Advanced Methods and Tools for ECG Data Analysis - Part 1 ppt

Advanced Methods and Toolsfor ECG Data Analysis... Advanced Methods and Toolsfor ECG Data Analysis Gari D.. 13.3 Unsupervised Learning Techniques and Their Applications in ECG13.3.4 Appl

Advanced Methods and Tools

for ECG Data Analysis

This book is part of the Artech House Engineering in Medicine & Biology Series,Martin L Yarmush and Christopher J James, Series Editors For a listing of recent

related Artech House titles, turn to the back of this book

Advanced Methods and Tools

for ECG Data Analysis

Gari D Clifford Francisco Azuaje Patrick E McSharry

Editors

a r t e c h h o u s e c o m

Library of Congress Cataloging-in-Publication Data

A catalog record for this book is available from this U.S Library of Congress

British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library

ISBN-10: 1-58053-966-1

ISBN-13: 978-1-58053-966-1

Cover design by Michael Moretti

© 2006 ARTECH HOUSE, INC.

685 Canton Street Norwood, MA 02062

All rights reserved Printed and bound in the United States of America No part of thisbook may be reproduced or utilized in any form or by any means, electronic or mechanical,including photocopying, recording, or by any information storage and retrieval system,without permission in writing from the publisher

All terms mentioned in this book that are known to be trademarks or service marks havebeen appropriately capitalized Artech House cannot attest to the accuracy of this infor-mation Use of a term in this book should not be regarded as affecting the validity of anytrademark or service mark

10 9 8 7 6 5 4 3 2 1

CHAPTER 1

1.3 Introduction to Clinical Electrocardiography: Abnormal Patterns 121.3.1 The Normal Determinants of Heart Rate: The Autonomic

1.3.3 Conduction Blocks, Bradycardia, and Escape Rhythms 181.3.4 Cardiac Ischemia, Other Metabolic Disturbances,

2.5.3 Primary Common-Mode Noise Reduction: Active Grounding

2.5.4 Increasing Input Impedance: CMOS Buffer Stage 43

v

vi Contents

2.5.9 Hardware Design Issues: Sampling Frequency Choice 472.5.10 Hardware Testing, Patient Safety, and Standards 48

3.5.2 Arrhythmia Classification from Power-Frequency Analysis 683.5.3 Arrhythmia Classification from Beat-to-Beat Statistics 68

3.7.6 The Effect of Ectopy and Artifact and How to Deal with It 803.7.7 Choosing an Experimental Protocol: Activity-Related Changes 81

3.8.1 Nonstationary HRV Metrics and Fractal Scaling 84

Contents vii

4.2.6 Coupled Oscillators and Phase Synchronization 110

5.3.2 The Discrete Wavelet Transform and Filter Banks 142

5.4.3 Independent Component Analysis for Source Separation

6.4.2 State Space Independent Component Analysis 181

7.5.2 Short-Term Fourier Transform–Based Methods 2027.5.3 Interpretation of Spectral TWA Test Results 203

7.6 Tailoring Analysis of TWA to Its Pathophysiology 207

7.6.3 Fluctuating Heart Rates and Nonstationary TWA 2097.6.4 Rhythm Discontinuities, Nonstationary TWA, and TWA Phase 210

8.4 EDR Algorithms Based on Both Beat Morphology and HR 229

Contents ix

9.3 Derivation of Diagnostic and Morphologic Feature Vectors 2519.3.1 Derivation of Orthonormal Function Model Transform–Based

9.3.2 Derivation of Time-Domain Diagnostic and Morphologic

10.3.2 Correction of Reference ST Segment Level 274

10.4.2 Comparison of Performance of ST Analyzers 284

13.3 Unsupervised Learning Techniques and Their Applications in ECG

13.3.4 Application of Unsupervised Learning in ECG Classification 346

13.3.6 Evaluation of Unsupervised Classification Models: Cluster

13.4 GSOM-Based Approaches to ECG Cluster Discovery and

xii Preface

rate variability studies However, the enormity of this problem has led to apervasive analysis of beat-to-beat intervals based upon the QRS complex as afiducial marker

• Reliable QT interval estimation Similarly, despite the relatively large tude of the QRS complex and T wave, the onset of the Q wave and offset of

ampli-the T wave are difficult features to measure (even for highly trained experts).Changes in the QRS complex, ST segment, and T wave morphology due toheart rate and sympathetic nervous system changes make this problem par-ticularly acute

• Distinguishing ischemic from nonischemic ST changes Even in subjects who

are known to have myocardial ischemia, ST changes are not considered a basisfor definitive diagnosis of individual episodes of ischemia This is because the

ST segment changes with a subject’s body position due to the movement ofthe ECG electrodes relative to the heart Recent attempts to identify suchartefacts are promising (see the Computers in Cardiology Challenge 2003),but the road to a full working system (particularly for silent ischemia) is notclear

• Reliable beat classification in Holter monitoring Ambulatory monitoring

presents many challenges since the data collection process is essentially pervised, and evolving problems in the acquisition often go undetected (or atleast are not corrected) One major problem is the high level of in-band noiseencountered (usually muscle- or movement-related) Another problem for am-bulatory monitoring is the degradation in electrode contact over time, leading

unsu-to a lower signal-unsu-to-noise ratio Since the P wave is usually a low-amplitudefeature, reliable detection of the P wave is particularly problematic in thisenvironment

• Robust, reliable in-band signal filtering or source separation (such as muscle

noise removal and fetal-maternal separation) Principal and independent ponent analysis has shown promise in this area, but problems persist withnonstationary mixing and lead position inaccuracies/changes Model-basedfiltering methods have also shown promise in this area

com-• Identification of lead position misplacements or sensor shifts Most classifiers

assume that the label for the clinical lead being recorded is known and make sumptions about waveform morphology, amplitude, and polarity based uponthe lead label If the lead is misplaced, or shifts during recording, then misclas-sifications or misdetections will probably occur In particular, the maternal-fetal signal mixing is problematic in this respect In this case two cardiacsources are moving with respect to each other, sometimes without correla-tion, and sometimes entrained on one or more scales

as-• Reliable confidence measures The ability of an algorithm to report back a

“level of trust” associated with its parameter estimates is very important,particularly if the algorithm’s output is fed to another data fusion algorithm

• The inverse problem It is well known that no unique solution exists for the

inverse problem in ECG mapping Attempts to reconstruct the dipole moment

in the ECG have had some degree of success, and recent models for the ECGhave proved useful in this respect

Preface xiii

• ECG modeling and parameter fitting To date there exist many accurate

rep-resentations for the ECG from a cellular level up to more phenomenologicalmodels However, at best, each of these models can accurately reproduce theECG waveform for only a short period of time Although models also ex-ist for variations in the beat-to-beat timings of the heart on both short andlong scales, there are no known models that can reproduce realistic dynamicactivity on all scales, together with an accurate or realistic resultant ECG.Recent work in the fitting of real data to ECG and pulsatile models is cer-tainly promising, but much work is needed in order to ascertain whether theresultant parameters will yield any more information than traditional ECGmetrics

• The mapping of diagnostic ECG parameters to disease classifications or dictive metrics Neural networks have shown promise in this area due to their

pre-ability to extrapolate from small training sets However, neural networks arehighly sensitive to outliers and the size and distribution of the training set Ifonly a few artifacts, or mislabeled patterns, creep into the training set, the per-formance of the classifier is significantly reduced Conversely, over-restrictingthe training set to include too few patterns from a class results in over-training.Another known problem from the use of neural networks is that it is difficult

to extract meaning from the output; a classification does not obviously mapback to the etiology

• Global context pattern analysis Factoring in patient-specific history (from

minutes, to hours, to years) for feature recognition and classification isrequired if a full emulation of the clinical diagnostic procedure is to bereproduced Hidden Markov models, extended Kalman filters, and Bayesianclassifiers are likely candidates for such a problem

• The development of closed-loop systems It is always problematic for a

clin-ician to relinquish the task of classification and intervention for personal,legal, and ethical reasons However, with the increasing accuracy of classifiersand the decreasing costs of machines relative to human experts, it is almostinevitable that closed-loop devices will become more pervasive To date, fewsuch systems exist beyond the classic internal cardioverters that have been

in use for several decades Despite promising advances in atrial fibrillationdetection, which may soon make the closed-loop injection of drugs for thistype of condition a reality, further work is needed to ensure patient safety

• Sensor fusion Information that can be derived from the ECG is insufficient to

effectively solve many of the above problems It is likely that the combination

of information derived from other sensors (such as blood pressure ers, accelerometers, and pulse oximeters) will be required The paradigm ofmultidimensional signal analysis is well known to the ECG signal analyst,and parallel analysis of the ECG (or ECG-derived parameters) almost alwaysenhances an algorithm’s performance For instance, blood pressure waves con-tain information that is highly correlated with the ECG, and analysis of thesechanges can help reduce false arrhythmia alarms The ECG is also highly cor-related with respiration and can be used to improve respiration rate estimates

transduc-or to facilitate sleep analysis However, when the associated signals do not

xiv Preface

present in a highly correlated manner, the units of measurement are different(so that a 20-mmHg change does not mean the same as a 20-bpm change, forexample) and their associated distributions differ, the task at hand is far more

difficult In fact, normalizing for these differences, and building trust metrics

to differentiate artifactual changes from real changes in such signals, is one ofthe more difficult challenges in ECG signal processing today

In order to address these issues, we have attempted to detail many of the keyrelevant advances in signal processing Chapter 1 describes the physiological back-ground and the specific autonomic mechanisms which regulate the beat-to-beatchanges of timing and morphology in the ECG, together with the cause and effect

of breakdowns in this mechanism Chapter 2 presents an overview of the primaryissues that should be taken into account when designing an ECG collection system.Chapter 3 presents an overview of the relevant mathematical descriptors ofthe ECG such as clinical metrics, spectral characteristics, and beat-to-beat variabil-ity indices Chapter 4 presents an overview of simple, practical ECG and beat-to-beat models, together with methods for applying these models to ECG analysis.Chapter 5 describes a unified framework for linear filtering techniques includingwavelets, principal component analysis, neural networks, and independent compo-nent analysis Chapter 6 discusses methods and pitfalls of nonlinear ECG analysis,with a practical emphasis on filtering techniques

Chapter 7 provides an overview of T wave alternan methodologies, andChapter 8 presents a comparative study of ECG derived respiration techniques.Chapter 9 presents advanced techniques for extracting relevant features from theECG, and Chapter 10 uses these techniques to describe a robust ST-analyzer.Chapter 11 presents a wavelet and hidden Markov model–based procedure forrobust QT-analysis Chapter 12 describes techniques for supervised classificationand hybrid techniques for classifying ECG metrics, where the data labels are alreadyknown, and Chapter 13 presents unsupervised learning techniques for ECG patterndiscovery and classification

Although many of the basics of ECG analysis are presented in Chapter 1, this issimply to draw the reader’s attention the etiology of many of the problems we areattempting to solve As a thorough grounding in the basics of ECG signal processing,

the reader is referred to Chapters 7 and 8 in S ¨ornmo and Laguna’s recent book electric Signal Processing in Cardiac and Neurological Applications (Elsevier, 2005).

Bio-The reader is assumed to be familiar with the basics of signal processing and cation techniques Furthermore, these techniques are necessarily implemented using

classifi-a knowledge of computclassifi-ationclassifi-al progrclassifi-amming This book follows the open-sourcephilosophy that the development of robust signal processing algorithms is best done

by making them freely available, together with the labeled data on which they wereevaluated Many of the algorithms and data sets described in this book are availablefrom the following URLs: http://www.ecgtools.org, http://www.physionet.org, andhttp://alum.mit.edu/www/gari/ecgbook

Most of these algorithms have been written either in C or Matlab Additionally,Java applet versions of selected algorithms are also available Libraries for readingthese databases are also freely available We hope that through these URLs this

Gari D Clifford Cambridge, Massachusetts

Francisco Azuaje Jordanstown, United Kingdom

Patrick E McSharry Oxford, United Kingdom

Editors September 2006

2 The Physiological Basis of the Electrocardiogram

Figure 1.1 A typical action potential from a ventricular myocardial cell Phases 0 through 4 are

marked (From: [2] c 2004 MIT OCW Reprinted with permission.)

(e.g., nerves and skeletal muscle), the myocardial cell at rest has a typical

trans-membrane potential, V m, of about−80 to −90 mV with respect to surroundingextracellular fluid.2The cell membrane controls permeability to a number of ions,including sodium, potassium, calcium, and chloride These ions pass across themembrane through specific ion channels that can open (become activated) and

close (become inactivated) These channels are therefore said to be gated channels

and their opening and closing can occur in response to voltage changes (voltagegated channels) or through the activation of receptors (receptor gated channels).The variation of membrane conductance due to the opening and closing of ionchannels generates changes in the transmembrane (action) potential over time Thetime course of this potential as it depolarizes and repolarizes is illustrated for a ven-tricular cell in Figure 1.1, with the five conventional phases (0 through 4) marked

When cardiac cells are depolarized to a threshold voltage of about−70 mV (e.g.,

by another conducted action potential), there is a rapid depolarization (phase 0 —the rapid upstroke of the action potential) that is caused by a transient increase

in fast sodium channel conductance Phase 1 represents an initial repolarizationthat is caused by the opening of a potassium channel During phase 2 there is anapproximate balance between inward-going calcium current and outward-goingpotassium current, causing a plateau in the action potential and a delay in repolar-ization This inward calcium movement is through long-lasting calcium channelsthat open up when the membrane potential depolarizes to about−40 mV Repo-larization (phase 3) is a complex process and several mechanisms are thought to

be important The potassium conductance increases, tending to repolarize the cellvia a potassium-mediated outward current In addition, there is a time-dependent

2 Cardiac potentials may be recorded by means of microelectrodes.