Analysis of cardiac and epileptic signals using higher order spectra

ANALYSIS OF CARDIAC AND EPILEPTIC SIGNALS USING HIGHER ORDER SPECTRA by Chua Kuang Chua B.Eng Hons, MSc Dist March 2010... The use of nonlinear features motivated by the higher order s

ANALYSIS OF CARDIAC AND EPILEPTIC SIGNALS USING HIGHER ORDER SPECTRA

by

Chua Kuang Chua

B.Eng (Hons), MSc (Dist)

March 2010

Abstract

The theory of nonlinear dyamic systems provides some new methods to handle complex systems Chaos theory offers new concepts, algorithms and methods for processing, enhancing and analyzing the measured signals In recent years, researchers are applying the concepts from this theory to bio-signal analysis In this work, the complex dynamics of the bio-signals such as electrocardiogram (ECG) and electroencephalogram (EEG) are analyzed using the tools of nonlinear systems theory

In the modern industrialized countries every year several hundred thousands of people die due to sudden cardiac death The Electrocardiogram (ECG) is an important bio-signal representing the sum total of millions of cardiac cell depolarization potentials It contains important insight into the state of health and nature of the disease afflicting the heart Heart rate variability (HRV) refers to the regulation of the sinoatrial node, the natural pacemaker of the heart by the sympathetic and parasympathetic branches of the autonomic nervous system Heart rate variability analysis is an important tool to observe the heart's ability to respond to normal regulatory impulses that affect its rhythm A computer-based intelligent system for analysis of cardiac states is very useful in diagnostics and disease management Like many bio-signals, HRV signals are non-linear in nature Higher order spectral analysis (HOS) is known to be a good tool for the analysis of non-linear systems and provides good noise immunity In this work, we studied the HOS of the HRV signals of normal heartbeat and four classes of arrhythmia This thesis presents some general characteristics for each of these classes of HRV signals in the bispectrum and bicoherence

(ANOVA) test The results are very promising for cardiac arrhythmia classification with a number of features yielding a p-value < 0.02 in the ANOVA test An automated intelligent system for the identification of cardiac health is very useful in healthcare technology In this work, seven features were extracted from the heart rate signals using HOS and fed to a support vector machine (SVM) for classification The performance evaluation protocol in this thesis uses 330 subjects consisting of five different kinds of cardiac disease conditions The classifier achieved a sensitivity of 90% and a specificity of 89% This system is ready to run

on larger data sets

In EEG analysis, the search for hidden information for identification of seizures has a long history Epilepsy is a pathological condition characterized by spontaneous and unforeseeable occurrence of seizures, during which the perception or behavior of patients is disturbed An automatic early detection of the seizure onsets would help the patients and observers to take appropriate precautions Various methods have been proposed to predict the onset of seizures based on EEG recordings The use of nonlinear features motivated by the higher order spectra (HOS) has been reported to be a promising approach to differentiate

between normal, background (pre-ictal) and epileptic EEG signals In this work, these

features are used to train both a Gaussian mixture model (GMM) classifier and a Support Vector Machine (SVM) classifier Results show that the classifiers were able to achieve 93.11% and 92.67% classification accuracy, respectively, with selected HOS based features About 2 hours of EEG recordings from 10 patients were used in this study

This thesis introduces unique bispectrum and bicoherence plots for various cardiac conditions and for normal, background and epileptic EEG signals These plots reveal distinct

understanding of spectral analysis such as medical practitioners It includes original contributions in extracting features from HRV and EEG signals using HOS and entropy, in analyzing the statistical properties of such features on real data and in automated classification using these features with GMM and SVM classifiers

ABSTRACT I CONTENTS IV LIST OF FIGURES IX LIST OF TABLES XII STATEMENT OF AUTHORSHIP XIV ACKNOWLEDGEMENTS XV PUBLICATIONS XVI

CHAPTER 1 INTRODUCTION 1

1.1 Introduction 1

1.2 Motivation 3

1.3 Objectives 5

1.4 Contributions 5

CHAPTER 2 BIOSIGNALS USED (HEART RATE AND EEG SIGNALS) 7

2.1 General 7

2.2 Electrocardiography 8

2.2.1 Data Acquisition 9

2.2.2 Steps in ECG Analysis 12

2.2.3 Preprocessing Of ECG 13

2.2.4 Noise Filtering Technique 18

2.2.7 Band Pass Integer Filter 22

2.2.8 Low Pass Integer Filter 23

2.2.9 High Pass Integer Filter 23

2.2.10 Derivative 24

2.2.11 Squaring Function 24

2.2.12 Moving Window Integral 25

2.2.13 QRS Detection Using Adaptive Thresholds 25

2.2.14 Cardiac Abnormalities 26

2.2.15 Heart Rate Variability (HRV) 32

2.3 Electroencephalogram 33

2.3.1 EEG Recording Methods 34

2.3.2 Advantages of monopolar recording 35

2.3.3 Advantages of bipolar recording 35

2.3.4 EEG Lead Positioning 35

2.3.5 Classification of EEG Rhythms 36

2.3.6 Delta Waves 36

2.3.7 Theta Waves 36

2.3.8 Alpha Waves 37

2.3.9 Beta Waves 37

2.3.10 Uses of EEG 37

2.3.11 Epileptic EEG Signal 38

2.3.12 Different types of EEG Signal 41

CHAPTER 3 LITERATURE REVIEW 44

3.1 Introduction 44

3.2 HOS and features derived from HOS 47

3.2.1 Higher order spectra 47

3.2.2 Frequency Domain Definition and Properties 51

3.3 ANALYSIS USING HOS FEATURES 55

3.3.1 Bispectrum, Bicoherence and quadratic phase coupling 55

3.4 Application of HOS on various signals 64

3.4.1 Electroencephalogram (EEG) analysis 64

3.4.2 ECG and HRV analysis 66

3.5 Summary 68

CHAPTER 4 CLASSIFIERS 69

4.1 Gaussian mixture models (GMM): 69

4.2 Support Vector Machine (SVM): 71

CHAPTER 5 CARDIAC STATE DIAGNOSIS USING HIGHER ORDER SPECTRA 75

5.1 Introduction 75

5.2 Data and Classes 78

5.3 Methods used for analysis 78

5.4 Statistical Analysis 81

5.5 Results 81

5.6 Discussion 90

5.6.1 The scope of the study 91

5.7 Conclusion 95

CHAPTER 6 ANALYSIS OF EPILEPTIC EEG SIGNALS USING HIGHER

ORDER SPECTRA 96

6.1 Data 96

6.2 Methods 97

6.3 Quantitative Analysis 97

6.4 Results 98

6.5 Discussion 105

6.6 Conclusion 107

CHAPTER 7 CARDIAC HEALTH DIAGNOSIS USING HIGHER ORDER SPECTRA AND SUPPORT VECTOR MACHINE 108

7.1 Introduction 108

7.2 Data Acquisition Process 111

7.2.1 Preprocessing 112

7.3 Methods Used 113

7.4 Quantitative analysis 113

7.5 Support Vector Machine (SVM) Classifier 113

7.6 Principal Component Analysis 114

7.7 Test vector generation 115

7.8 Results 115

7.9 Discussion 126

7.10 Conclusion 130

CHAPTER 8 AUTOMATIC IDENTIFICATION OF EPILEPTIC EEG

SIGNALS USING HIGHER ORDER SPECTRA 131

8.1 Introduction 131

8.2 Data 134

8.3 HOS based features 134

8.4 Classifiers 136

8.4.1 Test vector generation 136

8.4.2 Performance comparison and Receiver Operating Characteristics (ROC) 136

8.5 Results 137

8.6 Discussion 149

8.7 Conclusion 155

CHAPTER 9 CONCLUSION AND FUTURE WORK 156

9.1 Conclusion 156

9.2 Future work 157

REFERENCES 159

List of Figures

Figure 2.1 Genesis of ECG [Athayde 1981] 10

Figure 2.2 ECG lead configuration [ Carr et al., 2000] 11

Figure 2.3 The ECG signal .13

Figure 2.4 Result of power line interference removal (a) Original signal with power line interference (b) Output of power line interference filter .15

Figure 2.5 Algorithm to remove the baseline wander 17

Figure 2.6 Results of baseline wander algorithm (a) Original signal (b) Output of baseline wander filter .18

Figure 2.7 Block diagram of the QRS detector z(n) is the time averaged signal y(n) is the band passed ECG x(n) is the differentiated ECG 21

Figure 2.8 High pass filter .23

Figure 2.9 Normal ECG signal .28

Figure 2.10 Congestive heart failure ECG Signal .28

Figure 2.11 Atrial fibrillation ECG signal .29

Figure 2.12 Ventricular fibrillation ECG signal .29

Figure 2.13 Preventricular contraction ECG signal 29

Figure 2.14 Left bundle branch block ECG signal 30

Figure 2.15 Complete heart block ECG signal .30

Figure 2.16 Ischemic/Dilated cardiomyopathy ECG signal .31

Figure 2.17 Sick Sinus Syndrome ECG signal 31

Figure 2.18 EEG lead positioning system .36

Figure 2.19 A taxonomy of seizures based upon classifications from the International League Against Epilepsy (1985) .38

Figure 2.20 Scheme of intracranial electrodes implanted for presurgical evaluation of epilepsy patients Depth electrodes were implanted symmetrically into the hippocampal formations (top) These are the locations where the epileptic EEGs were collected .42

Figure 2.21 EEG Signals (a) normal (b) pre-ictal and (c) seizure .42

Figure 3.1 Non-redundant region of computation of the bispectrum of a discrete-time signal assuming that the sampling interval is 1 and the Nyquist

frequency is thus  radians/second .54

Figure 3.2 Bicoherence noise plots (a) Gaussian (b) Generalized extreme value noise with 100 blocks of data and each block 256 samples .58

Figure 3.3 Typical heart rate signal of normal subject .59

Figure 3.4 Plots of (a) bispectrum (b) bicoherence of Figure 3.3 60

Figure 3.5 Region of computation of the bispectrum for real signals Features are calculated by integrating the bispectrum along the dashed line with slope=a Frequencies are shown normalized by the Nyquist frequency .63

Figure 4.1 Working of SVM algorithm 71

Figure 5.1 Bispectrum and bicoherence of Normal heart rate .86

Figure 5.2 Bispectrum and bicoherence of heart rate with Ectopic beat .86

Figure 5.3 Bispectrum and bicoherence of heart rate with AF .87

Figure 5.4 Bispectrum and bicoherence of heart rate with CHB .87

Figure 5.5 Bispectrum and bicoherence of heart rate with SSS-III 88

Figure 5.6 Bispectrum and bicoherence of heart rate with Isc\Dil Cardiomyopathy 88

Figure 5.7 Bispectrum of heart rate with LBBB 89

Figure 5.8 Bispectrum and bicoherence of heart rate with VF .89

Figure 5.9 Mean magnitudes of the Bispectrum and one standard deviation for each class 90

Figure 5.10 Bicoherence plots of different records of heart rate with PVC 93

Figure 5.11 Bicoherence plots of different records of heart rate with Isc/Dil Cardiomyopathy 94

Figure 6.1 Average bispectrum, bicoherence(b95% 0.015) and contour plots of 100 normal EEG signals .101

Figure 6.2 Average bispectrum, bicoherence (b95% 0.015) and contour plots of 100 pre-ictal EEG signals .103

Figure 6.3 Average bispectrum, bicoherence(b95% 0.015) and contour plots of 100 ictal EEG signals 104

Figure 6.4 Cluster plot of 3 features (P1, P2 and Mave ) 105 Figure 7.1 Distributions of various features extracted from the bispectrum

of the logarithmic amplitudes of the bispectrum (H1), (d) sum of

logarithmic amplitudes of diagonal elements (H2) 119 Figure 7.1 Distributions of various features extracted from the bispectrum(e) first order spectral moment of amplitudes of diagonal elements inthe bispectrum (H3), (f) the weighted centre of bispectrum frequencyindex 1 (f1m), (g) the weighted centre of bispectrum frequency index

2 (f2m) 120 Figure 7.2 (a) PCA analysis of invariant features P(a) projected unto 3 main

principal components 123 Figure 7.2 (b) PCA analysis of line integral B2 projected unto 3 main principal components (PCs) 124 Figure 7.2 (c) PCA analysis of HOS features projected unto 3 main principal

components (PCs) 125 Figure 8.1 A cluster plot of 3 features (P1, P2and Mave) 135 Figure 8.2 Comparison of ROC curves for the two classifiers: (a) ROC curves of GMM classifier with different combinations of features (b) ROC curves of SVM classifier with different combinations of features (c) ROC curves of GMM classifier for Normal, Epileptic and Pre-ictal EEG, and (d) ROC

curves of SVM classifier for Normal, Epileptic and Pre-ictal EEG 145 Figure 8.3 Percentage of classification accuracy for different classifiers .148

List of Tables

Table 5.1 ECG Data for different cardiac health states .84 Table 5.2 Invariant features, Mean Bispectrum Magnitude and Phase entropy,

Bispectral Entropy1 and 2 Entries in columns other than the last two

correspond to mean and standard deviation .85 Table 6.1 Results of various HOS parameters for normal, pre ictal and

epilepticEEG signals 100 Table 7.1 Numbers of ECG data sets for different cardiac health states .112 Table 7.2 Results of ANOVA on various bispectral features Entries in the

columns other than the last correspond to mean and standard deviation .117 Table 7.3 Confusion matrix for five different classes of arrhythmia with aSVM classifier 118 Table 7.4 Classification accuracy for five different classes of arrhythmia

with a SVM classifier 118 Table 7.5 Sensitivity, Specificity, and Positive Predictive Value for the

SVM classifier Entries to the left are the numbers of True and False

negatives and the True and False positives .121 Table 7.6 Comparison of arrhythmia classification with non-linear feature .128 Table 8.1 Values of various HOS parameters (in mean ± standard deviation) for

normal, pre-ictal and epileptic EEG signals (p<0.0001) .138

Table 8.2 Percentage of classification of GMM and SVM classifiers with

different combination of features (700 training data and 300 test data

in each case) 138 Table 8.3 Results of ROC for the two classifiers with features P1, P2and

Mave 141 Table 8.4 Basic mathematical formulae for classifier analysis .141 Table 8.5 Values of various spectral based features (in mean±standard

deviation) for normal, pre ictal and epileptic EEG signals (p<0.0001) 146 Table 8.6 Percentage of classification accuracy of classifiers with power spectral features (P1, P2, Mave) .147

Table 8.7 Percentage of classification accuracy of classifiers with HOS features (P1, P2, Mave) .147 Table 8.8 Percentage of classification accuracy for different classifiers in mean ± standard deviation ( p<0.0001) 148 Table 8.9 Comparison of some of the recent works in epilepsy detection .152

Statement of Authorship

The work in this thesis has not been previously submitted for a degree or diploma at any other higher education institution To the best of my knowledge and belief, the thesis contains no material previously published or written by another person except where due reference is made

Signed:

Acknowledgements

I would like to express my gratitude to all those who gave me the possibility to pursue this research I wish to express my deep gratitude to my supervisor Prof Vinod Chandran for his invaluable guidance, encouragement and direction throughout this work I would also like to thank Dr Lim Choo Min, director, Electronic and Computer Engineering (ECE) division, NgeeAnn Polytechnic for his help, support, interest and valuable suggestions Many thanks for Dr Rajendra Acharya U for many of his interesting discussions and suggestions towards my research Thanks for Dr Duncan Campbell for proof reading my thesis and his invaluable suggestions to improve the thesis Special thanks to my wife, Jenny Pang for her constant support and patience throughout my thesis I would not have been able to complete this, but for her co-operation Finally, I want to thank, everyone who have, in one way or another, helped me to conduct this research

Publications

Journal Papers

[1] Chua KC, Chandran V, Acharya UR, Lim CM Application of higher order

spectra to identify epileptic EEG Journal of Medical Systems, Springer 2010 (in Press)

[2] Chua KC, Chandran V, Acharya UR, Lim CM Cardiac health diagnosis using

higher order spectra and support vector machine Open Access Medical Informatics Journal 2009 (in Press)

[3] Chua KC, Chandran V, Acharya UR, Lim CM Automatic Identification Of

Epileptic EEG signals Using Higher Order Spectra International Journal of Engineering in Medicine 2009; 223(4): 485-495

[4] Chua KC, Chandran V, Acharya UR, Lim CM Analysis of epileptic EEG signals

using higher order spectra Journal of Medical & Engineering Technology, UK, 33(1), 2009, 42-50

[5] Chua KC, Chandran V, Acharya UR, Lim CM Computer- based analysis of

cardiac state using entropies, recurrence plots and Poincare geometry Journal of Medical & Engineering Technology 2008; 32(4): 263-272

[6] Chua KC, Chandran V, Acharya UR, Lim CM, Cardiac State Diagnosis Using

Higher Order Spectra of Heart Rate Variability, Journal of Medical & Engineering Technology 2008; 32(2): 145-155

[7] Chua KC, Chandran V, Acharya UR, Lim CM, Application Of Higher order

Statistics/spectra for Biomedical Signals – A Review , Journal of Medical

Engineering & Physics (in Press)

Conference Papers

[1] Chua KC, Chandran V, Acharya UR, Lim CM Higher Order Spectra Analysis for

Heart Rate Variability Proceedings of 15th International conference on mechanics in medicine and biology, Singapore 2006; 262-267

[2] Chua KC, Chandran V, Acharya UR, Lim CM Analysis of epileptic EEG signals

using higher order spetra Proceedings of 15th International conference on mechanics in medicine and biology, Singapore 2006; 303-308

[3] Chua KC, Chandran V, Acharya UR, Lim CM, Higher Order Spectral (HOS)

Analysis Of Epileptic EEG Signals, 29th IEEE-EMBS-2007, Lyon, August 2007

[4] Chua KC, Chandran V, Acharya UR, Lim CM, Automatic identification of

epilepsy by HOS and power spectrum parameters using EEG signals: A comparative study, 29th IEEE-EMBS-2008, Vancouver, August 2008

[5] Chua KC, Chandran V, Acharya UR and Lim CM, Higher Order Spectra based

Support Vector Machine for Arrhythmia Classification, ICBME 2008 3–6 December 2008 Singapore

A great deal of attention has been focused recently on the extraction of dynamical information from chaotic time series [Broomhead et al., 1986; Simm et al., 1987; Denker

et al., 1986] Chaos is the state in which a nonlinear dynamical system exhibits bounded motion, with exponential sensitivity to initial conditions, in that initially neighboring states of a chaotic system diverge exponentially as the system evolves forward in time [Guckenheimer et al., 1983] Chaotic time series analysis has greatly enhanced the understanding of chaos in experimental systems by allowing multidimensional dynamical information to be recovered from a time series of measurements of a single variable [Broomhead et al., 1986; Simm et al., 1987; Denker et al., 1986] This is usually done using the method of time delay embedding, which allows the recovery of information

1986] This allows the strange attractor of a chaotic dynamical system to be extracted from a time series of measurements of a single variable The simplicity of the technique and the accessibility of experimental time series have encouraged the rapid exploration of numerous fields as varied as plasma fluctuations [Simm et al., 1987], climatic variations [Essex et al., 1990] and non-equilibrium chemical systems [Roux et al., 1983]

Recently, the nonlinear techniques have been used to analyze physiological signals: heart rate, nerve activity, renal blood flow, arterial pressure, EEG and respiratory signals [Kannathal et al, 2004, Acharya et al, 2004a, Garret et al, 2003, Yuru et al, 2004]

To investigate the time-varying spectral characteristics of the underlying process, most of the methods often begin by computing the time variation of the common statistical properties of the process [Roberto et al., 1995; Kaplan, 1999; Laurent et al., 1998] However, these methods all assume that piecewise first-order or second-order stationarity

is satisfied for each segment of the observation after segmentation In practice, many medical signals show significant nonlinear and non-Gaussian characteristics, such as the presences of nonlinear effects of phase coupling among the signal frequency components [Shen et al., 2000; Ning, 1993; Ning et al., 1990; Husar et al., 1997] The methods based

on spectral analysis fail to properly deal with the nonlinearity and non-Gaussianity of the processes, but higher order spectra (HOS) allow us to effectively process these signals to obtain their higher-order statistics Bispectral estimation has been shown to be a very useful tool for extracting the degree of quadratic phase coupling (QPC) between frequency components of the process

In this work, HOS are applied to biosignals such as electrocardiogram (ECG) and electroencephalogram (EEG) The ECG signal is the electrical signal generated by the heart’s muscle measured on the skin surface of the body The EEG signal represents the

function of time These bio-signals are essentially non-stationary signals; they display a fractal like self-similarity They may contain indicators of current disease, or even warnings about impending diseases The indicators may be present at all times or may occur at random in the time scale However, to (study and) pinpoint anomalies in voluminous data collected over several hours is strenuous and time consuming Therefore, computer based analytical tools for in-depth study and classification of data over day long intervals can be very useful in diagnostics

1.2 MOTIVATION

ECG has a basic role in cardiology since it consists of effective simple noninvasive low cost procedures for the diagnosis of cardiac disorders that have high epidemiological incidence and are very relevant for their impact on patient life and social costs Pathological alterations observable by ECG are Cardiac rhythm disturbances (or arrhythmia), Dysfunction of myocardial blood perfusion (or cardiac ischemia), Chronic alteration of the mechanical structure of the heart Cardiac rhythm disturbances are considered to lead to life threatening conditions Thus the detection of abnormalities in intensive care patients is very essential and critical Recently a lot of research is being done for automating the abnormality detection applying various engineering methods and non-conventional techniques Especially in the scenario of continuous monitoring of ECG

in intensive care units, automatic analysis of ECG and abnormality detection is very helpful, as it will be an aid to the clinical staff in the absence of the doctor It will also help the doctor to diagnose and act faster in case of emergency conditions Designing low cost, high performance, simple to use and portable equipment for ECG offering a combination of diagnostic features seem to be a global pursuit Such equipment should embed and integrate several techniques of data analysis such as signal processing, pattern

detection and recognition, decision support and human computer interaction Thus computerized methods are to be applied for detection and classification of abnormalities

Epilepsy is a pathological condition characterized by spiky patterns in continuous EEG and seizure at times Approximately one percent of the world’s population has epilepsy, one third of whom have seizures not controlled by medications [Hauser et al., 1990; Kandel et al., 1991] Individuals with epilepsy suffer considerable disability from seizures and resulting injuries, the stigma and social isolation attached to having seizures, and from side effects of medical and other therapies Some patients, whose seizures reliably begin in one discrete region, usually in the mesial (middle) temporal lobe, may be cured by epilepsy surgery This requires removing large volumes of brain tissue, due to the lack of a reliable method for pinpointing the location of seizure onset and the pathways through which seizures spread Successful surgical treatment of focal epilepsies requires exact localization of the epileptic focus and its delineation from functionally relevant areas For this purpose, different pre-surgical evaluation methodologies are currently in use [Engel et al., 1998] Neurological and neuropsychological examinations are complemented by neuro-imaging techniques that try to identify potential morphological correlates Currently for localization of the epileptic focus is to record the patient’s spontaneous habitual seizure using electroencephalography Depending on the individual occurrence of seizures this task requires long lasting and continuous recordings

of EEG In case of ambiguous scalp EEG findings, invasive recordings of electrocorticogram and stereo-EEG via implanted depth electrodes are used This procedure is time consuming and offers greater risk to the patient Thus reliable EEG analysis techniques are required to localize and to demarcate the epileptic focus

1.3 OBJECTIVES

The present work is to perform nonlinear time series analysis on biosignals and use support vector machine and Gaussian Mixtures model techniques to classify and model the biosignals In this work, the biosignals considered are ECG and EEG Various milestones in this work are

 To establish an appropriate and relevant set of HOS features to detect various cardiac abnormalities from the ECG signals

 To analyze EEG signals and to identify set of HOS features that distinguish different types of EEG, specifically the normal, background (preictal) and epileptic (ictal) EEG

 To identify suitable classifier architectures to classify the biosignals for the abnormalities based on the feature set chosen

 To propose unique HOS plots for different cardiac diseases and epileptic, background and normal EEG signals that aid visual interpretation

1.4 CONTRIBUTIONS

The contributions derived from this research are summarized as follows:

 The implementation of an automatic approach to achieve highly reliable detection of cardiac abnormalities, which entails feature extraction, feature selection, feature fusion, event classification and assessment

 A new set of features based on HOS and entropy for the different cardiac diseases, background and epileptic EEG signals

 Evaluation of features extracted using HOS analysis techniques for detection

of cardiac abnormalities

 Evaluation of the performance of suitable classifier architectures and classifier inputs in the detection of various cardiac abnormalities

 Characterization of normal, background and epileptic signals using HOS features

Chapter 2 Biosignals Used (Heart rate and EEG Signals)

to be a reliable replacement of clinician in applications involving routine and tiresome work such as patient monitoring in Intensive Coronary Care Unit (ICCU) or processing of large amount of information as in holter data analysis [Holter, 1961] Quantitative assessment of diagnostic results is very helpful in clinical therapy, as subjective interpretations are prone to a wide range of inconsistencies and inaccuracies Thus, in areas heavily dependent upon quantitative assessment, accuracy and speed, computer based analysis is very useful in diagnostics The advances made in the fields of electrocardiography, sonography, imaging, laser technology and many others have given

a new dimension to health care in the 20th century and have a bigger role to play

in the coming years

In this thesis, some new results are presented on Heart Rate Variability (HRV) analysis and classification The heart rate of a normal healthy person rarely remains constant due to imbalance between the sympathetic and parasympathetic autonomic nervous systems The result of these counteractions is a continuous variation of the heart

investigating the state of the autonomic control of the cardiovascular system In this thesis, a method of analysis for HRV signal using HOS is developed, that provides patterns on the computer screen, which can be used by the clinician to diagnose the nature

of affliction The same HRV data is also used for classification using support vector machine (SVM) In the latter techniques, the computer tool itself substitutes the clinician

in diagnosing the disease, which may have about 85% accuracy In addition to classification, a visualization technique is developed for hierarchical display of data recorded over a period of several hours The technique would enable the clinician to get a comprehensive picture of the patient status, with a facility to get further detail for shorter intervals at the click of the computer mouse Thus, the thesis focuses on a comprehensive development of analysis, classification and display of HRV data useful in the diagnostics of cardiac diseases

A brief introduction to the related topics such as data acquisition, analysis and classification tools is presented below

Electrocardiography deals with the electrical activity of the heart Monitored

by placing sensors at the limb extremities of the subject, electrocardiogram (ECG) is a faithful record of the origin and propagation of the electric potential through cardiac muscles It is considered a representative signal of cardiac physiology, useful in diagnosing cardiac disorders

The cardiac cycle mainly consists of three electrical components representing the

activation and deactivation of the atria and ventricles, and of the blood pumping

chambers of the heart Figure 2.1(a) illustrates the electrical conduction system of the

human heart, which consists of a Bundle of His, which divides into left and right bundle

over the respective ventricles As shown in Figure 2.1(b), the electrical impulse originates

at the sino-atrial (SA) node and propagates along the walls of atria, depolarising the cells

on its way, before reaching the atrio-ventricular (A-V) node The SA node is located at the upper tip of the right atrium, while the A-V node is located at the junction between the atria and ventricles The resultant depolarisation potential of the atria shows up as a ‘P

wave’ in the ECG, illustrated in the Figure 2.1(c) The electrical impulse then conducts

through the Bundle of His, splits into right and left bundle branches and reaches the respective ventricles through the Purkinje fibres The sharp QRS complex in the ECG represents the resultant depolarisation potential of the ventricles The collective potential

of the ventricle depolarisation results in the ‘T wave’, which is sometimes followed by a

low amplitude ‘U wave’ as shown in Figure 2.1(c) The amplitude and duration of the component waves, such as PR, ST segments represent the strength of contraction and propagation delay of the electrical impulse through the heart muscles

2.2.1 Data Acquisition

Computer analysis of ECG begins with the acquisition and digitisation of data The recording of ECG is carried out in routine clinical check up of normal subjects The body potential is sensed through a set of electrodes placed on the torso and limb extremities Various configurations of electrode placements associated with 12 lead scalar ECG measurements are illustrated in Figure 2.2 In an extension to this approach called

vector-cardiography, electrical activity of the heart in three perpendicular planes, viz,

frontal, transverse and sagittal, derived from a set of scalar ECG signals, is displayed as vectors on an oscilloscope Body surface potential mapping is another emerging technique for monitoring cardiac functioning

Figure 2.2 ECG lead configuration [ Carr et al., 2000 ]

While recording the ECG, various kinds of noise and artifacts may corrupt or

degrade the data Electromyogram (EMG) due to muscle tremor, baseline wander due to

electrode displacement and mains interference are some of the noise sources usually encountered Differential amplifiers with a gain of 1000, high input impedance and large common mode rejection ratio (CMRR) are used to enhance the signal to noise ratio (SNR) The electrocardiogram is passed through a low pass anti-aliasing filter with a cut-off frequency of 100 Hz to suppress the EMG, and high pass filter with a cut-off frequency of 0.05 Hz to remove the baseline wander and respiratory interference The sampling rate has been fixed at 360 samples per second (sps) while the analog to digital converter (ADC) resolution is set at 12 bits

For post-digitisation removal of noise from ECG, a wide range of filters including moving average and recursive types has been proposed and these are implemented in software [Pryor, 1971] The interference by 50 Hz supply mains is usually eliminated with the help of a notch filter centred on the power line frequency A reference noise signal adaptively weighted by the filter is subtracted from the noise-contaminated data to yield a clean signal This least mean square (LMS) filter tracks the changes in noise frequency

2.2.2 Steps in ECG Analysis

The major steps in the analysis of the ECG signals are:

 Noise elimination from ECG using noise filtering techniques

 Cardiac cycle detection by detecting QRS complex

 Detection of significant characteristic points in ECG signal

 Formulation of characteristic feature set.

Noise filter removes and reduces the noise components from various sources in the ECG

signal

Cardiac cycle detection involves detecting the QRS complex peak corresponding to

each beat QRS Complex detection is implemented using Tompkins QRS complex detection algorithm [Pan , et al, 1985]

ECG characteristic points detection involves determining of significant points on the

ECG for feature extraction It includes the detection of QRS complex onset and offset,

ST segment detection and T peak detection

Feature set formulation includes formulation and selection of characteristics features

such that they significantly relate to the abnormalities Additional features are extracted

by performing complexity analysis on the signal

2.2.3 Preprocessing Of ECG

Figure 2.3 The ECG signal

The ECG consists of three basic waves, P, QRS and T These waves correspond

to the far field induced by specific electrical phenomena on the cardiac surface, namely

ventricular repolarization (T wave) The ECG does not look the same in all the leads of the standard 12 lead system used in clinical practice The polarity and the shape of the ECG constituent waves are different depending on the lead that is used A sample ECG wave measured at lead II is shown in Figure 2.3

In a normal cardiac cycle, the P wave occurs first, followed by the QRS complex and the T wave The sections of the ECG between the waves and complexes are called segments The ECG is characterized by three segments namely the PR segment, the ST segment and the TP segment The characteristic time periods in the ECG wave are the

PR interval, the RT interval, and the R-R interval

Usually ECG signals are contaminated by various kinds of noise Various types

of noise contaminating the ECG are described next

Power Line Interference

Power line interference consists of 60/50 Hz pickup and harmonics that can be

modelled as sinusoids and combination of sinusoids According to Friesen et al [Friesen

et al.,1990], the frequency content of this kind of noise is 60/50 Hz with harmonics and the amplitude is 50% of peak-to-peak ECG amplitude

Let us consider the presence of a periodic artifact with the fundamental frequency

of 60 Hz and odd harmonics at 180 Hz, 300 Hz, and 420 Hz Let the sampling frequency (fs) be 1000 Hz, and assume the absence of any aliasing error Zeros are then desired at

60 Hz, 180 Hz, 300 Hz, and 420 Hz, which translate to ±21.6û, ±64.8û, ±151.2û, with 360û corresponding to 1000 Hz The coordinates of the zeros are 0.9297 ± j0.3681, 0.4257 ± j0.9048, -0.3090 ± j0.9510, and -0.8763 ± j 0.4817 The transfer function of the filter is

6180.018515

.018595

.11)

(z  G  z z  z z  z z

H

where G is the desired gain or scaling factor With G computed so as to set the gain at DC to be unity, the filter transfer function becomes

8 7

6 5

4 3

2 1

6310.02149.01512.01288.0

1227.01288.01512.02149.06310.0)

z z

z z

z z

z

H

(2.2)

Figure 2.4 Result of power line interference removal (a) Original signal with power line interference (b) Output of power line interference filter

Electrode Contact Noise

Electrode contact noise is transient interference caused by loss of contact between the electrode and the skin, which can be permanent or intermittent The switching action can result in large artifacts since the ECG signal is usually capacitive coupled to the system This type of noise can be modeled as a randomly occurring rapid baseline transition that decays exponentially to the base line and has a superimposed 60Hz

(b)

(a)

signal is 1 sec and the amplitude is the maximum-recorded output with the frequency of

60 Hz

Motion Artifact

Motion artifacts are transient base line changes in the electrode skin impedance with electrode motion The shape of the base line disturbance caused by the motion artifacts can be assumed to be a biphasic signal resembling one cycle of a sine wave The peak amplitude and duration of the artifacts are variables The duration of this kind of noise signal is 100-500 ms with amplitude of 500% peak-to-peak ECG amplitude

Muscle Contraction

Muscle contractions cause artifact millivolt level potentials to be generated They can be assumed to be transient bursts of zero mean, band limited Gaussian noise The variance of the distribution may be estimated from the variation and duration of the bursts The standard deviation of this kind of noise is about 10% of the peak-to-peak ECG amplitude, with duration of 50 ms and the frequency content being DC to 10 kHz

Base line wander

The baseline wander of the ECG signals causes problems in the detection of peaks For example, due to the wander, the T peak could be higher than R peak, and it is detected as an R peak instead Low frequency wander of the ECG signal can be caused by respiration or patient movement The drift of the baseline with respiration can be represented as a sinusoidal component and the frequency of respiration added to the ECG signal The variation could be reproduced by amplitude modulation of the ECG by the sinusoidal component that is added to the base line The amplitude variation is 15% of peak-to-peak ECG amplitude and the base line variation is 15% of ECG amplitude at 0.15

These noise artifacts should be removed from ECG before extracting the characteristic features Noise removal is accomplished by passing the cardiovascular signal through a filter whose cutoff frequency is a function of the noise frequency

To solve baseline wander, median filtering can be used The steps involved in the implementation of baseline wander are shown in figure 2.5 First, the first 200ms of samples were extracted and sorted out in ascending order, then its median was calculated Then for every 200ms of samples till the end of the ECG signal, the same procedure was carried out Now, these samples are fed as input to the 600 ms window median filtering Later, the median value is evaluated for every 600ms of samples Then these median values were subtracted from the original waveform to remove the baseline wander of the ECG signal Figure 2.6 shows the result of the baseline wander filter

Figure 2.5 Algorithm to remove the baseline wander

Original Signal

Baseline Removed

Figure 2.6 Results of baseline wander algorithm (a) Original signal (b) Output of baseline wander filter

2.2.4 Noise Filtering Technique

The first step in noise filtering is determining the frequency corresponding to the significant characteristics of the ECG signal and the noises From the Fourier transform

of human ECG signal it has been found that P and T wave frequency generally lie between 0.5-10 Hz and QRS complex frequency ranges between 4-20 Hz [Thakor et al.,

1984] The P or T wave sometimes coincides with the baseline noise having a low frequency range of 0-0.8 Hz Hence it is very essential to eliminate the baseline noise from ECG signals like that of patient ECG having abnormalities where P and T wave play

In order to attenuate noise, the signal is passed through a band pass filter composed of cascaded high-pass and low-pass integer filters For removing the base line wander and the AC power noise, the algorithm of Alste and Schilder is implemented [Alste et al., 1985] In this algorithm, a non recursive finite impulse response (FIR) filter has been used with the reduced number of taps The basic principle behind this filtering process is that the designed frequency response has been defined with small stop band notches at 0 Hz to remove base line wander, as well as 60 Hz and its higher harmonics to remove power line interference In this algorithm, in order to reduce the long computational time caused by the large number of multiplication involved in the filtering

in the time domain the property of the discrete Fourier transform and symmetrical nature

of the impulse response has been used This is achieved by the property that the phase is

a linear function of frequency that corresponds with an exact delay time The convolution sum for the FIR filter is

.2

1

21

)))1(((

))((

)

()

(

2 ) 3 (

M h

T ki M

k n x T ki n x ikT h nT

T = sampling interval of both the input and output signal

With Sampling frequency of 360 Hz, T is 2.78 ms

kT = Time interval between successive impulse response

h = filter impulse response coefficients

M = number of filter coefficients

k

f = stop band cut off frequency

p

t = time period of the impulse response

This filter removes the baseline wander and AC power frequency noise But some high frequency noises like muscle noise or EMG are still present in the ECG signal These high frequency noises are eliminated by using another low pass FIR filter with a cutoff frequency higher than the frequency of the QRS complex

2.2.5 QRS Complex Detection

All the required features from ECG are extracted from the filtered ECG signal The basic and essential component for feature extraction is the detection of the QRS complex i.e locating the R point for each beat of the signal Once the R point is determined, all other characteristic points on the wave are determined with reference to the R point Thus an accurate detection of the QRS complex of the ECG is an important task in ECG analysis

2.2.6 QRS Detection Algorithm

The algorithm used to detect QRS complexes is an adaptation of the commonly

used real-time QRS detection algorithm developed by Pan et al [Pan et al., 1985] and