PCA and the kernel methods for

LDA and PCA.

Chapter 6

Empirical Comparison of Subspace

Methods

6.1 Introduction

Although the electrocardiogram (ECG) has been a reliable diagnostic tool for decades, its deployment in

the context of biometrics is relatively recent. Its robustness to falsification, the evidence it carries about

aliveness and its rich feature space has encouraged the deployment of ECG based biometrics in real life

applications. The rich feature space contains unique information such as characteristics which reflect

the underlying physiological properties of the heart. The principal goal of this study is to quantitatively

evaluate the information content of different feature spaces in terms of their effect on verification and

identification accuracy.

To the author’s knowledge, no systematic study has been published reporting the relative ability of

different subspace methods to sufficiently represent the ECG signal.

In this chapter we evaluate several subspace methods, the linear methods are PCA, LDA, ICA while

the nonlinear methods are the kernel methods KDLDA and KPCA. Since the effort is on comparing

the performance of different subspace methods, we did not explore different classification and feature

spaces for optimizing their performances. Heartbeats extracted from the ECG signal is used as the input

feature space and the Log Likelihood Ratio Test (LLRT) for classification (except for ICA). The results

are reported in both verification and identification modes of operations.

61

Chapter 6. Empirical Comparison of Subspace Methods 62

6.1.1 Preprocessing

Similar to Chapter 3, the ECG signal of the subjects were first filtered by using a 4th order Butterworth

filter with cut-off frequencies at 0.5Hz and 40Hz. In this evaluation we use time-samples of heartbeats

as the input feature space. By using the Pan and Tomkins [67] QRS detection algorithm, the heartbeats

were extracted and aligned from their respective R peaks. Then each heartbeat, hb, was normalized to

be zero mean and have a dynamic range of one:

hbnormalized =hb− E{hb}

max(hb)−min(hb)(6.1)

Because the heartbeat duration is different among subjects, it was decided to fix the heartbeat durations

to 550 msec with 200 msec before the R peak. This duration ensures that each heartbeat contains all

the main waves (P, QRS and T waves) and also does not include any waves from adjacent heartbeats.

The outlier heartbeats were detected and excluded from the dataset. For each heartbeat, its Euclidean

distance from the median heartbeat was estimated and the heartbeats with a distance greater than a

threshold τ were excluded. The threshold τ = 1 was selected empirically which ensures the exclusion of

heartbeats that are corrupted by a large amount of noise.

There are two main sources for outliers, the first is due to noise and motion artifacts that cause

corruption in the signal. The second source is due to error in the QRS detection algorithm. For instance,

subjects that have strong T-waves relative to their QRS-complex, the algorithm may mistakenly detect

the T-wave as the R-peak.

6.1.2 Classification

Log Likelihood Ratio Test (LLRT) classifier used in Odinaka et al.’s [64] ECG biometric algorithm is

found to outperform the Euclidean classifier for almost all of the subspace methods except for ICA.

Since the focus of this chapter is not on improving the classification methods, we use LLRT classifier for

evaluating the performance of most of the methods and the Euclidean classifier for ICA.

Each projected heartbeat in a subspace was modelled as a multivariate Gaussian distribution. Using

Maximum Likelihood Estimation (MLE), the sample mean and sample covariance matrix was used to

estimate the mean vector µj and covariance matrix Cj for subject j.

µj =1

N

N∑i=1

xi,j, Cj =1

N − 1

N∑i=1

xi,j.xTi,j, (6.2)

where x denotes the projected heartbeat in a subspace and N is the total number of training examples


per subject.

Let uj represent an unknown vector with a claimed identity of subject i for j = 1, . . . , T . Then LLRT

matching score is computed as:

Λ =

T∑j=1

log

[p(uj|θi)p(uj|θ0)

]

where θi ∼ N (µi,Ci) is the distribution model of subject i, and θ0 ∼ N (µ0,C0) represents the nominal

distribution model in the database. Let K be the total number of subjects in the database, then the

nominal model is:

µ0 =1

N

K∑j=1

N∑i=1

xi,j, C0 =1

(K ×N)− 1

K∑j=1

N∑i=1

xi,j.xTi,j

In the verification mode of operation the final decision of the system (Accept/Reject) can be regulated by

a threshold τ . However in the identification mode, the system finds an identity match for the unknown

signal by calculating the matching scores, Λi, for every subject i in the database. The person with the

highest match score is then selected as the best match. As EER and ROC curves can be used to report

the algorithms performance in verification mode, Identification Rate (IR) ,which is the percentage of

correctly identifying subjects, can be used to evaluate the performance in the identification mode of

operation.

6.2 Experimental Results

6.2.1 Experiment Setup

As we saw in chapter 3, the accuracy of the systems under single-session analysis are very optimistic

and the obtained results are very different from the across-session analysis which is the case in most

real life application. Therefore in this chapter we focus on evaluating the subspace methods based on

across-session analysis where the enrol data and test data are from different sessions.

In the simulations, three sets of ECG recordings was extracted from the UofTDB: training set,

enrolment set and the testing set. The training set is used to train each subspace method and obtain

the projection matrices, while the enrolment and testing sets are used for evaluating the accuracy of the

methods. As it is very common to use the enrolment set for training, we also consider the effect of using

the training set and the enrolment set for training the algorithms.

From the single-session portion of UofTDB, a training set of 225 subjects with a total of 4500 heart-

beats was selected. There is no overlap between the individuals in the training set and the individuals


in the enrolment and the testing set. In this sense we are testing the ability of the methods to recognize

new subjects who were not part of the subspace computation with the training set.

There are 47 subjects in the UofTDB that have recordings in sitting posture from different sessions

(session two, three, four and six). For each subject there are 40 heartbeats for enrolment and 40

heartbeats for testing. Based on the four sessions available, there are 12 different testing combinations

where in each simulation the recordings from one session are used as the enrolment set and the recordings

from another session for testing. All the parameters of the subspace methods were fine tuned with data

from the first combination (enrolment set from session two and testing set from session three) and the

reported results are based on the average accuracy of the methods from all the 12 combinations.

Figure 6.3 shows the average ROC curve of the subspace methods namely the PCA, ICA, DLDA,

KPCA and KDLDA. The following sections provide the detailed parameters and performance of each

method.

6.2.2 Robust ECG Features for Across Sessions Variability

A heartbeat is composed of three different waves: P, QRS, and T waves. Although they all carry

important biometric information, not all of them are robust to across-session variabilities. Figure 6.1

shows the average verification accuracy when different waves are employed. The results are based on

using different parts of a heartbeat as the feature space. Note that the T-wave has the worst performance

among all, which can be due to its sensitivity to day to day heart rate changes whereas the QRS-complex

is the most robust feature to across-session variabilities. In fact a window of 355 msec (71 samples) with

150 msec before the R-peak resulted in an average EER of 16.13% which will be used as the input space

for the subspace methods. Also note that with the input dimension of 71 and 4500 training samples,

there are nearly 64 times as many the training samples as the data dimensionality. Thus the parameter

estimations for PCA, LDA, ICA, KPCA and KLDA were properly over-constrained.

6.2.3 PCA-based Recognition

The performance of PCA algorithm is used as the baseline for ECG recognition experiments. Figure

6.2 shows the verification performance of the system when different numbers of principal components

are used. The subspace dimensionality of k = 20 accounts for 72% of the total variance captured in

the subspace and beyond this point the verification accuracy does not improve. Therefore with PCA

dimensionality reduction from 71 to 20 results in no performance degradation. Based on the 12 trials,

an average EER of 16.21% was achieved and the mean IR was 63.44%. The PCA projection matrix was


Figure 6.1: The ECG verification accuracy under across-session analysis when different waves are used.

computed with the training set only and the enrolment and testing sets are from different sessions as

explained previously. However it was found that using the enrolment set for training does not improve

the performance of the system (mean EER = 16.72%, mean IR = 64.96%).

6.2.4 ICA-based Recognition

For ICA-based recognition a fixed point algorithm for ICA called FastICA was used which is based on

negentropy approximation. A PCA whitening step preceded the core ICA decomposition and the input

dimensionality was reduced to 20. Given the heartbeat observations, the objective is to compute the

separating matrix W in order to find the independent sources for each heartbeats. Therefore the ICA

subspace projection of the test set was obtained using y = Wx.

ICA assumes the sources are non-Gaussian and are statistically independent. However our classifier

assumes Gaussian distributions which contradicts with the assumption of non-Gaussian ICA sources.

For this reason, we used an Euclidean distance classifier and compared with the LLRT classifier. The

Euclidean distance classifier yields a mean identification rate of 76.51% with the highest IR of 81.82%

whereas it is 25% for the LLRT classifier which is very low. However the average EER of the LLRT

classifier is 18.78% and for the Euclidean classifier it is 18.77%.

Based on the reported results we can conclude that ICA outperforms PCA in identification by more

than 10% whereas there is no advantage of using ICA in the verification mode of operation. Similar to


Figure 6.2: The accuracy of ECG biometrics when different number of principal components are used. Thetop 20 principal components capture approximately 72% of the variance from the original space.

PCA-based recognition, training the ICA from the enrolment dataset did not improve the performance

of the system neither in verification nor in identification mode.

6.2.5 LDA-based Recognition

For LDA-based recognition, we use the classical Linear Discriminant Analysis (LDA) and compare it with

Direct LDA (DLDA) since it was found that the DLDA algorithm generally results in better performance

than RLDA, PCA-LDA and complete PCA+LDA methods discussed in the previous chapter. Based on

the training dataset, the within-class scatter matrix Sw is not singular since the input dimensionality

is much smaller than the number of training heartbeats (71 << 4500). However when training is done

with the enrolment dataset, Sw becomes singular and computation of the projection matrix based on

the classical LDA is not possible therefore we use the DLDA algorithm.

Using the training set for learning the subspace projection matrix, LDA yields a mean EER of 11.81%

whereas DLDA results in an average EER of 10.39% and an average IR of 78.59%. Although there is no

overlap between the training set and the enrolment set, the ”class separability” criterion of LDA methods

improved the overall verification performance significantly. We can conclude that for ECG recognition,

the direction of maximum variance (PCA) does not necessary result in optimal class separability and

there may be coordinates with small variances that contain important class discriminative information.


Furthermore when the enrolment set is used for training, the average EER of DLDA reduced to

8.67% however no improvement was found in the average identification rate. Unlike PCA and ICA,

the performance of DLDA enhances when the method is trained with the enrolment dataset which is

expected because it is a supervised dimensionality reduction technique and the subspace projection aims

to improve class separability.

It was found that DLDA provides a high performance up to a reduced dimension of 13, however for

PCA and ICA the minimum dimension was 20. DLDA based recognition not only outperforms the PCA

and ICA subspace methods, it also results in a more compact subspace representation of heartbeats.

6.2.6 Kernel-based Recognition

For KPCA and KDLDA, the parameters of polynomial, sigmoid and Gaussian kernels were first fine

tuned with the first combination (enrolment set from session two and testing set from session three).

For both KPCA and KDLDA, the Gaussian kernels were found to provide the best performance . For

each combination the kernel matrix was computed from the training dataset and the performances are

based on the LLRT classification.

For KPCA, based on a 20-dimensional subspace, the average EER was 14.35% and the average IR

was 67.8%. For KDLDA, also based on a 20-dimensional subspace, the average EER was 9.09% and

average IR was 75.75%. Both of the kernel methods improved the performance of the linear methods

to some degree however the performance improvement is not that significant for us to conclude that

nonlinear subspaces provide a better representation of the data.

6.3 Effectiveness of Template Updating

We further investigate the effectiveness of template updating on the performance of the subspace meth-

ods. Similarly the projection matrices are computed from the training set and the enrolment set is

composed of recordings from multiple sessions. For instance, the one session updating, the enrolment

data is from two sessions and the testing data is from a different session. In this case there are also 12

different combinations and for the two session updating case there are four combinations. Figure 6.4

and 6.5 show the verification and identification performances of the subspace methods when different

number of sessions are used for enrolment. The performance of all the methods enhances by template

updating and the performance of the DLDA method reaches an average 5.27% and an average IR of

92.61% under two session updating.

Clearly in all of the simulations, the DLDA method outperforms the other subspace methods and in


Figure 6.3: The ROC of different subspace methods under across-session analysis.

most cases outperforms the KDLDA method. Finally when the enrolment sets are used for training the

methods, only the performance of the DLDA method improves on average by 0.97% in EER and 2.59%

in IR.

6.4 Effect of Posture and Exercise

As we saw in chapter 3, there are other factors such as exercise and posture changes that affect the

performance of the system.

Figure 6.6 shows the performance summary of the DLDA method along with the baseline method

PCA under different conditions. Different posture analysis is based on a population of 54 subjects

from UofTDB under sitting, standing, supine and tripod postures. Similarly the training set was used

for training the methods and the enrolment set and testing set were from different postures. We also

considered the case where the enrolment and test data are from different postures and different sessions.

For post-exercise analysis, the enrolment set was from sitting posture and the testing set from the

post-exercise data. Figure 6.6 results are based on average EER. Clearly there is a large gap between the

performance under ECG variability and single-session single-posture ECG. The performance of DLDA

method under different postures and across sessions is approximately 12.62% and under different sessions


Figure 6.4: The verification performance of subspace methods with template updating.

Figure 6.5: The identification performance of subspace methods with template updating.


is 10.39%.

Figure 6.6: Performance summary of DLDA and PCA subspace methods.

6.5 Conclusion

In summary, this chapter provides a performance comparison of different subspace methods namely

the PCA, LDA, KPCA and KDLDA for ECG recognition. The simulations were performed on the

across-session portion of UofTDB in order to mimic real life scenarios. Direct-LDA (DLDA) method

achieved an average EER of 10.39% and average IR of 78.59% which was found to be the top performing

method. Under across-session template updating, the mean EER of DLDA reduced to 5.27% and mean

IR improved to 92.61%.

Based on the reported results we can conclude that ECG biometric templates need to be sampled

from different sessions. Although template updating boosts the performance of all biometric systems,

for ECG recognition this improvement is significant which may be due to characteristics of the ECG

signal. In other words, the ECG biometric templates that are sampled from one session are not reliable

which can be a fundamental limitation of ECG biometric systems.

Under posture variability combined with across-session variability, DLDA resulted in an EER of

12.62% and heart rate fluctuations due to exercise increased the average EER to 16.92%.

Chapter 7

Conclusion and Future Directions

This work presented a systematic and comprehensive analyses of the challenges associated with the

deployment of ECG biometric systems. While the field is gaining much attention, there has been limited

considerations to the practical aspects of using ECG biometric systems in the real-world.

Central to this work are the physiological factors that affect the performance of ECG biometric

systems. A new database (UofTDB) for the evaluation of such systems is presented to address the

shortcomings of the prior art. This database includes signals from 1020 individuals, recorded over

multiple experimental sessions and under various postures and bodily conditions. Furthermore, the

signals in this database have been acquired from the volunteers’ fingers thereby mimicking real-world

application settings.

Emphasis is placed on the differentiation between biometric performance reported for within-session

and across-session analysis. This work advocates that the practise of evaluating an ECG biometric

system using signals that are collected within a single recording session may not be indicative of the

system accuracy. In addition, this work illustrates a dependency of the accuracy to the body posture

and condition under which the ECG signals are captured. The analysis includes ECG samples under

sitting, standing, supine, tripod and exercise conditions and demonstrates that while the biometric

performance may be compromised when testing in such uncontrolled settings, there are treatments that

can be considered to potentially address this problem.

One of the main challenges of working with the ECG signal as a biometric is susceptibility to physical

activity. This paper investigated the robustness of ECG biometric authentication with body posture

changes. To our knowledge there are no articles in the ECG biometric research devoted to studying

the impact of body posture changes, which may occur frequently in practical use. As shown in this

71

Chapter 7. Conclusion and Future Directions 72

work when the posture at enrolment and verification are not identical, the performance of the system

drops significantly. We proposed a novel posture-detection verification system to reduce this drop in

performance. This system first detects the posture of the test signal and then verifies the identity of

the test signal. In this research, the multilevel-discrete wavelet coefficients are used in conjunction with

SVM for classification as a state-detection method. Further, this posture detection method is used in

the first step of the posture-detection verification system. In the next step the system is trained for each

posture to distinguish the subjects in the different postures.

We finally evaluated different linear and nonlinear subspace methods in the context of ECG recog-

nition. Based on different sets of simulations, it was concluded that the Direct Discriminant Analysis

(DLDA) method provides a better subspace representation of the ECG signal in terms of recognition ac-

curacy. It was also found that the QRS segment of a heartbeat is the most robust feature in a heartbeat

to across-session analysis.

7.1 Future Directions and Improvements

Based on the findings of this work the following are possible future directions that needs to be studied:

• While the present work demonstrates that there is great potential in utilizing ECG for biometric

recognition, it is prominent that future analyses are performed on even larger datasets of ECG

signals and across multiple years of recordings.

• There are other factors that affect the ECG signal that need to be addressed in future work. Mental

conditions, stress, diet, alcohol and physical training are a few to name.

• As we saw in chapter 6 ECG signal can be represented by a subspace of intrinsically low dimension

and this dimension was 13 for DLDA subspace method. In the field of biometrics this dimensional-

ity is very low whereas for fingerprint ,for instance, at least 150 dimensions are used for matching

(Bozorth3 fingerprint matching algorithm). This low cost of ECG matching can be very encourag-

ing for many applications such as the fusion of ECG and fingerprint specifically in the identification

mode of operation. For example the ECG biometric can select a subset of possible identity matches

and then fingerprint matching is done on a smaller subset. This indeed can increase the overall

identification speed.

Bibliography

[1] M. G. Adams and B. J. Drew. Body position effects on the ECG: implication for ischemia moni-

toring. J. Electrocardiol., 30(4):285–291, 1997.

[2] F. Agrafioti, F. M. Bui, and D. Hatzinakos. Medical biometrics: The perils of ignoring time

dependency. In IEEE 3rd Int. Conf. on Biometrics: Theory, Applications, and Systems, pages

1–6, Sept 2009.

[3] F. Agrafioti and D. Hatzinakos. ECG based recognition using second order statistics. In Proc.

Communication Networks and Services Research Conf. 2008, pages 82–87, 2008.

[4] F. Agrafioti and D. Hatzinakos. Fusion of ECG sources for human identification. In Proc. 3rd Int.

Symp. Communications, Control and Signal Processing (ISCCSP 2008), pages 1542–1547, 2008.

[5] F. Agrafioti and D. Hatzinakos. ECG biometric analysis in cardiac irregularity conditions. Signal,

Image and Video Process., 3(4):329–343, 2009.

[6] F. Agrafioti and D. Hatzinakos. Signal validation for cardiac biometrics. In Proc. 2010 IEEE Int.

Conf. Acoustics Speech and Signal Processing (ICASSP), pages 1734–1737, 2010.

[7] F. Agrafioti, D. Hatzinakos, and A. K. Anderson. ECG pattern analysis for emotion detection.

IEEE Trans. on Affect. Comput., 3(1):102–115, 2012.

[8] M. Astrom, J. Garcia, P. Laguna, O. Pahlm, and L. Sornmo. Detection of body position changes us-

ing the surface electrocardiogram. Medical and Biological Engineering and Computing, 41(2):164–

171, 2003.

[9] G. Baudat and F. Anouar. Generalized discriminant analysis using a kernel approach, 2000.

[10] P. N. Belhumeur, J. P. Hespanha, and D. J. Kriegman. Eigenfaces vs. fisherfaces: Recognition using

class specific linear projection. IEEE Transactions on Pattern Analysis and Machine Intelligence,

19(7):711–720, 1997.

73

Bibliography 74

[11] A. J. Bell and T. J. Sejnowski. An information-maximization approach to blind separation and

blind deconvolution. Neural Comput., 7(6):1129–1159, November 1995.

[12] L. Biel, O. Pettersson, L. Philipson, and P. Wide. ECG analysis: a new approach in human

identification. IEEE Trans. Instrum. Meas., 50(3):808–812, 2001.

[13] C. M. Bishop. Pattern recognition and machine learning. springer, 2006.

[14] O. Boumbarov, Y. Velchev, and S. Sokolov. ECG personal identification in subspaces using radial

basis neural networks. In IEEE Int. Workshop on Intelligent Data Acquisition and Advanced

Computing Systems: Technology and Applications(IDAACS 2009), pages 446–451, Sept 2009.

[15] F. M. Bui, F. Agrafioti, and D. Hatzinakos. Electrocardiogram (ECG) Biometric for Robust Iden-

tification and Secure Communication, pages 383–427. John Wiley & Sons, Inc., 2009.

[16] J. F. Cardoso. High-order contrasts for independent component analysis. Neural Comput.,

11(1):157–192, January 1999.

[17] A. D. C. Chan, M. M. Hamdy, A. Badre, and V. Badee. Wavelet distance measure for person

identification using electrocardiograms. IEEE Trans. Instrum. Meas., 57(2):248–253, 2008.

[18] Z. ChenXing, T. Wysocki, F. Agrafioti, and D. Hatzinakos. Securing handheld devices and finger-

print readers with ECG biometrics. In 2012 IEEE Fifth International Conference on Biometrics:

Theory, Applications and Systems (BTAS), pages 150–155, Sept 2012.

[19] C. C Chiu, C. M Chuang, and C. Y Hsu. A novel personal identity verification approach using

a discrete wavelet transform of the ECG signal. In 2nd Int. Conf. Multimedia and Ubiquitous

Engineering, 2008.

[20] P. Comon. Independent component analysis, a new concept? Signal Processing, 36(3):287 – 314,

1994. Higher Order Statistics.

[21] R. Courant and D. Hilbert. Methods of mathematical physics, volume 1. CUP Archive, 1966.

[22] D. P. Coutinho, A. L. N. Fred, and M. A. T. Figueiredo. One-lead ECG-based personal identifica-

tion using Ziv-Merhav cross parsing. In Proc. 2010 20th Int. Conf. Pattern Recognition (ICPR),

pages 3858–3861, 2010.

[23] T. M. Cover and J. A. Thomas. Elements of Information Theory. John Wiley & Sons, 1991.

Bibliography 75

[24] H. P. Da Silva, A. Fred, A. Lourenco, and A. K. Jain. Finger ECG signal for user authentica-

tion: Usability and performance. In IEEE Sixth International Conference on Biometrics: Theory,

Applications and Systems (BTAS), pages 1–8, Sept 2013.

[25] H. W. Draper, C. J. Peffer, F. W. Stallmann, D. Littmann, and H. V. Pipberger. The corrected

orthogonal electrocardiogram and vectorcardiogram in 510 normal men (frank lead system). Cir-

culation, 30(6):853–864, 1964.

[26] W. Einthoven, G. Fahr, and A. De Waart. On the direction and manifest size of the variations of

potential in the human heart and on the influence of the position of the heart on the form of the

electrocardiogram. American Heart Journal, 40(2):163 – 211, 1950.

[27] S. C. Fang and H. L. Chan. Human identification by quantifying similarity and dissimilarity in

electrocardiogram phase space. Pattern Recognition, 42(9):1824 – 1831, 2009.

[28] Y. Gahi, M. Lamrani, A. Zoglat, M. Guennoun, B. Kapralos, and K. El-Khatib. Biometric iden-

tification system based on electrocardiogram data. In New Technologies, Mobility and Security,

pages 1–5, Nov 2008.

[29] J. Garcıa, M. Astrom, J. Mendive, P. Laguna, and L. Sornmo. ECG-based detection of body po-

sition changes in ischemia monitoring. IEEE Transactions on Biomedical Engineering, 50(6):677–

685, 2003.

[30] N. Ghofrani and R. Bostani. Reliable features for an ECG-based biometric system. In 2010 17th

Iranian Conf. Biomedical Engineering (ICBME), pages 1–5, Nov 2010.

[31] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E.

Mietus, G. B. Moody, C. K. Peng, and H. E. Stanley. Physiobank, Physiotoolkit, and Physionet:

Components of a new research resource for complex physiologic signals. Circulation, 101(23):e215–

e220, 2000.

[32] M. Guennoun, N. Abbad, J. Talom, M. Rahman, and K. El-Khatib. Continuous authentication by

electrocardiogram data. In 2009 IEEE Toronto Int. Conf. Science and Technology for Humanity

(TIC-STH), pages 40–42, Sept 2009.

[33] Aapo H. and O. Erkki. A fast fixed-point algorithm for independent component analysis. Neural

Computation, 9:1483–1492, 1997.

Bibliography 76

[34] R. Hoekema, G. J. H. Uijen, and A. Van Oosterom. Geometrical aspects of the interindividual

variability of multilead ECG recordings. IEEE Transactions on Biomedical Engineering, 48(5):551–

559, 2001.

[35] M. Homer, J. M. Irvine, and S. Wendelken. A model-based approach to human identification using

ECG. In Proc. SPIE, volume 7306, 2009.

[36] C. W. Hsu and C. J. Lin. A comparison of methods for multiclass support vector machines. IEEE

Transactions on Neural Networks, 13(2):415–425, Mar 2002.

[37] P. J. Huber. Projection pursuit. The Annals of Statistics, 13(2):435 – 475, 1985.

[38] A. Hyvarinen. New approximations of differential entropy for independent component analysis

and projection pursuit. In Proceedings of the 1997 Conference on Advances in Neural Information

Processing Systems 10, NIPS ’97, pages 273–279, Cambridge, MA, USA, 1998. MIT Press.

[39] A. Hyvarinen. Fast and robust fixed-point algorithms for independent component analysis. IEEE

Transactions on Neural Networks, 10(3):626–634, May 1999.

[40] J. Irvine, S. Israel, M. Wiederhold, and B. Wiederhold. A new biometric: Human identification

from circulatory function. In Joint Statistical Meetings of the American Statistical Association,

2003.

[41] J. Irvine, B. Wiederhold, L. Gavshon, S. Israel, S. McGehee, R. Meyer, and M. Wiederhold. Heart

rate variability: A new biometric for human identification. In Proc. Int. Conf. Artificial Intelligence

ICAI’, 2001.

[42] J. M. Irvine and S. A. Israel. A sequential procedure for individual identity verification using ECG.

EURASIP J. Adv. Signal Process, pages 3:1–3:13, January 2009.

[43] J. M. Irvine, S. A. Israel, W. T. Scruggs, and W. J. Worek. eigenpulse: Robust human identification

from cardiovascular function. Pattern Recognit., 41:3427 – 3435, 2008.

[44] S. A. Israel, J. M. Irvine, A. Cheng, M. D. Wiederhold, and B. K. Wiederhold. ECG to identify

individuals. Pattern Recognition, 38(1):133 – 142, 2005.

[45] D. Jang, S. Wendelken, and J. M. Irvine. Robust human identification using ECG: eigenpulse

revisited. In Proc. SPIE, volume 7667, 2010.

Bibliography 77

[46] T. Jernberg, B. Lindahl, M. Hogberg, and L. Wallentin. Effects on QRS-waveforms and ST-T-

segment by changes in body position during continuous 12-lead ECG: A preliminary report. In

Computers in Cardiology, pages 461–464. IEEE, 1997.

[47] I. Jolliffe. Principal component analysis. Wiley Online Library, 2002.

[48] A. Y. M. Jones, C. Kam, K. W. Lai, H. Y. Lee, H. T. Chow, S. F. Lau, L. M. Wong, and

J. He. Changes in heart rate and R-wave amplitude with posture. Chinese Journal of Physiology,

46(2):63–70, 2003.

[49] K. S. Kim, T. H. Yoon, Lee J. W., D. J. Kim, and H. S. Koo. A robust human identification

by normalized time-domain features of electrocardiogram. In Proc. 27th Ann. Conf. IEEE Eng.

Medicine and Biology, pages 1114–1117, 2005.

[50] M. A. Kramer. Nonlinear principal component analysis using autoassociative neural networks.

AIChE journal, 37(2):233–243, 1991.

[51] M. Kyoso. A technique for avoiding false acceptance in ECG identification. In IEEE EMBS

Asian-Pacific Conference on Biomedical Engineering, pages 190–191, Oct 2003.

[52] M. Kyoso and A. Uchiyama. Development of an ECG identification system. In Proceedings of the

23rd Annual International Conference of the IEEE Engineering in Medicine and Biology Society,

volume 4, pages 3721–3723 vol.4, 2001.

[53] H. J. Lee, S. H. Hwang, S. M. Lee, Y. G. Lim, and K. S. Park. Estimation of body postures on bed

using unconstrained ECG measurements. IEEE Journal of Biomedical and Health Informatics,

17(6):985–993, 2013.

[54] A. Leon-Garcia. Probability and Random Processes For EE’s (3rd Edition). Prentice-Hall, Inc.,

2007.

[55] M. Li and S. Narayanan. Robust ECG biometrics by fusing temporal and cepstral information. In

2010 20th International Conference on Pattern Recognition (ICPR), pages 1326–1329, Aug 2010.

[56] J. L. C. Loong, R. Subari, K. S.and Besar, and M. K. Abdullah. A new approach to ECG biometric

systems: A comparative study between LPC and WPD systems. In World Academy of Science,

Engineering and Technology, pages 759–764, Aug 2010.

[57] A. Lourenco, H. Silva, and A. Fred. Unveiling the biometric potential of finger-based ECG signals.

Computat. Intell. Neurosci., 2011(720971):5:1–5:8, 2011.

Bibliography 78

[58] J. Lu, K. N. Plataniotis, and A. N. Venetsanopoulos. Face recognition using kernel direct discrim-

inant analysis algorithms. IEEE Transactions on Neural Networks, 14(1):117–126, Jan 2003.

[59] J. P. Martinez, R. Almeida, S. Olmos, A. P. Rocha, and P. Laguna. A wavelet-based ECG

delineator: evaluation on standard databases. IEEE Transactions on Biomedical Engineering,

51(4):570–581, April 2004.

[60] R. Matta, J. K. H. Lau, F. Agrafioti, and D. Hatzinakos. Real-time continuous identification system

using ECG signals. In 2011 24th Canadian Conference on Electrical and Computer Engineering

(CCECE), pages 001313–001316, 2011.

[61] A. Minchole, L. Sornmo, and P. Laguna. ECG-based detection of body position changes using

a laplacian noise model. In 2011 Annual International Conference of the IEEE Engineering in

Medicine and Biology Society (EMBC), pages 6931–6934. IEEE, 2011.

[62] G. G. Molina, F. Bruekers, C. Presura, M. Damstra, and M. V. Veen. Morphological synthesis of

ECG signals for person authentication. In Eur. Signal Processing Conf., 2007.

[63] M. H. Nathanson. Electrocardiographic study of movements of the heart with change of posture.

Experimental Biology and Medicine, 28(7):766–770, 1931.

[64] I. Odinaka, P. H. Lai, A. D. Kaplan, J. A. O’Sullivan, E. J. Sirevaag, S. D. Kristjansson, A. K.

Sheffield, and J. W. Rohrbaugh. ECG biometrics: A robust short-time frequency analysis. In

Proc. 2010 IEEE Int. Workshop on Information Forensics and Security, pages 1–6, 2010.

[65] I. Odinaka, P. H. Lai, A. D. Kaplan, J. A. O’Sullivan, E. J. Sirevaag, and J. W. Rohrbaugh. ECG

biometric recognition: A comparative analysis. IEEE Trans. Inf. Forensics Security, 7(6):1812–

1824, 2012.

[66] R. Palaniappan and S. M. Krishnan. Identifying individuals using ECG beats. In 2004 International

Conference on Signal Processing and Communications, pages 569–572, Dec 2004.

[67] J. Pan and W. J. Tompkins. A real-time QRS detection algorithm. IEEE Trans. Biomed. Eng.,

BME-32(3):230–236, 1985.

[68] S. Pathoumvanh, S. Airphaiboon, and K. Hamamoto. Robustness study of ECG biometric iden-

tification in heart rate variability conditions. IEEJ Transactions on Electrical and Electronic

Engineering, 9(3):294–301, 2014.

Bibliography 79

[69] T. Pawar, N. S. Anantakrishnan, S. Chaudhuri, and S. P. Duttagupta. Impact analysis of body

movement in ambulatory ECG. In 29th Annual International Conference of the IEEE Engineering

in Medicine and Biology Society (EMBS), pages 5453–5456. IEEE, 2007.

[70] K. N. Plataniotis, D. Hatzinakos, and J. K. M. Lee. ECG biometric recognition without fiducial

detection. In Biometrics Symp. : Special Session on Research at the Biometric Consortium Conf.,

pages 1–6, Sept 2006.

[71] S. Pouryayevali, S. Wahabi, S. Hari, and D. Hatzinakos. On establishing evaluation standards

for ECG biometrics. In 2014 IEEE International Conference on Acoustics, Speech and Signal

Processing (ICASSP), pages 3774–3778. IEEE, 2014.

[72] S. J. Raudys and A. K. Jain. Small sample size effects in statistical pattern recognition: rec-

ommendations for practitioners. IEEE Trans. Pattern Anal. Mach. Intell., 13(3):252–264, Mar

1991.

[73] S. Saechia, J. Koseeyaporn, and P. Wardkein. Human identification system based ECG signal. In

TENCON 2005 2005 IEEE Region 10, pages 1–4, Nov 2005.

[74] B. Scholkopf and K. R. Mullert. Fisher discriminant analysis with kernels. Neural networks for

signal processing IX, 1:1, 1999.

[75] B. Scholkopf, A. Smola, and K. R. Muller. Nonlinear component analysis as a kernel eigenvalue

problem. Neural computation, 10(5):1299–1319, 1998.

[76] T. W. Shen, F. C. Liu, Y. T. Tsao, and S. C. Chang. A body position detection method by fusing

heterogeneous information from surface ECG. In Computing in Cardiology, pages 517–520. IEEE,

2010.

[77] T. W. Shen, W. J. Tompkins, and Y. H. Hu. One-lead ECG for identity verification. In Proceed-

ings of the Second Joint 24th Annual Conference and the Annual Fall Meeting of the Biomedical

Engineering Society EMBS/BMES Conference Engineering in Medicine and Biology, volume 1,

pages 62–63. IEEE, 2002.

[78] T. W. D. Shen, W. J. Tompkins, and Y. H. Hu. Implementation of a one-lead ECG human

identification system on a normal population. J. Eng. Comput. Innovations, 2:12–21, 2011.

[79] L. H. Sigler. Electrocardiographic changes occurring with alterations of posture from recumbent

to standing positions. American Heart Journal, 15(2):146 – 157, 1938.

Bibliography 80

[80] M. L. Simoons and P. G. Hugenholtz. Gradual changes of ECG waveform during and after exercise

in normal subjects. Circulation, 52(4):570–7, 1975.

[81] Y. N. Singh and P. Gupta. ECG to individual identification. In 2nd IEEE International Conference

on Biometrics: Theory, Applications and Systems, pages 1–8, 2008.

[82] Y. N. Singh and P. Gupta. Correlation-based classification of heartbeats for individual identifica-

tion. Soft Computing, 15(3):449–460, 2011.

[83] L. Sornmo and P. Laguna. Bioelectrical Signal Processing in Cardiac and Neurological Applications.

Elsevier Academic Press, June 2005.

[84] D. J. Sutherland, D. D. McPherson, C. A. Spencer, C. S Armstrong, B. M. Horacek, and T. J.

Montague. Effects of posture and respiration on body surface electrocardiogram. The American

Journal of Cardiology, 52(5):595 – 600, 1983.

[85] M. M. Tawfik and T. Kamal. Human identification using QT signal and QRS complex of the ECG.

Online J. Electron. Elect. Eng., 3(1):383–387, 2011.

[86] N. V. Thakor and Y. S. Zhu. Applications of adaptive filtering to ECG analysis: noise cancellation

and arrhythmia detection. IEEE Trans. Biomed. Eng., 38(8):785–794, 1991.

[87] N. Venkatesh and S. Jayaraman. Human electrocardiogram for biometrics using DTW and FLDA.

In 2010 20th Int. Conf. Pattern Recognition (ICPR), pages 3838–3841, Aug 2010.

[88] S. Wahabi, S. Pouryayevali, S. Hari, and D. Hatzinakos. On evaluating ECG biometric systems:

Session-dependence and body posture. IEEE Transactions on Information Forensics and Security,

9(11):2002–2013, 2014.

[89] S. Wahabi, S. Pouryayevali, S. Hari, and D. Hatzinakos. Posture-invariant ECG recognition with

posture detection. In 2015 IEEE International Conference on Acoustics, Speech and Signal Pro-

cessing (ICASSP). IEEE, 2015.

[90] Y. Wan and J. Yao. A neural network to identify human subjects with electrocardiogram signals.

In World Congress on Engineering and Computer Science, 2008.

[91] Y. Wang, F. Agrafioti, D. Hatzinakos, and K. N. Plataniotis. Analysis of human electrocardiogram

for biometric recognition. EURASIP J. Adv. Signal Process, 2008, January 2008.

[92] G. Wubbeler, M. Stavridis, D. Kreiseler, R. D. Bousseljot, and C. Elster. Verification of humans

using the electrocardiogram. Pattern Recognition Letters, 28:1172 – 1175, 2007.

Bibliography 81

[93] C. M. Yang, C. M. Chou, S. G. Chen, B. J. Li, S. H. Hung, C. H. Yang, C. C. Wu, M. C. Hsieh,

and T. L. Yang. Sleeping ECG and body position monitoring system. In Annual International

Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pages 1856–1859.

IEEE, 2009.

[94] C. M. Yang, C. C. Wu, C. M. Chou, and T. L. Yang. Vehicle driver’s ECG and sitting posture

monitoring system. In 9th International Conference on Information Technology and Applications

in Biomedicine (ITAB), pages 1–4. IEEE, 2009.

[95] J. Yang, Z. Jin, J. Y. Yang, D. Zhang, and A. F. Frangi. Essence of kernel fisher discriminant:

{KPCA} plus {LDA}. Pattern Recognition, 37(10):2097 – 2100, 2004.

[96] J. Yang and J. Y. Yang. Why can lda be performed in pca transformed space? Pattern recognition,

36(2):563–566, 2003.

[97] J. Yao and Y. Wan. A wavelet method for biometric identification using wearable ECG sensors. In

Proc. 5th Int. Summer School and Symp. Medical Devices and Biosensors, pages 297–300, 2008.

[98] J. Yao and Y. Wan. Improving computing efficiency of a wavelet method using ECG as a biometric

modality. Int. J. Comput. Netw. Security,, 2(1):15–20, Jan 2010.

[99] C. Ye, M. T. Coimbra, and B. V. K. V. Kumar. Investigation of human identification using two-lead

electrocardiogram (ECG) signals. In 2010 Fourth IEEE International Conference on Biometrics:

Theory Applications and Systems (BTAS), pages 1–8, Sept 2010.

[100] H. Yu and J. Yang. A direct lda algorithm for high-dimensional data with application to face

recognition. Pattern Recognition, 34:2067–2070, 2001.

[101] H. Yu and J. Yang. A direct LDA algorithm for high-dimensional datawith application to face

recognition. Pattern recognition, 34(10):2067–2070, 2001.

[102] Z. Zhang and D. Wei. A new ECG identification method using bayes’ teorem. In Proc. IEEE

Region 10 Conf., pages 1–4, 2006.

[103] W. Zhao, A. Krishnaswamy, R. Chellappa, D. L. Swets, and J. Weng. Discriminant analysis of

principal components for face recognition. In Face Recognition, pages 73–85. Springer, 1998.

[104] Z. Zhao, L. Yang, D. Chen, and Y. Luo. A human ECG identification system based on ensemble

empirical mode decomposition. Sensors, 13(5):6832–6864, 2013.

by saeid wahabi - university of toronto t-space · abstract variability in ecg biometrics: state of...

Documents