Ansys Cardiac Arrhythmia Model

Provided by the author(s) and NUI Galway in accordance with publisher policies. Please cite the published version when available.

Downloaded 2022-03-04T12:03:36Z

Some rights reserved. For more information, please see the item record link above.

Title

Improved cardiac arrhythmia prediction based on heart rate

College Of Science And Engineering,

In fulfilment of the requirements for the degree of Doctor of Philosophy.

Contents

Contents .......................................................................................... i Declaration of Originality .............................................................. iv Abstract ......................................................................................... v Acknowledgments .......................................................................... vi List of Figures ............................................................................. viii List of Tables ................................................................................. xi List of Acronyms ......................................................................... xiv Chapter 1 Introduction ........................................................ 1 1.1 Motivation ..................................................................................................... 2 1.2 Thesis Contributions ...................................................................................... 6 1.3 Thesis Structure ............................................................................................. 7 1.4 Publications ................................................................................................... 8

1.4.1

Peer-reviewed Journal Publications .................................................... 8

1.4.2

Peer-reviewed Conference Publications .............................................. 8 Chapter 2 Background .......................................................... 9 2.1 Cardiac Anatomy and Electrophysiology ....................................................... 9

2.1.1

Anatomy of Human Heart .................................................................. 9

2.1.2

Electrocardiogram Signal .................................................................. 11

2.1.3

Heart Rate Variability ...................................................................... 15 2.2 Cardiac Arrhythmia ...................................................................................... 16

2.2.1

Supraventricular Arrhythmia ............................................................ 18

2.2.2

Ventricular Arrhythmia .................................................................... 19 2.3 Cardiac Rhythm Management Approaches ................................................... 21

2.3.1

Automated External Defibrillator ..................................................... 21

2.3.2

Implantable Cardioverter Defibrillator .............................................. 21 2.4 Summary ....................................................................................................... 22 Chapter 3 Heart Rate Variability Analysis ........................ 23 3.1 Introduction to HRV ..................................................................................... 23 3.2 Time Domain HRV Analysis ......................................................................... 26 3.3 Frequency Domain HRV Analysis ................................................................. 27 3.4 Bispectrum Analysis...................................................................................... 28

3.4.1

Magnitude and Power Average ......................................................... 29

3.4.2

Bispectrum Entropies ........................................................................ 30

3.4.3

Logarithmic Bispectrum Features ..................................................... 30

3.4.4

Weight Centre of Bispectrum ............................................................ 31 3.5 Nonlinear and Dynamic Analysis .................................................................. 32 3.6 Discussion and Concluding Remarks ............................................................. 35 Chapter 4 Ventricular Arrhythmia Prediction using Heart Rate Variability Analysis .................................. 37 4.1 Overview ....................................................................................................... 38 4.2 Databases ...................................................................................................... 42

4.2.1

Medtronic Version 1.0 ....................................................................... 43 4.3 General Ventricular Arrhythmia Prediction Model ....................................... 45

4.3.1

Preprocessing and Noise Reduction ................................................... 46

4.3.2

Feature Extraction ............................................................................ 46

4.3.3

Statistical Feature Analysis............................................................... 47

4.3.4

Signal Classification and Performance Metrics .................................. 49 4.4 Experimental Results .................................................................................... 49

4.4.1

Time Domain Results ........................................................................ 50

4.4.2

Frequency Domain Result ................................................................. 52

4.4.3

Bispectrum Analysis Results ............................................................. 53

4.4.4

Nonlinear and Dynamic Analysis Results .......................................... 55 4.5 Discussion and Conclusions ........................................................................... 57 Chapter 5 Optimizing Heart Rate Variability Features for Ventricular Arrhythmia Prediction .................. 59 5.1 Overview ....................................................................................................... 60 5.2 HRV Feature Selection Methods ................................................................... 63

5.2.1

Student’s t-Test ................................................................................ 64

5.2.2

Minimal Redundancy-Maximal Relevance ......................................... 65

5.2.3

Infinite Latent Feature Selection ....................................................... 66 5.4 Proposed Feature Selection Method .............................................................. 67

5.4.1

Proposed Dataset Division Algorithm ............................................... 69

5.4.2

Improved Signal Pre-processing and Noise Reduction ....................... 69

5.4.3

Signal Classification and Cross-validation ......................................... 71 5.6 Experimental Results .................................................................................... 75

5.6.1

Learning Phase Results ..................................................................... 76

5.6.2

Evaluation Phase Results .................................................................. 81 5.7 Discussion ..................................................................................................... 84 5.8 Conclusions ................................................................................................... 86 Chapter 6 Atrial Fibrillation Prediction using Heart Rate Variability Analysis ........................................... 87 6.1 Literature Review ......................................................................................... 87

6.1.1

PAC Detection Methods ................................................................... 88

6.1.2

HRV Analysis Methods ..................................................................... 90 6.2 Proposed HRV Features ................................................................................ 93

6.2.1

Poincaré Plot Representation ............................................................ 93

6.2.2

Difference Map .................................................................................. 95

6.2.3

Kernel Density Estimation ................................................................ 95

6.2.4

Univariate KDE Features.................................................................. 99

6.2.5

Bivariate KDE Features .................................................................... 99 6.3 Atrial Fibrillation Prediction Database ....................................................... 100 6.4 Proposed Evaluation Method ...................................................................... 101

6.4.1

Prepossessing and Noise Reduction ................................................. 102

6.4.2

Feature Extraction .......................................................................... 105

6.4.3

PAF Prediction ............................................................................... 107 6.5 Experimental Results and Discussion .......................................................... 108

6.5.1

Results of Feature Ranking ............................................................. 108

6.5.2

PAF Prediction: Results of Feature Comparison............................. 110

6.5.3

Comparison with Previous Work .................................................... 111 6.6 Conclusion................................................................................................... 112 Chapter 7 Conclusions ....................................................... 113 7.1 Summary and Conclusions .......................................................................... 113 7.2 7.2. Summary of Contributions ................................................................... 116 7.3 Future Work ............................................................................................... 117 Bibliography ................................................................................119

Declaration Of Originality

I the Candidate Ashkan Parsi, certify that this thesis entitled “Improved Cardiac Arrhythmia Prediction Based on Heart Rate Variability



has not been previously submitted for any degree or qualification at



and where any work in this thesis was conducted in collaboration, appropriate reference to published work by my collaborators has been made and the nature and extent of my contribution has been clearly stated.

Abstract

Many types of ventricular and atrial cardiac arrhythmias have been discovered in clinical practice in the past 100 years, and these arrhythmias are a major contributor to sudden cardiac death. Ventricular tachycardia, ventricular fibrillation, and paroxysmal atrial fibrillation are the most commonly-occurring and dangerous arrhythmias, therefore early detection is crucial to prevent any further complications and reduce fatalities. Implantable devices such as pacemakers are commonly used in patients at high risk of sudden cardiac death. While great advances have been made in medical technology, there remain significant challenges in effective management of common arrhythmias. This thesis proposes novel arrhythmia detection and prediction methods to differentiate cardiac arrhythmias from non-life-threatening cardiac events, to increase the likelihood of detecting events that may lead to mortality, as well as reduce the incidence of unnecessary therapeutic intervention. The methods are based on detailed analysis of Heart Rate Variability (HRV) information.

A range of features based on HRV analysis are investigated, including features from time, frequency, bispectrum and nonlinear analysis, and a range of classification techniques is used for prediction based on these features. Firstly, a thorough review of existing methods for ventricular arrhythmia prediction is conducted, including a detailed comparative experimental study. Following this, feature extraction for ventricular arrhythmia prediction is investigated using feature ranking methods based on mutual information. Using classification techniques such as support vector machine, k-nearest neighbour and random forests, the proposed approaches compare well with related work in the literature using different signal analysis durations. Furthermore, for paroxysmal atrial fibrillation prediction, seven novel features are proposed and investigated using a number of standard classification techniques. Using only the seven newly-proposed features, classification performance outperforms those of the classical state-of-the-art feature set, and the results further improve when the features are combined with several of the classical features.

The results of the work show good performance of the proposed methods and support the potential for their deployment in resource-constrained devices for ventricular and atrial arrhythmia prediction, such as implantable pacemakers and defibrillators.

Acknowledgments

I would like to thank the following people, without whom I would not have been able to complete this research. First and foremost, I would like to thank my primary supervisor, Prof. Edward Jones, and co-supervisor Prof. Martin Glavin for all their help, direction, and guidance throughout my PhD. Your continued interest in my work and professional development is greatly appreciated and your unwavering enthusiasm was invaluable to me. Thank you for believing in me.

Without the support of Irish Research Council (IRC), this project would not have been possible, so I thank them for backing up my work (Grant number GOIPG/2016/1604). I would also like to thank all the members of my Graduate Research Committee at NUI Galway, Prof. Peter Corcoran, Dr Maeve Duffy, and Dr Ted Vaughan. To all the staff of Electrical and Electronic Engineering Mary Costello, Myles Meehan, Martin Burke, Liam Kilmartin, Dr Adnan Elahi, Prof. Martin O'Halloran, and Prof. Gearóid Ó Laighin, thank you for always supporting me.

I would like to express my gratitude to my wife, SaBa, for listening to my boring daily challenges, providing guidance, and helping to step one at a time every day. Without her tremendous understanding and encouragement in the past few years, it would not be impossible for me to complete my study.

Special thanks to Dr Dallan Byrne and Dr Declan O'Loughlin for fruitful collaboration and support in the conduct of this research. Thanks also to Dr Darragh Mullins and Dr Bárbara Luz Oliveira for asking the right questions, for challenging my thought through my research in the best possible way, and their help along the way.

I would like to thank Dr Hossein Javidnia and Dr Shabab Bazrafkan for all the support, guidance, and laughter specially during the start of my PhD journey. You guys made it much more enjoyable for me, I finally did it! I have been lucky to be surrounded by so many great friends who made brunch, lunch, and tea such a pleasure, not least. I am particularly grateful for it and would like to thank Dara, Dave, Eoghan, Esteban, Hao, Jibran, Jordan, Joseph, Kevin, Luke, Richie, Shane, Stephen, Tim, Tomás, Toto and many

Vii

past and current researchers in the Alice Perry Engineering Building. Many thanks to everyone in Galway Climbing Co-op and moreover, to all my friends from throughout my time in Galway including Adib, Afrooz, Alessio, Amin, Amir, Amirali, Anna, Alireza, Behrouz, Charlie, Christopher, Colm, Cynthia, David, Davood, Dennis, Hoda, Jajo, Jakub, Julia, Kambiz, Laura, Luison, Mahmoud, Mahshid, Malcolm, Mariusz, Mona, Neda, Sahar, Santi, Sara, Shadan, Shideh, Shirin, Padhraic, Pat, Patrick, Ronan, Tarlan, Théo, Tohid.

You have made my life better, and I am lucky to have you guys here! Finally, to my mother, Nabegheh, my father, Sadegh, my sister, Niusha, my brother Kasra and extended family, thank you for being there for me and for supporting me in every possible way during this time. I truly cannot thank you enough for your endless support, encouragement, and love.

Ashkan Parsi

Spring of 2021.

List Of Figures

Figure 1.1. Rate of inappropriate detection of SVT (totally 34.5% of all detection in both types of ICDs, single and dual-chamber), with break-down by arrhythmia subtype (AF, atrial flutter/atrial tachycardia, sinus tachycardia, and the other. “Other” arrhythmias include junctional tachycardia, AV nodal re-entrant tachycardia, and AV re-entrant tachycardia.

Subtype classification was based on blinded episode reviewer . ................ 4 Figure 2.1. Anatomical structure of the heart . .................................... 10 Figure 2.2. Heart conduction system. ......................................................... 11 Figure 2.3. Main wave, interval and segments in normal ECG signal. ....... 12 Figure 2.4. The interval between an R wave and the next R wave; normal resting heart rate is between 60 and 100 bpm. R-R Interval duration generally varies between 0.6-1.2 seconds . .............................................................. 15 Figure 2.5. Examples of ventricular arrhythmias showing misplaced electrical impulses in lower side of the heart. ............................................................... 20 Figure 3.1. (a) RRI for normal, VT and VF conditions for a typical cardiac patient ; (b) Corresponding IHR. ............................................................ 25 Figure 3.2. Segmentation of the bispectrum into sub band regions (LL, LH, and HH), and the ROI for bispectrum feature extraction, indicated by the red (with dash-dot line) triangle. ........................................................................ 29 Figure 4.1. Overview of arrhythmia prediction methods based on cardiac signal activities. ............................................................................................ 45 Figure 5.1. Flowchart describing the decision-making flow inside the ICD and where VT-VF prediction takes place in this flow. Having accurate VT-VF prediction make the ICD to deliver ATP, which can finally lead cardiac system to go to normal heartbeat. Normal heartbeat would fail tachycardia reconfirmation step which makes the ICD skip VT-VF detection and avoiding un-necessary electrical shock, while saving battery life and improve patient’s life experience and quality. ........................................................................... 61 Figure 5.2. Best results from recent studies on VT-VF prediction. ............ 63

Ix

Figure 5.3. Overview of proposed VT-VF prediction learning (a) and evaluation (b) phases. Using the proposed method, learning the correct features, and evaluation of prediction performance are separated. ............... 68 version 1.0 dataset, including noise reduction and resampling. Note the removal of high-frequency complnents evident in (a) by filtering. Figure 3.1 shows the same patient’s R-R interval and IHR signals without any pre-processing. ... 70 Figure 5.5. The technique of K-fold cross-validation has illustrated here for the case of 𝐾= 3. It involves taking the available data, shuffle it randomly and splitting it into 𝐾 groups (in the simplest case these are of equal size). For each unique group use the group as test dataset while the other 𝐾−1 of the groups are used to fit and train a model. Retain the evaluation score and discard the model. This procedure is then repeated for all 𝐾 possible choices for the held-out group, indicated here by the red blocks, and the performance scores from the 𝐾 runs are then averaged. .................................................... 74 Figure 5.6. Sample ROC curve with Reference Line .................................. 76 Figure 5.7. Leave-one-patient-out cross-validation method results using 5- minute of R-R interval signals on Learning Dataset using mRMR ((a), (b), and (c)) and ILFS ((d), (e), and (f)) method ranking, with optimal number of features for evaluation phase indicated by vertical green dotted lines. ......... 77 Figure 5.8. Leave-one-patient-out cross-validation method results using 1- minute of R-R interval signals on Learning Dataset using mRMR ((a), (b), and (c)) and ILFS ((d), (e), and (f)) method ranking, with optimal number of features for evaluation phase indicated by vertical green dotted lines. ......... 78 Figure 5.9. ROC results for both 1 and 5-minute of R-R interval signals analysis. ........................................................................................................ 83 Figure 6.1. The return map (left) and the difference map (right) of R-R intervals for a normal ECG signal (top row) and for onset of PAF event dispersion along minor and major axis of the fitted ellipse. .......................... 94 Figure 6.2. The difference map probability density of one axis of the patient PAF for this patient. .................................................................................... 97 row shows the normal event and bottom shows the onset of PAF for this patient. .......................................................................................................... 98 Figure 6.4. PAF prediction proposed algorithm block diagram. Proposed features along with other commonly used features are ranked using mRMR and ILFS based on their ability to generate the highest degree of differentiation

X

between the two classes. With different combinations of features, prediction performance is calculated using four different classification algorithms (SVM, k-NN, RF, and MLP neural network), using cross-validation to allow fairer comparison between different results. .......................................................... 102 ..................................................................................................................... 103

List Of Tables

Table 2.1. Typical ECG durations for healthy person , . ................ 14 Table 3.1. Summary of HRV features in time domain. ............................... 26 Table 3.2. Summary of selected HRV features in the frequency domain. ... 27 Table 3.3. Summary of bispectral HRV features. ....................................... 31 Table 3.4. Selected HRV features in nonlinear and dynamic analysis. ....... 34 Table 4.1. Summary of previous studies conducted in prediction of sudden cardiac death using R-R interval signals from ICD devices. ......................... 40 Table 4.2. Summary of databases of R-R interval signals from ICD devices, and their usage in the literature. .................................................................. 42 Table 4.3. Patient metadata in Medtronic version 1.0 dataset. .................. 44 Table 4.4. Summary of features considered in time domain (6 features), frequency domain (4 features), bispectrum analysis (31 features), and nonlinear analysis (9 features) , . ....................................................................... 47 Table 4.5. Mean, standard deviation (SD), and p-value of different feature categories using 1-minute windows of IHR signals from Medtronic version 1.0 dateset........................................................................................................... 48 Table 4.6. Mean, standard deviation (SD), and p-value of different feature categories using 5-minute windows of IHR signals from Medtronic version 1.0 dateset........................................................................................................... 48 Table 4.7. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of time domain features using 1-minute windows of R-R interval signals from Medtronic version 1.0 dateset. .............................................................. 51 Table 4.8. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of time domain features using 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset. .............................................................. 51 Table 4.9. Sensitivity (SN), specificity (SP), accuracy (ACC), and AUC of time domain features together using 1- and 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset. .................................................. 51

Xii

Table 4.10. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of frequency domain features using 1-minute windows of R-R interval signals from Medtronic version 1.0 dateset. .................................................. 52 Table 4.11. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of frequency domain features using 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset. .................................................. 52 Table 4.12. Sensitivity (SN), specificity (SP), accuracy (ACC), and AUC of frequency domain features together using 1- and 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset. ..................................... 52 Table 4.13. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of bispectrum analysis features using 1-minute windows of R-R interval signals from Medtronic version 1.0 dateset. .................................................. 53 Table 4.14. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of bispectrum analysis features using 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset. .................................................. 54 Table 4.15. Sensitivity (SN), specificity (SP), accuracy (ACC), and AUC of bispectrum analysis features together using 1- and 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset. ..................................... 54 Table 4.16. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of nonlinear and dynamic analysis features using 1-minute windows of R- R interval signals from Medtronic version 1.0 dateset. ................................. 55 Table 4.17. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of nonlinear and dynamic analysis features using 5-minute windows of R- R interval signals from Medtronic version 1.0 dateset. ................................. 56 Table 4.18. Sensitivity (SN), specificity (SP), accuracy (ACC), and AUC of nonlinear and dynamic analysis features together using 1- and 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset............ 56 Table 4.19. Summary of experimental results from this chapter using Medtronic version 1.0 dataset. ...................................................................... 57 Table 4.20. Summary of experimental results from previous research, along with results from this chapter (including results obtained with Medtronic version 1.0 dataset). ...................................................................................... 58 Table 5.1. The p-value, Student’s t-test (with probability less than 0.005), and selected features based on mRMR and ILFS methods for each classifier (SVM, k-NN and RF) using 5-minute of R-R interval signals from Medtronic version 1.0 dateset. ....................................................................................... 80 Table 5.2. Summaries of test results conducted in prediction of VT-VF using SVM, k-NN, and RF classifiers based on mRMR and ILFS feature ranking

Xiii

methods on using 5-minute of R-R interval signals from Medtronic version 1.0 dataset. ......................................................................................................... 81 Table 5.3. Summaries of test results conducted in prediction of VT-VF using SVM, k-NN, and RF classifiers based on mRMR and ILFS feature ranking methods on using 1-minute of R-R interval signals from Medtronic version 1.0 dataset. ......................................................................................................... 81 Table 5.4. Summary of results from previous research with Medtronic version 1.0 dataset, and results presented in this chapter. ........................................ 85 Table 5.5. Summaries of top ten features from mRMR ranking processing 1 and 5-minute of R-R interval signals. Shaded rows are the ones which are common between 1 and 5-minute analysis. ................................................... 85 Table 6.1. Performance comparison between the previous works using PAC analysis on AFPDB. ..................................................................................... 90 Table 6.2. Performance comparison between the previous works using PAC analysis on AFPDB. ..................................................................................... 92 Table 6.3. Proposed features extracted from difference map. .................... 100 Table 6.4. Existing features used for PAF prediction, along with their descriptions , . ................................................................................... 106 Table 6.5. The Lasso, mRMR, and ILFS features ranking and selection processing on 5 minutes signals from AFPDB database. Shaded rows are the ones which are selected using each methodology and numbers under mRMR and ILFS columns represent feature ranking. .............................................. 109 Table 6.6. Summaries of experimental results using different feature sets on AFPDB using 10-fold cross-validation method. ........................................... 110 Table 6.7. Comparison of experimental results in the leading literature using different classifiers on 5-minue of AFPDB ECG signals using 10-fold cross- validation method. ....................................................................................... 111

Introduction

Cardiovascular diseases, especially those which can result in sudden cardiac death, remain a major cause of mortality globally . Up to 80% of sudden cardiac deaths are caused by ventricular arrhythmias , and the accurate and reliable prediction of ventricular arrhythmias has the potential to tangibly impact the survival rate in patients with cardiovascular disease; in particular, the successful prediction and management of both ventricular tachycardia (VT) and ventricle fibrillation (VF) will increase the survival rate in patients with cardiac issues. Their distinctive characteristics motivate the use of machine learning to distinguish these rhythms from normal cardiac activity.

Heart rate variability (HRV) calculated from cardiac electrical activities is a key indicator of an individual’s cardiovascular condition. Assessment of HRV has been shown to aid diagnosis and treatment strategies and is widely used in clinical practice. However, the variety of HRV estimation and analysis methods in cardiac arrhythmia prediction and detection indicate that there is a need for a more rigorous investigation of HRV features. This thesis investigates the development of appropriate HRV signal processing techniques in the context of ventricular and atrial arrhythmias prediction, with a particular focus on techniques that may be appropriate for use in implantable defibrillators.

Motivation

Currently, automated external defibrillators (AEDs) and implantable cardiac defibrillators (ICDs) are the most effective therapeutic approaches for those at high risk of sudden cardiac death due to cardiac arrhythmias –. An AED is a portable device which includes a control algorithm that checks the surface electrocardiogram (ECG) and can send an electrical shock in order to restore normal cardiac rhythm if either VT or VF are detected by its algorithm .

ICDs are battery-operated devices placed under the skin in the chest that constantly monitor cardiac activity through thin leads directly lodged into the right ventricle, which produce an electrogram (EGM). ICDs are the more usual treatment for patients with substantial daily risk of severe arrhythmia and are highly effective in decreasing mortality due to cardiac arrhythmias in high-risk patients , . By processing intra-cardiac electrical activities, ICDs deliver treatments such as electrical shock, if necessary, to restore normal heart rhythm. ICDs have been proven to reduce the mortality of patients with life- threatening ventricular arrhythmias and have become the therapy of choice for many patients, however, delivery of inappropriate shocks caused by misclassification of events is problematic. For example, mistaking rapidly conducted supraventricular tachycardia (SVT) affects 8% to 40% of patients –. Inappropriate shocks can lead to pain, anxiety, depression, impaired quality of life, proarrhythmia1, and poor tolerance of life-saving ICD therapy .

ICDs may apply one of two treatment mechanisms in a situation of

Arrhythmia Event Recognition:

1. Direct current defibrillation (electrical shock) 2. Antitachycardia pacing (ATP). Because of the somewhat traumatic nature of electric shock therapy needed to address arrhythmias; it is important to ensure that such therapy is only applied when needed. Appropriate therapy delivery is dependent upon accurately classifying a patient’s rhythm in ICDs, however, as noted above, misclassification of supraventricular arrhythmias like SVT have been shown to lead to inappropriate therapy delivery in ICD recipients. For example, studies described in , indicated that “inappropriate” detection of SVT can be as high as 34.5% of all detection in both types of ICDs, single and dual- 1 Proarrhythmia is a new or more frequent occurrence of pre-existing arrhythmias, paradoxically precipitated by antiarrhythmic therapy, which means it is a side effect associated with the administration of some existing antiarrhythmic drugs, as well as drugs for other indications.

3

chamber, of which approximately 53% were atrial fibrillation (AF) or more specifically paroxysmal atrial fibrillation (PAF) episodes. The MADIT II study in shows inappropriate electric shocks were applied on 11.5% of 719 patients, constituting 31.2% of all electrical shocks applied in therapy overall . While there is no specific guidance on what the minimum required detection accuracy for clinical practice is, it is clear that current ICD usage results in unacceptably high rates of false positives, leading to inappropriate therapy, therefore, a significant increase in detection accuracy should result in improved outcomes. A robust method would be one that combines accurate prediction with detection, and where appropriate, this may be used to trigger ATP which is a less aggressive therapy that controls heartbeats. In particular, this thesis considers ventricular arrhythmias which could be prevented by nonaggressive treatment like ATP, with more invasive electrotherapy reserved only for cases that absolutely require it. For example, the PainFree I and II trials showed the effectiveness and safety of the application of ATPs (2 sequences of 8-pulse burst pacing train at 88% of the VT cycle length) before shock delivery even on fast VT with a heart rate of 188-250 beats per minute –.

On the other hand, a major drawback of therapy with an implantable defibrillator is the low specificity of detection . In principle, adding atrial sensing information to a decision algorithm could improve specificity of detection especially with current advances in dial chamber ICDs which can record atrial and ventricular activities separately. As shown in Figure 1.1, about 50% of misclassification is due to AF. AF occurs when the heart beats in a disorganized and irregular way, which can lead to a range of symptoms and potential complications, and this irregular beating could be misclassified as VT-VF arrhythmias by ICDs.

Chapter 1

Figure 1.1. Rate of inappropriate detection of SVT (totally 34.5% of all detection in both types of ICDs, single and dual-chamber), with break-down by arrhythmia subtype (AF, atrial flutter/atrial tachycardia, sinus tachycardia, and the other.

“Other” arrhythmias include junctional tachycardia, AV nodal re-entrant tachycardia, and AV re-entrant tachycardia. Subtype classification was based on blinded episode reviewer .

The most desirable method of limiting inappropriate therapies is appropriate rhythm prediction and discrimination by the ICD . Many cardiac event detector algorithms have been presented to distinguish VT-VF from AF, SVT, and normal sinus rhythm. Generally, these algorithms are sub- divided into two main categories: algorithms based on the EGM morphology, and algorithms based on heart rate interval statistics. Algorithms based on EGM morphology use either direct/figurative models of heartbeat (QRS complex) processed using wavelet transform , , covariance-based model , spatial projection of tachycardia , or an indirect/implicit model of heartbeat such as similarity measures, which are extracted from information theory concepts . Unlike morphological algorithms, which are generally more intensive computationally, the second category of algorithms utilize computationally-lighter features such as rate, onset, stability and entropy from HRV signal , –. All proposed algorithms have strengths and weaknesses however, to date, no consensus exists concerning the detection zones (different levels of heart rate) of VT-VF arrhythmias to decreasing inappropriate shocks due to misclassification of SVT and VT-VF arrhythmias.

The second category of algorithms, based on heart rate, is the primary focus of this thesis.

5

Heart rate (HR) represents the number of contractions per minute that occurs in the heart. Changes in heart rate, which is called heart rate variability (HRV), occurs because of the rising and falling of heart rate caused by various factors such as human activity or cardiovascular related diseases. The irregularity of the time interval between human heartbeats was first noted in the early 1600s, however, its physiological importance was not fully appreciated until 1965 when Hon and Lee found that the changes in pattern of foetal HR were an indicator of distress. In the late 1980s, the clinical importance of HRV became apparent when it was confirmed that the HRV was a strong and independent predictor of mortality following an acute myocardial infarction . Today, there is much research using HRV as a biological marker to diagnosis various cardiovascular related diseases such as arrhythmia detection or prediction , , diabetes , , and heart failure . Such interest arises because researchers have shown that HRV is one of the most promising markers to assess autonomic nervous system (ANS) activity, which can be directly correlated to cardiovascular disease . At the same time, with massive improvements and advances in consumer and wearable technology, HRV data can be calculated using smart watches or small portable chest straps, which makes it more accessible for diagnosis , . Existing works on HRV analysis and their limitations will be addressed in Chapter 3.

In the context of the field outlined above, this thesis explores techniques for signal processing of HRV from motivation to application, in an attempt to develop robust methods for HRV analysis that may be applied in implanted devices. A particular focus of the work is on minimising the number of features and HRV parameters to predict mainly VT, VF, and AF arrhythmias.

Thesis Contributions

The thesis aims to contribute first to an improved understanding of different types of ventricular and atrial arrhythmias, in both ischemic and nonischaemic heart disease and secondly to develop an HRV-based analysis method to predict and differentiate them from normal cardiac activities. This will contribute to the development of novel, more effective therapies for implanted devices such as ICDs. A further benefit of automatic prediction and detection of cardiac arrhythmias is the ability to develop time- and cost-effective clinical management procedures during a patient’s short-time visit to clinics. To achieve these targets, optimized HRV-based prediction methods have been proposed to overcome weaknesses of existing arrhythmia prediction studies.

Before listing contributions of this thesis, it is important to point out the scope

Of This Work:

1. The main cardiac arrhythmias which have been included in this thesis are onset of VT, VF and AF. 2. The data that have used in this work are coming from standard publicly- available database where a fair comparison with other work is possible.

3. HRV is the main input of all proposed prediction and detection methods. 4. All algorithms have been implemented in MATLAB 2018 using additional toolboxes and libraries.

The primary contributions of the work are as follows: 1. A comprehensive comparative study on the performance of HRV features in time domain, frequency domain, bispectrum, and nonlinear analysis in VT-VF prediction. As many as fifty HRV features and their performance in VT-VF prediction system are analysed, and a prediction study based on mixtures of these feature is carried out, achieving competitive results in comparison with the literature.

2. The application of feature selection techniques to HRV signals to predict VT and VF with a reduced set of features, with performance comparable to using an exhaustive set of features. This work proposes low-complexity pre-processing using a mix of wavelet transform and median filter. The features proposed in this study achieved a more balanced distribution of false positive and false negative results than related methods in the literature, using as few as six features which is the lowest published within this field.

3. Seven novel HRV features for onset of PAF detection are proposed, and evaluated in comparison with the state of the art. The method proposed in this work provides an improvement of 3% in accuracy over the leading published results in literature using the same classifier.

Thesis Structure

The structure of the rest of the thesis is as follows. Chapter 2 presents human heart anatomy, electrical functionality, and cardiac arrhythmia as heart abnormal behaviour. Chapter 3 summarises HRV and classical features in time domain, frequency domain, bispectrum, and nonlinear analysis.

Chapter 4 presents a full review of existing VT-VF prediction methods based on HRV signals from ICDs, highlights current issues, and discusses some directions and motivations for the research presented later in this thesis. It describes current research approaches, features and databases used to address the problem. A comparative study on the prediction performance of a range of features, using a common database, is also presented.

Chapter 5 proposes a feature selection technique to reduce the computational overhead of prediction. Signal buffer lengths of 1-minute and 5- minute duration are examined, feature extraction and selection are applied, and evaluation is carried out. Learning and evaluation results using cross- validation are compared with other studies in the literature.

Detection of PAF, a cardiac arrhythmia that can eventually lead to heart failure, is presented in Chapter 6. Early detection of PAF is crucial to prevent any further fatal cardiac arrhythmia such as VT or VF. Seven novel HRV- based features are presented in this chapter and analysed using feature ranking methods. To evaluate these features, they are applied to four standard classification techniques and their performance is compared to the existing state of the art from the literature. Finally, in Chapter 7 the summary, conclusion, and future work of this thesis are presented.

Peer-Reviewed Journal Publications

A. Parsi, M. Glavin, E. Jones, and D. Byrne, “Prediction of Paroxysmal Atrial Fibrillation using New Heart Rate Variability Features,” Comput.

Biol.

Med.,

p.

104367,

Apr.

2021,

10.1016/j.compbiomed.2021.104367. A. Parsi, D. Byrne, M. Glavin, and E. Jones, “Heart Rate Variability Feature Selection Method for Automated Prediction of Sudden Cardiac 10.1016/j.bspc.2020.102310.

A. Parsi, D. O’Loughlin, M. Glavin, and E. Jones, “Prediction of Sudden Cardiac Death in Implantable Cardioverter Defibrillators: A Review

Peer-Reviewed Conference Publications

A. Parsi, D. O’Loughlin, M. Glavin, and E. Jones, “Heart Rate Variability Analysis to Predict Onset of Ventricular Tachyarrhythmias in Implantable Cardioverter Defibrillators,” in 41st Annual International

Background

In this chapter, the fundamental basis of human cardiac signals relevant to this thesis is discussed. Firstly, an overview of the anatomy and electrophysiology of the human heart is presented. This includes consideration of the ECG and HRV, which are widely used in clinical practice for cardiac health monitoring , . Following this, cardiac arrhythmias which have been studied in this research are covered. Finally, in the last part of the chapter, two commonly-used cardiac rhythm management systems, automated external defibrillator, and implantable cardioverter defibrillator, are discussed.

Platforms like this are the main intended target application of the prediction algorithms developed in this thesis.

Cardiac Anatomy And Electrophysiology

The human cardiovascular system is responsible for circulating blood throughout the human body. The heart is located in the mediastinum, a region that extends from the sternum (breast-bone) to the vertebral column in the dorsoventral aspect, from the first rib to the diaphragm in the anterior- posterior aspect, and between the lungs. Heart has an electrical conduction system which makes your heart pump blood around your body.

The cardiovascular system is a closed circulatory system mainly composed of the heart, blood vessels and blood. Its function is to transport nutrients, oxygen, carbon dioxide, hormones, and blood cells throughout the organism in order to provide nourishment, protect from diseases, regulate body temperature and pH, and maintain homeostasis.

The Heart Is Composed Of Four Chambers :

 Two superior receiving chambers called atria (left and right), which receive blood coming back to the heart from all over the body.

Chapter 2

 Two inferior pumping chambers called ventricles (left and right) that pump blood out of the heart for supply to all parts of the body. Generally, the right chambers, the right atrium and right ventricle are responsible for collecting deoxygenated (oxygen-poor) blood from all parts of the body and supplying it to the lungs for oxygenation. The left chambers are responsible for collecting oxygenated (oxygen-rich) blood from the lungs through the pulmonary vein and supplying it all over the body through the aorta. The right atrium receives deoxygenated blood from the body through the superior vena cava and inferior vena cava; while oxygenated blood arrives to the left atrium coming back from the lungs, through the pulmonary veins.

Figure 2.1. Anatomical structure of the heart . Figure 2.1 displays the main anatomical structures of the heart. Atrioventricular valves connect atria and ventricles, ensuring unidirectional blood flows. The tricuspid valve separates the right atrium from the right ventricle, whereas the mitral valve separates the left atrium from the left ventricle. Two additional semilunar valves are located at the exit of each ventricle to prevent blood backflow, one on the aorta for the left ventricle and one on the pulmonary artery from the right ventricle. The flow of blood in and out of these chambers is controlled by the valves opening and closing in response to the pressure changes as the heart contracts and relaxes during a heartbeat. This contraction of the heart muscle is in turn modulated by the flow of electrical signals through the muscle .

Electrocardiogram Signal

The electrocardiogram signal captures the electrical activity of the heart over a period of time. The cardiac electrical system is based on the ability of cardiac cells to contract in response to the depolarization of the cells. At rest, the cardiac cell has less positive charge inside than outside so the net electric potential in the cell is negative. As the cell is stimulated, it goes into an excited state and in response it allows sodium ions (Na+), among a series of other positively charged ions, to diffuse into the cells; this entry of positive ions creates a positive charge inside the cell. This permeation process is known as depolarization and as depolarization occurs, the cardiac cell contracts. There is selective permeability of ions back and forth through cardiac tissue which have special ion channels. To repeatedly depolarize, cardiac tissue must return to its resting state; this return is called the repolarization .

The continuous cycle of depolarization and repolarization is the essential part of the cardiac electrical system. This cycle is conducted from cell to cell, causing an electric current to propagate. Depolarization of the atria and ventricles creates the contraction of the atria and ventricles. This wave of depolarization goes through the heart in the following path (Figure 2.2):

4. Purkinje Fibres

Figure 2.2. Heart conduction system.

Chapter 2

On every heartbeat, the electrical conduction begins in the SA node in the right atrium which acts as a natural pacemaker. From here it is transferred to the AV node. From AV node the impulse propagates through the bundle of His (the only electric connection between the atria and the ventricles) and then into specialized Purkinje fibres, spreading throughout the ventricles muscle . This conduction pathway is illustrated in Figure 2.2.

This cellular depolarization and subsequent repolarization or conduction are recorded as the ECG. To measure ECG, typically several electrodes are placed at several strategic positions over the heart. The electrodes detect the tiny electrical changes that occur because of the depolarization of the heart muscles during each heartbeat. The output from these electrodes is recorded over time and the output graph of electrical voltage (y-axis) versus time (x-axis) is the electrocardiogram (Figure 2.3).

Figure 2.3. Main wave, interval and segments in normal ECG signal. A typical ECG is depicted in Figure 2.3 and is composed of three main parts: P-wave (corresponding to atrial depolarization), the QRS complex (ventricular depolarization) and the T wave (ventricular repolarization). The ECG reflects the transfer of ions through the myocardium, which changes in each heartbeat. The isoelectric line is the baseline voltage of ECG which is traced following the T wave and preceding the next P-wave. The P-wave is a

13

small upward deflection on the ECG. The P-wave represents atrial depolarization, which arises in the SA node and spreads through contractile fibres in both atria. After the P-wave begins, the atria contract. During the plateau period of steady depolarization, the ECG tracing is flat. The QRS complex as a second wave, represents rapid ventricular depolarization, as the action potential reaches the AV node and spreads through the bundle branches and Purkinje fibres to all the ventricular contractile fibres. This causes the ventricles to contract shortly after the QRS complex appears and continues in the ST segment. At the same time, atrial repolarization is occurring, but it is not usually evident in an ECG because the larger QRS complex masks it. The final wave, which is called the T wave, is produced in the resting state of the ventricles. It indicates ventricular repolarization and occurs just as the ventricles are starting to relax. The T wave is smaller and wider than the QRS complex because repolarization occurs more slowly than depolarization. The repetitive cycle of the electrical activity of heart is represented by the P-QRS- T sequences. The ECG is a key diagnostic tool used to assess the health conditions of the heart. In reading an ECG, the amplitude of the waves and time intervals between the occurrence of the waves can provide clues to abnormalities. A summary description and typical durations for the different components of the ECG waveform is presented in Table 2.1.

Chapter 2

Table 2.1. Typical ECG durations for healthy person , .

P-Wave

During normal atrial depolarization, the main electrical vector is directed from the SA node toward the AV node and spreads from the right atrium to the left atrium. This turns into the P-wave on the ECG.

Interval

The P-R interval is measured from the beginning of the P-wave to the beginning of the QRS complex. The P-R interval reflects the time the electrical impulse takes to

Travel From The Sa Node Through The Av Node And

entering the ventricles. The P-R interval is, therefore, a good estimate of AV node function.

The Pr Segment Connects The P-Wave And The Qrs

complex. The impulse vector is from the AV node to the bundle of His to the bundle branches and then to the Purkinje fibres. This electrical activity does not produce a contraction directly and is merely traveling down toward the ventricles, and this shows up flat on the ECG.

The P-R interval is more clinically relevant.

Complex

The QRS complex reflects the rapid depolarization of the right and left ventricles. The ventricles have a large muscle mass compared to the atria, so the QRS complex usually has a much larger amplitude than the P-wave.

The St Segment Connects The Qrs Complex And The T

wave. The ST segment represents the period when the ventricles are depolarized. It is isoelectric.

Interval

The Q-T interval is measured from the beginning of the QRS complex to the end of the T wave. A prolonged Q- T interval is a risk factor for ventricular arrhythmias and sudden cardiac death. It varies with heart rate and needs a correction for clinical relevance.

T Wave

The T wave represents the repolarization (or recovery) of the ventricles. The interval from the beginning of the QRS complex to the apex of the T wave is referred to as the absolute refractory period. The last half of the T wave is referred to as the relative refractory period (or vulnerable period).

Heart Rate Variability

Heart rate variability (HRV) was first used clinically in 1965 when Hon and Lee noted that foetal distress was accompanied by changes in beat-to-beat variation of the foetal heart, even before there was detectable change in the heart rate. Heart rate (HR) is the number of heartbeats per minute, while HRV captures the temporal variation between sequences of adjacent heartbeats. It measures the time interval between instantaneous heartbeats calculated from cardiac R-R intervals (Figure 2.4). On a standard ECG, the maximum upwards deflection of a normal QRS complex is at the peak of the R wave (Figure 2.3), and the duration between two adjacent R wave peaks is termed the R-R interval. R-R intervals may reflect normal or pathological behaviour. The ECG signal requires editing before HRV analysis can be performed, a process requiring the removal of all non-sinus-node-originating beats. The period between adjacent QRS complexes resulting from normal sinus node depolarizations is termed the normal-to-normal (N-N) interval. HRV is the measurement of the variability of the N-N intervals .

Figure 2.4. The interval between an R wave and the next R wave; normal resting heart rate is between 60 and 100 bpm. R-R Interval duration generally varies between 0.6-1.2 seconds .

HRV is a measure of the balance between sympathetic mediators of HR (i.e. the effect of epinephrine and norepinephrine, released from sympathetic nerve fibres, acting on the SA and AV nodes), which increase the rate of cardiac contraction and facilitate conduction at the AV node, and parasympathetic mediators of HR (i.e. the influence of acetylcholine, released by the parasympathetic nerve fibres, acting on the SA and AV nodes), leading to a

Chapter 2

decrease in the HR and a slowing of conduction at the AV node. HRV operate on different time scales to help us adapt to environmental and psychological challenges. It also reflects regulation of autonomic balance, blood pressure, gas exchange, gut, heart, and vascular tone, which refers to the diameter of the blood vessels that regulate blood pressure .

HRV analysis has been widely used in clinical practice for detecting or predicting life-threatening arrhythmias – for many years. In the 1970s, Ewing et al. used short-term HRV measurements as a marker of diabetic autonomic neuropathy. In 1977, Wolf et al. showed that patients with reduced HRV after a myocardial infarction had an increased mortality, and this was confirmed by other studies showing that HRV is an accurate predictor of mortality post-myocardial infarction , . HRV reduces within 2 to 3 days after myocardial infarction, begins to recover within a few weeks, and is maximally but not fully recovered by 6 to 12 months .

Patients with persisting low HRV values have mortality almost three times greater than those with a normal HRV . Higher HRV is not always better since pathological conditions can increase HRV. When cardiac conduction abnormalities elevate HRV measurements, this is strongly linked to increased risk of mortality. An optimal level of HRV is associated with health and self- regulatory capacity, and adaptability or resilience. Higher levels of resting HRV are linked to performance of executive functions like attention and emotional processing by the prefrontal cortex . This signal as a main source of data for cardiac arrhythmias in the current research will be explained and investigated in the next chapter.

Cardiac Arrhythmia

Cardiac arrhythmia affects millions of people worldwide, causing almost half of deaths due to cardiovascular disease , . A cardiac arrhythmia is an abnormal or irregular heart rhythm. The normal heart rhythm is called normal sinus rhythm where the triggering impulses from the SA node propagate throughout the four chambers of the heart in a coordinated manner. Cardiac arrhythmia is the disruption of the normal electrical activity caused by an abnormality in that electrical conduction system. It can make the heartbeat too slowly, too quickly, or in an irregular way. Some cardiac arrhythmias are more serious than others which may lead to cardiac arrest and sudden cardiac death if not detected and left untreated .

There are many reasons why a person may develop a cardiac arrhythmia. For example, they are more common in older people, and in people with a

17

heart condition such as coronary heart disease or heart valve disease. The symptoms of an arrhythmia will depend on what type of arrhythmia is present, and how it affects the functioning of the heart. A small change in cardiac normal sinus pace may lead to severe arrhythmia attacks, which can reduce the ability of the heart to pump blood and causes shorting of breath, pain in chest, dizziness, tiredness, and loss of consciousness. These are the most common symptoms of an arrhythmia.

Cardiac Arrhythmia Is Of Two Main Types:

1. Bradycardia1, where the heartbeat is slower than 60 beats/min 2. Tachycardia, where the heartbeat is faster than 100 beats/min In bradycardia, typically, no symptoms are visible till a heart rate of about 50 beats/min. Low heartbeat rates are common during rest and also among certain highly trained athletes. The heart muscle of athletes has become conditioned to have a higher stroke volume and, so, requires fewer contractions to circulate the same volume of blood. Bradycardia is less likely to be congenital and older patients are most often affected. Shortness of breath, fatigue, and dizziness are common symptoms associated with bradycardia, severe cases may result in fainting. Bradycardia does not usually need treatment . However, if it is due to an underlying medical condition – such as an underactive thyroid gland – treatment may be needed for that condition.

In the other group, tachycardia2 also called tachyarrhythmia is the one that needed to be detected and treated in a more immediate manner as it is potentially more serious. In general, a resting heart rate over 100 beats/min is classified as tachycardia in regular adults. As the heart rate is controlled by electrical signals sent across heart tissues, tachycardia occurs when an abnormality in the heart produces rapid electrical signals. When the rate of the heart is too rapid, it may not effectively pump blood to the rest of the body, including the heart itself which provides less blood flow. This ends up depriving an individual’s organs and tissues of oxygen. The high heart rate or tachycardia can cause symptoms such as, dizziness, shortness of breath, light headedness, rapid pulse rate, heart palpitations, chest pain, and fainting or syncope. Some people with tachycardia have no symptoms, and the condition is only discovered during a physical examination or with a heart-monitoring test such as ECG. There are many different structural abnormalities that can alter electrical signals and lead to faster heart rates. The common types of 1 Bradycardia is derived from Greek, “brady” means slow and “cardia” means heart.

2 Tachycardia is derived from Greek, “tachy-” means quick or rapid and “cardia” means heart.

Chapter 2

tachycardia which are the main focus of this thesis are covered in the following sub-sections.

Supraventricular Arrhythmia

Supraventricular arrhythmias start with fast heart rates. They begin in the areas above the ventricles, such as the upper chambers which are heart’s atria, or the atrial conduction pathways. Generally, supraventricular or “atrial arrhythmias” are not as serious as ventricular arrhythmias and sometimes, they do not even require any treatment. Different types of supraventricular arrhythmias are SVT, paroxysmal supraventricular tachycardia (PSVT). AF, atrial flutter, and Wolff-Parkinson-White (WPW) syndrome.

Supraventricular arrhythmias can happen in response to a number of things, including tobacco, alcohol, or caffeine consumption, and cough and cold medicines. Supraventricular arrhythmias can cause shortness of breath, heart palpitations, chest tightness, and a very fast pulse.

Supraventricular Tachycardia

Supraventricular3 tachycardia is an overall term for any fast heart rhythm that starts from above the ventricles. SVTs are often ‘paroxysmal’, which means occasionally or from time to time. SVT happens when electrical signals in the heart’s upper chambers fire abnormally, which interferes with electrical signals coming from the SA node, the heart’s natural pacemaker. The beats in the atria then speed up the heart rate. SVT is a rapid, regular heart rate where the heartbeats anywhere from 150-250 beats/minute in the atria. This type of arrhythmia is more common in infants and young people. It is also more likely to occur in women, anxious young people, and people who are extremely tired (fatigued). People who drink a lot of coffee or alcohol or who are heavy smokers also have a greater risk. SVTs are quite common, but are rarely life-threatening .

Atrial Fibrillation

Atrial fibrillation – or AF for short – is a fast, irregular rhythm where single muscle fibres in the heart twitch or contract. AF happens when different places in and around the atria fire off electrical impulses in an uncoordinated way . These random electrical impulses result in rapid, uncoordinated, and weak contractions of the atria. The chaotic electrical signals also affect the AV node, 3 Supra means above.

19

and this usually results in an irregular rhythm of the ventricles. The origin of this chaotic activity is around the pulmonary veins . AF is the most common type of arrhythmia affecting adults of any age .

In Europe and North America alone, about 3% of the population are affected by AF . Approximately 4% of the population over 65 years old are affected by AF, and 10% of those older than 80 years old. People usually need treatment to control their AF. The type of treatment depends on several factors, including the type of AF. AF may cause blood to pool in the heart’s upper chambers.

The pooled blood can lead to the formation of blood clots. A stroke can occur if a blood clot travels from the heart and blocks a smaller artery4 in the brain.

Ventricular Arrhythmia

Ventricular arrhythmias are fast, abnormal heart rhythms that start from the ventricles. Most ventricular arrhythmias are caused by underlying heart disease, and can often be life-threatening . The two main types of ventricular arrhythmias are ventricular tachycardia (VT for short) and ventricular fibrillation (VF for short). VT and VF are the major causes of sudden cardiac death (up to 80%) in patients with structural heart disease .

Sudden cardiac death is defined as death from an unsuspected circulatory arrest, usually due to an arrhythmia occurring within an hour of the onset of symptoms .

Ventricular Tachycardia

Ventricular tachycardia - or VT for short - is a rapid heart rate that originates with abnormality of electrical impulses in the lower chambers or ventricles of the heart. This rapid heart rate does not allow the ventricles to completely fill with blood as well as contract fully to pump enough blood to the body (Figure 2.5). Since the new electrical impulses in the lower chambers do not move through the heart muscle along the regular route, the heart muscle does not beat normally.

VT is technically defined as a series of three or more consecutive premature ventricular impulses with QRS duration > 120 milliseconds at rate between 100 and 300 beats/minute. In this state, the QRS complex is not clearly present in the ECG signal. VT may be monomorphic or polymorphic. Monomorphic VT consists of a stable (organized) single QRS morphology often with a left bundle or right bundle block pattern. Polymorphic or multiform VT exhibits an irregularly changing QRS configuration with variable cycle lengths between

Chapter 2

600 and 180 milliseconds. VT may be sustained (duration > 30 seconds) and/or require termination due to hemodynamic compromise in less than 30 seconds or non-sustained (three or more beats but less than < 30 seconds) . VT is often a life-threatening medical emergency and without heart rhythms treatment it will cause death within minutes .

Ventricular Fibrillation

Ventricular fibrillation - or VF for short - occurs when rapid, chaotic electrical impulses cause the ventricles to quiver ineffectively and so they do not function effectively as pumping chambers. This lack of pumping means the body and brain do not receive sufficient blood in time. VF produces a completely disorganized tachyarrhythmia usually faster than 300 beats/minute with random, asynchronous electrical activity of the ventricles . There are no discrete QRS complexes.

Tachycardia brought on by VF can be very serious and can even be fatal unless treated very quickly. This is the most serious arrhythmia, causing cardiac arrest and most people who experience VF have an underlying heart disease or have experienced serious trauma . In comparison with VT, instead of one misplaced beat from the ventricles, VF may have several impulses that begin at the same time from different locations, all telling the heart to beat (Figure 2.5). VF is more serious than VT and without immediate medical attention and treatment like appropriate electrical shock, it causes sudden cardiac death .

One Misplaced Electrical Impulse In

the left ventricle.

Several Misplaced Electrical Impulses

in the left and right ventricle. Figure 2.5. Examples of ventricular arrhythmias showing misplaced electrical impulses in lower side of the heart.

Cardiac Rhythm Management Approaches

Increasing incidences of arrhythmia and heart failure have fostered rapid advancements in cardiac rhythm management (CRM) devices and technologies, and associated clinical practice , . Currently, automated external defibrillator (AEDs) as an out-of-hospital cardiac arrest treatment, and single-chamber pacemakers and/or implantable cardioverter defibrillator (ICD) as a permanent treatment, are the most efficient therapy for sudden cardiac death prevention –.

The use of CRM devices has substantially surged over the past few decades due to increasing life expectancy and constant technological advancements that have extended their scope and abilities. With rising indications and accessibility to CRM around the world, the number of patients requiring such devices is subsequently increasing . Such technologies have transformed clinical electrophysiology into one of the most rapidly expanding cardiology subspecialties. Pacemakers are the accepted standard of care for those with complex atrial and ventricular arrhythmias , . In addition, technological innovations have resulted in the emergence of remote monitoring as an effective alternative to routine in-office patient care.

Automated External Defibrillator

An AED is portable device which includes a shock advice algorithm that checks the surface ECG and can send electrical shock if either VT or VF are detected by its algorithm . These defibrillators are able to automatically detect the presence of VT and VF and perform cardioversion to cardiac arrest victims.

When a cardiac arrest case is suspected, an AED operator can place electrodes on the patient's chest. The AED will detect the patient's condition, confirm the presence of a cardiac arrest, and decide if electric shock intervention is needed. The AED will also instruct the operator to carry out necessary steps for automated external defibrillation of the patient .

Implantable Cardioverter Defibrillator

An ICD is a battery-operated device placed under the skin in the chest that constantly monitor cardiac activities by using thin leads directly lodged into the right ventricle, and are the usual treatment for patients with substantial daily risk of severe arrhythmia . By processing intra-cardiac electrical activities internally (called EGM), ICDs can decide to deliver electrical shock, if necessary, to restore normal heart rhythm. Cardioversion and defibrillation

Chapter 2

are both forms of high-energy shocks that stop dangerous arrhythmias and restore a normal heart rhythm. If the patient is conscious (awake) at the time of the shock, it is painful and usually described informally as feeling like a “kick in the chest”. Most of the time, however, a shock for a dangerous arrhythmia is delivered after the patient has lost consciousness and is therefore unable to feel the shock.

Antitachycardia pacing – or ATP for short – is an alternative method of stopping ventricular arrhythmias and is available in most ICDs . It involves delivering a short series (e.g., 5 to 10) of paced beats. This is not painful and may be unnoticed by the patient, although some patients may feel a brief burst of palpitations. ATP can be very effective for some slower ventricular arrhythmias (e.g., 150 to 200 beats/minute), and is often programmed as the initial therapy for these arrhythmias. High-energy shocks are often programmed as the initial therapy for very fast rhythms (e.g., more than 200 to 220 beats/minute), or as a rescue therapy if ATP treatment fails.

Nowadays, developing differential algorithms to recognize VT and VF arrhythmias from other cardiac rhythms is one of the most critical issues in ICD patients. The main goal of this process is to develop algorithms with highest possible sensitivity to avoid untreated arrhythmias and highest possible specificity to avoid inappropriate ICD shocks . Nearly 80% of inappropriate shocks are caused by misclassification of supraventricular arrhythmias as a VT or VF , . Supraventricular arrhythmias include SVT, AF, and atrial flutter, which do not need any immediate therapy.

Summary

This chapter has discussed some important aspects of the cardiac system and their relevance in the context of this thesis. Cardiac anatomy and electrophysiology were first discussed, including a description of the ECG waveform and HRV, which are central to this thesis. Following this, the main types of cardiac arrhythmia were presented. Finally, some current practice in the treatment of cardiac problems through external and implanted defibrillation devices were briefly considered. Such platforms can be considered a target application for the techniques developed later in this thesis.

The next chapter continues consideration of HRV analysis, through a detailed discussion of the main features that are typically used for quantitative analysis and diagnosis of arrhythmias.

Heart Rate Variability Analysis

This chapter introduces HRV analysis principles, measurements, and features that have been used in cardiac arrhythmia prediction and detection research. HRV is a reliable reflection of the many physiological factors modulating the normal rhythm of the heart. It shows that the structure generating the signal is not only simply linear, but also involves nonlinear contributions. Over the last three decades, alterations in HRV have been found in patients with many cardiovascular conditions . Since the introduction of CRM devices such as different type of ICDs, many studies have examined HRV processing, analysing effects such as the circadian rhythm in HRV signals to analyse significant changes immediately prior to an episode of VTA , –. Some of the material in this chapter has been published in Parsi et al., “Prediction of Sudden Cardiac Death in Implantable Cardioverter Defibrillators: A Review Variability Analysis to Predict Onset of Ventricular Tachyarrhythmias in Implantable Cardioverter Defibrillators,” in 41st Annual International Jul. 2019, pp. 6770–6775 .

Introduction To Hrv

HRV measures variations in the heart rate and can reflect both the sympathetic and the parasympathetic components of the ANS. It is defined as the physiological variation in the duration of intervals between sinus beats and serves as a measurable indicator of cardiovascular integrity and prognosis. The instantaneous heart rate (IHR) signal is calculated from the R-R intervals recorded by e.g., an ICD using the following equation:

Chapter 3

Where 𝑅𝑅𝐼 (R-R intervals) is the time in milliseconds between instantaneous heartbeats as measured by any CRM devices, or as described in Chapter 2, the period between adjacent QRS complexes resulting from sinus node depolarizations, which is termed the N-N interval. The unit of IHR is beat/minutes (bpm). HRV features could be calculated from measurements of the R-R intervals (N-N intervals) or IHR signal (in this thesis, mostly IHR signals were used). By way of example, Figure 3.1 (a) shows the variations of the R-R intervals of normal (which is called CON) and abnormal (VT and VF) samples of a typical patient from the Medtronic version 1.0 dataset which will be described in depth in Chapter 4. Figure 3.1(b) shows the corresponding IHR derived from these RRI traces. The significant variation in IHR for normal and abnormal conditions is evident6.

Formal criteria for HRV and comparison of variables were developed by a joint task force between the European Society of Cardiology and the North American Society of Pacing and Electrophysiology in 1996 and updated in 2015 . Generally, the major features in HRV analysis can be divided in four main groups as described in the following subsections.

6 VT and VF samples in these charts show cardiac activities 2 minutes before related tachyarrhythmias happening. The “CON” sample, on the other hand, was captured on a normal patient visit and check-up routine with specialist supervision.

25

Medtronic version 1.0 dataset. Medtronic version 1.0 dataset. Figure 3.1. (a) RRI for normal, VT and VF conditions for a typical cardiac patient ; (b) Corresponding IHR.

Time Domain Hrv Analysis

Standard time domain parameters of HRV analysis are the simplest to extract and are valuable measures of variation of the R-R intervals . The most commonly used time domain measures are the mean of N-N intervals or instantaneous heartbeats (MeanNN) and standard deviation (SDNN).

Standard deviation is the square root of variance. Since variance is mathematically equal to total power of spectrum, SDNN reflects all the cyclic components responsible for variability in the period of recording . Other time domain features include the root mean square of successive R-R intervals differences (RMSSD), number of adjacent R-R intervals differing by more than 50 milliseconds (NN50), and pNN50, the number obtained by dividing NN50 by the total number of N-N intervals , –. HRV triangular index (HRVTri), which is a total number of N-N intervals used in histogram analysis divided by the maximum of the signal histogram, is employed in HRV analysis to extract and evaluate geometrical feature , , . As the value of HRVTri is dependent on the size of the bin in the histogram, the size should be specified if the HRV signal is resampled to different frequencies. Summaries of time domain HRV features have presented in Table 3.1.

Table 3.1. Summary of HRV features in time domain.

Pnn50

Percentage of successive N-N intervals that differ

Frequency Domain Hrv Analysis

Although the time domain methods are computationally simple, many cannot discriminate between the sympathetic and the parasympathetic contributions in the HRV signals. Sympathetic activity is associated with the low-frequency modulation of the heart rate (0.04-0.15 Hz) while parasympathetic activity is associated with the higher frequency modulation (0.15-0.4 Hz). Therefore, frequency domain features can be used to separately estimate the sympathetic and the parasympathetic contributions in the HRV signals. This facilitates preventive intervention, such as ATP, before the onset of sudden cardiac death .

Many studies used average power of HRV signals across the entire frequency band in addition to the average power for specific frequency bands. The power in a given band is typically computed by integrating the power spectral density (PSD) estimate in the frequency range of interest, such as very low frequency (VLF) band (0.0-0.04 Hz), low frequency (LF) band (0.04-0.15 Hz), high frequency (HF) band (0.15-0.4 Hz) and comparing different ratio variation like LF/HF or VLF/HF , , –. These frequency domain features (0.04- 0.4 Hz) can estimate the sympathetic and parasympathetic contributions to HRV signal, commonly reflected in changes in energy distribution prior to sudden cardiac death .

For example, studies have shown that there is an increase in spectral power at lower frequencies that can help predict the onset of VT and VF, and also that the spectrum power increases just prior to breakdown of normal rhythm to a ventricular arrhythmia . Similarly, spectral power analysis using either Welch’s method for periodogram spectrum estimates , – or fast Fourier transform, found similar changes prior to both VT and VF , , , , , regardless of processing time and computational complexity that each algorithm has , . Selected HRV features in frequency domain are presented in Table 3.2.

Table 3.2. Summary of selected HRV features in the frequency domain.

Bispectrum Analysis

Poly-spectral analysis using different higher-order spectral (HOS) features uses spectral representations of higher order moments or cumulants of a signal . HOS, up to third-order cumulant, was employed to estimate the bispectrum from IHR data in . The HOS features described in this section consider the phase of the signal in addition to magnitude from the spectral power described in the previous section . Since the HRV signal is nonlinear and non-Gaussian in nature , , the bispectrum which employs up to the third order statistics can be used to reveal information not present in the spectral domain. Also, this feature has been employed to reveal the hidden information that cannot otherwise be observed by using regular power spectrum calculated from the second-order moment of the signal and to detect quadratic phase coupled harmonics arising from nonlinearities of the HRV signal7, in the recent HRV analysis based research , , , , .

Bispectrum analysis may be carried out as follows. Let 𝑥(𝑛) represent a discrete, zero-mean random signal. The third order cumulant sequence 𝑅(𝑚, 𝑛) of the signal is identical to its third-moment sequence, and it may be calculated

(3.2)

Where 𝑘 is sample index, 𝑚, 𝑛 are shift values in samples and 𝐸{… } denotes the expectation operation. The bispectrum 𝐵(𝜔ଵ, 𝜔ଶ) of 𝑥(𝑛) is defined as the two-dimensional Fourier transform of 𝑅(𝑚, 𝑛) as a follow:

(3.3)

with the condition: |𝜔ଵ|, |𝜔ଶ| ≤𝜋 for 𝜔= 2𝜋𝑓 . As the third-order moment and cumulant are identical, a bispectrum is the spectrum of these values.

The bispectrum is a symmetric function which has 12 symmetric regions however, there are three sub-bands inside the region of interest (ROI) of the bispectrum which can discriminate sympathetic and parasympathetic contents of HRV signals . Figure 3.2 below indicates these sub-bands (inside the triangular area indicated with red line) as LF-LF (LL), (LF-HF) (LH), and HF-HF (HH). In previous work, features have been extracted from these three 7 Studies shows mechanisms involved in the regulation of cardiovascular system interact with each other in a nonlinear way.

29

regions and the entire ROI to characterize HRV. The most commonly used features derived from the bispectrum are covered in the following sub-sections. Figure 3.2. Segmentation of the bispectrum into sub band regions (LL, LH, and HH), and the ROI for bispectrum feature extraction, indicated by the red (with dash-dot line) triangle.

Magnitude And Power Average

Two heuristic features are the average magnitude (𝑀௔௩௚) and the average power (𝑃௔௩௚) in LL, LH, HH sub-bands, and the entire ROI. They are defined

(3.5)

where Γ represent one of four regions previously described (LL, LH, HH, and ROI in Figure 3.2) and 𝑁 is the number of points within Γ. These values represent fundamental bispectral properties of specific region.

Bispectrum Entropies

The mean magnitude of the bispectrum can be useful in discriminating between processes with similar power spectra but different third order statistics. However, it is sensitive to amplitude changes . Entropy has been shown to be useful as a measure of “randomness” in signal analysis and bispectral entropies have been successfully exploited to measure the regularity of nonlinear signals like EEG . For any region Γ from LL, LH, HH, and ROI, the normalized bispectrum entropy (𝐸௡௕) is defined as:

(3.6)

and the squared normalized bispectrum entropy (𝐸௦௡௕) is defined as:

(3.7)

Both normalized entropy measures 𝐸௡௕ and 𝐸௦௡௕ lie between 0 and 1.

Logarithmic Bispectrum Features

Linear amplitude scaling is often used to estimate the relative importance of spectral features, however, use of linear scaling can mask small, but potentially important features. Logarithmic scaling of the bispectrum has been shown to be effective in characterizing clinical EEG . Two main logarithmic features have been used in HRV classification studies , , , , namely the sum of the logarithmic magnitude (𝐿௠) and the sum of the logarithmic magnitudes of the diagonal elements of the bispectrum (𝐿ௗ௠), as an attempt to emphasize both larger and smaller components. These features are defined

(3.9)

where D represents the diagonal elements of the log bispectrum of a region Γ (one of LL, HH, and DOI; there are no diagonal elements in LH).

Weight Centre Of Bispectrum

The weight centre of bispectrum (WCOB) represents the weight centre of the contour map of a bispectrum. This feature has been used in recent studies as a valuable parameter in HRV event detection methods , . Researchers in ischemic EEG analysis have reported differences in the morphology of ischemic and normal EEG bispectral contour maps . As cardiac abnormality might also show differences in the HRV bispectral contour map, WCOB has been adapted for use in previous studies by two frequency axes

(3.11)

for any Γ in LL, LH, HH, and ROI where 𝑖, 𝑗 are the first and second frequency indices, respectively. Summaries of HRV bispectrum analysis features have presented in Table 3.3.

Table 3.3. Summary of bispectral HRV features.

Nonlinear And Dynamic Analysis

Viewing HRV as an indirect electrical measure of autonomic heart rate modulations, influenced mainly by the permanent interplay between the two branches of the ANS and considering this as the output of a nonlinear system, can lead to better understanding of cardiac system dynamics –.

Studies have also stressed the importance of nonlinear techniques to study HRV in issues related to both cardiovascular health and disease , . Among them, the Poincaré plot has been demonstrated to be a simple and effective method for discriminating cardiovascular diseases based on HRV . The Poincaré plot displays each R-R interval as a function of the previous R-R interval, which can indicate the degree of heart failure in a subject . An ellipse can be generated to fit these data points. The width (𝑆𝐷ଵ) of this ellipse (minor axis), which is related to the fast beat-to-beat variability in the data, and the length (𝑆𝐷ଶ) of the ellipse (major axis) relate to the longer-term variability of those data . These values can be directly calculated from two of the time domain features: 1) the standard deviation of differences between adjacent R-R intervals (𝑆𝐷𝑆𝐷), and 2) the mean of the

(3.13)

There are more features that have been extracted from Poincaré plot 𝑆𝐷ଵ/𝑆𝐷ଶ for detecting VTAs one hour before occurrence , as well as HRV density analysis (Dyx) based on the Poincaré plot . Other symbolic dynamics to distinguish between different states of the autonomic interactions , like Polvar10 (measures the incidence of 6 beat segments during which the beat-to-beat difference of R-R intervals is less than 10 ms) and Fitgra9 (finite time growth rates, dimension 9) regarding the complexity of the sinus node activity modulation system have been proposed and investigated in . The results showed significant differences in the dynamic behaviour before onset of

Vt.

Entropy as a measure of signal complexity has also been applied. Sample entropy (SampEn) and approximate entropy (ApEn) have been investigated to more strongly show HRV complexity and regularity changes in onset of ventricular arrhythmias . The symbolic parameters, FWShannon, ForbiddenWords and Hjorth’s parameters have also been applied .

33

FWShannon denotes the Shannon entropy of the word distribution whereas Forbidden Words stands for the number of words, which never or very seldom occur. Larger values of FWShannon reflect higher complexity whereas a high number of Forbidden Words reflect regular behaviour in the signal .

Conclusions can be drawn that when a ventricular arrhythmia approaches, the sympathetic tone of the patients is increased, and the complexity of their R-R intervals immediately before the onset of VT-VF events is lower than normal sinus rate and hence the values of entropy-based measures are significantly smaller for VT-VF episodes than those for normal series. Hjorth’s parameters are used to capture the statistical properties of nonlinear time-series signals.

Different Hjorth’s parameters namely activity, mobility and complexity are used which relate to the signal power, mean frequency, and change in frequency respectively .

Selected HRV features in nonlinear and dynamic analysis are presented in Table 3.4.

Chapter 3

Table 3.4. Selected HRV features in nonlinear and dynamic analysis.

𝑆𝐷ଵ/𝑆𝐷ଶ

Using two time domain measures, the standard deviation of differences between adjacent R-R intervals (𝑆𝐷𝑆𝐷) and the mean of the standard deviation of all R-R intervals (𝑆𝐷𝑅𝑅), 𝑆𝐷ଵ and 𝑆𝐷ଶ are calculated as presented in formula (3.12) and (3.13).

Entropy

Assume 𝑋(𝑡) = {𝑥ଵ, 𝑥ଶ… , 𝑥௡} represent the signal. The template vector of lenth 𝑚 defines such that 𝑋௠(𝑡) = {𝑥௜, 𝑥௜ାଵ… , 𝑥௜ା௠ିଵ} and the distance function 𝑑[𝑋௠(𝑖), 𝑋௠(𝑗)] is to be Chebyshev distance.

⁄ , Where

𝐴 = number of temple vector pair having 𝑑[𝑋௠ାଵ(𝑖), 𝑋௠ାଵ(𝑗)] < 𝑟 𝐵 = number of temple vector pair having 𝑑[𝑋௠(𝑖), 𝑋௠(𝑗)] < 𝑟 Generally, 𝑚= 2 and the value of 𝑟 would be 0.2 × 𝑆𝐷 where 𝑆𝐷 represent standard deviation of the input signal.

Entropy

The Rényi entropy of order 𝛼 where 𝛼≥0 and 𝛼≠1 is defined as:

), Where 𝑋(𝑡) = {𝑥ଵ, 𝑥ଶ… , 𝑥௡}

represent the signal and corresponding probabilities 𝑝௜= Pr (𝑋=

௜ୀଵ

= 1.

Entropy

The Tsallis entropy of order 𝛼 where 𝛼≥0 and 𝛼≠1 is defined

𝑋(𝑡) =

{𝑥ଵ, 𝑥ଶ… , 𝑥௡} represent the signal and corresponding probabilities 𝑝௜= Pr (𝑋= 𝑥௜) for 𝑖= 1, 2, … , 𝑛with the condition ∑

௜ୀଵ

= 1.

ൗ

which represent the signal power, mean frequency, and change in frequency respectively, where 𝑥(𝑡) represents the signal and 𝑣𝑎𝑟 is the variance of time function.

Discussion And Concluding Remarks

Originally, the aim of capturing HRV signals and investigating changes in HR were its use in the prediction of long-term survival in patients who had suffered myocardial infarction or had valvular or congestive heart disease .

Detecting HRV changes over a period of hours or even days requires a large volume of ECG data to be collected and analysed. This process traditionally was difficult and limiting for patients as Holter devices that record the ECG or in hospital monitoring were the only options. Also capturing dynamic changes in HRV in the period prior to arrhythmias events is harder to attain, due to the relative infrequency of such events. More recently, work has concentrated on attempts to predict the timing of onset of fatal arrhythmias like VT and VF as more CRM devices like advanced ICDs have been developed , . Most of the cardiac arrhythmias such as ventricular arrhythmias have a circadian rhythm with increased frequency during the early morning (7–11 am) and early evening (6–7 pm) . This is consistent with findings of a circadian distribution of cardiac arrests, with a low incidence at night, and peaks between 0800 - 1100 hours and 1600 - 1900 hours . The evening peak may be attributable to VF, while the morning peak may be attributable to patients in non-VF and non-VT rhythms , . The characteristics of HRV immediately prior to the onset of VTA, however, need more complex methods to be detected. Some studies report significant changes in HRV in the period immediately preceding a ventricular arrhythmia, and HR increasing prior to an episode of VT- VF is common , –, .

Studies also show a significant change in HR, SDNN, and power in VLF, HF, and HF in patients who developed VT-VF sometimes one hour prior to the event , , and an alteration in the interaction between the sympathetic and parasympathetic nervous system prior to the onset of ventricular arrhythmias . The effect on HR and HRV variables is more often heterogeneous and affected by individual patient characteristics, which make the use of advanced machine learning methods necessary to fully capture this variability.HRV as a useful tool to assess sympathetic and parasympathetic influences on disease states and is thus a useful application in assessing autonomic dysfunction in patients with AF , . Recent study shows there is a modest correlation for a baseline HR < 60 beats/minute with an increased risk of AF. The study looked at data from a large database where subjects were followed up for an average of 20 years, allowing them to identify factors such as HRV and resting HR which might predict the chances of contracting AF years later . One of the goals of this research is to combine different HRV measurements in time domain, frequency domain, bispectrum, and nonlinear analysis with the state-of-the-art machine learning

Chapter 3

algorithms for the prediction or detection of life threatening arrhythmias. As a step towards this, the next chapter in this thesis presents a detailed evaluation of a range of the features described in this chapter for this purpose. This includes a review of the relevant literature in the field, databases used in previous work, and discussion of relevant experimental results from the literature.

Heart Rate Variability Analysis

ICDs and other technology platforms have been widely deployed in the last four decades to reduce sudden cardiac death risk in patients with a history of life-threatening arrhythmia. By continuous monitoring of the heart rate, ICDs can use decision algorithms to distinguish normal cardiac sinus rhythm or SVT from abnormal cardiac rhythms like VT and VF and deliver appropriate therapy such as an electrical stimulus. Despite the success of ICDs, more research is still needed, particularly in decision-making algorithms. Because of low specificity in practical devices, patients with ICDs still receive inappropriate shocks which may lead to inadvertent mortality and reduction of quality of life. At the same time, higher sensitivity can lead to the use of newer tiered therapies.

This chapter presents a review of existing methods for ventricular arrhythmia prediction in ICDs based on HRV analysis, highlights current issues, and discusses some directions for future research. Specifically, the varied features which have been used for VT-VF prediction are reviewed and experimentally evaluated using a classification algorithm, and results compared with results from the literature. More than fifty different features to address heart rate changes before sudden cardiac death are considered and a general methodology for experimental evaluation is proposed, based on a variety of studies of ICD functionality. A comparative study on the prediction performance of these features, using a publicly-available database, is presented, and the results compared with those from the literature where possible. The work described in this chapter has been published in Parsi et al., “Prediction of Sudden Cardiac Death in Implantable Cardioverter Defibrillators: A Review

Overview

This section reviews relevant topics for HRV analysis for arrhythmia prediction and detection. It starts with a brief overview of the clinical usage of this technology, as this provides motivation for the functional requirements of lower-level signal processing algorithms; on the one hand, sensitivity is important but at the same time, specificity is also important, not least in order to avoid inappropriate therapy. Two common types of arrhythmias are VT and VF, both of which have the potential to cause sudden cardiac death. As presented in Chapter 2, VT and VF both show fast hart rate sometimes even more than 300 beats/minute with chaotic and asynchronous electrical activity in the ventricle. Without immediate and appropriate treatment such as electrical shock, VF can result in sudden cardiac death and its distinctive characteristics motivate the use of machine learning to distinguish these rhythms from normal cardiac activity. ICDs are an indicated treatment for patients at high risk of sudden cardiac death due to VT or VF . These battery-operated devices are implanted in the body and constantly monitor cardiac activity through thin leads lodged into the right ventricle. ICDs are typically indicated for patients with substantial daily risk of severe arrhythmia . ICDs have been proven to reduce the mortality of patients with life- threatening ventricular arrhythmias and have become the therapy of choice for patients at high risk of sudden cardiac death.

By processing intra-cardiac electrical activities between electrodes, called an electrogram (EGM), ICDs may deliver one of two treatments if either VT or VF is predicted: (1) direct current defibrillation (electrical shock) or (2) ATP, to restore normal heart rhythm and prevent sudden cardiac death occurring , , . A critical issue for ICD design is the development of algorithms to predict VT and VF arrhythmias and distinguish VT and VF from the healthy cardiac activity. Both sensitivity and specificity are important: highly sensitive algorithms minimize the risk of untreated arrhythmias and hence sudden cardiac death, but high specificity is also important to avoid invasive interventions such as electrical shock where inappropriate or unnecessary .

Unnecessary and inappropriate interventions are associated with increased mortality, and more than 11.5% of 719 preventative individual ICD patients received one or more inappropriate electrical shocks based on MADIT II research in 2008, constituting 31.2% of all electrical shocks applied in therapy overall . Nearly 80% of inappropriate electrical shock interventions are caused by the misclassification of SVT as a VT or VF . SVT is an abnormally fast heart rhythm, however, SVT in itself does not require any immediate therapy .

39

Recent ICDs offer tiered therapy and include ATP to treat VT and prevent sudden cardiac death. ATP requires a longer duration of analysis and hence better VT prediction, but the use of ATP reduces the incidence of electrical shock, causes less patient discomfort and requires less energy . Furthermore, by analysing longer cardiac signals and predicting VT-VF events before they occur, ATP can be considered the most appropriate treatment at this early stage. Should ATP be ineffective, more aggressive therapies, typically electrical shock, are administered immediately after unsuccessful ATP or if a high probability of VF is detected.

Historically, EGM signals were used to extract time or the R-R interval, which is the time between instantaneous heartbeats. The R-R interval can be converted to HRV, which is used for diagnosis. Advances in embedded device computation power, programmability, memory functions, device-related history features, therapy options and battery life have enabled EGM signal morphology to be used in ICDs . SVT, VT, and VF usually activate the ventricle in a distinct manner, and as a result, different shapes are expected from ventricle electrogram morphology as well, which can be used to distinguish these cases. Algorithms based on differences in EGM morphology use either direct models of heartbeat (QRS complex) or models using wavelet features , , covariance-based model , spatial projection of tachycardia or an indirect/implicit model of heartbeat such as similarity measures which are based on information theory concepts . In contrast, interval-based algorithms are less computationally intensive, utilizing information such as rate, onset, stability and entropy from the HRV signal , –. No consensus exists as to the optimal algorithms for onset of VT and VF detection, nor as to methods to decrease the inappropriate electrical shock interventions due to misclassification of SVT, VT and VF .

A number of aspects need to be considered when reviewing prediction algorithms used in ICDs: 1) feature extraction and classification methods, which are computationally complex; 2) accuracy (sensitivity and specificity); 3) window length required; and finally, 4) expected time remaining until sudden cardiac death, which affects the available therapeutic options for the ICD. Numerous machine learning methods have been proposed to address these concerns, however, in each case, different subsets of features derived from HRV have been used to predict sudden cardiac death. The focus in this thesis is on interval-based algorithms. Chapter 3 presented a detailed description of the individual features. Table 4.1 summarises previous research from the literature that makes use of these features in various combinations. This table includes the features used in each study, the classification method, as well as a summary of the main outcome from the study.

Chapter 4

Table 4.1. Summary of previous studies conducted in prediction of sudden cardiac death using R-R interval signals from ICD devices.

Work

Used Features, Classification Methods and Outcomes

1999

Features: time domain: MeanNN, frequency domain: spectral power analysis in different frequency band from 0-0.1 to 0.9-1.0 Hz

Classification: Simple Thresholding

Outcome: best result in 0.8-0.9 Hz frequency band with time window = 100 seconds just before sudden cardiac death.

Al., 2000

Features: time domain: MeanNN, frequency domain: spectral power analysis in VLF, LF, HF, and total power frequency (TPF)

Band (0.0033-0.4 Hz), And Lf/Hf Ratio

Outcomes: decrease in MeanNN before VT onset and increase in LF/HF component before VT onset.

Al., 2002

Features: time domain: MeanNN, SDNN, pNN50, RMSSD, frequency domain: spectral power analysis in VLF, LF, HF, TPF, and following ratios: VLF/TF, LH/HF, HF/P, LFn (LF power in normalised units: 𝑃𝑜𝑤𝑒𝑟௅ி(𝑃𝑜𝑤𝑒𝑟்௉−𝑃𝑜𝑤𝑒𝑟௏௅ி)

100), Dynamic Analysis: Polvar10, Fitgra9

Outcomes: significant differences in the dynamic behaviour, significant MeanNN difference before onset of VT and the incidence of VT is higher during the daytime than at night.

2007

Features: time domain: HR and acceleration pattern Classification: detecting two patterns (each 50 R-R intervals) in 1.8 hour with heart rate exceed 86 bpm (700 ms) (rule-based

Decision)

Outcomes: the basic acceleration pattern was found during sinus rhythm in 1.8-hour period prior to 83% of VT episodes.

2008

Features: time domain: MeanNN, SDNN, RMSSD, pNN50,

Classification: Simple Thresholding

Outcomes: obvious increase in HR of the patients, the values of two entropy measures are significantly smaller for VT-VF episodes than those for normal series.

2008

Features: time domain: MeanNN, SDNN,, RMSSD frequency domain: spectral power analysis in LF, HF, and LF/HF, LFn, and LHn ratio, nonlinear analysis: FWShannon, Forbidden Words, Polvar10, along with other features: PVCs (premature ventricular complexes in %), VTCI (VT coupling interval - CI for short - in millisecond), RRn-1 (in millisecond), CI/RRn-1, VT cycle

Length (In Millisecond)

Outcome: Polvar10 gave significant changes with respect to controlled data as early as 20 min prior to the VT onset.

Work

Used Features, Classification Methods and Outcomes

2009

Features: frequency domain: spectral power analysis of sub-

Bands In Lf Base-Bands And Hf Base-Bands

Classification: multilayer perceptron (MLP) neural network Outcome: most dominant frequency range in LF is [0.039-0.085] and in HF is [0.195-0.281], however, there is no result for general sudden cardiac death prediction.

2011

Features: time domain: MeanNN, SDNN, frequency domain: frequency marginal spectrum and Hilbert spectrum in LF, HF VHF,

And Lf/Hf Ratio

Outcome: the amplitude (energy) of Hilbert marginal spectrum of HRV signal preceding the onset of VT-VF event significantly increases.

, ,

Features: time domain: MeanNN, SDNN, RMSSD, pNN50, frequency domain: spectral power analysis in VLF, LF, HF, and LF/HF ratio, nonlinear analysis: 𝑆𝐷ଵ, 𝑆𝐷ଶ, 𝑆𝐷ଵ/𝑆𝐷ଶ

Classification: Artificial Neural Network

Outcomes: in 2011: using PCA as features selection method, in 2012: best prediction result for VT-VF events using artificial neural network, 10 seconds before sudden cardiac death.

Features: Nonlinear Analysis: Dyx Parameter

Classification: Dyx value of less than 3 or greater than 6 in eight consecutive R-R intervals smaller than 600 milliseconds (rule-based

Decision)

Outcomes: by using multipole method, they showed that Dyx parameters dynamic became typically pathological between 800 and 5000 R-R intervals (10-60 minutes) before onset of ventricular arrhythmias and used it for sudden cardiac death prediction.

Al., 2015

Features: time domain: MeanNN, SDNN, RMSSD, frequency domain: spectral power analysis in LF, HF, and LF/HF ratio

Classification: Regression Tree Analysis

Outcomes: improving the sensitivity using value of different feature as a node of the regression tree before onset of VT-VF, and early and gradual increase in HR 20 minutes before VT-VF.

Features: Time Domain: Meannn, Sdnn, Rmssd, Nn50,

pNN50, HRVTri, frequency domain: spectral power analysis in LF, HF, and LF/HF ratio, bispectrum analysis, nonlinear

Classification: Svm

Outcomes: using genetic algorithm as a feature selection method to get the best result just before VT-VF events.

Databases

In this subsection a number of ICD signal databases are reviewed to identify the most suitable test cases for this comparative study. A number of proprietary and publicly-available datasets have been used in the literature ,

, , , , , , , –, –, And Are Summarized In

Table 4.2. Six of these datasets –, , , are private hospital records specifically for VT and VF prediction using mostly Medtronic ICDs (Medtronic 7220 to 7223 models), Biotronik Phylax XM and Biotronik Mycro Phylax single chamber ICDs, and there is little publicly-available information on these databases.

Table 4.2. Summary of databases of R-R interval signals from ICD devices, and their usage in the literature.

7218); 126 Controlled And 135 Pre-Vt-Vf

events.

Devices (Minneapolis, Minnesota) From

September 1, 1997 to December 1, 1998.

Stored By Biotronik Phylax Xm And

Mycro Phylax single chamber ICDs.

Medtronic Version 1.0

The experimental data used in this thesis for ventricular arrhythmia prediction come from a public database called “spontaneous ventricular tachyarrhythmia database (SVTD) - Medtronic version 1.0” from Medtronic, Inc. , available in PhysioNet signal archives . This database was recorded from 78 subjects (63 males and 15 females, aged 20.7-75.3 years) in United States and Canada who were implanted between February 1995 and January 1997 by Medtronic Jewel Plus 7218 ICDs (Minneapolis, MN, USA).

Medtronic version 1.0 consists of the following records providing R-R interval data with 10 milliseconds resolution: 106 pre-VT records, 29 pre-VF records, and 126 controlled (CON) records (there were 135 controlled records,

Chapter 4

but 9 records were duplicated). Each sample of 135 spontaneous episodes of VT or VF included R-R intervals prior to the corresponding event. All patients experienced at least one electrical shock after VT-VF was detected by ICD.

VT-VF records contain 1024 of most recently R-R interval measures (before ICD treatments), which the ICD has recorded when a tachyarrhythmia is detected by the device; these signals are recorded while patients were carrying ICDs in the course of their normal daily activities outside the hospital. The CON records were obtained from the patients at the time of interrogation of the ICD during scheduled check-ups and were the most recent R-R intervals of normal rhythm available for recording. This database has been generated by a market leader in the production of ICDs and has been used in many studies.

Therefore, it is a useful choice to use for this research as it permits comparison of results obtained here with results from the literature. Nonetheless, it is noted that the amount of control data is necessarily limited, and this may have some impact on the generalizability of the results. Patient metadata including gender, age, cardiac history, and arrhythmia history are available as shown in Table 4.3.

Table 4.3. Patient metadata in Medtronic version 1.0 dataset.

126

VT = ventricular tachycardia; VF = ventricular fibrillation; CON = controlled data

Model

Many researchers have developed algorithms for prediction of ventricular arrhythmia before onset using HRV signals from different models of ICDs on the market. HRV analysis has been performed on the R-R intervals from 1 minute to 108 minutes either to improve prediction accuracy or to simplify the computational complexity of the algorithm for practical implementations on ICDs. Note that while many researchers have carried out detailed analysis of features, only a subset of these works have actually implemented computer- aided prediction systems. A summary of some previous work in this area is presented in Table 4.1, including works that have specifically used the Medtronic version 1.0 dataset, and have implemented automated prediction systems. A general methodology for ventricular arrhythmia prediction (used for experimental work in this thesis) is explained in this section including statistical analysis and performance metrics.

Figure 4.1. Overview of arrhythmia prediction methods based on cardiac signal activities. Most studies have used a similar approach to that shown in Figure 4.1.

Firstly, the database is labelled and divided into training and test datasets. In many cases, train and test cases are chosen randomly. This may mean that cardiac rhythms recorded from a single patient may appear in both the train and testing sets. In practice, a better approach is to have no overlap between training and test sets.

Pre-processing (e.g., noise reduction or resampling) is typically applied to obtain signals that are more suitable for feature extraction. Systems may also include feature selection or statistical analysis either independently of, or as part of the training phase, to get the best subset of features. Many algorithms have been proposed for sudden cardiac death prediction, including artificial neural networks , , , , , , regression tree , and SVM , , , , , , –. Here, we present prediction results from a comparison of some of the feature categories in time domain, frequency

Chapter 4

domain, bispectrum and nonlinear analysis described in Chapter 3 using the Medtronic version 1.0 dataset as a common dataset for the comparison. The system of Figure 4.1 is used as a basis for the evaluation. Many of the system components in Figure 4.1 are well known and have been extensively reviewed in previous work; however, some of the blocks used in this research that are less commonly used in previous work will be described in more detail.

Preprocessing And Noise Reduction

Data from the Medtronic version 1.0 dataset are interpolated by using cubic

Spline Interpolation As Used In –, , , , , , ,

and resampled at 16 Hz and converted to IHR signals based on formula (3.1). Afterwards, ectopic beats are removed using a median filter . Median filter as a nonlinear filtering technique is widely used in digital signal and image processing. Given a sequence in ascending order as a 𝑥ଵ, 𝑥ଶ, … , 𝑥௡ the 𝑛௧௛ order

Feature Extraction

In this chapter fifty HRV features in four different categories are considered: time domain, frequency domain, bispectrum and nonlinear analysis. The features have already been described in Chapter 3 and summarised in their individual categories in tables in that chapter; for completeness, the full set is listed again in Table 4.4.

47

Table 4.4. Summary of features considered in time domain (6 features), frequency domain (4 features), bispectrum analysis (31 features), and nonlinear analysis (9 features) , .

Statistical Feature Analysis

Student’s t-test method is one of the most common methods for feature ranking applied to analyse features extracted from HRV analysis in , , , , , , . This test can be used to determine whether the mean of two datasets are different or not . The test gives the p-value and t-values for the features extracted for two groups of data. Statistically, a low p-value is preferred (p < 0.05) , and the higher the t-value, the better the ranking.

The Student’s t-test in this experiment was implemented using p-value = 0.05 where both tails were included. In Table 4.5 and Table 4.6, comparisons of different features derived from Medtronic version 1.0, and ranked using Student’s t-test, are presented. Features are divided into their four main

Chapter 4

categories and statistical values have been calculated for one and five minutes before VT-VF and also for CON signals (Table 4.5 and Table 4.6). Although in all categories we see better p-values for 1-minute R-R interval lengths, statistical analysis cannot readily be used as a main feature selector for VT- VF prediction as the analysis duration is too short. Moving from 5-minute to 1-minute R-R intervals shows these features are statistically more discriminative as we get closer to a sudden cardiac death event. In many studies the time and frequency domain categories have been used over 5-minute windows to achieve better prediction rate , , . In the other two categories, bispectrum analysis and nonlinear analysis, p-values are more discriminative with a 1-minute window, which suggests that prediction of VT- VF is better using shorter signal length.

Table 4.5. Mean, standard deviation (SD), and p-value of different feature categories using 1-minute windows of IHR signals from Medtronic version 1.0 dateset.

0.0258

Table 4.6. Mean, standard deviation (SD), and p-value of different feature categories using 5-minute windows of IHR signals from Medtronic version 1.0 dateset.

Signal Classification And Performance Metrics

Results from different studies for detection of onset of VT-VF have mainly been presented in terms of sensitivity (true positive rate), specificity (true negative rate), accuracy and area under the Receiver Operating Curve (AUC).

(4.4)

where TP, TN, FP and FN stand for true positive, true negative, false positive and false negative respectively. For AUC calculations, the receiver operator characteristic (ROC) curves are first computed. For the results presented here, an SVM classifier was used for detection, with features extracted from signals of length 1 minute and 5 minutes. For each series of experiments, as well as a decision on detection (or not) the classifier can also output the confidence level (𝐶) for each test, reflected in the output value of the classifier before binary decision, which lies in the range (−1 ≤𝐶≤1). The AUC measures the two-dimensional area underneath the ROC curve to represent how well predictions are ranked, rather than their absolute values. With the value between 0 and 1, it also measures the quality of the model's predictions irrespective of what classification threshold is chosen .

Experimental Results

In the comparison described here, we measured acute changes in the HRV features extracted from signals beginning 1 or 5 minutes before a point 10 seconds prior to the onset of ventricular arrhythmias recorded from patients with an ICD, so with correct detection, the device still has time to take effective treatment action. The features have been used according to the categories in Table 4.4, i.e., time domain, frequency domain, bispectrum, and nonlinear analysis. The SVM has been implemented in MATLAB with Gaussian function

Chapter 4

as a kernel. After applying Bayesian optimization algorithm1 on SVM in the training phase, the kernel scale has been set to 5 for 1-minute HRV analysis and 10 for 5-minute HRV analysis with the box constraint equal to 2 for both signal lengths. Broadly speaking, there is a trade-off between the signal length and performance. Longer analysis windows include more normal data which may lead to better specificities so the ICD device can detect normal (CON) events more accurately, though not with all feature categories.

Generally-speaking, adding additional categories of feature gives a performance improvement, especially in specificity, when calculated using the longer analysis window. On the whole, frequency domain features display the lowest performance when used on their own, while time domain features perform best when used on their own. A number of observations have been made which provide insight into feature selection and new feature development.

Time Domain Results

Table 4.7 and Table 4.8 show mean, standard deviation, sensitivity, and specificity of time domain features where 1 and 5-minute windows of R-R intervals have been processed. Considering time domain features individually, all achieve comparatively high specificity but at the cost of very poor sensitivity, with the exception of HRVTri. This suggests that future time domain feature development should focus on achieving higher sensitivity. In Table 4.9 the performance metrics of SVM have been presented with both 1 and 5-minute windows. The results indicate that time domain features work better with a longer analysis window.

1 The Bayesian optimization algorithm attempts to minimize a scalar objective function in a bounded domain. The function can be deterministic or stochastic, meaning it can return different results when evaluated at the same input. The components of input can be continuous reals, integers, or categorical. It is usually employed to optimize expensive-to-evaluate functions .

51

Table 4.7. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of time domain features using 1-minute windows of R-R interval signals from Medtronic version 1.0 dateset.

1.21 E+00

Table 4.8. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of time domain features using 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset.

1.18 E+00

Table 4.9. Sensitivity (SN), specificity (SP), accuracy (ACC), and AUC of time domain features together using 1- and 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset.

Frequency Domain Result

Table 4.10 and Table 4.11 show mean, standard deviation, sensitivity, and specificity of frequency domain features while 1 and 5-minute R-R intervals have been processed. Similar to time domain analysis, higher specificity than sensitivity is observed for frequency domain features. None of the four frequency domain features evaluated achieve both sensitivity and specificity higher than 50%. In Table 4.12 the performance metrics of are presented for 1 and 5-minute windows. Based on the presented results, frequency domain features display better performance with a shorter analysis window.

Table 4.10. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of frequency domain features using 1-minute windows of R-R interval signals from Medtronic version 1.0 dateset.

2.27 E+00

Table 4.11. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of frequency domain features using 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset.

1.81 E+00

Table 4.12. Sensitivity (SN), specificity (SP), accuracy (ACC), and AUC of frequency domain features together using 1- and 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset.

Bispectrum Analysis Results

Table 4.13 and Table 4.14 shows mean, standard deviation, sensitivity, and specificity of bispectrum features while 1 and 5-minute R-R intervals have been processed. Higher variance is observed in the bispectral features. The first two types achieve high specificity but at the cost of effectively 0% sensitivity. In contrast, the latter six types are more balanced, achieving comparable levels of both sensitivity and specificity, suggesting that these features may be suitable. In Table 4.15 he performance metrics of SVM are presented while 1 and 5-minute R-R intervals have been used.

Table 4.13. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of bispectrum analysis features using 1-minute windows of R-R interval signals from Medtronic version 1.0 dateset.

Chapter 4

Table 4.14. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of bispectrum analysis features using 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset.

4.84 E-02

Table 4.15. Sensitivity (SN), specificity (SP), accuracy (ACC), and AUC of bispectrum analysis features together using 1- and 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset.

Nonlinear And Dynamic Analysis Results

Table 4.16 and Table 4.17 show mean, standard deviation, sensitivity, and specificity of nonlinear and dynamic analysis features while 1 and 5-minute R- R intervals have been processed. Some nonlinear features also show balanced performance, for example, both entropy-based features. These results suggest there is benefit in further feature development to improve both the sensitivity and specificity of prediction. In the Table 4.18 the performance metrics of SVM are presented while 1 and 5-minute R-R intervals have been used.

Table 4.16. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of nonlinear and dynamic analysis features using 1-minute windows of R-R interval signals from Medtronic version 1.0 dateset.

Chapter 4

Table 4.17. Sensitivity (SN), specificity (SP), mean, and standard deviation (SD) of nonlinear and dynamic analysis features using 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset.

1.71 E+01

Table 4.18. Sensitivity (SN), specificity (SP), accuracy (ACC), and AUC of nonlinear and dynamic analysis features together using 1- and 5-minute windows of R-R interval signals from Medtronic version 1.0 dateset.

Discussion And Conclusions

Modern ICDs allow for tiered therapy, permitting up to three rate detection ranges, each with a specific therapy function. By employing strategies such as burst pacing, adaptive burst pacing, and ramp pacing, many ventricular tachyarrhythmia events can be prevented from developing into sudden cardiac death through ATP , . Detecting ventricular arrhythmias at early stages and delivering ATP as a non-aggressive inhibitor therapy can both improve patient quality of life and the lifespan of the ICD. Moreover, neither appropriate nor inappropriate ATP was associated with a significant mortality increase and ATP has been shown to be effective in terminating approximately 80% to 90% of spontaneous VT episodes , , , .

As many as fifty features have been proposed for sudden cardiac death prediction and are reviewed in this chapter. Moreover, these features are compared using a common database and classification approach. This comparison suggests that time domain features can achieve high performance alone, but the addition of more features using different approaches can increase performance, particularly with the shorter time window of one minute duration. A summary of the experimental results in different feature categories, and when all 50 features are combined, is presented in Table 4.19.

Table 4.19. Summary of experimental results from this chapter using Medtronic version 1.0 dataset.

0.85

SN = sensitivity; SP = specificity; ACC = accuracy; AUC = area under curve

Chapter 4

Table 4.20. Summary of experimental results from previous research, along with results from this chapter (including results obtained with Medtronic version 1.0 dataset).

1-Minute

SN = sensitivity; SP = specificity; ACC = accuracy Table 4.20 presents a comparison of results obtained in previous works, as well as results obtained in the research presented in this chapter. Although windows of duration 4 to 5 minutes prior to ventricular arrhythmias are commonly used in previous works, very long intervals have also been examined, e.g. more than half an hour in , with the best result being 69% sensitivity and 91% specificity. In the study presented in this chapter, the best result was achieved with 85.7% in sensitivity, specificity, and accuracy using all features considered, and with analysis window of 5-minute. In this comparative study, time domain features were found to have the highest AUC when used alone compared to the other feature categories reviewed, suggesting that these features achieve the best balance. However, addition of other feature categories was found to increase performance, suggesting that further development with appropriately chosen features is needed. Furthermore, using a lower number of features decrease the possibility of overfitting as in the current method, number of features is too high considering having only 78 patients in presented study. The next chapter extends this analysis to consider the most differentiating features in ventricular arrhythmia prediction.

Prediction

As noted earlier, approximately 80% of occurrences of sudden cardiac can be attributed to ventricular arrhythmias (VT-VF). The previous chapter carried out a comprehensive evaluation of HRV features for VT-VF prediction. This chapter extends this feature analysis with the goal of reducing unnecessary computational overhead by selecting only the most salient features to accurately predict a ventricular arrhythmia episode. The motivation for this is to facilitate more efficient implementation of arrhythmia prediction on resource-constrained devices such as ICDs. In order to reduce the computational overhead of the detection method, a feature selection technique is applied to reduce the number of features from an initial set. Different signal buffer lengths are examined, where 1-minute and 5-minute signal durations are processed. The features are then applied to classifiers.

To ensure there is a robust evaluation of the feature selection technique, the dataset is split into two, non-overlapping subsections. The first subset of data is used to rank and select the appropriate features and extract model tuning parameters. The second subset is then used to evaluate models using features chosen from the first section. The proposed method is evaluated using R-R interval patient data. All tests are carried out using a leave-patient-out cross-validation method, where the same patient’s data cannot exist in both learning and evaluation data. This is done in order to have better generalisation in the proposed machine learning model for unseen data. Results from a selection of classification algorithms are compared. The work described in this chapter has been published in Parsi et al., “Heart Rate Variability Feature Selection Method for Automated Prediction of Sudden Cardiac Death,”

Overview

Currently, there are several challenges to integrate a reliable rhythm management system in an ICD. If an ICD is configured incorrectly, it can administer inappropriate device therapy and endanger the patient and in some cases could in itself be the cause of a VT episode . Misclassification of VT- VF episodes can result in false-positive detections in healthy patients and lead to improper ICD shock events . Algorithms can be used to regulate ICD treatment by predicting and differentiating VT and VF arrhythmias from other cardiac rhythms in ICD patients. Such algorithms typically extract features from electrical signals measured by an ICD and use these features to make diagnostic decisions that in turn decide which treatment the ICD should apply.

A key aim for any ICD detection technique is to register high sensitivity metric scores to ensure all arrhythmias are addressed as well as high specificity to reduce the number of inappropriate ICD therapies such as electrical shocks . This, in practice, is a difficult balance to achieve.

In the last decade, ICD technology has advanced considerably. Now, multiple programming selection, long battery life, device history features and therapy control options are all standard features. Recently EGM signal shape analysis has been added . SVT and VT-VF usually activate the ventricle in a distinct way, and as a result, different shapes and rhythms are expected from ventricle electrogram morphology as well. Many cardiac event detector algorithms have been applied to distinguish VT-VF from SVT and normal sinus rhythm. As mentioned in Chapter 4, these algorithms are sub-divided into two main categories: 1) morphological/EGM-based algorithms which use in/direct models of heartbeat signal and QRS complex –, and 2) those that rely on temporal intervals within the signal , , , , , .

Unlike morphological algorithms which have a significant computational cost, temporal classification methods use features with a lower computational overhead, such as rate, onset, stability and entropy from HRV signals , – . A review of algorithms from the literature was presented in and summarized in Table 4.1. As discussed in Chapter 4 , each algorithm has its strengths and weaknesses, however, no consensus yet exists concerning the number of detection zones (based on HR changes), nor the optimal duration of signal used for analysis, for detection in order to decrease inappropriate shocks due to misclassification of SVT and VT-VF arrhythmias . Interval based methods attempt to predict onset of an event by analysing a time window of cardiac signal that can range from 10 seconds , , , up to as high as sixty minutes , in advance of a VT-VF event. Ventricular

61

arrhythmia prediction/detection at an early stage can allow ICDs to deliver ATP as a non-aggressive inhibitor therapy, improving the patient’s quality of life and the ICD lifespan.

An overview of ventricular arrhythmia detection and treatment by ICD is presented in Figure 5.1. The detection and treatment algorithms typically involve a sequence of event detections. By adding an effective ventricular arrhythmia prediction method into an ICD there are opportunities to prevent unnecessary shocks . Daubert et al. in have found that the occurrence of false positives was not associated with a significant mortality increase. On the other hand, as mentioned in Chapter 4, ATP that is only applied when needed (true positive) has been effective in stopping approximately 80% to 90% of spontaneous VT episodes , , , . However, there is a trade-off between early prediction and the buffering overhead available on a computationally limited ICD.

Figure 5.1. Flowchart describing the decision-making flow inside the ICD and where VT-VF prediction takes place in this flow. Having accurate VT-VF prediction make the ICD to deliver ATP, which can finally lead cardiac system to go to normal heartbeat. Normal heartbeat would fail tachycardia reconfirmation step which makes the ICD skip VT-VF detection and avoiding un-necessary electrical shock, while saving battery life and improve patient’s life experience and quality.

An effective VT-VF prediction method can be used to complement detection and may be used to trigger ATP as a less aggressive therapy that controls heartbeats. The prediction method proposed in this section targets ventricular arrhythmias which could be prevented by nonaggressive treatment like ATP, with more invasive electrotherapy reserved only for cases that absolutely require it. For example, the PainFree I and II trials showed the effectiveness and safety of the application of ATPs (2 sequences of an 8-pulse burst pacing train at 88% of the VT cycle length) before shock delivery even on fast VT

Chapter 5

with a heart rate of 188-250 beats per minute –. Daubert et al. have found that the occurrence of false positives was not associated with a significant mortality increase. On the other hand, ATP that is only applied when needed (true positive) has been effective in stopping approximately 80% to 90% of spontaneous VT episodes , , , . However, there is a trade-off between early prediction and the buffering overhead available on a computationally limited ICD. Sensitivity (percentage of VT-VF candidate patients who are correctly identified as such) and specificity (percentage of healthy people who are identified as healthy) metrics are used to assess VT- VF interval prediction methods. A detection method normally attempts to maximize both values and achieve an appropriate balance between them depending on the clinical requirements. Boon et al. in used a genetic algorithm to achieve 77.94% sensitivity and 80.88% specificity using a subset of features derived from an original set of 53 extracted features. Rozen et al.

in used a decision rule based on multipole analysis and reported over 94% sensitivity, however specificity was significantly lower at 51%. With 11 features, Lee et al. in recorded 70.6% sensitivity and 76.5% specificity. The work in Chapter 4 achieved 85.7% sensitivity and 85.7% specificity using 50 HRV features. While the metric results are high, the significant number of feature calculations are a significant challenge for an ICD with limited computational and storage resources.

Feature selection to search for the best subset of attributes in observed data may be an important and advantageous factor when applying machine learning methods on portable sensing devices. A minimum but representative set of features can simplify classification models, reduce the dimensionality, and enhance generalization. More recent machine learning models have reduced data and storage requirements, and improved prediction performances .

Many approaches, from genetic algorithm feature selection to principal component analysis, have been investigated to find a suitable HRV feature set , . For example, Boon et al. in used 20 features from a larger set of 53 features to predict onset of VT-VF. Lee et al. in selected 14 most representative features for congestive heart failure recognition out of 42 features. Age and gender were among the selected fourteen features. Figure 5.2 is a radar diagram summarising the results of some previously-reported studies, along a number of different performance axes.

Parsi Et Al., 2020 , 5-Minute Window

Figure 5.2. Best results from recent studies on VT-VF prediction.

Hrv Feature Selection Methods

As each R-R interval signal is presented in terms of 𝑑 features (or measurements), it is viewed as point in 𝑑-dimensional space. The main goal here is to choose those features that present patterns from different classes in a well separated way . The effectiveness of the representation space also is determined by how well classifiers could establish decision boundaries in the feature space which separate patterns belonging to different classes. There are many potential benefits of feature selection including facilitating data understanding, reducing the measurement and storage requirements, reducing training and utilization times, and last but not least addressing the “curse of dimensionality” to improve prediction performance .

Feature selection techniques can be divided into filters, wrappers, and embedded methods . Wrappers utilize the machine learning algorithm of interest as a black box to score subsets of variables according to their predictive power. Filters select subsets of variables as a pre-processing step, independently

Related Journal Articles & DOI Links

Selected peer-reviewed publications relevant to 12 Lead ECG Acquisition. Click the DOI to access the full paper (may require institutional access).

Why Choose Us?

Bangalore guidance for robotics, Spectre and autonomous systems projects.