Processing of The Phonocardiographic Signal Methods For The Intelligent Stethoscope

Linköping Studies in Science and Technology
Thesis No. 1253
Processing of the Phonocardiographic Signal −

Methods for the Intelligent Stethoscope
Christer Ahlström
LiU-TEK-LIC-2006: 34
Department of Biomedical Engineering
Linköpings universitet, SE-58185 Linköping, Sweden
http://www.imt.liu.se
In cooperation with Biomedical Engineering,

Örebro County Council, Sweden
Linköping, May 2006

Processing of the Phonocardiographic Signal −
Methods for the Intelligent Stethoscope
© 2006 Christer Ahlström
Department of Biomedical Engineering

Linköpings universitet
SE-58185 Linköping
Sweden
ISBN: 91-85523-59-3 ISSN: 0280-7971
Printed in Sweden by LiU-Tryck, Linköping 2006

Abstract
Phonocardiographic signals contain bioacoustic information reflecting the
operation of the heart. Normally there are two heart sounds, and additional sounds
indicate disease. If a third heart sound is present it could be a sign of heart failure
whereas a murmur indicates defective valves or an orifice in the septal wall. The
primary aim of this thesis is to use signal processing tools to improve the
diagnostic value of this information. More specifically, three different methods
have been developed:
• A nonlinear change detection method has been applied to automatically
detect heart sounds. The first and the second heart sounds can be found
using recurrence times of the first kind while the third heart sound can be
found using recurrence times of the second kind. Most third heart sound
occurrences were detected (98 %), but the amount of false extra detections
was rather high (7 % of the heart cycles).
• Heart sounds obscure the interpretation of lung sounds. A new method
based on nonlinear prediction has been developed to remove this undesired
disturbance. High similarity was obtained when comparing actual lung
sounds with lung sounds after removal of heart sounds.
• Analysis methods such as Shannon energy, wavelets and recurrence
quantification analysis were used to extract information from the
phonocardiographic signal. The most prominent features, determined by a
feature selection method, were used to create a new feature set for heart
murmur classification. The classification result was 86 % when separating
patients with aortic stenosis, mitral insufficiency and physiological
murmurs.
The derived methods give reasonable results, and they all provide a step forward
in the quest for an intelligent stethoscope, a universal phonocardiography tool able
to enhance auscultation by improving sound quality, emphasizing abnormal events
in the heart cycle and distinguishing different heart murmurs.
i
ii
List of Publications
This thesis is based on three papers, which will be referred to in the text by their
roman numerals.
I. Ahlstrom C, Liljefelt O, Hult P, Ask P: Heart Sound Cancellation from
Lung Sound Recordings using Recurrence Time Statistics and Nonlinear
Prediction. IEEE Signal Processing Letters. 2005. 12:812-815.
II. Ahlstrom C, Hult P, Ask P: Detection of the 3rd Heart Sound using
Recurrence Time Statistics. Proc. 31st IEEE Int. Conf. on Acoustics,
Speech and Signal Processing, Toulouse, France, 2006.
III. Ahlstrom C, Hult P, Rask P, Karlsson J-E, Nylander E, Dahlström U, Ask
P: Feature Extraction for Systolic Heart Murmur Classification.
Submitted.
Related publications not included in the thesis.

• Ahlstrom C, Johansson A, Hult P, Ask P: Chaotic Dynamics of
Respiratory Sounds. Chaos, Solitons and Fractals. 2006. 29:1054-1062.
• Johansson A, Ahlstrom C, Länne T, Ask P: Pulse wave transit time for
monitoring respiration rate. Accepted for publication in Medical &
Biological Engineering & Computing. 2006.
• Ahlstrom C, Johansson A, Länne T, Ask P: Non-invasive Investigation of
Blood Pressure Changes using Pulse Wave Transit Time: a novel
approach in the monitoring of dialysis patients. Journal of Artificial
Organs. 2005. 8:192-197.
• Ahlstrom C, Hult P, Ask P: Thresholding Distance Plots using True
Recurrence Points. Proc. 31st IEEE Int. Conf. on Acoustics, Speech and
Signal Processing, Toulouse, France, 2006.
• Ahlstrom C, Hult P, Ask P: Wheeze analysis and detection with non-
linear phase space embedding. Proc. 13th Nordic Baltic Conf. in
Biomedical Eng. and Med. Physics, Umeå, 2005.
• Hult P, Ahlstrom C, Rattfält L, Hagström C, Pettersson NE, Ask P: The
intelligent stethoscope. 3rd European Med. Biol. Eng. Conf., Prague,
Czech Republic, 2005
iii
• Ahlstrom C, Johansson A, Länne T, Ask P: A Respiration Monitor Based
on Electrocardiographic and Photoplethysmographic Sensor Fusion. Proc.
26th Ann. Int. Conf. IEEE Eng. Med. Biol., San Francisco, US, 2004.
The following M. Sc. theses have also contributed to this thesis.

• Hasfjord F: Heart Sound Analysis with Time-Dependent Fractal
Dimensions.
LiU-IMT—EX—358, 2004
• Nilsson E: Development of an Application for Visualising Heart Sounds.
LiTH-IMT/FMT20-EX—04/397—SE, 2004
• Liljefeldt O: Heart Sound Cancellation from Lung Sounds using Non-
linear Prediction.
LiTH-IMT/FMT20-EX—05/388—SE, 2005
iv
Preface
The intelligent stethoscope has occupied my mind for three years now, this dragon
which consumes my time and drowns me in endless riddles. Many times have I
looked at its mysteries in despair, but once and again the dusk disperses. Perhaps
you can liken the process with the butterfly effect, where a butterfly flapping its
wings over the beautiful island of Gotland can cause a hurricane in Canada.
Similarly, the seed of an idea can be planted in the most unlikely ways; while
hanging on the edge of a cliff or when biking along a swaying trail, while waiting
in line at the local grocery store and sometimes even at work. A fragment of a
thought suddenly starts to make sense, the idea tries to break free but gets lost and
sinks into nothingness. Somewhere in a hidden corner the seed lay fallow, waiting
for a new opportunity to rise. This could happen any day, any week or any year. In
the end, you can only hope that the idea is unleashed while it still makes sense.
After all, what would I write in my book if the seed decided not to grow?
Linköping, April 2006
v
vi
Acknowledgements
To all of my friends, thank you for still being my friends. Especially Markus, who
kept up with my complaints, and Jonas, for telling me when I was working too
much (I believe the exact phrase was: “You’re about as funny as a genital
lobotomy”).
My mother, father, brother and sister, the mainstay of my life, I’ll try to visit you
more often now when this work is behind me.
All colleagues at the Department of Biomedical Engineering, particularly Amir,
my gangsta brotha in arms, my faithful office-mate and my personal music
provider. Your company has been most appreciated over the last couple of years.
Linda, I owe you more than you’d like to admit, thanks again!
Anders Brun and Eva Nylander helped proof-reading the manuscript.
My supervisors; Per Ask, Peter Hult and Anders Johansson. Especially Per for
having faith in my ideas, Peter for introducing me to the intelligent stethoscope
and Anders for guiding me in scientific methodology and research ethics, for
endless patience and most of all, for also being a friend.
Many colleagues have helped me in my works; your contributions will not be
forgotten. To name but a few, Olle Liljefeldt, Fredrik Hasfjord, Erik Nilsson, Jan-
Erik Karlsson, Peter Rask, Birgitta Schmekel, Christina Svensson, Björn
Svensson, AnnSofie Sommer, Ulf Dahlström, January Gnitecki, Per Sveider,
Bengt Ragnemalm, Solveig Carlsson, Susanne Skytt and Nils-Erik Pettersson with
staff at Biomedical Engineering, Örebro County Council.
The artwork on the cover was kindly provided by Nancy Munford. The idea
behind the image comes from the heart muscle being spiral in shape, but as it turns
out, also the flow in the heart advances along spiral pathways (or rather, vortices
or eddies). With a little good will it looks a bit like a signal moving around in its
state space as well.
Finally, my beloved Anneli, thank you for letting me use your time.
This work was supported by grants from the Swedish Agency for Innovation Systems, the Health
Research Council in the South-East of Sweden, the Swedish Research Council, the Swedish
National Centre of Excellence for Non-invasive Medical Measurements and CORTECH (Swedish
universities in cooperation for new cardiovascular technology).
vii
viii
Abbreviations
Abbreviations have been avoided as much as possible, but every now and then
they tend to sneak in anyhow.
AR Autoregressive
ARMA Autoregressive moving average
AS Aortic stenosis
MA Moving average
MI Mitral insufficiency
PM Physiological murmur
ROC Receiver operating curve
RP Recurrence plot
RQA Recurrence quantification analysis
S1 The first heart sound
S2 The second heart sound
S3 The third heart sound
S4 The fourth heart sound
ST Stockwell transform
STFT Short time Fourier transform
TFR Time Frequency Representation
VFD Variance fractal dimension
WT Wavelet transform
ix
x
Table of Contents
ABSTRACT ................................................................................................................................................... I
LIST OF PUBLICATIONS....................................................................................................................... III
PREFACE.....................................................................................................................................................V
ACKNOWLEDGEMENTS......................................................................................................................VII
ABBREVIATIONS .................................................................................................................................... IX
1. INTRODUCTION ...............................................................................................................................1
1.1. AIM OF THE THESIS .......................................................................................................................3
1.2. THESIS OUTLINE ...........................................................................................................................4
2. PRELIMINARIES ON HEART SOUNDS AND HEART MURMURS.........................................5
2.1. PHYSICS OF SOUND .......................................................................................................................5
2.2. PHYSIOLOGY OF THE HEART .........................................................................................................6
2.3. HEART SOUNDS ............................................................................................................................7
2.4. HEART MURMURS .........................................................................................................................8
2.5. AUSCULTATION AND THE PHONOCARDIOGRAM ..........................................................................10
2.6. ACQUISITION OF PHONOCARDIOGRAPHIC SIGNALS .....................................................................10
2.6.1. Sensors ..................................................................................................................................11
2.6.2. Pre-processing, digitalization and storage............................................................................11
3. SIGNAL ANALYSIS FRAMEWORK ............................................................................................13
3.1. MEASURING CHARACTERISTICS THAT VARY IN TIME .................................................................15
3.1.1. Intensity .................................................................................................................................15
3.1.2. Frequency..............................................................................................................................17
3.2. NONLINEAR SYSTEMS AND EMBEDOLOGY ..................................................................................18
3.3. NONLINEAR ANALYSIS TOOLS ....................................................................................................21
3.3.1. Non-integer dimensions.........................................................................................................21
3.3.2. Recurrence quantification analysis .......................................................................................23
3.3.3. Higher order statistics...........................................................................................................25
3.4. NONLINEAR PREDICTION ............................................................................................................27
4. PROPERTIES OF PHONOCARDIOGRAPHIC SIGNALS ........................................................31
4.1. TIME AND FREQUENCY ...............................................................................................................31
4.1.1. Murmurs from stenotic semilunar valves ..............................................................................34
4.1.2. Murmurs from regurgitant atrioventricular valves ...............................................................34
4.1.3. Murmurs caused by septal defects.........................................................................................35
4.1.4. Quantifying the results ..........................................................................................................36
4.2. HIGHER ORDER STATISTICS ........................................................................................................37
4.3. RECONSTRUCTED STATE SPACES ................................................................................................40
4.3.1. Quantifying the reconstructed state space.............................................................................41
4.3.2. Recurrence time statistics......................................................................................................42
4.4. FRACTAL DIMENSION .................................................................................................................43
5. APPLICATIONS IN PHONOCARDIOGRAPHIC SIGNAL PROCESSING ...........................47
5.1. SEGMENTATION OF THE PHONOCARDIOGRAPHIC SIGNAL ...........................................................47
xi
5.2. FINDING S3 .................................................................................................................................49
5.3. FILTERING OUT SIGNAL COMPONENTS ........................................................................................50
5.4. CLASSIFICATION OF MURMURS ...................................................................................................52
5.4.1. Feature extraction .................................................................................................................53
5.4.2. Finding relevant features ......................................................................................................55
5.4.3. Classifying murmurs..............................................................................................................57
6. DISCUSSION.....................................................................................................................................59
6.1. CONTEXT OF THE PAPERS............................................................................................................59
6.2. PATIENTS AND DATA SETS ..........................................................................................................60
6.2.1. Measurement noise................................................................................................................62
6.3. METHODOLOGY ..........................................................................................................................62
6.4. FUTURE WORK............................................................................................................................64
6.4.1. Clinical validation.................................................................................................................64
6.4.2. Multi-sensor approach ..........................................................................................................64
6.4.3. Dimension reduction .............................................................................................................65
6.4.4. Choosing an appropriate classifier .......................................................................................65
7. REVIEW OF PAPERS......................................................................................................................67
7.1. PAPER I, HEART SOUND CANCELLATION ....................................................................................67
7.2. PAPER II, DETECTION OF THE 3RD HEART SOUND ........................................................................67
7.3. PAPER III, FEATURE EXTRACTION FROM SYSTOLIC MURMURS ..................................................68
REFERENCES ............................................................................................................................................69
xii
1. Introduction
“The way to the heart is through the ears.”
Katie Hurley
The history of auscultation, or listening to the sounds of the body, is easily

described by a few evolutionary leaps. Hippocrates (460-377 BC) provided the
foundation for auscultation when he put his ear against the chest of a patient and
described the sounds he could hear from the heart. The next leap was made by
Robert Hooke (1635-1703) who realized the diagnostic use of cardiac
auscultation:
"I have been able to hear very plainly the beating of a man's heart…Who knows, I
say, but that it may be possible to discover the motion of the internal parts of
bodies…by the sound they make; one may discover the works performed in several
offices and shops of a man's body and thereby discover what instrument is out of
order."
The biggest breakthrough in auscultation came in 1816 when René Laennec
(1781-1826) invented the stethoscope. Laennec was about to examine a woman
with the symptoms of heart disease, but due to her sex and age, direct auscultation
was inappropriate. Also percussion and palpation gave little information on
account of the patient’s obesity [1]. Consequently, Laennec used a roll of paper to
avoid physical contact during the examination. As a spin-off, he found that heart
and lung sounds were amplified and previously unheard sounds emerged. The
invention of the stethoscope resulted in, without precedent, the most widely spread
diagnostic instrument in the history of biomedical engineering. The stethoscope
has evolved over the years, but the underlying technology remains the same.
Attempts have been made to take the stethoscope into the IT age, but the success
has so far been limited. A selection of stethoscopes from different eras is presented
in Figure 1.
Mechanical working processes in the body produce sounds which indicate the
health status of the individual. This information is valuable in the diagnosis of
patients, and it has been widely used since the days of Hippocrates. In modern
health care, auscultation has found its primary role in primary or in home health
1
Processing of the Phonocardiographic Signal
care, when deciding which patients need special care. The most important body
sounds are heart sounds and lung sounds, but sounds from swallowing,
micturition, muscles and arteries are also of clinical relevance. The main sources
for production of body sounds are acceleration or deceleration of organs or fluids,
friction rubs and turbulent flow of fluids or gases.
The auscultatory skills amongst physicians demonstrate a negative trend. The loss
has occurred despite new teaching aids such as multimedia tutorials, and the
reasons are the availability of new diagnostic tools such as echocardiography and
magnetic resonance imaging, a lack of confidence and increased concern about
litigations [2]. The art of auscultation is often described as quite difficult, partly
because of the fact that only a portion of the cardiohemic vibrations are audible,
see Figure 2.
Figure 1. Early monaural stethoscopes (top left), Cummanns and Allisons stethoscopes (lower left), a
modern binaural stethoscope (middle) and a modern electronic stethoscope, Meditron M30 (right).
Acoustic stethoscopes transmit sound mechanically from a chest-piece via air-

filled hollow tubes to the listener's ears. The diaphragm and the bell work as two
filters, transmitting higher frequency sounds and lower frequency sounds,
respectively. Electronic stethoscopes function in a similar way, but the sound is
converted to an electronic signal which is transmitted to the listener by wire.
Functionalities often included in electronic stethoscopes are amplification of the
signal, filters imitating the function of the diaphragm and the bell and in some
cases recording abilities to allow storage of data.
Heart sounds and murmurs are of relatively low intensity and are band limited to
about 10–1000 Hz. Meanwhile, the human hearing is adapted to speech. This
explains why physicians sometimes have easier to detect heart sounds by palpation
than by auscultation. Part of this problem can be avoided by amplification of the
sound, but much information contained in the phonocardiographic signal is hard to
reach using acoustic stethoscopes as well as electronic stethoscopes. An intelligent
stethoscope could make use of this extra information. Possible tasks for an
intelligent stethoscope are to classify different murmurs, to detect differences in
the heart sounds (such as the splitting of the second heart sound) and to detect
additional sounds (such as the third heart sound and ejection clicks).
2
Chapter 1. Introduction
Audible heart sounds

and murmurs
10
1
Speech
-1
10
-2
10
-3
10
Heart sounds
10 -4 and murmurs Threshold of
audibility
-5
10
8 16 32 64 128 256 512 1024 2048 4096
Frequency (Hz)
Figure 2. The frequency content of heart sounds and murmurs in relation to the human threshold of
audibility. Drawn from [3]. Note that without amplification, the area representing the audible part of
the phonocardiographic signal is very small.
50-80% of the population has murmurs during childhood, whereas only about 1%
of the murmurs are pathological [2]. A simple tool able to screen murmurs would
be both time- and cost-saving while relieving many patients from needless anxiety.
The sensor technology in such a tool is basically simple and the parameters
obtained are directly related to mechanical processes within the body (in contrast
to ECG which measures the heart’s electrical activity). In the new field of
telemedicine and home care, bioacoustics is definitely a suitable method.
1.1. Aim of the Thesis

The all-embracing goal of bioacoustic research is to establish a relationship
between mechanical events within the body and the sounds these events give rise
to. The medical use of this knowledge is of course to link sounds that diverge from
normality to certain pathological conditions. Clearly, there is valuable information
hidden in the bioacoustic signal, and the primary aim of this thesis is to make use
of signal processing tools to emphasize and extract this information. Clinical value
has been a guiding-star throughout the work, and the goal has been to develop
methods for an intelligent stethoscope. More specifically, the aims of this thesis
were to:
• identify and compare signal analysis tools suitable for phonocardiographic
signals.
• emphasize audibility of bioacoustic signals (Paper I emphasizes lung
sounds by removal of heart sounds).
• extract specific components in the phonocardiographic signal (Paper II
describes a method able to detect the third heart sound).
• extract information suitable for classification of heart diseases (Paper III
presents a feature set for classification of systolic murmurs).
3
1.2. Thesis Outline

The thesis consists of two parts. The first part, chapter 1-7, contains an
introduction while the second part consists of three papers.
Chapter 1 introduces the problem at hand and defines the aim of the thesis.
Chapter 2 contains a brief summary on the physiology of the heart and the origin
of the phonocardiographic signal.
Chapter 3 describes methods for nonstationary and nonlinear time series analysis
and introduces dynamical systems theory within the context of signal processing.
Chapter 4 covers the characteristics of phonocardiographic signals in different
domains.
Chapter 5 discusses applications of phonocardiographic signal processing using
the methods and results from chapter 3 and 4.
Chapter 6 contains discussion and conclusions, including notes on future work.
Chapter 7 is a review of the papers presented in the second part of the thesis.
4
2. Preliminaries on Heart Sounds and Heart Murmurs
"The heart is of such density that fire can scarcely damage it."
Leonardo da Vinci (1452-1519)
This chapter sets the scene for up-coming sections. The physics of sound is
introduced followed by a review of the operation of the heart and the associated
terminology. The genesis of heart sounds and heart murmurs is discussed and
finally a short presentation of auscultation techniques and signal acquisition is
given.
2.1. Physics of Sound

It would be a mistake for a study on heart sounds to leave out an introduction to
the acoustic phenomena where everything actually starts. A sound is generated by
a vibrating object and propagates as waves of alternating pressure. The vibrating
source sets particles in motion, and if the sound is a pure tone, the individual
particle moves back and forth with the frequency of that tone. Each particle is thus
moving around its resting point, but as it pushes nearby particles they are also set
in motion and this chain effect results in areas of compression and rarefactions.
The alternating areas of compression and rarefaction constitute a pressure wave
that moves away from the sound source, see Figure 3. These pressure variations
can be detected via the mechanical effect they exert on some membrane (the
diaphragm of the stethoscope, the tympanic membrane etc.). If the sound source
vibrates in a more irregular manner, the resulting sound wave will be more
complicated. Usually, sound is described by its intensity, duration, frequency and
velocity [4]. If the sound is nonstationary, these measures have to be time varying
to give relevant information. Time varying analysis techniques are described in
section 3.1.
The number of vibrations per second, or frequency, is a physical entity. What
humans perceive as frequency is however called pitch. The two are closely related,
but the relationship is not linear. Up to 1 kHz, the measured frequency and the
perceived pitch are fairly the same. Above 1 kHz, a larger increase in frequency is
required to create an equal perceived change in pitch.
5
1 A B C D E F G H I
2 AB C D E F G H I
3 A BC D E F G H I
4 A B CD E F G H I
5 AB C DE F G H I
6 ABC D EF G H I
Time
7 A B CD E FG H I
8 A B CDE F GH I
9 AB C DE F G HI
10 ABC D EFG H I
11 A B CD E F GH I
12 A B CDE F GHI
Figure 3. The left figure is a schematic drawing of nine particles in simple harmonic motion at twelwe
moments in time. The sound source is located on the left side and the pressure wave, indicated by
clustering of three adjacent particles, moves from left to right. Note that each particle moves
relatively little around a rest position. In the right figure a pressure wave emanating from a sound
source (black circle) is illustrated. Drawn from [4].
2.2. Physiology of the Heart

The primary task of the heart is to serve as a pump propelling blood around the
circulatory system. When the heart contract, blood is forced through the valves,
from the atria to the ventricles and eventually out through the body, see Figure 4.
There are four heart chambers; right and left atria and right and left ventricles. The
two atria mainly act as collecting reservoirs for blood returning to the heart while
the two ventricles act as pumps to eject the blood to the body. Four valves prevent
backflow of blood; the atrioventricular valves (the mitral and tricuspid valve)
prevent blood from flowing back from the ventricles to the atria and the semilunar
valves (aortic and pulmonary valves) prevent blood from flowing back into the
ventricles once being pumped into the aorta and the pulmonary artery.
Deoxygenated blood from the body enters the right atrium, passes into the right
ventricle and is ejected into the pulmonary artery on the way to the lungs.
Oxygenated blood from the lungs re-enter the heart in the left atrium, passes into
the left ventricle and is then ejected into the aorta.
The blood pressure within a chamber increases as the heart contracts, generating a
flow from higher pressure areas towards lower pressure areas. During the rapid
filling phase (atrial and ventricular diastole), venous blood from the body and from
the lungs enters the atria and flows into the ventricles. As the pressure gradient
between the atria and the ventricles level out (reduced filling phase), a final
volume of blood is forced into the ventricles by atrial contraction (atrial systole).
In the beginning of ventricular systole, all the valves are closed resulting in an
isovolumic contraction. When the pressure in the ventricles exceeds the pressure
6
Chapter 2. Preliminaries on Heart Sounds and Heart Murmurs
in the blood vessels, the semilunar valves open allowing blood to eject out through
the aorta and the pulmonary trunk. As the ventricles relax the pressure gradient
reverses, the semilunar valves close and a new heart cycle begins.
Arch of Aorta
Superior Vena Cava Pulmonary Trunk
Left Atrium
Mitral Valve
Right Atrium
Aortic Semilunar Valve
Pulmonary Semilunar
Valve
Left Ventricle
Tricuspid Valve
Interventricular Septum
Right Ventricle
Inferior Vena Cava
Figure 4. Anatomy of the heart (left) and the blood flow pathways through left and right heart
(right).
2.3. Heart Sounds

The relationship between blood volumes, pressures and flows within the heart
determines the opening and closing of the heart valves. Normal heart sounds occur
during the closure of the valves, but how they are actually generated is still
debated. The valvular theory states that heart sounds emanate from a point sources
located near the valves, but this assumption is probably an oversimplification [5].
In the cardiohemic theory the heart and the blood represent an interdependent
system that vibrates as a whole [5]. Both these theories originate from a time when
the physiological picture was based on a one-dimensional conception of flow.
Recent research provides means to visualize the actual three-dimensional flow
patterns in the heart [6], and this new knowledge will probably clarify our view on
the underlying mechanisms of heart sounds. An example of a visualisation
technique called particle trace is shown in Figure 5. The blood’s pathway through
the heart is far from fully understood, but the induced vortices seem optimized to
facilitate flow and thereby increase the efficiency of the heart as a pump. The
impact of this new knowledge on the understanding of heart sounds and their
origin is yet to be investigated. Awaiting this new insight, the cardiohemic theory
will be assumed valid.
Normally, there are two heart sounds, see Figure 6. The first sound (S1) is heard in
relation to the closing of the atrioventricular valves, and is believed to include four
major components [3]. The initial vibrations occur when the first contraction of the
ventricle move blood towards the atria, closing the AV-valves. The second
component is caused by the abrupt tension of the closed AV-valves, decelerating
the blood. The third component involves oscillation of blood between the root of
the aorta and the ventricular walls, and the fourth component represents the
7
vibrations caused by turbulence in the ejected blood flowing into aorta. The
second sound (S2) signals the end of systole and the beginning of diastole, and is
heard at the time of the closing of the aortic and pulmonary valves [7]. S2 is
probably the result of oscillations in the cardiohemic system caused by
deceleration and reversal of flow into the aorta and the pulmonary artery [5].
Left ventricle
Left atrium Mitral valve
Figure 5. Particle trace (path line) visualization of intra-cardiac blood flow. The colour coding
represents velocity according to the legend to the right. The image was adapted from [6].
There is also a third and a fourth heart sound (S3 and S4). They are both connected
with the diastolic filling period. The rapid filling phase starts with the opening of
the semilunar valves. Most investigators attribute S3 to the energy released with
the sudden deceleration of blood that enters the ventricle throughout this period
[8]. A fourth heart sound may occur during atrial systole where blood is forced
into the ventricles. If the ventricle is stiff, the force of blood entering the ventricle
is more vigorous, and the result is an impact sound in late diastole, S4 [7]. There
are also sounds such as friction rubs and opening snaps, but they will not be
treated further.
2.4. Heart Murmurs

Murmurs are produced by turbulent blood flow as a result of narrowing or leaking
valves or from the presence of abnormal passages in the heart. More specifically,
heart murmurs occur when the blood flow is accelerated above the Reynolds
number. The resulting blood flow induces non-stationary random vibrations,
which are transmitted through the cardiac and thoracic tissues up to the surface of
the thorax. There are five main factors involved in the production of murmurs [7]:
• High rates of flow through the valves.
• Flow through a constricted valve (stenosis).
8
• Backward flow through an incompetent valve (insufficiency or

regurgitation).
• Abnormal shunts between the left and right side of the heart (septal
defects).
• Decreased viscosity, which causes increased turbulence.
Heart murmurs are graded by intensity from I to VI. Grade I is very faint and
heard only with special effort while grade VI is extremely loud and accompanied
by a palpable thrill. Grade VI murmurs are even heard with the stethoscope
slightly removed from the chest. When the intensity of systolic murmurs is
crescendo-decrescendo shaped and ends before one or both of the components of
S2, it is assumed to be an ejection murmur (S2 is composed of two components,
one from the aortic valve and one from the pulmonary valve). Murmurs due to
backward flow across the atrioventricular valves are of more even intensity
throughout systole and reach one or both components of S2. If the regurgitant
systolic murmur starts with S1 it is called holosystolic and if it begins in mid- or
late systole it is called a late systolic regurgitant murmur. Besides murmurs,
ejection clicks might also be heard in systole. They are often caused by
abnormalities in the pulmonary or aortic valves. Different murmurs, snaps, knocks
and plops can also be heard in diastole, but such diastolic sounds are beyond the
scope of this thesis.
Diastole Systole Diastole
Left ventricular
pressure
Aortic pressure
Pressure
Left atrial
R pressure
T
P
Amplitude
ECG
Q S
Heart sounds
S4 S1 S2 S3 S4
Mitral valve Open Open
Aortic valve Open
Figure 6. The four heart sounds in relation to various hemodynamic events and the ECG. All units
are arbitrary.
9
2.5. Auscultation and the Phonocardiogram

Auscultation is the technical term for listening to the internal sounds of the body.
The loudness of different components varies with the measurement location. For
instance, when listening over the apex, S1 is louder than S2. Also, the location of a
heart murmur often indicates its origin, e.g. mitral valve murmurs are usually
loudest at the mitral auscultation area. The traditional areas of auscultation, see
Figure 7, are defined as [7]:
• Mitral area: The cardiac apex.
• Tricuspid area: The fourth and fifth intercostal space along the left sternal
border.
• Aortic area: The second intercostal space along the right sternal border.
• Pulmonic area: The second intercostal space along the left sternal border.
Even though the definition of these areas came to life long before we had much
understanding of the physiology of the heart, they are still good starting points.
Revised areas of auscultation, allowing more degrees of freedom, have however
been adopted [7].
A
P
T
M
Figure 7. Traditional areas of auscultation (M refers to the mitral area, T the tricuspid area, P the
pulmonic area, and A the aortic area).
A graphical printing of the waveform of cardiac sounds is called a

phonocardiogram, PCG. An example of a phonocardiogram was shown in Figure
6. To obtain the phonocardiogram, a microphone is placed on the patient’s chest
and the signal is plotted on a printer, similar to the ones used in ECG recordings.
This technique promotes visual analysis of cardiac sounds, thus allowing thorough
investigation of temporal dependencies between mechanical processes of the heart
and the sounds produced.
2.6. Acquisition of Phonocardiographic Signals

The audio recording chain involves a sequence of transformations of the signal: a
sensor to convert sound or vibrations to electricity, a pre-amplifier to amplify the
signal, a prefilter to avoid aliasing and an analogue to digital converter to convert
the signal to digital form which can be stored permanently. In the setting of the
intelligent stethoscope, this chain is complemented with an analysis step and an
information presentation step.
10
2.6.1. Sensors
Microphones and accelerometers are the natural choice of sensor when recording
sound. These sensors have a high-frequency response that is quite adequate for
body sounds. Rather, it is the low-frequency region that might cause problems [9].
There are mainly two different kinds of sensors, microphones and accelerometers.
The microphone is an air coupled sensor that measure pressure waves induced by
chest-wall movements while accelerometers are contact sensors which directly
measures chest-wall movements [10]. For recording of body sounds, both kinds
can be used. More precisely, condenser microphones and piezoelectric
accelerometers have been recommended [11].
Electronic stethoscopes make use of sensors specially designed to suit cardiac
sounds. Compared to classic stethoscopes, electronic stethoscopes tries to make
heart and lung sounds more clearly audible using different filters and amplifiers.
Some also allow storage and the possibility to connect the stethoscope to a
computer for further analysis of the recorded sounds. The leading suppliers of
electronic stethoscopes are Thinklabs, Welch-Allyn and 3M. Thinklabs uses a
novel electronic diaphragm detection system to directly convert sounds into
electronic signals. Welch-Allyn Meditron uses a piezo-electric sensor on a metal
shaft inside the chest piece, while 3M uses a conventional microphone.
The studies included in this thesis have used two different sensors; the Siemens
Elema EMT25C contact accelerometer and an electronic stethoscope from Welch-
Allyn Meditron (the Stethoscope, Meditron ASA, Oslo, Norway).
2.6.2. Pre-processing, digitalization and storage

The preamplifier amplifies low level signals from transducer to line level. This is
important to be able to use the full range of the analogue to digital converter and
thus minimizing quantization errors. Another matter concerning the digitalization
of signals is aliasing which will occur unless the Nyquist-Shannon sampling
theorem is fulfilled.
In this work, when using EMT25C, a custom-built replica of a phonocardiography
amplifier (Siemens Elema E285E) was used. This amplifier includes a lowpass
filter with a cut-off frequency of 2 kHz. The signal was digitized with 12-bits per
sample using analogue to digital converters from National Instruments (see paper I
and II for specific details). Acquisition of the data was conducted in a Labview-
application (National Instruments, Austin, Texas, US) after which the data were
stored on a personal computer.
For the electronic stethoscope, the matching acquisition equipment and software
were used (Analyzer, Meditron ASA, Oslo, Norway). According to the
manufacturer, the digital recordings are stored without pre-filtering. An excessive
sampling frequency of 44.1 kHz was thus used to avoid aliasing (and with the idea
of post-filtering in mind). The signals were stored in a database on a personal
computer. This approach was used in paper III.
11
12
3. Signal Analysis Framework
“Calling a science nonlinear is like calling
zoology the study of non-human animals.”
Stanislaw Ulam
The underlying assumption of many signal processing tools is that the signals are
Gaussian, stationary and linear. This chapter will introduce methods suitable for
analysing signals that do not fall into these categories. Two short examples will
precede the theoretical treatment to illustrate the problem at hand.
Example 1, Characteristics that vary with time: A sinusoid with changing mean,
amplitude and frequency is shown in Figure 8. Using the Fourier transform to
investigate the signal’s frequency content, it can be seen that the signal contains
frequencies up to about 35 Hz. However, much more information can be obtained
by investigating how the frequency content varies over time. Methods able to
investigate how a certain signal property varies over time are suitable for
nonstationary signal analysis. A number of such methods are introduced in section
3.1.
(a) (b) (c)
350 50
10
300
40
Frequency (Hz)
FFT magnitude
5 250
Amplitude
200 30
0 150 20
100
−5 10
50
0 0
2000 4000 6000 8000 0 20 40 2000 4000 6000 8000
Time (ms) Frequency (Hz) Time (ms)
Figure 8. A sinusoid with changing mean, amplitude and frequency plotted as a waveform in the time
domain (a), as a frequency spectrum in the frequency domain (b) and in a combined time-frequency
domain (c).
Example 2, Distinguishing signals with similar spectra (example adapted from

[12]): In many traditional linear methods it is assumed that the important signal
characteristics are contained in the frequency power spectrum. From a stochastic
process perspective, the first and second order statistics of the signal are
13
represented by this power spectral information. However, there are many types of
signals, both theoretical and experimental, for which a frequency domain
representation is insufficient to distinguish two signals from each other. For
example, signals generated through nonlinear differential or difference equations
typically exhibit broadband spectral characteristics that are difficult to interpret
and compare. Two signals with indistinguishable power spectra are presented in
Figure 9.
(a) (b)
1.5 1.5
1 1
Amplitude
Amplitude
0.5 0.5
0 0
−0.5 −0.5
200 400 600 800 1000 200 400 600 800 1000
Time (sample) Time (sample)
Power/frequency (dB/rad/sample)
Power/frequency (dB/rad/sample)
(c) (d)
5 5
0 0
−5 −5
−10 −10
−15 −15
−20
0 0.5 1 0 0.5 1
Normalized Frequency (×π rad/sample)Normalized Frequency (×π rad/sample)
(e) (f)
1 1.5
s(t+2)
s(t+2)
0.5 0.5
0 −0.5
1 1.5
0.5 1 0.5 1.5
0.5 0.5
s(t+1) 0 0 s(t) s(t+1)−0.5 −0.5 s(t)
Figure 9. The logistic map, s(t+1) = c•s(t)[1-s(t)], where c is a constant, is often used as a model in
population studies. Here a logistic map with c=4 and s(0) = 0.1 is presented in (a) and a phase
randomized correspondence is shown in (b). Their respective frequency spectra, which are almost
identical, are shown in (c) and (d). Finally, in (e) and (f) their corresponding phase portraits are
shown.
The first signal is the logistic map and the second signal is its phase randomized
correspondence. Even though they have the same (but rather obscure) frequency
spectrum, the logistic map has structure in its phase portrait while the phase
14
Chapter 3. Signal Analysis Framework
randomized signal does not (the phase portrait is just the signal plotted against a
time delayed version of itself). To distinguish between the two, or for that matter,
to find the structure in the logistic map, it is obviously not enough to study their
spectra. Methods for nonlinear signal analysis will be introduced in section 3.3.
3.1. Measuring Characteristics that Vary in Time

Much information can be gained by investigating how certain signal properties
vary over time. Some properties possess natural time dependence, such as the
envelope when calculated as the energy of a signal. In other cases, there is a
typical trade-off between time resolution and accuracy in the sought parameter
(the typical example being the short time Fourier transform, STFT). The property
under investigation is then either a function of some quantity other than time (such
as frequency) or a scalar value (such as the mean value of the data). In either way,
time dependence can be introduced by calculating the wanted characteristic with a
sliding window.
3.1.1. Intensity
The textbook approach to extract a signal’s envelope, E(t), is via the analytic
signal [13, 14]. The continuous analytic signal is composed by the original signal
and its Hilbert transform according to equation (3.1) where H(t) is the Hilbert
transform (equation (3.2)).
s A ( t ) = s ( t ) + i ⋅ sH ( t ) (3.1)
1
∞
s (τ )
sH ( t ) = ∫ τ − t dτ (3.2)
π −∞
The Hilbert transform can be interpreted as a convolution between the signal and -
1/ t, or as a rotation of the argument with /2 for positive frequencies and – /2 for
negative frequencies. Similarly, the analytic signal can be obtained by removing
the negative frequencies and multiplying the positive frequencies by two [13].
Amongst several interesting properties of the analytic signal, the desired envelope
can be found as:
Analytic signal envelope: s A ( t ) = s ( t ) + sH ( t )

2 2
(3.3)
Other envelope-like measures are the absolute value or the square of the signal,
see equations (3.4)-(3.5). The absolute value gives equal weight to all samples
regardless of the signal’s intensity. The energy (square) on the other hand colours
the measure by emphasizing higher intensities compared to lower intensities. Two
other envelope-like measures are the Shannon entropy and the Shannon energy
[15], see equation (3.6)-(3.7). These measures give greater weight to medium
intensity signal components, thus attenuating low intensity noise and high intensity
disturbances. A practical issue with these approaches is that the envelope becomes
rather jagged. This is usually dealt with by low-pass filtering E(t) [13, 15]. A
15
method developed by Teager, equation (3.8), results in an on-the-fly envelope-

estimate that is very useful for analyzing signals from an energy point of view
[16].
Absolute value: E ( t ) = s ( t ) (3.4)
Energy (square): E ( t ) = s ( t )
2
(3.5)
Shannon entropy: E ( t ) = − s ( t ) ⋅ log s ( t ) (3.6)
Shannon energy: E ( t ) = − s ( t ) ⋅ log s ( t )

2 2
(3.7)
Teager’s energy operator: E ( t ) = s ( t ) − s ( t − 1) s ( t + 1)

2
(3.8)
A comparison of the methods can be seen in Figure 10. The test signal was a 200
Hz sinusoid with amplitude ranging from 0 to 1 and sampled with 10 kHz. In this
comparison, all outputs (but Teager), were low pass filtered by a 5th order
Butterworth filter with the cut off frequency 150 Hz.
(a) (b) (c)
1 1 0.7
|s(t)|
|s|
0.8 0.5
0.6 s(t)2
0.6 0 −s(t)2⋅log(s(t)2)
0 0.2 0.4 0.6 0.8 1
1 0.5 −|s(t)|⋅log|s(t)|
Amplitude of "energy"
0.4
s(t)2−s(t−1)*s(t+1)
s2
0.5 0.4
0.2
Amplitude
0
0 0 0.2 0.4 0.6 0.8 1 0.3
0.4
−s2⋅log(s2)
−0.2
0.2 0.2
−0.4
0
0 0.2 0.4 0.6 0.8 1 0.1
−0.6 0.4
−|s|⋅log|s|
0
−0.8 0.2
−1 0 −0.1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s) Amplitude of normalized signal
Figure 10. Comparison of different envelope estimation methods. The test signal is presented in (a)
and the result of various envelope-measures are shown in (b). The smoothed envelope measures are
presented in (c).
Another way to extract a signal’s envelope is through homomorphic signal

processing. Algebraically combined signals are here mapped into a domain where
linear signal processing tools can be used [17]. The envelope can be seen as an
amplitude modulation, where the slow varying envelope is multiplied with a
higher frequency signal. By taking the logarithm of the signal, equation (3.9), the
non-linear multiplication changes into a linear addition, equation (3.10). The low
frequency contribution l(t) may then be extracted from the high frequency
contribution h(t) by a low pass filter. Exponentiation takes the result back to the
original signal domain, equation (3.11). If the low pass filter is properly chosen,
l(t) will be a good estimate of the envelope.
s (t ) = l (t ) h (t ) (3.9)
log s ( t ) = log l ( t ) + log h ( t ) (3.10)
e { ( ) ( )} = e { ( )} { ( )} ≈ e { ( )} ≈ l ( t )
LP log l t + log h t LP log l t + LP log h t LP log l t
(3.11)
16
3.1.2. Frequency
The Fourier transform can be used to produce a time-averaged frequency
spectrum. However, it is often desirable to study how the frequency content of a
signal varies over time. There are many techniques available to perform such
analyses, and they are generally named time-frequency representations (TFR). The
simplest approach is probably the short time Fourier transform (STFT), which is
merely a windowed Fourier transform:
N kt
−2π i
STFT ( m, k ) = ∑ s ( t ) w ( t − m ) e N
(3.12)
t =1
where w denotes the time window, m is the translation parameter and k the
frequency parameter. If w is chosen as the Gaussian function in equation (3.13),
the obtained TFR is called the Gabor transform [18].
t2
1 − 2
g (t ) = e 2σ (3.13)
σ 2π
denotes the variance and is related to the width of the analyzing window. The
STFT does, however, suffer from the uncertainty principle. This means that the
frequency resolution decreases as the time resolution increases and the other way
around (since time is the dual of frequency). One way to obtain better resolution is
to use shorter windows for higher frequencies and longer windows for lower
frequencies. One such approach is the wavelet transform (WT) [13]:
1 N
⎛t−m⎞
WT ( m, a ) =
a
∑ s ( t ) w ⎜⎝
t =1
⎟
a ⎠
(3.14)
where m is a translation parameter and a is a scale parameter. The main idea is that
any signal can be decomposed into a series of dilatations or compressions of a
mother wavelet denoted w(t). The mother wavelet should resemble interesting
parts of the signal, and the choice is important for the results. An issue with
wavelets is that the link to local frequency is lost (hence the term scale is preferred
instead of frequency). A similar but phase corrected transform, able to maintain
the notion of frequency, is the S-transform (ST) [19]:
N ( t − m )2 k 2 kt
k − −2π i
ST ( m, k ) = ∑ s ( t ) e 2
e N
(3.15)
t =1 2π
Compared to STFT, the window function is chosen as a Gaussian function where
the variance is allowed to vary proportionally with the period of the analyzing
sinusoid. When changing the variance, the width of the window is altered giving a
multi-resolution description of the signal. An example comparing STFT, WT and
ST is given in Figure 11.
17
There are many other approaches available for joint time-frequency analysis. The
methods just described belong to the linear nonparametric group. The quadratic
nonparametric group, the parametric group etc., will not be treated in this thesis.
(a) (b)
0.5
1 0.4
Frequency
Amplitude
0.3
0
0.2
−1
0.1
0
200 400 600 800 1000 200 400 600 800 1000
(c) (d)
200 200
150 150
Frequency
Frequency
100 100
50 50
200 400 600 800 1000 200 400 600 800 1000
Figure 11. An example signal consisting of two well separated sinusoids (sample 1-500) and two
chirps, one ascending and one descending (sample 600-1100) is given in (a). TFR-plots calculated
with STFT, WT (with a Daubechie 2 mother wavelet) and ST are given in (b-d), respectively. The
frequency axes are in arbitrary units.
3.2. Nonlinear Systems and Embedology

Dynamical systems theory is an important ingredient in nonlinear signal
processing. A dynamical system is a system whose state changes with time. In
continuous time, the system is described by differential equations and in discrete
time by iterated maps. Finding explicit solutions for these equations is not the
purpose of this chapter. Instead, the main goal is to improve knowledge about the
system by deducing quantitative information. Since sampled data is used, only
iterative maps will be considered.
The dynamics of a time discrete system is determined by its possible states in a
multivariate vector space (called state space or phase space). The transitions
between the states are described by vectors, and these vectors form a trajectory
describing the time evolution of the system according to equation (3.16).
x ( t + 1) = ϕ ( x ( t ) ) (3.16)
where x(t) is the state of the system, t is the time index, is a mapping function
such that :M M and M is the true state space. A geometrical display of the
trajectory, such as in Figure 12, provides a phase portrait of the system. If the
18
trajectory is drawn to a particular set this set is called an attractor. Examples of

different attractors are given in Figure 12.
(a) (b) (c)
Figure 12. Examples of a fixed point attractor (a), a limit cycle (b) and a strange attractor from a
Lorenz system (c). A physical example of a fix point attractor is a pendulum, where all initial states
will converge to a single point. Modifying this example so that the pendulum has a driving force, thus
creating a simple oscillation, a periodic attractor is obtained. Chaotic systems like the Lorenz system
have been used to describe weather, and give rise to strange attractors, where the trajectories never
cross or touch each other.
The true state space (M) thus contains the true states x, whose time evolution is
described by the map , x(t)= t(x(0)). Now suppose that the only information
available about this system is what we can find in a scalar measure s(t)=h(x(t)),
where h:M → . If s(t) is a projection from the true (multivariate) state space M,
then it might be possible to undo the projection, see Figure 13. That is, given a
measured signal s(t) in , is there a way to create a map from an unknown state
x(t) in M to a corresponding point y(t) in a reconstructed state space in d?
Takens’ theorem provides us with such a map [20]:
F : M → d
(3.17)
x(t ) → y ( t ) = F ( x ( t ) ) = ⎡⎣ s ( t ) , s ( t + τ ) ,..., s ( t + ( d − 1)τ ) ⎤⎦
where is a delay parameter, d is the embedding dimension and F is the map from
the true state space to the reconstructed state space. The selection of and d affects
how accurately the embedding reconstructs the system’s state space. These issues
are important, but there are no bullet-proof ways to determine and d. In this
thesis will be determined using average mutual information [21] and d will be
chosen based on Cao’s method [22].
What Takens actually proved was that the reconstructed state space d is a
dynamical and topological equivalent to M. Since the dynamics of the
reconstructed state space contains the same topological information as the original
state space, characterization and prediction based on the reconstructed state space
is as valid as if it was made in the true state space.
19
Table 1. Comparison of linear and nonlinear signal processing techniques. The table is adapted from
[21].
Linear signal processing Nonlinear signal processing
Finding the signal: Finding the signal:
Separate broadband noise from Separate broadband signal from broadband
narrowband signal using spectral noise using the deterministic nature of the
characteristics. Method: Matched filter signal. Method: Manifold decomposition
in frequency domain. or statistics on the attractor.
Finding the space: Finding the space:

Use Fourier space methods to turn Time lagged variables form coordinates for
difference equations into algebraic a reconstructed state space in d
forms. dimensions.
s ( t ) is observed and y ( t ) = ⎡⎣ s ( t ) , s ( t + τ ) ,..., s ( t + ( d − 1)τ ) ⎤⎦
S ( f ) = ∑ s ( t ) ei 2π tf is used. where d and are determined by false
nearest neighbours and mutual
information.
Classify the signal: Classify the signal:

• Sharp spectral peaks • Lyapunov exponents
• Resonant frequencies of the • Fractal dimension measures
system
• Unstable fixed points
• Recurrence quantification
• Statistical distribution of the
attractor
Making models, predict: Making models, predict:

s (t + 1) = ∑ α k s (t − k ) y (t ) → y (t + 1)
Find parameters k consistent with y (t + 1) = F ⎡⎣ y (t ), a1 , a2 ,… , a p ⎤⎦

invariant classifiers – location of
spectral peaks. Find parameters aj consistent with
invariant classifier – Lyapunov exponents,
fractal dimensions.
20
Figure 13. Delay reconstruction of states from a scalar time series (example using the Lorenz system).
Redrawn from [20].
3.3. Nonlinear Analysis Tools

Nonlinear analysis tools are rather different from their linear analogue. A brief
comparison between linear and nonlinear methods can be found in Table 1.
3.3.1. Non-integer dimensions

The concept of non-integer dimensions may sound abstract, but it can be
intuitively motivated. For example, is the Henon map one or two dimensional? It
seems two-dimensional when looking at it from a broad scale, and, it never breaks
down into a one-dimensional line no matter how much it is magnified, see Figure
14. The answer to the question whether the Henon map is one or two-dimensional
seems to be that it is “somewhere in between”, i.e. it has a fractal dimension.
Strange attractors are fractal, and their fractal dimension is less than the dimension
of the state space it lives in.
(a) (b) (c)
0.4
0.2
−0.2
−0.4
−2 0 2
Figure 14. Zooming into the Henon map reveals new levels of complexity. No matter how much the
figure is magnified it will never collapse into a one-dimensional line, nor does it fill the two-
dimensional space in (a). Instead, the Henon map has a dimension somewhere in between one and
two, i.e. a fractal dimension.
21
There are two types of approaches to estimate the fractal dimension; those that
operate directly on the waveform and those that operate in the reconstructed state
space. Note that the dimension of the attractor (measured in the reconstructed state
space) is normally different from the waveform fractal dimension (measured on
the projected signal s(t), and thus limited to the range 1 dim 2. There are a
number of problems when determining the fractal dimension of an attractor in
state space, one being the computational burden [23]. For this reason, we will only
consider waveform fractal dimensions. In this setting, the signal is looked upon as
a planar set in 2, where the waveform is considered a geometric figure. Even
though there are many ways to estimate the fractal dimension of a waveform, we
focus on the variance fractal dimension (VFD) due to its robustness to noise [24].
The calculations are based on a power law relation between the variance of the
amplitude increments of the signal and the corresponding time increments, see
equation (3.18).
Var ( s ( t2 ) − s ( t1 ) ) ∼ t2 − t1
2H
(3.18)
where H is the Hurst exponent, a measure of the smoothness of a signal. The Hurst
exponent can be calculated by taking the logarithm of (3.18):
(
log Var ( s ( t2 ) − s ( t1 ) ) = 2 H ⋅ log ( t2 − t1 ) ) (3.19)
Plotting log(Var(s(t2)-s(t1))) against log(|t2-t1|) for different time increments in a

log-log plot, H is determined as the slope of the linear regression line, see Figure
15.
(a)
Distance from origin
600
400
200
0
2 4 6 8 10
Time (sample) 4
x 10
(b)
log( var( s(t2)−s(t1)))
1
slope = 2H
0
0 1 2 3 4
log(|t2−t1|)
Figure 15. An example showing Brownian motion (a) and the corresponding log-log plot (b). H is
calculated to 0.4983 which is very close to the theoretical answer 0.5. The variance fractal dimension
is VFD = Ed+1-H = 1+1-0.5 = 1.5.
22
The variance fractal dimension is related to H as VFD = Ed+1-H, where Ed is the

Euclidian dimension (Ed = 1 for time series). The choice of time increments is
reflected in the VFD trajectory for the part of the signal where signal exists. If the
time increments are chosen in a dyadic manner (1, 2, 4, 8,…), differences between
various signal components are emphasized whereas if the time increments are
chosen by unit decimation (1, 2, 3, 4,…), segmentation of signal from noise is
favourable [24]. Since VFD is a scalar value calculated from the samples at hand,
a sliding window approach has to be used to describe the complexity over time.
3.3.2. Recurrence quantification analysis

The state space of a system is often high-dimensional, especially when
reconstructed from experimental data where noise tends to inflate the dimension.
Its phase portrait can therefore only be visualized by projection into two or three
dimensions. This operation does however fold the attractor, and by doing so,
destroys its structure. A recurrence plot (RP) is a way to visually investigate the d-
dimensional state space trajectory through a two-dimensional representation. RPs
can be used on rather short time series and represent the recurrence of states of a
system (i.e. how often a small region in phase space is visited). Unlike other
methods such as Fourier, Wigner-Ville or wavelets, recurrence is a simple relation,
which can be used for both linear and nonlinear data [25]. An RP is derived from
the distance plot, which is a symmetric NxN matrix where a point (i, j) represents
some distance between y(i) and y(j). Thresholding the distance plot at a certain
cut-off value transforms it into an RP:
(
RP(i, j ) = Θ ε − y ( i ) − y ( j ) ) (3.20)
where i,j=1,…,N, is a cut-off distance, • is some norm and (•) is the

Heaviside function. An example of a recurrence plot is shown in Figure 16. States
that are close to each other in the reconstructed state space are represented by
black dots in the recurrence plot.
Time series Phase portrait Recurrence plot
1 1 800
Timek+10 (sample)
Timek (sample)
0.5 0.5 600

Amplitude
0 0
400
−0.5 −0.5
200
−1 −1
200 400 600 800 −1 0 1 200 400 600 800
Time (sample) Timek (sample) Timek (sample)
Figure 16. A noisy sinusoid represented with its waveform, its phase portrait (in a reconstructed state
space) and by its recurrence plot. Dots are positioned on the waveform near amplitude values of 0.5,
red dots for increasing amplitude and blue dots for decreasing amplitude. The recurrence plot shows
clear diagonal lines which arise when trajectories in the state space run in parallel for some time
period. The red and blue dots end up on these lines, and it can be seen that the distance between two
diagonal lines is the period of the sinusoid.
23
There are seven parameters affecting the outcome of an RP; the embedding
dimension d, the time delay , the range (or length) of the time series under
investigation, the norm • , the possibility to rescale the distance matrix, the cut-
off distance and the minimal number of adjacent samples to be counted as a line
(minimum line length) [26]. The last parameter is not used when creating RPs, but
rather when trying to quantify them (recurrence quantification analysis, RQA).
Measures used for RQA are often based on diagonal structures, vertical structures
and time statistics. Isolated recurrence points occur if states are rare, if they do not
persist for any time or if they fluctuate heavily. Diagonal lines occur when a
segment of the trajectory runs in parallel with another segment, i.e. when the
trajectory visits the same region of the phase space at different times, see Figure
16. Vertical (horizontal) lines mark a time length in which a state does not change
or changes very slowly. The most common RQA-parameters are [27-29]:
• Recurrence rate: The percentage of recurrence points (black dots) in the
recurrence matrix.
• Determinism: The percentage of the recurrence points that form diagonal
lines. Diagonal lines are associated with deterministic patterns in the
dynamic, hence determinism.
• Laver: The average length of the diagonal lines.
• Lmax: The length of the longest diagonal line. Lmax is inversely proportional
to the largest Lyapunov exponent which describes how fast trajectories
diverge in the reconstructed state space.
• Entropy: The Shannon entropy of the distribution of the diagonal line
lengths. Measures the complexity of the signal.
• Laminarity: The percentage of recurrence points which form vertical lines.
• Trapping time: The average length of the vertical lines.
• Vmax: The length of the longest vertical line.
• T1: Recurrence time of the first kind, see below.
• T2: Recurrence time of the second kind, see below.
Detection of changes based on recurrence times

Detection of changes in signals is traditionally fitted into a residual based
framework [30]. The main idea is to make a time varying model of the signal. If
the model is correct, the residuals are expected to be white (or ideally zero). When
the model and the signal no longer correspond to each other a change is indicated.
A change in the dynamics of a signal may also be detected as a change in the
trajectories in the reconstructed state space. Distance measures of such changes
often involve some count of nearest neighbours since the neighbours indicate
recurrence of states. Actually, the RQA-parameters T1 and T2, recurrence times of
the first and second kind, are such measures. Nearest neighbours in the
reconstructed state space can be divided into true recurrence points and sojourn
points [27], see Figure 17, where T1 is all the points and T2 is the black points.
24
y(ref)
Figure 17. Recurrence points of the second kind (solid circles) and the sojourn points (open circles) in
B (y(ref)). Recurrence points of the first kind comprise all circles in the set.
More formally, an arbitrary state, yref, is chosen somewhere on the trajectory

whereupon all neighbouring states within a hypersphere of radius are selected.
{
Bε ( y ( ref ) ) = y ( t ) : y ( t ) − y ( ref ) ≤ ε } ∀t (3.21)
The recurrence points of the first kind (T1) are defined as all the points within the
hypersphere (i.e. the entire set B ). Since the trajectory stays within the
neighbourhood for a while (thus generating a whole sequence of points), T1
doesn’t really reflect the recurrence of states. Therefore, the recurrence points of
the second kind (T2) are defined as the first states entering the neighbourhood in
each sequence (these points are commonly called true recurrence points). T2 is
hence the set of points constituted by B (y(ref)) excluding the sojourn points, see
Figure 17. Both T1 and T2 are related to the information dimension via a power
law, motivating their ability to detect weak signal transitions based on amplitude,
period, dimension and complexity [31]. Specifically, T2 is able to detect very
weak transitions with high accuracy, both in clean and noisy environments while
T1 has the distinguished merit of being more robust to the noise level and not
sensitive to the choice of . A mathematically more rigorous definition of T1 and
T2 can be found in [31]. A sliding window approach is necessary to obtain time
resolution.
3.3.3. Higher order statistics

Standard methods in signal processing are based on second-order statistics, but
second-order measures contain no phase information. As a consequence, non-
minimum phase signals and certain types of phase coupling (associated with
nonlinearities) cannot be correctly identified. In contrast to second-order statistics,
higher order statistics are based on averages over products of three or more
samples of the signal, thus allowing nonlinear dependencies among multiple signal
samples to be evaluated. Assuming zero mean signals and limiting the survey to
order 4, the order moments and their corresponding cumulants are defined as [32]:
25
ms(1) = cs(1) = E {s ( t )} = 0
ms(2) (τ ) = cs(2) (τ ) = E {s ( t ) s ( t + τ )}
ms(3) (τ 1 ,τ 2 ) = cs(3) (τ 1 ,τ 2 ) = E {s ( t ) s ( t + τ 1 ) s ( t + τ 2 )} (3.22)
ms(4) (τ 1 ,τ 2 ,τ 3 ) = E {s ( t ) s ( t + τ ) s ( t + τ 2 ) s ( t + τ 3 )}
cs(4) (τ 1 ,τ 2 ,τ 3 ) = E {s ( t ) s ( t + τ ) s ( t + τ 2 ) s ( t + τ 3 )} − 3 ⎡⎣ E {s ( t ) s ( t + τ )}⎤⎦
2
where E represents the expected value. Interesting special cases are cs(1)(0),
cs(2)(0,0) and cs(3)(0,0,0) which represent the variance, skewness and kurtosis of
s(t). The Fourier transforms of cumulants are called polyspectra, and are defined
according to equation (3.23). An example of a simple bispectrum, the Fourier
transform of the third order cumulant, is shown in Figure 18.
(a) (b)
100
600
(ω1,ω2)|
80
FFT magnitude
400
200
|C(3)
60
S
40
30
20 20
10
0 30
0 10 20 30 40 ω2 10
20
Frequency (ω) ω1
Figure 18. An example of phase coupling. The frequency spectrum of a signal composed of three
sinusoids with frequencies 1, 2 and 3 = 1 + 2 is shown in (a). The corresponding bispectrum is
shown in (b). Since 3 is caused by phase coupling between 1 and 2, a peak due to the phase
coupling will appear in the bispectrum at 1 = 1, 2 = 2 (another peak will also emerge at 1 = 2,
2 = 1).
{
Cs(2) (ω ) = FT cs(2) (τ ) } Power spectrum
C s (ω1 , ω2 ) = FT {c (τ 1 ,τ 2 )} Bispectrum
(3) (3)
s (3.23)
Cs(4) (ω1 , ω2 , ω3 ) = FT {cs(4) (τ 1 ,τ 2 ,τ 3 )} Trispectrum
Complete characterisation of a stochastic process requires knowledge of all

moments. Generally speaking, moments correspond to correlations and cumulants
correspond to covariances. Even though both measures contain the same statistical
information, cumulants are preferred in practice since [33]:
1. Cumulant spectra of order n > 2 are zero for Gaussian signals and their
polyspectra provide a measure of the extent of non-Gaussianity.
2. The covariance function of white noise is an impulse function and its
spectrum is flat. Similarly, cumulants of white noise are multidimensional
impulse functions with multidimensionally flat polyspectra.
26
3. The cumulant of two independent random processes equals the sum of the
cumulants of the individual random processes.
Higher order cumulants provide a measure of how much a random vector deviates
from a Gaussian random vector with an identical mean and covariance matrix.
This property can be used for extracting the nongaussian part of a signal (one
application is removal of Gaussian noise). Other interesting properties are that the
bispectrum is zero for a Gaussian signal and that the bicoherence (normalized
bispectra) is constant for a linear signal.
3.4. Nonlinear Prediction

There are different sources of predictability in a time series. If the signal contain
linear correlations in time, then linear models are suitable (moving average (MA)
models, autoregressive (AR) models, autoregressive moving average (ARMA)
models etc.). MA models can be used if the spectrum of the signal behaves like
coloured noise while AR models are preferable if the spectrum is dominated by a
few distinct peaks. The ARMA model is a natural extension of AR and MA. The
AR model is obtained as a sum of weighted previous outputs s(t-k) and the
innovation signal e(t):
N
s (t ) = ∑α k s (t − k ) + e (t ) (3.24)
k =1
where k are the linear weights. For prediction, only the weighting coefficients are
important and the prediction is obtained by ignoring the unknown innovation. This
model can be expanded to allow nonlinear dependencies between previous outputs
s(t-k). Actually, a very general framework for predicting time series is given in
Ljung [34] ( may include all available signal samples including multivariate
inputs and outputs):
N
s ( t θ ) = ∑ α k g k (ϕ )
k =1
θ = [α1 , α 2 ,..., α n ]
T
(3.25)
g k ( ϕ ) = κ ( β k (ϕ − γ k ) )
ϕ = ⎣⎡ s ( t − k ) ,..., s ( t − 1) ⎦⎤
where all the gk are formed from dilated and translated versions of a mother basis
function . is a vector of weights and is a vector of known signal samples.
are the coordinates or weights, are the scale or dilation parameters and are the
location or translation parameters. A few examples of how this model framework
can be used are:
Autoregressive model: set most of the parameters to unity.
Sigmoid Neural Network: is a ridge construction such as the sigmoid function.
Radial basis networks: is a radial construction such as the Gaussian bell.
27
Turning back to the reconstructed state space setting, it can be seen that in
equation (3.25) is very similar to a reconstructed coordinate. Toss in a delay
parameter , or set = 1, and turns into y (see equation (3.17)). A way to look at
the model in (3.25) is thus as a function describing the whole attractor. Usually, all
parameters but the :s are design parameters that either vary in a predetermined
way or are fixed. Inserted into a cost function, (3.25) leads to linear equations
when estimating the :s, thus simplifying their determination [23]. Since this
modelling attempt tries to model the whole attractor, it is called a global model.
That being said about global models, we will abandon them altogether and focus
on local methods working directly in the reconstructed state space. Similar
trajectories in state space share the same waveform characteristics in time domain,
and a way of predicting the future is thus to mimic the evolution of neighbouring
trajectories, see Figure 19. If the data is sampled with high frequency, most of the
discovered nearest neighbours will probably be samples adjacent to each other in
the time series. A considerable improvement could thus be obtained by using
nearest trajectories instead of nearest neighbours see Figure 20.
^
y(t+1)
y(t)
Figure 19. Three trajectory segments and a (forward) predicted trajectory in a two-dimensional
phase space. The average change between the nearest neighbouring trajectory points (black stars)
and their successors (white circles) are used to predict the next point (white square).
y(t)
Figure 20. Many of the nearest neighbours to y(t), stars, are actually phoneys due to the high
sampling rate. Using a nearest trajectory algorithm instead of a nearest neighbour algorithm is one
solution.
28
There are two approaches to predict p steps ahead, either using iterated prediction
or direct prediction. If the prediction is iterated, the algorithm predicts one step
ahead p times (the predicted values will then be used as a starting point in the next
iteration). In direct prediction, the evolutions of the nearest neighbours are
modelled and the resulting function maps p steps into the future. It is empirically
shown that iterated prediction is better on short term forecasts for a variety of non-
linear models. However, iterated predictions do not take the accumulated errors in
the input vector into account, and these errors grow exponentially [35].
More options are available. In averaged prediction, the average of the neighbours’
successors (white circles) locations are chosen as the predicted value while in
integrated prediction the next point is estimated as the current point plus the
average change amongst the neighbours. If the trajectory that is to be predicted is
an outlier; the mean of the nearest neighbours will always be misleading.
To conclude, local models can give excellent short-term prediction results, they
are conceptually simple but may require a large computational burden due to the
dependence of nearest neighbour calculations.
29
4. Properties of Phonocardiographic Signals
“Let your heart guide you. It whispers, so listen closely…”
The land before time (1988)
During auscultation, identification of heart sounds and murmurs is primarily based

on pitch and timing of occurrences. In phonocardiography, information about
morphology, and to some extent frequency content, is also included in the
diagnosis. There is however more knowledge to extract. Physicians talk about
auscultation as an art. The diagnosis is often based on a sensation that is hard to
explain and even harder to implement in a computer. To automatically extract
information that is even close to this sensation, every possible grain of information
has to be exploited, even though it means going beyond the well-known properties
of time and frequency.
Some examples in this chapter are based on data (used with permission) from
FamilyPractice.com (http://www.familypractice.com/) and The Auscultation
Assistant (http://www.med.ucla.edu/wilkes/). These data were chosen to give
distinct and illustrative figures. The data were originally intended for educational
purposes, and consists of text-book examples of recorded phonocardiographic
signals and in some cases even simulated sounds. Other examples and all the
presented results are however obtained using data from paper I-III. This means
that noisy signals with, in some cases, large interpatient variability are used in the
calculations.
4.1. Time and Frequency

Time and frequency properties are the most important features when physicians
perform auscultation on a patient. In healthy subjects, the frequency spectrum of
S1 contains a peak in the low frequency range (10-50 Hz) and in the medium
frequency range (50-140 Hz) [36]. S2 contains peaks in low- (10-80 Hz), medium-
(80-220 Hz) and high-frequency ranges (220-400 Hz) [37]. S2 is composed of two
components, one originating from aortic valve closure and one originating from
pulmonary valve closure. Normally, the aortic component (A2) is of higher
frequency than the pulmonary component (P2) [38]. The peaks probably arise as a
result of the elastic properties of the heart muscle and the dynamic events that
31
causes the various components of S1 and S2 [36, 37]. S3 and S4 are believed to
originate from vibrations in the left ventricle and surrounding structures powered
by the acceleration and deceleration of blood flow. 75 % of the total energy in S3
is contained below 60 Hz [39] while S4 mainly contain frequencies below 45 Hz
[40]. The time and frequency properties of heart sounds are summarized in
Table 2 and examples of two heart sounds and their frequency spectra are
illustrated in Figure 21. The different heart sounds are affected by various heart
diseases, and the main changes are described in sections 4.1.1 - 4.1.3.
Table 2. Time and frequency properties for the heart sounds.

Sound Location (ms) Duration Main frequency
(ms) range (Hz)
S1 10-50 after R-peak in ECG 100-160 10-140
S2 280-360 after R-peak in ECG 80-140 10-400
S3 440-460 after R-peak in ECG or 40-80 15-60
120-180 after closure of semilunar
valves
S4 40-120 after beginning of P-wave in 30-60 15-45
ECG
Phonocardiogram S1 S2 S3
S1 S2
FFT magnitude
FFT magnitude
FFT magnitude
Amplitude
S3
Subject 1
0.2 0.4 0.6 0 100 200 0 100 200 0 100 200

Time (s) Frequency (Hz) Frequency (Hz) Frequency (Hz)
S1
FFT magnitude
FFT magnitude
FFT magnitude
S2
Amplitude
S3
Subject 2
0.2 0.4 0.6 0 100 200 0 100 200 0 100 200

Time (s) Frequency (Hz) Frequency (Hz) Frequency (Hz)
Figure 21. Heart sounds and their respective frequency spectra from a 13 year old girl (top row) and
a 36 year old male (bottom row). Data obtained from paper I and II.
There is a small delay between the aortic component and the pulmonary
component causing a splitting of S2 (since right ventricular ejection terminates
after left ventricular ejection). Normally, the splitting increases with inspiration
due to increased blood return to the right heart, increased vascular capacitance of
the pulmonary bed and decreased blood return to the left heart [7]. That is, the
32
Chapter 4. Properties of Phonocardiographic Signals
aortic component occurs earlier and the pulmonary component occurs later during
inspiration. In certain heart diseases, this splitting can become wide, fixed or
reversed (see sections 4.1.1 - 4.1.3). FFT analysis does not take timing into
consideration, so it cannot reveal which of the two valves closes first. Meanwhile,
it is hard to notice any difference between the two components in the time domain.
A tool able to investigate how the signal’s frequency content varies over time is
thus called for. Such methods were introduced in section 3.1.2, and an example
showing the four heart sounds is presented in Figure 22. Taking a closer look at S2
in Figure 22, it can be seen that the two components are merged together, but it is
also clear that the higher frequency aortic component precede the lower frequency
pulmonary component.
S1 S2
300 300
Frequency (Hz)
250 250
200 200
150 150
100 100
50 50
50 100 150 50 100 150
S3 S4
300 300
Frequency (Hz)
250 250
200 200
150 150
100 100
50 50
20 40 60 80 100 120 20 40 60 80
Time (s) Time (s)
Figure 22. Example of TFR contour plots of S1, S2, S3 and S4 (note the different scaling of the x-
axis). Stockwell’s method was used to calculate the TFR. Data obtained from paper II.
An exposé of systolic murmurs is given in sections 4.1.1 - 4.1.3, where the

emphasis is on the changes imposed by pathologies on the phonocardiogram.
More information can be found in [7]. Basic layout sketches describe the time
domain properties, while frequency domain information is illustrated with TFR-
plots (since heart murmurs are nonstationary [41], useful frequency investigations
have to be conducted using joint time frequency analysis tools).
33
4.1.1. Murmurs from stenotic semilunar valves

If the aortic or pulmonary valve becomes narrowed or constricted (stenotic), blood
has to be forced through the valve opening. The arising turbulent blood flow
causes vibrations in the cardiac structure which are transmitted through the tissue
and perceived as a murmur. The murmur peaks in mid-systole at the time of
maximal ejection and produces a crescendo-decrescendo shape in the
phonocardiographic signal. The severity of the stenosis influences the shape of the
murmur, where the intensity will increase and the peak will occur later in systole
as the stenosis becomes more severe. An ejection click may occur if the valves are
brought to an abrupt halt instead of opening completely when moving from their
closed position in diastole to their open position in systole. Since a large boost
from the atrium might be necessary to help build up the pressure in the ventricle, a
fourth heart sound may be present. The appearance of a murmur caused by
stenosis in the semilunar valves is illustrated in Figure 23.
1000
Frequency (Hz)
800
600
S1 S2
400 S2
EC
S1
200
100 200 300 400 500

Time (ms)
Figure 23. A basic layout sketch of the phonocardiographic signal from a murmur caused by stenosis
in the semilunar valves is presented in the left plot while an example TFR (showing pulmonary
stenosis, calculated by Stockwell’s transform) is illustrated in the right plot. EC = Ejection click.
In aortic stenosis, there will be a paradoxical splitting of S2 that increases with

expiration. If the stenosis is severe the aortic component is attenuated or even
missing. Aortic stenosis is usually caused by congenital aortic valve disease,
rheumatic fever or degenerative calcification.
In pulmonary stenosis, the splitting of S2 is caused by increased capacitance in the
dilated pulmonary trunk. In severe stenosis, S2 is widely split but difficult to hear
since the pulmonary component is faint and the aortic component is obscured by
the murmur. The degree of stenosis correlates well with the width of S2. In mild
stenosis the murmur ends before the aortic component of S2 while it is stretched
up to the pulmonary component in severe cases. Pulmonary stenosis is usually
caused by congenital fusion of the pulmonary valve cusps.
4.1.2. Murmurs from regurgitant atrioventricular valves

Backward flow through the mitral or tricuspid valves causes a murmur that begins
as soon as the atrioventricular valves closes and continues up to the semilunar
valve closure. Because the pressure gradient between ventricle and atrium is large
throughout systole, the murmur tends to have a constant intensity throughout
34
systole (holosystolic). The appearance of a murmur caused by insufficient

atrioventricular valves is illustrated in Figure 24.
In mitral insufficiency (MI), the murmur begins with systole and continues as long
as left ventricular pressure exceeds that of the enlarged left atrium. The murmur
engulfs the aortic component of S2 but stops before the pulmonary component. In
acute mitral insufficiency the murmur is diamond-shaped, loud (grade IV or more)
and does not necessarily extend to the end of systole. Due to the decrease in the
rate of rise of the ventricular pressure, S1 is weak. A third heart sound is often
present, although it doesn’t necessarily imply systolic dysfunction or elevated
filling pressure as it usually does. The mitral valve apparatus is composed of five
different structures; mitral annulus, leaflets, chordae tendinae, papillary muscles
and the free wall of the left ventricle. Malfunction in any of these may result in
mitral insufficiency. Typical examples are rheumatic heart disease in the mitral
valves, papillary muscle dysfunction or calcified mitral annulus.
In tricuspid insufficiency, the holosystolic murmur increases with inspiration.
Unlike mitral insufficiency the murmur of tricuspid insufficiency persists through
S2 and engulfs the pulmonary component. In mild cases, S4 may be present.
Tricuspid insufficiency is often found in severe right heart failure.
1000
Frequency (Hz)
800
600
S1 S2
400 S1 S2
200
100 200 300 400 500

Time (ms)
Figure 24. A basic layout sketch of the phonocardiogram from a murmur caused by a regurgitant
atrioventricular valve is presented in the left plot while an example TFR (mitral insufficiency,
calculated by Stockwell’s transform) is illustrated in the right plot.
4.1.3. Murmurs caused by septal defects

A septal defect is an opening between the left and the right side of the heart. In
atrial septal defect the opening connects the two atria and in ventricular septal
defect the opening connects the two ventricles. Septal defects are often congenital,
and ventricular septal defects are in fact the most common congenital heart
defects.
In ventricular septal defect, already oxygen-rich blood is pumped back to the
lungs, causing a dilation of the heart. During systole, a jet from the left to the right
ventricle causes a holosystolic murmur that sometimes engulfs S2. The aortic
component of S2 is early due to the short ejection period (caused by low
impedance in the systemic circulation system). This results in an S2-split that
increases with inspiration.
35
In atrial septal defect, recently oxygenated blood leaks back to the right atrium
where it is, again, sent through the pulmonary circulation system. The increased
flow through the pulmonary valve produces a soft mid-systolic ejection murmur.
S2 has a large fix split caused by decreased resistance in the pulmonary aorta
which delays the pulmonary component of S2.
4.1.4. Quantifying the results

It is a well known fact that systolic ejection murmurs like aortic stenosis are
crescendo-decrescendo shaped while systolic murmurs due to backflow, like
mitral insufficiency, are holosystolic (do not vary in intensity). This is not always
obvious when looking at actual recorded signals. Using normalized Shannon
energy as a measure of intensity, the shape of the murmur in aortic stenosis (AS),
mitral insufficiency (MI) and physiological murmurs (PM) are shown in Figure
25. The data used is from the 36 patients in paper III. The classical shapes are
indicated, but having the standard deviation in mind, this conclusion is not evident.
In the AS case, the large standard deviations are probably due to the progress of
the disease, which is ranging from moderate to severe.
The nine presented instants were selected at times before S1, peak S1, after S1, ¼
into systole, ½ into systole, ¾ into systole, before S2, peak S2 and finally after S2.
The data were derived as mean values of each heart cycle in one patient. These
mean-values were then used to create the plots in Figure 25. Similarly, the mean
and standard deviation TFRs of AS, MI and PM were calculated. An average TFR
was derived for each patient using all available heart cycles. These averaged TFRs
were then used to create the mean and standard deviation TFRs in Figure 26.
Distinct areas in the standard deviation plots are clearly areas where there are great
differences between patients (this might be a hint on which time-frequency areas
that are stable to use in a murmur classifier). From the figure, it is obvious that
there is great interpatient variability. This is unfortunate when the aim is to
classify patients into various categories; ideally all patients with the same disease
behave similarly. Focusing on differences between diseases, it can be seen that the
murmur in MI has frequency content up to about 160 Hz and that the murmur
doesn’t change much over systole (holosystolic). The murmur in AS on the other
hand seems very unstable. Since the murmur in AS is crescendo-decrescendo
shaped where the peak depends on the severity of the stenosis, it makes sense that
the mean of several AS murmurs deviates between patients. The frequency content
reaches up to about 180 Hz (as can be seen in the standard deviation plot). Finally,
PM are known to be of low frequency, which is verified by the figure.
The TFR technique is a valuable tool when visually inspecting signals, but the
amount of information is immense. In a classification setting we need to find a
compact representation consisting of a manageable number of informative
features. Paper III describes a number of methods able to quantify time-frequency
information.
36
AS MI PM
1.5 1.5 1.5
Shannon energy
Shannon energy
Shannon energy
S1 S2 S1 S2 S1 S2
1 1 1
0.5 0.5 0.5
0 0 0
0 5 10 0 5 10 0 5 10
Feature number Feature number Feature number
Figure 25. Mean value of the Shannon energy calculated at nine time instants in systole, the whiskers
show the standard deviation. Data obtained from paper III.
AS, mean MI, mean PM, mean

200 200 200
Frequency (Hz)
150 150 150
100 100 100
50 50 50
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1
AS, standard deviation MI, standard deviation PM, standard deviation
200 200 200
Frequency (Hz)
150 150 150
100 100 100
50 50 50
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1
Time Time Time
Figure 26. Mean (top) and standard deviation (bottom) TFRs (calculated by Stockwell’s transform)
of aortic stenosis, mitral insufficiency and physiological murmurs. The time scale was resampled to
2048 samples after calculating the TFR, and is here represented in arbitrary normalized units. Data
obtained from paper III.
4.2. Higher Order Statistics

Since methods based on second-order statistics do not take nonlinearity and non
Gaussianity into account, higher order statistics might provide more information
about the phonocardiographic signal. Bispectra for signals from various systolic
murmurs are presented in Figure 27. First of all it can be seen that the bispectrum
is not zero, as it would be if the signals were Gaussian. Secondly, the main
frequency content is well below 300 Hz and exhibits distinct peaks. Thirdly,
considerable phase coupling exists between different frequencies. Finally it is also
seen that the patterns revealed in the bispectra differ between various pathologies.
It has previously been indicated that phonocardiographic signals are non Gaussian
37
[42, 43], but it has not been explicitly stated that this is the case. When performing
Hinich's Gaussianity test on each heart cycle in the data from paper III, it turns out
that each and every one of the 445 heart cycles have zero skewness with
probability <0.05. This strongly suggests that the data are non Gaussian (nonzero
skewness) and that investigations of the higher-order statistics of
phonocardiographic signals are relevant. Similarly, a hypothesis regarding
linearity could be rejected using Hinich's linearity test (for a nonlinear process, an
estimated statistic may be expected to be much larger than a theoretical statistic,
and in this case the estimated value is, on average, 3.4 times larger). This
motivates the use of nonlinear techniques in the two subsequent sections, 0-4.3.
Both Hinich’s tests are described in [44].
Normal Aortic stenosis Pulmonary stenosis
−100 −100 −100

0 0 0
100 100 100
−100 0 100 −100 0 100 −100 0 100

Mitral insufficiency Mitral valve prolapse Tricuspid insufficiency
−100 −100 −100

0 0 0
100 100 100
−100 0 100 −100 0 100 −100 0 100

Ventricular septal defect Atrial septal defect
−100 −100
0 0
100 100
−100 0 100 −100 0 100
Figure 27. Examples of bispectra from one heart cycle in different heart diseases. One heart cycle
here roughly corresponds start S1 to stop S2. All axes represent frequency in Hz.
The bispectra in Figure 27 are all very nice to look at, but we need to quantify
them somehow. To make the number of quantifying units manageable, the
bispectrum can be discretized [45], see Figure 28. Due to symmetry, it is enough
to investigate the first nonredundant region [33]. Using data from paper III, box
and whisker plots were derived for these 16 features, see Figure 29 are presented
in. Unfortunately the features overlap and are more or less useless for
classification purposes (t-tests show significant differences (p<0.05) for feature 2
between AS PM, feature 3 between AS MI and AS PM, feature 7-8, 11
between MI PM). The bispectrum is however a useful tool and Figure 27 does
reveal a lot of information. If nothing more, it could be used as a visualisation
technique to support the physician’s decision.
38
There are distinct differences between the various heart valve diseases in Figure
27. Obviously these differences are lost in the discretization when attempting to
reduce the information into a manageable feature set. A different approach is thus
needed to extract this information. A few ideas are Gaussian mixture models or
perhaps some parametric models like the non Gaussian AR model, but these issues
are left for future studies.
(a) (b)
300 150
200
100 100 14 15
Frequency (Hz)
Frequency (Hz)
0 13 16
−100 50 5 9
2 3 6 7 10 11
−200
1 4 8 12
−300 0
−200 0 200 0 50 100 150 200 250
Frequency (Hz) Frequency (Hz)
Figure 28. Example of bispectrum from a patient with aortic stenosis. The different regions of the
bispectrum is plotted in (a) where the bold triangle shows the first non-redundant region. In (b) the
region of interest is highlighted. The smaller triangles indicate the 16 features obtained from the
bispectrum, where each feature is calculated as the mean intensity of each triangle.
Feature 1 Feature 2 Feature 3 Feature 4

8
10 6 6 4
4 4
5 2
2 2
0 0 0 0
AS MI PM AS MI PM AS MI PM AS MI PM
10 6
4 3
4
5 2
2 2 1
0 0 0 0
3 2
4 2
2
2 1 1
1
0 0 0 0
10 10 8 6
6 4
5 5 4
2 2
0 0 0 0
Figure 29. Box and whisker plots showing results from the bispectral analysis. The boxes have lines
at the lower quartile, median, and upper quartile values. The whiskers show the extent of the data.
Outliers (+) are data with values beyond the ends of the whiskers. Data obtained from paper III.
39
4.3. Reconstructed State Spaces

Phonocardiographic waveforms consist of a large variety of types, ranging from
impulses (snaps and clicks) through turbulence induced sounds (murmurs) to
nearly periodic oscillations (heart sounds). The transition between these types
could be described by switching between different linear models, but using a
nonlinear setting, such transitions occur naturally as bifurcations [46].
Hypothesising that the blood flow is a dynamic system which is observed via the
recorded phonocardiographic signal, then the reconstructed state space would be
an attempt to recreate the characteristics of the flow (compare with Figure 13).
Since turbulence is a nonlinear phenomenon with strong interaction between the
flow and the associated acoustic field [47], the theoretical foundation for the
hypothesis seems valid.
Before pursuing any attempts to use nonlinear analysis tools, one should execute
some tests to see whether the data really behave in a nonlinear fashion. Two such
tests were performed on the data in paper III; Hinich's linearity test (see section
4.2) and phase randomized surrogate data [23]. Both tests indicated nonlinearity
by rejecting the hypothesis of linearity.
(a) (b)
6
5 1
Mutual Information
4 0.8
Cao’s Method
3 0.6
2 0.4
1 0.2
0 0
0 100 200 300 400 500 2 4 6 8 10 12
Time delay (sample) Embedding dimension (d)
Figure 30. Average mutual information calculations are used to determine the time delay embedding
parameter (a). The first minimum of the mutual information function indicates a delay where the
signal contains little mutual information compared to a delayed version of itself (why the comb-
ination of the two provides as much information as possible). Cao’s method (b) is used to determine
the embedding dimension d. This method is similar to the common false nearest neighbour approach,
which make use of the fact that points are moved closer together in a reconstructed state space,
compared to the true state space, by folding. Data obtained from paper III.
Nonlinear analysis of measured data is generally based on the reconstruction of the

signal in a multidimensional state space. Proper embedding parameters were
calculated via mutual information and Cao’s method. The embedding dimension
was found to be d = 4, which can be seen by the clearly defined knee in Figure 30.
Determination of the delay parameter was however less obvious. The mean value
of first minima in the mutual information function was = 233 ± 72 samples.
40
Since roughly half of the patients had a minimum in the vicinity of = 150, while
the other half lacked an obvious minimum in the range = 1…500 samples, was
set to 150. These routines should not be used on nonstationary data, and to
minimize the damage, these values were determined on data consisting of only
murmur data. En example of a phonocardiographic signal embedded in three
dimensions is given in Figure 31. The heart sounds are clearly encircling the more
complex murmur.
4
2
0
−2
−4
0 −4
−2
0
2
−5 4
6
Figure 31. Example of embedded heart sound with d = 3 and = 200. Heart sounds (S1 and S2) are
plotted in red and the murmur (AS) in blue. The small bumps in the trajectory are due to
concatenation of segments. Data obtained from paper III.
4.3.1. Quantifying the reconstructed state space

Quantifying a four-dimensional phase portrait is not easy. There are a few
common dynamical invariants that can be used. One of them is the fractal
dimension which will be used in section 4.4. Another example of an invariant
measure is the largest Lyapunov exponent. Disadvantages with invariant measures
are that they are not sensitive to initial conditions or smooth transformations of the
space. Another approach for quantification of the reconstructed state space is by
direct modelling of its statistical distribution. A parametric model could be
constituted by a Gaussian mixture model (GMM). In paper III, a GMM with five
mixtures, see Figure 32, was fitted to the reconstructed state space using the
Expectation-Maximization (EM) algorithm.
The centres of the mixtures and the eigenvalues of their covariance matrices can
be used as a rather compact representation of the trajectory in the reconstructed
state space. However, the five mixtures used here to describe the four-dimensional
geometry require 40 parameters. Using only five mixtures to estimate the density
function of the trajectory is probably far from enough, but the number of
parameters increases rapidly with the number of mixtures. Due to the difficulty of
giving an overall summary of three groups (AS, MI and PM) and 40 parameters,
these results are omitted in this section. They are however used as features for
murmur classification in section 5.4 and in paper III.
41
y(n+150)
−2
−4
−6
−6 −4 −2 0 2 4
y(n)
Figure 32. A reconstructed state space (d = 2, = 150) of the systolic period from a patient with aortic
stenosis. The red ellipses symbolize a Gaussian mixture model with five mixtures. Note that d = 2 is
not enough to unfold the trajectory.
4.3.2. Recurrence time statistics

In previous sections, higher dimensional state spaces have been visualized by
projection into lower subspaces. Recurrence plots were introduced to avoid this
procedure by visualizing high dimensional trajectories through a two-dimensional
representation of its recurrences [28]. An example of recurrence plots is presented
in Figure 33. Figure 34 shows the RQA results using the data in paper III. It is
clear that several of these parameters are good at separating physiological
murmurs from pathological murmurs (determinism, longest diagonal line, longest
vertical line and trapping time). Unfortunately, the results obtained from AS and
MI are heavily overlapping.
Normal AS
0.1
Amplitude
Amplitude
0.1
0 0
−0.1 −0.1
500 1000 1500 500 1000 1500

Time Time
200 200
400 400
600 600
Time
Time
800 800
1000 1000
1200 1200
1400 1400
500 1000 1500 500 1000 1500
Time Time
Figure 33. Example of recurrence plots for a normal phonocardiographic signal and for an AS case.
The interpretation of recurrence plots was briefly explained in section 3.3.2.
42
Recurrence Rate Determinism Average Length of Diag. Lines Longest Diagonal Line
9 1500
0.9
0.1 8
0.8 1000
7
0.08 0.7
6 500
0.06 0.6
Shannon Entropy Laminarity Trapping Time Longest Vertical Line
0.06 12
0.95
80
0.9 10
0.04
0.85
8 60
0.02 0.8
0.75 6 40
Recurrence Time of First Kind Recurrence Time of Second Kind

14
120
12 100
10 80
60
8
40
AS MI PM AS MI PM
Figure 34. Box and whisker plots showing results from the recurrence quantification analysis. The
boxes have lines at the lower quartile, median, and upper quartile values. The whiskers show the
extent of the data. Outliers (+) are data with values beyond the ends of the whiskers. Data obtained
from paper III.
4.4. Fractal Dimension

Theories attempting to explain turbulence forecast the existence of eddies (vortices
with a characteristic size) at multiple scales [48], and this multiscale structure of
turbulence can in some cases be quantified by fractals. It has therefore been
suggested that turbulent flow is fractal in nature [48].
As illustrated in section 2.3, Figure 5, in vivo measures obtained with the particle
trace technique show that blood flow through the normal heart contains large
vortices. The resolution of the particle trace technique does not allow a thorough
investigation on a wide range of scales, but model studies show that the interaction
with the ventricle wall results in the generation of smaller secondary vortices
which may in turn interact with the ventricle wall and thus creating even smaller
vortices [49]. However, flow in the normal heart is predominantly laminar, even
though it may involve vortices on a large scale. This behaviour fits the fractal
theory with self similarity across scale, but only for large scales.
Stenosed or leaking valves give rise to turbulent flow in the heart or in the great
vessels leaving the heart. In turbulent flow, unsteady vortices appear on many
scales and interact with each other. Numerical studies on stenotic tube flow show
that the positions at which vortices are initiated, their size, and their life span are a
function of the Reynolds number [50]. There is thus a link connecting the
43
Reynolds number to the degree of turbulence which in turn is connected to fractal

behaviour. It has been shown that the smallest length scales of turbulence is about
three times larger than the size of a red blood cell [51], so self similarity can be
found from very small scales up to large scales bounded by the size of the
ventricle. Finally, the last piece of the puzzle is provided by the strong interaction
between the flow and its induced sound field [47]. It is thus reasonable to believe
that the turbulence can be quantified using the fractal dimension of the measured
acoustic signal.
Considering the mere waveform of bioacoustic time series, it appears that these
signals possess valid characteristics for pursuing fractal dimension calculations:
• The signals do not self-cross.
• The waveform is often self-affine, i.e. in order to scale the signal, a
different scaling factor is required for each axis. In physical systems, this
property is not strict but probabilistic, and there are minimum and
maximum scaling limits (depending on the accuracy of the measurement,
the sampling resolution etc.).
• The waveform exhibits clear quasiperiodicity (heart beats and breathing
specifically).
• The power spectral density is broad-band.
An example is given in Figure 35, where the acoustic waveform from a patient
with aortic stenosis is plotted along with its fractal dimension calculated over time.
Comparing heart sounds, murmurs and background noise, HS have a curtain
structure while murmurs are more complex and noise has no structure at all. It can
also be seen that the fractal dimension of the murmur is rather constant despite the
large amplitude variations (crescendo-decrescendo) in the time domain.
(a)
0.2
Amplitude (V)
0.1
0
−0.1
−0.2
200 400 600 800 1000 1200 1400 1600 1800
(b)
2.2
2
VFD
1.8
1.6
1.4
1.2
200 400 600 800 1000 1200 1400 1600 1800
Time (ms)
Figure 35. Example of aortic stenosis (a) showing the variance fractal dimension plotted over time
(b).
44
Variance fractal dimension for the data in paper III is shown in Figure 36. The
nine chosen instants were selected at times analogous to the Shannon energy plot
in Figure 25. Again, the variance is rather large, especially in the AS case. VFD of
the different murmurs are however quite well separated in their mean; AS = 1.202,
MI = 1.037 and PM = 1.336 (calculated as mean values for feature number 4-6 in
Figure 36, the values deviate from those in the figure due to normalization).
Hypothesis testing for the difference in mean between the groups (t-test) shows a
difference with significance p = 0.03, p = 0.07 and p = 0.01 when comparing
AS MI, AS PM and MI PM, respectively. Boxplots of the same data are
shown in Figure 37. Here the VFD was calculated using a concatenation of all S1
segments, all S2 segments and all murmur segments from each patient. Focusing
on the murmur, the trend from Figure 36 is recognised; MI has lowest dimension,
PM has highest dimension and AS is somewhere in between. The interpatient
variability is still a problem, especially in the AS case (probably due to the wide
range of mild to moderate AS).
AS MI PM
1.2 1.2 1.2

VFD
VFD
VFD
1 1 1
S1 S2 S1 S2 S1 S2
0.8 0.8 0.8
0 5 10 0 5 10 0 5 10
Feature number Feature number Feature number
Figure 36. Mean values of the variance fractal dimension at nine time instants in systole, the whiskers
show the standard deviation. The data were normalized so S1 had unit fractal dimension (for visual
appearance). Data obtained from paper III.
S1 Murmur S2
1.6 1.6 1.6
VFD (based on all data)
1.4 1.4 1.4
1.2 1.2 1.2
1 1 1
AS MI PM AS MI PM AS MI PM
Figure 37. Boxplots of the VFD when calculated for S1, murmur and S2 when concatenating all S1
data, all S2 data and all murmur data, respectively, within each patient. The boxes have lines at the
lower quartile, median, and upper quartile values. The whiskers show the extent of the data. Data
obtained from paper III.
45
5. Applications in Phonocardiographic
Signal Processing
“The modern age has a false sense of security because of the great
mass of data at its disposal. But the valid issue is the extent to which
people know how to form and master the material at their disposal.”
Johann Wolfgang von Goethe (1832)
This chapter presents some applications that make use of the knowledge gained in
previous chapters. Almost all phonocardiographic signal processing tasks are
dependent on accurate segmentation of the signal. The segmentation algorithms
that are described in 5.1 are later used in 5.2 - 5.4. Other treated applications
include detection of the third heart sound, denoising of lung sound signals and
classification or heart murmurs.
5.1. Segmentation of the Phonocardiographic Signal

Dividing the recorded signal into S1, systole, S2 and diastole is of great
importance in all phonocardiographic signal processing tasks. While this is
sometimes performed manually, several techniques are available to accomplish the
task automatically. A quite simple and robust way to do this is by ECG gating. The
ECG signal does usually have a better SNR than the sound signal, and numerous
algorithms exist for automatic detection of the rather distinct R-peak. Since it is
known that S1 follows shortly after the R wave in the ECG, segmentation into
heart cycles is easily obtained.
Using additional sensors is not an optimal approach in the intelligent stethoscope
setting, where simplicity is one of the key words. Present approaches for
segmentation of the phonocardiographic signal are often based on peak picking
after transformation into a domain where S1 and S2 are emphasized. Several
choices of this transformation has been presented; coefficients from an 8th order
AR model in a narrow sliding window [52], Shannon energy [15], wavelet
decomposition plus Shannon energy [53], homomorphic filtering [17], matching
pursuit [54] and fractal dimension trajectories [55, 56]. Typically, a threshold is
used to locate the heart sounds, without classification into S1 and S2. The actual
classification is done using interval statistics (the duration of systole is rather
47
constant and normally shorter than the duration of diastole, giving a binominal
distribution, see Figure 38). A problem with all of these approaches is that it is
hard to distinguish the heart sounds in noisy phonocardiographic signals. Noise in
this case could be heart murmurs, lung sounds and/or background noise such as
speech.
Systole
0.25
0.2
0.15
0.1
Diastole
0.05
0
0 200 400 600 800 1000 1200 1400
Time (ms)
Figure 38. Histogram showing the distribution of systolic and diastolic duration in a heart cycle
(based on 604 heart cycles from six subjects). It can be seen that the duration of systole is rather
constant and normally shorter than the duration of diastole, giving a binominal distribution. Data
obtained from paper I.
Different change detection methods have been employed for segmentation of heart
sounds (unpublished work). One-model, two-model as well as multiple-model
approaches were tried out, but neither of them was suitable for segmentation of
heart sounds. A reason could be that they were all based on a linear regression
framework, which was not able to separate changes due to heart sounds from
changes due to noise. Parts of paper I was devoted to a nonlinear change detection
method for the task of emphasizing S1 and S2 (as an alternative to the above
mentioned transformations and change detection approaches). Heart sounds have a
transient waveform that is superpositioned upon lung sounds and other
disturbances. Since the heart sounds and the noise originate from different sources,
they have different attractors, see Figure 39. These changes in signal dynamics can
be detected with the change detection scheme in section 3.3.2 (based on recurrence
times of the first kind, T1).
A sliding window was used to partition the phonocardiographic signal into
overlapping segments to obtain time resolution. T1 is plotted in Figure 40. Since
the application in paper I was to find and remove both S1 and S2, no attempts
were made to actually classify the two sounds. This method was used in both
paper I and paper II, and resulted in an error rate of 12.4 % and 0.5 %,
respectively. The big difference in detection accuracy depends on heavy breathing
in one of the provocation sequences in paper I.
48
Chapter 5. Applications in Phonocardiographic Signal Processing
Figure 39. State space trajectories (d = 3, = 12) of a sound signal with S1 and S2 cut out (a). In (b)
the whole signal including S1 and S2 is shown. The transition between the two attractors is reflected
in the recurrence time statistic, hence indicating when a heart sound is present.
(a) (b)
1
0.04
0.8
0.6
T1
0.12
ε
0.4
0.20 0.2
0
32 33 34 35 36 32 33 34 35 36
Time (s) Time (s)
Figure 40. An example showing how the recurrence time statistic indicate the location of heart
sounds. Note the obscuring noise (a deep breath) in the end of the signal. In (a) T1 is plotted over
time for various -values where the grey scale indicates the strength of T1. Superimposed in the
figure is the phonocardiographic signal (black waveform). T1( ) for one fixed -value is plotted in (b).
5.2. Finding S3
The third heart sound occurs normally in children but disappears with increasing
age. The sound can reappear in elderly persons and is clinically important because
of its established connection with heart failure [8, 57]. Compared to the task of
locating S1 and S2, finding S3 is harder due to its low amplitude, short duration
and low frequency.
Previous methods to detect S3 is limited to a matched wavelet approach, where the
mother wavelet was designed to have similar morphology as S3 [58, 59]. The idea
was to divide the signal into four frequency bands; 17, 35, 60 and 160 Hz. S1 and
49
S2 are present in all frequency bands while a potential S3 should be represented in

the three lower bands.
Paper II presents a novel technique based on the same method as the heart sound
locator in paper I. Recurrence points of the first kind, T1, are used to locate S1 and
S2 after which S3 is sought in time windows 100-300 ms after the two heart
sounds. To avoid the problems involved in discriminating between S1 and S2, S3
was sought for within the predetermined time window following both heart
sounds. As mentioned in section 3.3.2, T1 is more robust to noise while T2 is more
sensitive to changes in the signal. S3 is a very weak signal, so T2 statistics is the
better choice in this case.
Selecting the proper neighbourhood can be a problem. In Figure 41, T1 and T2 are
plotted for a range of -values to visualize the dependence on the neighbourhood
size. When looking for S1 and S2, a simple threshold can be used, but looking for
S3 (in T2) is somewhat harder. The chosen approach was to use a whole range of
-values to calculate a T2-matrix, see Figure 41d. The resulting 2D image can then
be converted to 1D by an edge detection algorithm (here implemented by lowpass
filtering and detection of the maximum value in each column). In the 1D signal,
occurrences of S3 can be found by looking for a maximum within the previously
defined time window. The detection rule could, for example, compare the
amplitude of the maximum and the amplitude of the base line level. In paper II this
rule states that the amplitude of the maximum should be one third larger than the
base line level. Comparing the described method with the matched wavelet
approach shows improved detection rate, 98 % compared to 93 %, at the expense
of more false detections, 7 % compared to 2 %. The comparison was based on data
from paper II consisting of ten children.
5.3. Filtering out Signal Components

Noise is a big problem in phonocardiography. The sensor, the sensor contact
surface, the patient’s position, the auscultation area, respiration phase and
background noise all influence the quality of the sound. In practice this means that
the recordings often contain noise such as friction rubs, rumbling sounds from the
stomach, respiratory sounds from the lungs and background noise from the clinical
environment. Most of these noise sources have frequency content in the same
range as the signal of interest.
A special case of noise cancellation techniques deals with removal of transient
noise. Potential use in a phonocardiographic setting is to remove disturbances such
as friction rubs, but in paper I we focus on a related field in respiratory sound
analysis, where the heart sounds themselves are the interfering noise.
50
(a)
0.5 S1 S2 S1 S2
Amplitude
S3 S3
0
−0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
(b)
0
ε
0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
(c)
1
T1(0.4)
0.5
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
(d)
0.1
ε
0.2
0.3
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
(e)
0.1
T2
0.2
0.3
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (s)
Figure 41. Example of a heart sound signal where S1, S2 and S3 are marked (a). T1, calculated for a
range of -values, is shown in (b) while a single T1 is shown in (c) for = 0.4. T1(0.4) is used to find S1
and S2. T2, calculated for a whole range of -values is shown in (d). An edge detection algorithm is
used to convert T2 to the 1D signal in (e) which is used to detect S3 (marked as arrows by the
detection algorithm).
There are many different methods available for heart sound cancellation from lung
sounds. Heart sounds and lung sounds have overlapping frequency spectra, and
even though high pass filtering is often employed to reduce the influence of heart
sounds, this results in loss of important signal information [60]. Previous
approaches to heart sound cancellation include wavelet based methods [60],
adaptive filtering techniques [61] and fourth-order statistics [62], all resulting in
reduced but still audible HS. Recent studies indicate that cutting out segments
containing HS followed by interpolation of the missing data yields promising
results [63, 64]. The method developed in paper I is based on work by Thomas et
al. [63, 64], but the used signal processing techniques are fundamentally different
and allow nonlinear behaviour in the lung sound signal. This is an important
difference since it has been indicated that lung sounds are indeed nonlinear [65-
68].
The method suggested in paper I uses the heart sound locator presented in section
5.1. The detections are simply cut out and the resulting gaps are filled with
predicted lung sound using the nonlinear prediction scheme described in section
51
3.4. Since the prediction error grows exponentially with prediction length [23],
both forward and backward prediction was used (hence dividing the missing
segment in two parts of half the size). To avoid discontinuities in the mid point,
the number of predicted points was allowed to exceed past half of the segment.
The two predictions were then merged in the time domain close to the midpoint at
an intersection where the slopes were similar.
Figure 42. Example of a recorded lung sound signal with heart sounds present (a) and reconstructed
lung sounds with heart sounds removed (b). The bars indicate heart sound detections. A zoomed in
version showing the predicted lung sound (solid) and lung sound including heart sounds (dashed), is
shown in (c).
The results are a bit hard to evaluate since the actual lung sound is unknown in the
segments that are predicted. However, the waveform similarity between predicted
segments and actual lung sound data is very high with a cross-correlation index of
CCI = 0.997±0.004. The spectral difference was 0.34±0.25 dB/Hz, 0.50±0.33
dB/Hz, 0.46±0.35 dB/Hz and 0.94±0.64 dB/Hz in the frequency bands 20 – 40 Hz,
40 – 70 Hz, 70 – 150 Hz and 150 – 300 Hz, respectively. Since the main objective
of the method was to give auditory high-quality results, a simple complementary
listening test was performed by a skilled primary health care physician. The
impression was that most heart sounds had been successfully replaced, but that
some predictions had a slightly higher pitch than pure lung sounds. An example of
heart sound cancellation is illustrated in Figure 42.
5.4. Classification of Murmurs

Common for all classification tasks is the importance of appropriate features, i.e.
measures that retain similarities within classes while revealing differences between
classes. Chapter 4 described characteristics of the phonocardiographic signal in
52
different domains. Although not mentioned explicitly, the underlying goal of the
characterization was to find data representations with distinct differences between
heart murmurs. This section is devoted to the problem of extracting features from
these data representations and selecting those features with best discriminative
power. Finally, the selected features are used to classify three different heart
murmurs; aortic stenosis, mitral insufficiency and physiological murmurs.
5.4.1. Feature extraction

Feature extraction is about quantifying available information into a few descriptive
measures. The different ways data were viewed upon in chapter 4 should hence be
summarized in a few informative features. In phonocardiographic classification,
the features are derived on a heart cycle basis. Knowledge about the accurate
timing of events in the heart cycle is thus of great importance. Segmentation into
the first heart sound (S1), systole, the second heart sound (S2) and diastole were
discussed in section 5.1. In paper III the ECG gating technique was used, mostly
because of its superior robustness when the phonocardiographic signal is noisy.
Once the features have been extracted, averaging each feature over available heart
cycles reduces the influence of noise.
A large amount of features were extracted in paper III. Some features provided
information about the phonocardiographic signal’s time varying behaviour.
Shannon energy was used to measure intensity and a wavelet detail was used to
measure intensity in a certain frequency interval. The variance fractal dimension
was used to measure the time progress of the signal’s complexity. An example of
these features is shown in Figure 43.
Stockwell’s TFR was used as a foundation when investigating how the frequency
content varied over time. These representations are information rich, but they also
contain a lot of data. Assuming a sample rate of 44.1 kHz and that one heart cycle
is one second long, then each TFR matrix will contain 1.9448·109 samples.
Clearly, considerable data reduction is required. The easiest way to do this is by
down-sampling as in Figure 44. A perhaps more refined method for data reduction
of TFR matrices use singular value decomposition. More details about this
approach can be found in paper III.
These measures are based on second order statistics. Expanding the view to third
order statistics, a measure of the non Gaussianity and nonlinearity of the signal can
be obtained. Similarly to TFR down-sampling, the same technique can be applied
to the bispectra, see Figure 28. Finally, features from the reconstructed state space
can be derived. This procedure follows section 4.3 and uses the RQA measures
and the GMM measures as features. These last additional features incorporate
nonlinear behaviour into the description of the signal.
53
(a)
Waveform
0.5
0
−0.5
5.4 5.6 5.8 6 6.2
(b)
Shannon energy
0.08
0.06
0.04
0.02
5.4 5.6 5.8 6 6.2

(c)
WT detail
0.02
0
−0.02
5.4 5.6 5.8 6 6.2
(d)
1.2
VFD
1.1
1
5.4 5.6 5.8 6 6.2
Time (s)
Figure 43. An example showing one heart cycle from a patient with aortic stenosis (a). In (b) the
signal’s envelope has been extracted (Shannon energy), the rings indicate the selected features. A
wavelet detail is illustrated in (c), where the vertical lines are time markers equidistantly distributed
over the region of interest. The absolute sum between each marker constitutes feature values. In (d)
the variance fractal dimension trajectory is plotted together with the selected features marked as
rings.
(a) (b)
150 4.5
4 13 14 15 16
3.5
100
Frequency (Hz)
3 9 10 11 12
2.5
2 5 6 7 8
50
1.5
1 1 2 3 4
0 0.5
5.3 5.4 5.5 5.6 5.7 5.8 1 2 3 4
Time (s)
Figure 44. Time frequency representation of one systolic heart beat from a patient with aortic
stenosis (a), S1 can be seen at 5.3 s and S2 at 5.8 seconds. In (b) the same data has been discretized
into a 4x4 map of features.
54
5.4.2. Finding relevant features

A large number of features were described in the previous section. However, too
many features often result in classifiers with low generality [69]. Actually, there
are many potential benefits in reducing the number of features; facilitating data
visualization, reducing the measurement and storage requirements, reducing
training and utilization times and defying the curse of dimensionality [70].
Auscultation is an old science where a lot of information and experience have been
gathered over the years. This domain knowledge should be incorporated in the
classification task, and as a matter of fact, it is already included in the features.
Many of the features in section 5.4.1 are based on the changes in timing, intensity
and pitch that physicians use to separate normal from abnormal
phonocardiographic signals. The first step in any feature selection process should
be to incorporate domain knowledge, but as this was done in the feature extraction
process, this important step can be omitted.
Scalar feature selection means that each feature is treated individually. A scoring
function is defined, and its outcome indicates the predictive power of the feature.
This way, all features can be ranked in decreasing order, and the best ones are
selected for the classification task. The scoring function could, for example, be the
distance from the feature to the centre of the distribution of the class it is supposed
to belong to. Selecting the most (individually) relevant features is usually
suboptimal for building a predictor though, particularly if the selected features are
correlated and thus contain redundant information [70].
A problem with scalar feature selection is that is does not account for
combinations of features that together have great predictive power. An optimal
selection of features requires an exhaustive search over all features, but this is
practically infeasible. Instead suboptimal search algorithms are employed, many of
which uses greedy hill climbing (hill climbing is a search algorithm where the
current path is extended with a successor node which is closer to the solution than
the end of the current path). A possible subset of features is then evaluated, and
other features are successively added or removed from this set to see if an
improvement can be achieved [70]. A simple way to do this is to start with one
feature (the one with highest ranking according to scalar feature selection), say x1.
Expand the set to contain two features by forming all possible pairs, say {x1, x2},
{x1, x3}, {x1, x4}. The pair that maximizes some class separability criterion is
selected as the new feature subset. More features are then progressively added into
larger and larger subsets until the desired number of features is reached. This
method is often used when the size of the final subset is supposed to be small
compared to the total amount of features. If the final subset is supposed to be
large, then all features could be included in a preliminary subset which is
progressively reduced. These methods are called sequential forward selection and
sequential backward selection. A common drawback for both of these is that once
a feature is included there is no way of getting rid of it (and vice versa in backward
selection). Pudil’s sequential floating forward selection is a workaround to this
55
problem, allowing features to be both included and excluded several times [71]. A
flow chart describing the algorithm is presented in Figure 45.
Let k = 0
Apply one step of

the sequential forward
selection algorithm
Let k = k + 1
Desired
Yes Let k = k -1
Stop number of features
reached?
No
Conditionally exclude Leave out the

Return the conditionally conditionally excluded
excluded feature one feature found by
applying one step of feature
the sequential backward
selection algorithm
Is this
No the best (k-1)- Yes
subset so far?
Figure 45. Flow chart of Pudil’s sequential floating forward selection method, where k is the current
number of features in the subset.
In paper III, Pudil’s sequential floating forward selection method was used to
reduce the number of features from 213 to 14. Inclusion or rejection of features
was based on the error estimate of a 1-nearest neighbour leave-one-out classifier
where the performance criterion equals 1 – the estimation error. The number of
features used in the final set was chosen to maximize the performance criterion
while keeping the number of features as low as possible.
14 features were selected, see Table 3, and the resulting set was denoted the SFFS
subset. Bearing in mind that the investigated murmurs are aortic stenosis, mitral
insufficiency and physiological murmurs, the selected features are actually very
reasonable. A wavelet detail represents the end of systole, where it can be used to
separate holosystolic mitral insufficiency murmurs from physiological murmurs
and aortic stenosis murmurs which are of crescendo-decrescendo shape. Three
Shannon energy measures represent the signal’s intensity in mid systole, thereby
describing the shape of the murmur in the time domain. A fractal dimension
measure represents the complexity of the murmur in relation to the heart sounds.
This measure can be seen as the amplitude normalized complexity of the murmur.
Another fractal dimension measure, located at S1, represents the change of S1 that
is associated with mitral insufficiency. Remaining features are a bit hard to explain
in a physiologically meaningful way.
56
Table 3. A brief summary of the features selected by Pudil’s sequential forward feature selection
method, the SFFS subset.
Wavelet detail: One feature representing the end of systole
Wavelet entropy: One feature describing the information content in the high frequency
range.
Shannon energy: Three features in mid systole able to describe the shape and intensity of
the murmur and one feature after S2 revealing the noise level.
Stockwell’s TFR: Two features giving a collected view of the low frequency content over
the heart cycle.
Bispectrum: One feature indicating phase coupling and frequency content for low
frequencies.
Reconstructed state space: Three features describing the width of the Gaussian mixture
model (probably located in the part of state space where the murmur lives), two of these
belong to the largest mixture.
Variance fractal dimension: Two features, one giving the amplitude normalized
complexity of the murmur and the other describing S1.
5.4.3. Classifying murmurs

The work in this thesis does not intend to cover the classification step. However,
to show the abilities of the feature set derived in the last section, a simple, off-the-
shelf classifier was however employed. A fully connected feed-forward neural
network was set up, with logarithmic sigmoid transfer functions and biased values
throughout. The number of input units was set to the nearest larger integer of the
square root of the number of features in the set, the number of units in the hidden
layer was set to three and the number of output units was set to two. The target
values were 00 (mitral insufficiency), 01 (aortic stenosis) or 10 (physiological
murmur). Each output from the network was thresholded at 0.5 and compared to
the results from a clinical echocardiography investigation. A leave-one-out
approach was used for training and testing due to the limited amount of patients.
Three additional subsets were used for comparison with the SFFS subset. The first
set consisted of the Shannon energy features, the second set consisted of the
wavelet detail features and was adapted from [17, 72] and the third set consisted of
a TFR decimated to 4x4 values (adapted from [73]).
Confusion matrices showing the classification results for the four tested feature
subsets are presented in Table 4. The percentage of correct classifications was
58%, 44%, 53% and 86% for Shannon energy, Wavelet detail, TFR and the SFFS
subset, respectively. If all pathological cases (mitral insufficiency and aortic
stenosis) are treated as one group, sensitivity and specificity can be specified. In
this setting, the sensitivity was 90%, 90%, 93% and 93% and the specificity was
28%, 14%, 57% and 100%, respectively. The number of patients with valve
57
pathology that were erroneously classified as physiological was comparable for all
feature subsets; 10%, 10%, 7% and 7%, respectively.
Table 4. Confusion matrices showing the classification results from four different feature subsets.
Target groups are presented horizontally while the predicted groups are presented vertically. Each
number represents number of patients (total number of patients in the study is 36).
Shannon Energy Wavelet detail TFR features SFFS
AS 17 4 4 15 5 2 14 4 3 19 1 0
MI 3 2 1 6 0 4 8 1 0 2 5 0
PM 3 0 2 2 1 1 1 1 4 2 0 7
58
6. Discussion
“Doubt is not a pleasant condition,
but certainty is absurd.”
Voltaire (1694 - 1778)
The focus of this thesis has been to investigate and develop new tools to facilitate
physicians’ daily work. Evaluation of patients with heart disease is a complex task,
where auscultation provides one piece of the puzzle. Therefore, our intelligent
stethoscope is not to be seen as a tool capable of replacing clinicians, but rather as
a provider of quantitative decision support. The stethoscope’s main usage will be
in the primary health care, when deciding who requires special care.
Consequently, it should neither be seen as a replacement of more advanced
techniques such as echocardiography.
The main tasks for the intelligent stethoscope are to improve sound quality, to
emphasize weak or abnormal events (such as reverse splitting of S2) and to
distinguish different heart murmurs from each other. In a small pilot study from
2002, interviews with nine primary health care physicians revealed that the most
interesting task for an intelligent stethoscope was classification of heart murmurs,
especially to distinguish physiological murmurs from pathological murmurs.
A danger with projects such as the intelligent stethoscope is that technology is
sometimes introduced for the sake of technology. Heart sound cancellation from
lung sounds (paper I) tends in this direction, see section 6.1. Detection of the third
heart sound is somewhat different since S3 can be very difficult to hear. Notifying
the physician that a third heart sound is present could thus be of great value (paper
II). When it comes to decision support and classification (paper III), the intended
use of the system becomes an important issue, see section 6.2.
6.1. Context of the Papers

The research on signal processing of heart sound recordings has been extensive
[74, 75], and several authors have investigated the possibility to automatically
classify cardiac murmurs. The contribution of paper III is partly a thorough survey
of features available for murmur classification and partly the addition of several
59
new features. The survey is based on features from the literature, ranging from
time domain characteristics [76-78], spectral characteristics [79-81] and frequency
representations with time resolution [17, 72, 73, 81-83]. The main contribution
compared to previous works is the incorporation of nonlinear and chaos based
features, a source of information that has not previously been explored in the
context of heart murmur classification.
There are a number of methods available for heart sound cancellation from lung
sound recordings. This is quite interesting since heart sound cancellation have
limited clinical use (physicians are able to more or less ignore heart sounds while
tuning in on lung sounds during auscultation). The problem at hand is a very
intriguing engineering problem though, and this is probably one reason for its
popularity. A justification of all these methods is that automatic classifiers seem to
be confused by the heart sounds. When trying to separate different lung diseases
based on lungs sounds, results tend to improve after removal of the heart sounds.
Some of the previous approaches to heart sound cancellation include wavelet
based methods [60], adaptive filtering techniques [61] and fourth-order statistics
[62], all resulting in reduced but still audible HS. The contribution of paper I is
that nonlinear behaviour in the lung sound signal is taken into account while
cancelling out the heart sounds. This is an important extension since it has been
indicted that lung sounds are indeed nonlinear [65-68].
In contrast to the other two papers, there is not much work available on detection
of the third heart sound. A matched wavelet approach giving good results has
previously been developed in our research group [58, 59]. The method presented
in paper II is based on a change detection scheme developed to find weak transient
components in signals. Compared to the wavelet approach, the change detection
method finds more third heart sounds at the expense of more false detections. It is
thus a complement rather than a replacement of the wavelet method, where the
new approach could be used to find the third heart sounds while the wavelet
approach could be used to exclude false detections.
6.2. Patients and Data Sets

Since this project is about developing of a new tool, a full clinical trial was not the
goal when assessing the developed methods. Nevertheless, to emulate the
environment where the intelligent stethoscope most likely will be used, most data
has been recorded in a clinical environment (paper I and III). Below is a run-
through of limitations implied by the selected study groups in paper I-III.
Paper I: This study was performed on six healthy male subjects (aged 28.5 ± 3.7
years). Limitations implied by the study population are that no adventitious lung
sounds are present. Future work should thus contain patients with various
pulmonary diseases. These limitations result in a few implications:
1. How would the heart sound detector perform in the presence of
adventitious sounds? In paper I, it is indicated that the detector performs
quite well even when the signal is obscured by heavy breathing. The
60
Chapter 6. Discussion
attractor of wheezing sounds has a similar morphology as the attractor of

heart sounds, and it is not inconceivable that wheezes will disturb the
detection algorithm. In paper I, the measurements contained several
friction rubs and one test subject had a distinct third heart sound. Most of
these friction rubs and third heart sounds were detected and removed along
with the other heart sounds. Explosive lung sounds like crackles will
probably be marked by the method as well. By including extra criteria,
such as interval statistics or the degree of impulsiveness, it is possible that
these false detections could be avoided.
2. How would the prediction scheme perform in the presence of adventitious
lung sounds? Wheezing sounds have a periodic structure and their attractor
is well defined (an example can be found in [84]). This would pose no
problem to the predictor. Crackles on the other hand occurs rather
stochastic (actually they come in avalanches following a fractal behaviour
[85]), and are thus harder to predict. This will most likely fail.
Paper II: Ten healthy children were used in this study (male = 5, female = 5, aged
5-13 years), mostly because third heart sounds with high signal quality are
common in this group. Investigation of pathological phonocardiograms is left for
future studies. In a validation study, patients with heart failure should be chosen.
After all, the intended use of the algorithm is to use S3 as a marker of heart failure.
Paper III: In total, 36 patients (19 male, 17 female, ages 69 ± 14 years, all with
native heart valves) were enrolled in the study. The prerequisite for including a
patient in the study was that a primary health care physician had identified a
murmur and that the patient had been sent to the cardiology clinic for further
investigations. As stated previously, physicians ask for a method able to separate
physiological murmurs from pathological murmurs. The usage of such a tool is
however more applicable in a young population. In a material with data from
elderly patients, it is more interesting to determine the actual source of the
murmur. This is exemplified in the following heart murmur classification
scenarios, where the demands on the system turn out to be quite different from
each other.
1. Screening in a young population. Physiological murmurs are very common
in children, and methods able to separate physiological murmurs from
pathological murmurs would be of great value. However, the performance
requirements are extremely high. For example, physiological murmurs have
to be separated from even very mild aortic stenosis due to the latter’s
implications on choice of profession, insurance issues and whether follow-
ups of a possible disease are needed or not. Due to the consequences of an
incorrect diagnosis, the tolerance for false positives and negatives is low.
The performance requirements of the system are huge and the interesting
question it all comes down to is: does the phonocardiographic signal
contain enough information?
61
2. Measuring the degree of stenosis in the elderly. Pathological changes in the

aortic valves are common in the elderly. Usually this change has little
physiological importance since the stenosis is mild. However, it is
important to find those patients who really have a significant narrowing of
the valve opening; partly because surgical correction improves the
prognosis of these patients and partly because, when surgery is out of the
question, some medications should be avoided. The classification task
would then be to measure the degree of the stenosis and decide whether the
stenosis is mild or moderate/severe. This scenario is easier so solve
compared to the previous scenario since the grey area between a
physiological murmur and a mild stenosis is not of interest.
6.2.1. Measurement noise

The data sets used in this thesis where all more or less affected by noise, and as a
matter of fact, the data in paper I and III were actually very noisy. In practice this
means that the recordings contained friction rubs, rumbling sounds from the
stomach, breathing sounds from the lungs (a necessity in paper I and an obstacle in
paper III) and background noise from the clinical environment. All of these
influence the recorded sound in a negative way, but a perhaps bigger problem is
handling the stethoscope. This is a setback since the whole idea behind the
intelligent stethoscope is that it should be easy to use (perhaps even in home care
by the patient herself). Firm application of the sensor cannot be stressed enough to
achieve high quality recordings.
6.3. Methodology
Most methods used in this thesis suffer from high computational burden. This is a
problem since the software is supposed to be implemented in a portable
stethoscope, preferably in real time. It is however difficult to assess the actual
performance limitations of the used methods because they were never designed to
be quick or efficient. A number of potential speed-ups come to mind.
• Matlab was used to implement all algorithms, but using a lower level
language would increase performance.
• Fast nearest neighbour routines are available, but currently a very simple
search routine is used.
• A sliding window approach is often used to gain time resolution. The
reconstructed state space is nearly identical between iterations due to the
overlap between segments, and this fact is not exploited.
A fundamentally different bottle-neck is the fact that the some calculations are
non-causal. For instance, many of the features in paper III were derived as
averages over all available heart cycles. In most cases this could be dealt with by
only using old data. Accumulated statistics could then be used to increase the
accuracy of the output as more data become available.
62
Significance tests could have been performed to statistically verify if there were
any differences between the different groups in paper III. The number of tests
would have been large, trying to separate three groups from each other in 213
cases (the total number of features). There are at least two reasons why these tests
were not performed. Firstly, when performing a great number of statistical tests,
the probability of getting significant differences by chance is rather high and
secondly, variables that are useless by themselves might be useful in combination
with others.
The greatest problems when using chaos based signal analysis tools are that the
results are almost always open for interpretation, that nearly noise free data is
required and that the amount of data should be large. Phonocardiographic data is
rather cyclo-stationary than nonstationary, so by concatenating stationary
segments, large data sets can be obtained. In this thesis, these segments were
simply concatenated in the time domain, while a better approach would have been
to append the reconstructed state space matrices to each other. An extra flag would
then be appended to each coordinate, keeping track of the last coordinate in each
segment. This way, false neighbours due to the concatenation procedure can be
excluded from the calculations. This addendum would have impact on all methods
where a reconstructed state space is constructed from concatenated data
(prediction in paper I and certain features in paper III).
Estimation of fractal dimension characteristics should also be based on large
enough data sets [86]. This implies a trade-off between time resolution and
accuracy in the estimation of the time dependent fractal dimension (similar to the
uncertainty principle when calculating time-frequency representations). As the
investigated signal segment does not possess self-similarity over an infinite range
of scales, the self-similar properties of the segment are lost if the sliding window is
too short. Similarly, if the window size is set too long, the different characteristics
of consecutive signal segments will be blurred together. Another reason for not
using too short windows is that the number of signal amplitude increments used to
calculate the variance in the variance fractal dimension algorithm must be greater
than 30 to be statistically valid [24].
Ideally, four different data sets should have been used in paper III. One set for
selecting analysis methods able to extract features, one set for selecting the most
important features, one set for training the classifier and a final set for validation.
This set-up requires an extensive database of phonocardiographic signals.
Unfortunately, this was unattainable within the scope of this project. A good
compromise when the number of data is limited is the leave-one-out approach.
Here the training is performed on N-1 samples and the test is carried out on the
excluded sample. This is repeated till all samples have been tested. The results
when using separate training and testing sets compared to the leave-one-out
approach are very similar [69]. However, the independence between the training
and testing sets are compromised since the same sets are used for both feature
selection and training of the classifier.
63
Selecting descriptive measures to demonstrate the performance of an algorithm

can be cumbersome. In paper I and II, receiver operating curves (ROC) would
have been informative to describe the dependence of thresholds. This was not
feasible since all detections had to be confirmed or rejected manually.
In paper I, the accuracy of the predictions was difficult to verify since the true
signal was not known. The spectral difference between the reconstructed lung
sound signal and a lung sound signal with cut-out heart sounds was small.
However, comparing spectral densities might be questionable as phase information
is lost. Another attempt is to compare the waveform similarity between predicted
segments and actual lung sound data. The results from this measure were very
good, but this is not surprising since the prediction scheme exploits that
trajectories in state space share the same waveform characteristics in time domain
(i.e. the prediction tries to reproduce past parts of the time series). Since both of
these measures are questionable, a simple complementary listening test was
performed by a physician. However, such tests are certainly not quantitative.
6.4. Future Work

6.4.1. Clinical validation
The methods presented in paper I-III have not been clinically evaluated, validation
on a large number of patients are thus necessary. The number of patients is
especially low in paper III, where the methods used for feature selection as well as
classification requires a lot of data to give reliable results. When recording a new
material, the clinical need and the selected patient groups should be re-evaluated
together with medical personnel. The two scenarios in section 6.2 are typical
examples of interesting study groups.
6.4.2. Multi-sensor approach

The magnitude of different components in the phonocardiographic signal varies
with the measurement location. For instance, listening over the apex, S1 is louder
than S2. Also, the location where a heart murmur is best heard often indicates its
origin. By using multiple sensors in parallel, this difference in intensity could be
used as a parameter in a classification system. Further uses could be to derive time
differences between the different signals and, using this information, calculate an
estimate of the location of the event. For instance, using S1 and S2 as reference
locations, the murmur location could be of diagnostic value. A third possible use
of multiple sensors is to use several sources of the signal when creating the
reconstructed state space. This would probably give better resistance to
measurement noise and, above all, a better embedding of the signal.
The major drawback with multiple sensors is that the simplicity of the stethoscope
would be afflicted. Nevertheless, if additional sensors are to be used, there are no
reasons not to use conceptually different techniques such as ECG and Doppler
ultrasound.
64
6.4.3. Dimension reduction

The recorded phonocardiographic signal is one-dimensional, and this signal can be
unfolded into a high-dimensional space (section 3.2). This is all very nice, but how
should the results be interpreted when they cannot even be visualized. One
possibility is to use recurrence plots (section 3.3.2), but there might be better ways.
Classical linear methods for dimension reduction include principal component
analysis and multidimensional scaling. A drawback with linear methods is that
data may not always be accurately summarized by linear combinations. An
example is a helix in a three dimensional space, whose one dimensional structure
cannot be discovered by linear methods.
Since the manifold on which the embedded phonocardiographic signal resides is
non-linear, linear methods are not adequate. Instead of finding the most important
linear subspace from a set of data points (like in principal component analysis),
non-linear parameterizations can be sought. Many of these manifold learning
techniques are based on a distance matrix (like multidimensional scaling, which
tries to preserve the distance between each data point even after the dimensionality
has been decreased). In principle, the difference compared to multidimensional
scaling is the way that distances are calculated. Instead of measuring a global
distance (a straight line through space), the distance is calculated by summing up
local distances between states as we move from one point to another (i.e. we are
only allowed to travel from point a to point b via other data points located in
between).
A problem associated with these techniques is whether the embedded
phonocardiogram really resides on a manifold and if so, whether the manifold is
sampled dense enough to be able to draw any valid conclusions.
6.4.4. Choosing an appropriate classifier

A proper set-up of the classifier in paper III, and, for that matter, the choice of the
actual classifier, was out of the scope of this thesis. The common picture is that the
actual type of classifier is not extremely important, at least not in comparison with
the importance of good features. Nevertheless, a good classifier should be used,
and some effort should be spent on its design. Looking at it from a practical point
of view, the chosen classifier might be adequate enough. After all, it works
properly for the application in question.
65
7. Review of Papers
“There is a coherent plan in the universe,
though I don't know what it's a plan for.”
Fred Hoyle (1915 - 2001)
This chapter introduces the papers which are included in the second part of this
thesis.
7.1. Paper I, Heart Sound Cancellation

This paper presents a new method to detect and remove heart sounds from
recorded lung sound signals. The idea is to locate the heart sounds, remove them
altogether and predict what should have been in their place using surrounding data.
The location of the heart sounds is found using a nonlinear change detection
method based on recurrence time statistics. Once the heart sounds have been
removed, the missing gap is filled in using a nonlinear prediction scheme.
The reason for using this rather odd change detection algorithm was that the data
was very noisy (high flow in the respiratory part of the signal). The error rate was
4% false positives and 8% false negatives using this method. Results from other
algorithms were not explicitly derived, but fast visual comparisons rejected
methods like Shannon energy. Similarly, the choice of a nonlinear local prediction
method was based on the fact that lung sounds are indeed nonlinear.
The proposed solution to the problem of heart sound cancellation was tested on six
subjects. The spectral difference between the denoised lung sound signal and a
lung sound signal with removed heart sounds was 0.34±0.25 dB/Hz, 0.50±0.33
dB/Hz, 0.46±0.35 dB/Hz and 0.94±0.64 dB/Hz in the frequency bands 20-40 Hz,
40-70 Hz, 70-150 Hz and 150-300 Hz, respectively. The cross-correlation index
was found to be 99.7%, indicating excellent similarity between actual lung sound
data and predicted lung sound data.
7.2. Paper II, Detection of the 3rd Heart Sound

An algorithm, actually based on the same recurrence statistics as the heart sound
locator in paper I, was developed for detection of the third heart sound. Recurrence
67
points of the first kind (T1) are known to be rather noise insensitive and robust,
while recurrence points of the second kind (T2) are better suited for finding very
weak signals. The two could be used as a pair where T1 is used on a coarse scale
to locate the first and second heart sounds, while T2 could be used on a finer scale
to locate the third heart sound within a predetermined window succeeding the
other heart sounds. The reason to look for S3 after both S1 and S2 was to avoid the
problem of distinguishing the two (statistics about their timing is hard to use on a
young population where the heart rate is high).
Since S3 is normally heard during auscultation of younger individuals, the method
was tested on ten children. Most S3 occurrences were detected (98 %), but the
amount of false extra detections was rather high (7% of the heart cycles).
7.3. Paper III, Feature Extraction from Systolic Murmurs

Available features for heart murmur classification were reviewed and new
features, mostly inspired by research in speech processing, dynamical systems and
chaos theory, were introduced. Techniques such as Shannon energy, wavelets,
fractal dimensions and recurrence quantification analysis were used to extract 213
different features. 163 of these features have not previously been used for heart
murmur classification. However, this many features often result in high
computational complexity, mutual correlation and classifiers with low generality.
Not to mention the curse of dimensionality. For this reason, a subset selection
method was employed to reduce the number of features. The derived feature
subset, maximizing a performance criterion while keeping the number of features
low, resulted in a subset of 14 features.
Heart sound data from 36 patients with aortic valve stenosis, mitral insufficiency
or physiological murmurs were used to test the new feature set. Using the results
from a neural network classifier, this new feature set was compared with three
other feature sets. The selected subset gave the best results with 86 % correct
classifications, compared to 58 % for the first runner-up. In conclusion, the
derived feature set was superior to those previously used, seems rather robust to
noisy data sets and will be tested in more clinically oriented studies in the future.
68
References
[1] L. A. Geddes, "Birth of the stethoscope," Engineering in Medicine and Biology

Magazine, IEEE, vol. 24, pp. 84, 2005.
[2] A. N. Pelech, "The physiology of cardiac auscultation," Pediatr Clin North Am,
vol. 51, pp. 1515-1535, 2004.
[3] S. Persson and J. Engqvist, Kardiologi: hjärtsjukdomar hos vuxna, 5., rev. och
utök. uppl. / ed. Lund: Studentlitteratur, 2003.
[4] G. J. Borden, K. S. Harris, and L. J. Raphael, Speech science primer: physiology,
acoustics, and perception of speech, 3. rev. ed. Baltimore: Williams & Wilkins,
1994.
[5] D. Smith and E. Craige, "Heart Sounds: Toward a Consensus Regarding their
Origin," Am. J. Noninvas. Cardiol., vol. 2, pp. 169-179, 1988.
[6] L. Wigstrom, T. Ebbers, A. Fyrenius, M. Karlsson, J. Engvall, B. Wranne, and A.
F. Bolger, "Particle trace visualization of intracardiac flow using time-resolved
3D phase contrast MRI," Magn Reson Med, vol. 41, pp. 793-799, 1999.
[7] A. G. Tilkian and M. B. Conover, Understanding heart sounds and murmurs:
with an introduction to lung sounds, 4. ed. Philadelphia: Saunders, 2001.
[8] N. J. Mehta and I. A. Khan, "Third heart sound: genesis and clinical importance,"
Int J Cardiol, vol. 97, pp. 183-186, 2004.
[9] R. B. Northrop, Noninvasive instrumentation and measurement in medical
diagnosis. Boca Raton, Fla., London: CRC; Chapman & Hall, 2002.
[10] H. Nygaard, "Evaluation of Heart Sounds and Murmurs - a Review with Special
Reference to Aortic Valve Stenosis," Department of Electrical Engineering,
College of Engineering, Aarhus, Denmark 1996.
[11] L. Vannuccini, J. E. Earis, P. Helistö, B. M. G. Cheetham, M. Rossi, A. R. A.
Sovijärvi, and J. Vanderschoot, "Capturing and preprocessing of respiratory
sounds," Eur Respir Rev, vol. 10, pp. 616-620, 2000.
[12] R. J. Povinelli, M. T. Johnson, A. C. Lindgren, F. M. Roberts, and J. Ye,
"Statistical Models of Reconstructed Phase Spaces for Signal Classification,"
IEEE Transactions on Signal Processing, vol. In press, pp. -.
[13] R. L. Allen and D. W. Mills, Signal analysis: time, frequency, scale and structure.
New York, Piscataway, N.J.: Wiley; IEEE Press, 2004.
[14] R. N. Bracewell, The Fourier transform and its applications, 3. ed. Boston:
McGraw Hill, 2000.
[15] H. Liang, S. Lukkarinen, and I. Hartimo, "Heart sound segmentation algorithm
based on heart sound envelogram," in Computers in Cardiology, Lund, Sweden,
1997, pp. 105-108.
[16] J. F. Kaiser, "Some useful properties of Teager's energy operators," in ICASSP-
93, 1993, pp. 149-152.
69
[17] C. N. Gupta, R. Palaniappan, S. Swaminathan, and S. M. Krishnan, "Neural

network classification of homomorphic segmented heart sounds," Applied Soft
Computing, vol. In Press, Corrected Proof.
[18] G. Livanos, N. Ranganathan, and J. Jiang, "Heart sound analysis using the S
transform," in Computers in Cardiology, 2000, pp. 587 - 590.
[19] R. G. Stockwell, L. Mansinha, and R. P. Lowe, "Localization of the complex
spectrum: The S transform," Ieee Transactions on Signal Processing, vol. 44, pp.
998-1001, 1996.
[20] U. Parlitz, "Nonlinear time-series analysis," in Nonlinear modeling: advanced
black-box techniques, J. A. K. Suykens and J. Vandewalle, Eds. Boston: Kluwer
Academic Publishers, 1998, pp. 256.
[21] H. D. I. Abarbanel, Analysis of observed chaotic data. New York: Springer-Vlg,
1996.
[22] L. Y. Cao, "Practical method for determining the minimum embedding dimension
of a scalar time series," Physica D, vol. 110, pp. 43-50, 1997.
[23] H. Kantz and T. Schreiber, Nonlinear Time Series Analysis, 2. ed. Cambridge:
Cambridge Univ. Press, 2004.
[24] W. Kinsner, "Batch and real-time computation of a fractal dimension based on
variance of a time series," Dept. of Electrical & Computer Eng., University of
Manitoba, Winnipeg, Canada DEL94-6, June 1994.
[25] J. P. Zbilut, N. Thomasson, and C. L. Webber, "Recurrence quantification
analysis as a tool for nonlinear exploration of nonstationary cardiac signals,"
Medical Engineering & Physics, vol. 24, pp. 53-60, 2002.
[26] C. L. Webber and J. P. Zbilut, "Recurrence quantification analysis of nonlinear
dynamical systems." In: Tutorials in Contemporary Nonlinear Methods for the
Behavioral Sciences: National Science Foundation, 2005.
[27] J. B. Gao, "Recurrence time statistics for chaotic systems and their applications,"
Phys. Rev. Lett., vol. 83, pp. 3178-3181, 1999.
[28] N. Marwan, N. Wessel, U. Meyerfeldt, A. Schirdewan, and J. Kurths,
"Recurrence-plot-based measures of complexity and their application to heart-
rate-variability data," Phys. Rev. E., vol. 66, pp. 1-8, 2002.
[29] C. L. Webber and J. P. Zbilut, "Dynamical assessment of physiological systems
and states using recurrence plot strategies," J. Appl. Physiol., vol. 76, pp. 965-973,
1994.
[30] F. Gustafsson, Adaptive filtering and change detection. Chichester: Wiley, 2000.
[31] J. B. Gao, Y. H. Cao, L. Y. Gu, J. G. Harris, and J. C. Principe, "Detection of
weak transitions in signal dynamics using recurrence time statistics," Physics
Letters A, vol. 317, pp. 64-72, 2003.
[32] A. Hyvärinen, J. Karhunen, and E. Oja, Independent component analysis. New
York: Wiley, 2001.
[33] C. L. Nikias and J. M. Mendel, "Signal processing with higher-order spectra,"
IEEE Signal Processing Magazine, vol. 10, pp. 10 - 37, 1993.
[34] L. Ljung, System identification: theory for the user, 2. ed. Upper Saddle River,
N.J.: Prentice Hall, 1999.
[35] J. McNames, "A nearest trajectory strategy for time series prediction," in Proc.
Int. Workshop on Advanced Black-Box Techniques for Nonlinear Modeling,
Leuven, Belgium, 1998, pp. 112-128.
70
References
[36] A. P. Yoganathan, R. Gupta, F. E. Udwadia, J. W. Miller, W. H. Corcoran, R.

Sarma, J. L. Johnson, and R. J. Bing, "Use of the fast Fourier transform for
frequency analysis of the first heart sound in normal man," Med Biol Eng, vol. 14,
pp. 69-73, 1976.
[37] A. P. Yoganathan, R. Gupta, F. E. Udwadia, W. H. Corcoran, R. Sarma, and R. J.
Bing, "Use of the fast Fourier transform in the frequency analysis of the second
heart sound in normal man," Med Biol Eng, vol. 14, pp. 455-460, 1976.
[38] M. S. Obaidat, "Phonocardiogram signal analysis: techniques and performance
comparison," J Med Eng Technol, vol. 17, pp. 221-227, 1993.
[39] C. Longhini, S. Aggio, E. Baracca, D. Mele, C. Fersini, and A. E. Aubert, "A
mass-spring model hypothesis of the genesis of the physiological third heart
sound," Jpn Heart J, vol. 30, pp. 265-273, 1989.
[40] E. Baracca, D. Scorzoni, M. C. Brunazzi, P. Sgobino, L. Longhini, D. Fratti, and
C. Longhini, "Genesis and acoustic quality of the physiological fourth heart
sound," Acta Cardiol, vol. 50, pp. 23-28, 1995.
[41] A. Haghighi-Mood and N. Torry, "Time-frequency analysis of systolic murmurs,"
in Computers in Cardiology 1997, 1997, pp. 113 - 116.
[42] L. J. Hadjileontiadis and S. M. Panas, "Discrimination of heart sounds using
higher-order statistics," in Proc. 19th Ann. Int. Conf. of the IEEE, EMBS, 1997,
pp. 1138-1141.
[43] S. Minfen and S. Lisha, "The analysis and classification of phonocardiogram
based on higher-order spectra," in Proc. of the IEEE-SP, Higher-Order Statistics.,
1997, pp. 29-33.
[44] A. Swami, J. M. Mendel, and C. L. Nikias, "Higher-Order Spectral Analysis
Toolbox Users Guide," 2001.
[45] Y. Xiang and S. K. Tso, "Detection and classification of flaws in concrete
structure using bispectra and neural networks," Ndt. & E. Int., vol. 35, pp. 19-27,
2002.
[46] G. Kubin, "Nonlinear processing of speech," in Speech coding and synthesis, W.
B. Kleijn and K. K. Paliwal, Eds. Amsterdam: Elsevier, 1995, pp. 557-610.
[47] J. Whitmire and S. Sarkar, "Validation of acoustic-analogy predictions for sound
radiated by turbulence," Physics of Fluids, vol. 12, pp. 381-391, 2000.
[48] P. Maragos and A. Potamianos, "Fractal dimensions of speech sounds:
Computation and application to automatic speech recognition," Journal of the
Acoustical Society of America, vol. 105, pp. 1925-1932, 1999.
[49] P. P. Vlachos, O. Pierrakos, A. Phillips, and D. P. Telionis, "Vorticity and
turbulence characteristics inside a transparent flexible left ventricle," in
Bioengineering Conference, ASME, 2001, pp. 493-494.
[50] D. Bluestein, C. Gutierrez, M. Londono, and R. T. Schoephoerster, "Vortex
shedding in steady flow through a model of an arterial stenosis and its relevance
to mural platelet deposition," Ann Biomed Eng, vol. 27, pp. 763-773, 1999.
[51] J. S. Liu, P. C. Lu, and S. H. Chu, "Turbulence characteristics downstream of
bileaflet aortic valve prostheses," J Biomech Eng, vol. 122, pp. 118-124, 2000.
[52] A. Iwata, N. Ishii, N. Suzumura, and K. Ikegaya, "Algorithm for detecting the
first and the second heart sounds by spectral tracking," Med Biol Eng Comput,
vol. 18, pp. 19-26, 1980.
71
[53] H. Liang, L. Sakari, and H. Iiro, "A heart sound segmentation algorithm using
wavelet decomposition and reconstruction," in Proc. 19th Ann. Int. Conf. of the
IEEE, EMBS, 1997, pp. 1630 - 1633.
[54] H. Sava, P. Pibarot, and L. G. Durand, "Application of the matching pursuit
method for structural decomposition and averaging of phonocardiographic
signals," Med Biol Eng Comput, vol. 36, pp. 302-308, 1998.
[55] J. Gnitecki and Z. Moussavi, "Variance fractal dimension trajectory as a tool for
hear sound localization in lung sounds recordings," in Proc. 25th Ann. Int. Conf.
IEEE EMBS, 2003, pp. 2420-2423.
[56] V. Nigam and R. Priemer, "Accessing heart dynamics to estimate durations of
heart sounds," Physiological Measurement, pp. 1005-1018, 2005.
[57] N. Joshi, "The third heart sound," South Med J, vol. 92, pp. 756-761, 1999.
[58] P. Hult, T. Fjallbrant, K. Hilden, U. Dahlstrom, B. Wranne, and P. Ask,
"Detection of the third heart sound using a tailored wavelet approach: method
verification," Med Biol Eng Comput, vol. 43, pp. 212-217, 2005.
[59] P. Hult, T. Fjallbrant, B. Wranne, and P. Ask, "Detection of the third heart sound
using a tailored wavelet approach," Med Biol Eng Comput, vol. 42, pp. 253-258,
2004.
[60] S. Charleston, M. R. Azimi-Sadjadi, and R. Gonzalez-Camarena, "Interference
cancellation in respiratory sounds via a multiresolution joint time-delay and
signal-estimation scheme," IEEE Trans Biomed Eng, vol. 44, pp. 1006-1019,
1997.
[61] S. Charleston and M. R. Azimi-Sadjadi, "Reduced order Kalman filtering for the
enhancement of respiratory sounds," IEEE Trans Biomed Eng, vol. 43, pp. 421-
424, 1996.
[62] L. J. Hadjileontiadis and S. M. Panas, "Adaptive reduction of heart sounds from
lung sounds using fourth-order statistics," IEEE Trans Biomed Eng, vol. 44, pp.
642, 1997.
[63] Z. K. Moussavi, D. Flores, and G. Thomas, "Heart sound cancellation based on
multiscale products and linear prediction," in Proc. 26th Annu. Int. Conf. IEEE
Engineering in Medicine and Biology Society, EMBC’04, San Francisco, USA,
2004, pp. 3840-3843.
[64] M. T. Pourazad, Z. K. Moussavi, and G. Thomas, "Heart sound cancellation from
lung sound recordings using adaptive threshold and 2D interpolation in time-
frequency domain," in Proc. 25th Annu. Int. Conf. IEEE Engineering in Medicine
and Biology Society, EMBC’03, Cancun, Mexico, 2003, pp. 2586-2589.
[65] C. Ahlstrom, A. Johansson, P. Hult, and P. Ask, "Chaotic dynamics of respiratory
sounds," Chaos, Solitons & Fractals, vol. In Press, Corrected Proof.
[66] J. Gnitecki and Z. Moussavi, "The fractality of lung sounds: A comparison of
three waveform fractal dimension algorithms," Chaos Solitons & Fractals, vol.
26, pp. 1065-1072, 2005.
[67] J. Gnitecki, Z. Moussavi, and H. Pasterkamp, " Geometrical and Dynamical State
Space Parameters of Lung Sounds," in 5th Int. Workshop on Biosignal
Interpretation, 2005, pp. 113-116.
[68] A. Vena, E. Conte, G. Perchiazzi, A. Federici, R. Giuliani, and J. P. Zbilut,
"Detection of physiological singularities in respiratory dynamics analyzed by
recurrence quantification analysis of tracheal sounds," Chaos Solitons & Fractals,
vol. 22, pp. 869-881, 2004.
72
References
[69] S. Theodoridis and K. Koutroumbas, Pattern Recognition, 2. ed. Amsterdam:

Academic Press, 2003.
[70] I. Guyon and A. Elisseeff, "An introduction to variable and feature selection," J.
Mach. Learn. Res., vol. 3, pp. 1157-1182, 2003.
[71] P. Pudil, J. Novovicova, and J. Kittler, "Floating search methods in feature-
selection," Patt. Recogn. Lett., vol. 15, pp. 1119-1125, 1994.
[72] T. Olmez and Z. Dokur, "Classification of heart sounds using an artificial neural
network," Pattern Recogn. Lett., vol. 24, pp. 617-629, 2003.
[73] T. S. Leung, P. R. White, W. B. Collis, E. Brown, and A. P. Salmon, "Analysing
paediatric heart murmurs with discriminant analysis," in Proc. 20th Ann. Int.
Conf. of the IEEE, EMBS., 1998, pp. 1628 - 1631.
[74] L. G. Durand and P. Pibarot, "Digital signal processing of the phonocardiogram:
review of the most recent advancements," Crit. Rev. Biomed. Eng., vol. 23, pp.
163-219, 1995.
[75] R. M. Rangayyan and R. J. Lehner, "Phonocardiogram signal analysis: a review,"
Crit Rev Biomed Eng, vol. 15, pp. 211-236, 1987.
[76] I. Cathers, "Neural network assisted cardiac auscultation," Artif. Intell. Med., vol.
7, pp. 53-66, 1995.
[77] S. A. Pavlopoulos, A. C. Stasis, and E. N. Loukis, "A decision tree--based method
for the differential diagnosis of aortic stenosis from mitral regurgitation using
heart sounds," Biomed. Eng. Online, vol. 3, pp. 21, 2004.
[78] Z. Sharif, M. S. Zainal, A. Z. Sha'ameri, and S. H. S. Salleh, "Analysis and
classification of heart sounds and murmurs based on the instantaneous energy and
frequency estimations," in Proc. of TENCON, Kuala Lumpur, 2000, pp. 130-134.
[79] C. G. DeGroff, S. Bhatikar, J. Hertzberg, R. Shandas, L. Valdes-Cruz, and R. L.
Mahajan, "Artificial neural network-based method of screening heart murmurs in
children," Circulation, vol. 103, pp. 2711-2716, 2001.
[80] H. Shino, H. Yoshida, K. Yana, K. Harada, J. Sudoh, and E. Harasewa,
"Detection and classification of systolic murmur for phonocardiogram screening,"
in Proc. 18th Ann. Int. Conf. of the IEEE, EMBS, 1996, pp. 123-124.
[81] A. Voss, A. Mix, and T. Hubner, "Diagnosing aortic valve stenosis by parameter
extraction of heart sound signals," Ann. Biomed. Eng., vol. 33, pp. 1167-1174,
2005.
[82] H. Liang and I. Hartimo, "A heart sound feature extraction algorithm based on
wavelet decomposition and reconstruction," in Proc. 20th Ann. Int. Conf. of the
IEEE, EMBS., 1998, pp. 1539 - 1542.
[83] I. Turkoglu, A. Arslan, and E. Ilkay, "An intelligent system for diagnosis of the
heart valve diseases with wavelet packet neural networks," Comput. Biol. Med.,
vol. 33, pp. 319-331, 2003.
[84] C. Ahlstrom, P. Hult, and P. Ask, "Wheeze Analysis and Detection with Non-
linear Phase Space Embedding," in 13th Nordic Baltic Conference Biomedical
Engineering and Medical Physics, Umeå, Sweden, 2005, pp. 305-306.
[85] B. Suki, "Fluctuations and power laws in pulmonary physiology," Am J Respir
Crit Care Med, vol. 166, pp. 133-137, 2002.
[86] R. Esteller, G. Vachtsevanos, J. Echauz, and B. Litt, "A comparison of waveform
fractal dimension algorithms," IEEE Transactions on Circuits and Systems I-
Fundamental Theory and Applications, vol. 48, pp. 177-183, 2001.
73
The phonocardiography factory.
75

Processing of The Phonocardiographic Signal Methods For The Intelligent Stethoscope

Uploaded by

Copyright:

Available Formats

Processing of The Phonocardiographic Signal Methods For The Intelligent Stethoscope

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Processing of The Phonocardiographic Signal Methods For The Intelligent Stethoscope

Uploaded by

Copyright:

Available Formats

Linköping Studies in Science and Technology

Thesis No. 1253

Processing of the Phonocardiographic Signal −

In cooperation with Biomedical Engineering,

Linköping, May 2006

© 2006 Christer Ahlström

Department of Biomedical Engineering

ISBN: 91-85523-59-3 ISSN: 0280-7971

Printed in Sweden by LiU-Tryck, Linköping 2006

Related publications not included in the thesis.

The following M. Sc. theses have also contributed to this thesis.

Linköping, April 2006

The history of auscultation, or listening to the sounds of the body, is easily

Acoustic stethoscopes transmit sound mechanically from a chest-piece via air-

Audible heart sounds

1.1. Aim of the Thesis

1.2. Thesis Outline

2.1. Physics of Sound

2.2. Physiology of the Heart

Superior Vena Cava Pulmonary Trunk

Inferior Vena Cava

2.3. Heart Sounds

Left atrium Mitral valve

2.4. Heart Murmurs

• Backward flow through an incompetent valve (insufficiency or

Mitral valve Open Open

Aortic valve Open

2.5. Auscultation and the Phonocardiogram

A graphical printing of the waveform of cardiac sounds is called a

2.6. Acquisition of Phonocardiographic Signals

2.6.2. Pre-processing, digitalization and storage

Example 2, Distinguishing signals with similar spectra (example adapted from

3.1. Measuring Characteristics that Vary in Time

Analytic signal envelope: s A ( t ) = s ( t ) + sH ( t )

method developed by Teager, equation (3.8), results in an on-the-fly envelope-

Shannon entropy: E ( t ) = − s ( t ) ⋅ log s ( t ) (3.6)

Shannon energy: E ( t ) = − s ( t ) ⋅ log s ( t )

Teager’s energy operator: E ( t ) = s ( t ) − s ( t − 1) s ( t + 1)

Another way to extract a signal’s envelope is through homomorphic signal

log s ( t ) = log l ( t ) + log h ( t ) (3.10)

3.2. Nonlinear Systems and Embedology

trajectory is drawn to a particular set this set is called an attractor. Examples of

(a) (b) (c)

Finding the space: Finding the space:

Classify the signal: Classify the signal:

Making models, predict: Making models, predict:

Find parameters k consistent with y (t + 1) = F ⎡⎣ y (t ), a1 , a2 ,… , a p ⎤⎦

3.3. Nonlinear Analysis Tools

3.3.1. Non-integer dimensions

Plotting log(Var(s(t2)-s(t1))) against log(|t2-t1|) for different time increments in a

The variance fractal dimension is related to H as VFD = Ed+1-H, where Ed is the

3.3.2. Recurrence quantification analysis

where i,j=1,…,N, is a cut-off distance, • is some norm and (•) is the

0.5 0.5 600

Detection of changes based on recurrence times

More formally, an arbitrary state, yref, is chosen somewhere on the trajectory

3.3.3. Higher order statistics

Complete characterisation of a stochastic process requires knowledge of all

3.4. Nonlinear Prediction