G Model

JJBE-1986; No. of Pages 6

ARTICLE IN PRESS
Medical Engineering & Physics xxx (2011) xxxxxx

Contents lists available at SciVerse ScienceDirect

Medical Engineering & Physics

journal homepage: www.elsevier.com/locate/medengphy

Matrix decomposition based feature extraction for murmur classication

Yuerong Chen, Shengyong Wang , Chia-Hsuan Shen, Fred K. Choy
Department of Mechanical Engineering, The University of Akron, Akron, OH 44325-3903, USA

a r t i c l e

i n f o

Article history:
Received 19 November 2010
Received in revised form
20 September 2011
Accepted 22 September 2011
Keywords:
Murmur differentiation
Wavelet transform
Singular value decomposition
QR decomposition
Shannon entropy
Gini index
Classication and regression tree

a b s t r a c t
Heart murmurs often indicate heart valvular disorders. However, not all heart murmurs are organic. For
example, musical murmurs detected in children are mostly innocent. Because of the challenges of mastering auscultation skills and reducing healthcare expenses, this study aims to discover new features for
distinguishing innocent murmurs from organic murmurs, with the ultimate objective of designing an
intelligent diagnostic system that could be used at home. Phonocardiographic signals that were recorded
in an auscultation training CD were used for analysis. Instead of the discrete wavelet transform that has
been used often in previous work, a continuous wavelet transform was applied on the heart sound data.
The matrix that was derived from the continuous wavelet transform was then processed via singular
value decomposition and QR decomposition, for feature extraction. Shannon entropy and the Gini index
were adopted to generate features. To reduce the number of features that were extracted, the feature
selection algorithm of sequential forward oating selection (SFFS) was utilized to select the most signicant features, with the selection criterion being the maximization of the average accuracy from a 10-fold
cross-validation of a classication algorithm called classication and regression trees (CART). An average
sensitivity of 94%, a specicity of 83%, and a classication accuracy of 90% were achieved. These favorable
results substantiate the effectiveness of the feature extraction methods based on the proposed matrix
decomposition method.
Published by Elsevier Ltd on behalf of IPEM.

1. Introduction
Heart disease is a major cause of death in the United States
[1,2]. Because they are the rst signs of heart valvular disease
[3,4], heart murmurs receive a substantial amount of attention
from healthcare providers. Heart murmurs usually result from turbulence in blood ow or the vibration of heart tissues. They can
occur in a systolic period or in a diastolic period, and can even
cover both periods. Heart murmurs frequently appear in children,
and fortunately, most heart murmurs that are detected in children are innocent [57]. To reduce unnecessary parental anxiety,
it is important to recognize an innocent murmur as soon as it is
found [8,9]. However, because of the difculty of mastering auscultation skills, innocent and organic heart murmurs cannot be
readily distinguished [1012]. As many as 87% of the patients who
are referred to cardiologists for evaluation have benign murmurs
[13]. Currently, echocardiography is the standard clinical method
for diagnosing cardiac disease. However, ordering an echocardiogram is very expensive and can cost between $750 and $1500
[14], and the waiting time for an echocardiogram appointment
may be long. It would be of great value to have an intelligent and

Corresponding author at: ASEC 101, Department of Mechanical Engineering, The

University of Akron, Akron, OH 44325-3903, USA. Tel.: +1 330 972 8616.
E-mail address: wangs@uakron.edu (S. Wang).

affordable device that is capable of automatically distinguishing

innocent murmurs from organic murmurs.
Much research effort has been exerted to address this challenging task [1536]. Thompson et al. [15] designed a system to
distinguish innocent murmurs from other murmurs with different
conditions such as aortic valve stenosis (AS), atrial septal defect
(ASD), pulmonic valve stenosis (PS), and ventricular septal defect
(VSD). These investigators examined the correlation between the
relative murmur intensity and the likelihood of pathology. Andrisevic et al. [16] proposed an algorithm that applied wavelet analysis
and articial neural networks for the detection of heart murmurs,
distinguishing normal heart murmurs from those associated with
AS, PS, and mitral stenosis (MS). Noponen et al. [17] conducted
phono-spectrographic analysis to differentiate innocent murmurs
from other murmurs that were caused by VSD, AS, PS, ASD, and
patent ductus arteriosus (PDA). Ahlstrom et al. [18] presented a
wide range of techniques to extract various features, from which a
suitable feature subset was selected for classifying murmurs into
normal murmurs, murmurs with AS, and murmurs with mitral
insufciency (MI). Reed et al. [19] developed a system for heart
sound analysis and evaluated the system using heart sounds with
ve different conditions: normal, mitral valve prolapse (MVP),
coarctation of the aorta (CA), VSD, and PS. lmez and Dokur [20]
used wavelet transforms and dynamic programming to extract and
analyze features, and applied articial neural networks to the classication of seven different heart sounds: AS, MI, MS, PS, aortic

1350-4533/$ see front matter. Published by Elsevier Ltd on behalf of IPEM.

doi:10.1016/j.medengphy.2011.09.020

Please cite this article in press as: Chen Y, et al. Matrix decomposition based feature extraction for murmur classication. Med Eng Phys (2011),
doi:10.1016/j.medengphy.2011.09.020

G Model
JJBE-1986; No. of Pages 6

ARTICLE IN PRESS
Y. Chen et al. / Medical Engineering & Physics xxx (2011) xxxxxx

regurgitation (AR), summation gallop, and normal heart sound.

Babaei and Geranmayeh [28] used wavelet transforms and articial
neural networks to discriminate AS, aortic insufciency (AI), and PS
from normal sounds. Uguz [29] classied heart sounds into normal
sounds, MS, and PS. They applied a discrete Fourier transform and
Burg autoregressive spectrum analysis for feature extraction, principle component analysis for dimension reduction, and articial
neural networks for classication. Choi and Jiang [30] differentiated
normal heart murmurs from those with six disorders: AF, AI, AS, MR,
MS, and split sounds, using normalized autoregressive power spectral density curves with support vector machine techniques. Sengur
[31] developed an expert system that was based on principle component analysis, articial immune systems, and fuzzy k-NNs for
diagnosing heart diseases. Additional related studies can be found
in the literature [3236].
In this paper, we aim to explore a set of novel features that
complement existing features for the differentiation of innocent
from organic murmurs. Because heart murmurs that occur in the
diastolic periods are mostly pathological [18], we consider only
systolic murmurs in this paper. Feature extraction methods in the
previous literature include time-frequency analysis [21], instantaneous energy and frequency estimation [22], discrete wavelet
transforms [23], recurrence quantication analysis [24], and other
methods [18]. Wavelet transforms provide both time and frequency content and are very suitable for non-stationary signals
such as heart sounds. Therefore, wavelet transforms are used in
this paper. Instead of the discrete wavelet transform, which has
been widely used in previous work, the continuous wavelet transform was applied in this paper. The techniques of singular value
decomposition and QR decomposition were then applied on the
time-frequency matrix that was obtained from continuous wavelet
transforms for feature extraction. Shannon entropy and the Gini
index were further computed on the decomposition results and
were taken as features. To reduce the number of features and
thereby the computational complexity, a feature selection algorithm of sequential forward oating selection (SFFS) was utilized.
Different from most of the previous work that used articial neural
networks for classication [16,18,20,26], we applied the technique
of classication and regression trees (CART). The maximization of
the average classication accuracy from a 10-fold cross-validation
of CART was taken as the criterion of the feature selection method.
The high classication accuracy of 90% that was obtained substantiates the effectiveness of the proposed feature extraction
methods.

Fig. 1. A heart sound recording segment.

segments were obtained, including 132 innocent segments and 248

organic segments.
The original heart sound recordings might contain noise and
could have inconsistent measurement units; thus, a preprocessing
step was necessary. First, we used the Daubechies-2 wavelet at the
sixth level to remove the possible noise. Second, to guarantee a consistent measurement unit, each murmur segment was normalized
according to the following normalization formula:
y(i) =

x(i) m
,
s

i = 1, . . . , N

where N is the number of data points in a murmur segment, x(i),

i = 1,...,N is the data in the murmur segment, m and s are the sample
mean and the sample standard deviation of x(i), i = 1,...,N, respectively, and y(i), i = 1,...,N is the normalized data that were used in
this study.
2.2. Methodology framework

2. Data and methods

Fig. 2 shows the methodology framework for developing an

intelligent system that would be used to differentiate innocent murmurs from organic murmurs. The inputs to the system
are heart sound recordings and the output is the diagnosis result (i.e. murmurs are innocent or organic). First, heart
sound recordings pass through the feature extraction process.
The favorable features should be capable of keeping similarities within classes while exposing differences between classes. In
general, in this stage a large number of features are extracted,
and they cause inconvenience in later classication steps. To
reduce the computational complexity, a smaller number of the
most signicant features must be selected. This step is the feature selection process. At the end is the classication process,
where innocent and organic heart murmurs are differentiated
via an appropriate classication method. The specic procedure
for each stage is described in detail in the following subsections.

2.1. Data description and preprocessing

2.3. Feature extraction

We considered 43 heart sound recordings from an auscultation skill training CD [37], including 15 innocent murmurs and
28 organic murmurs. The reason to select these 43 heart sound
recordings was that they contain murmurs that have been explicitly
specied as innocent or organic. Usually, a heart sound recording
includes a number of cardiac cycles (see Fig. 1). By denition, a
cardiac cycle is a heart sound segment that is between two consecutive S1s or S2s. The segment between S1 and S2 is called a systole
period, while the segment between S2 and the next S1 is called a
diastole period. Heart murmurs can occur in either period or can
even cover both periods. Murmurs that occur in a diastole period
are mostly pathological [18] and therefore are not of great value for
further analysis. Only systolic murmurs were taken into account.
For each of the 43 heart sound recordings, we randomly extracted
several segments of systolic murmurs. In total, 380 murmur

A heart sound is non-stationary. Therefore, we applied the

wavelet transform, which is a time-frequency representation
method that is very suitable for analyzing non-stationary signals. A wavelet transform is superior to other time-frequency
analysis methods such as the short-time Fourier transform
because, not only is it capable of providing time and frequency information simultaneously, it has a varying time and
frequency resolution. The discrete wavelet transform (DWT)
was widely used in the earlier literature of murmur classication. In this paper, instead of the DWT, we utilized the

Fig. 2. The diagram of the methodology framework.

Please cite this article in press as: Chen Y, et al. Matrix decomposition based feature extraction for murmur classication. Med Eng Phys (2011),
doi:10.1016/j.medengphy.2011.09.020

ARTICLE IN PRESS

G Model
JJBE-1986; No. of Pages 6

Y. Chen et al. / Medical Engineering & Physics xxx (2011) xxxxxx

continuous wavelet transform (CWT). The adopted wavelet

was the Daubechies-2 wavelet, which showed excellent performance for heart sound classication in previous studies
[20,27,38].
To apply the CWT, the normalized heart sound signal y() was
regarded to be continuous by using a piecewise constant interpolation, and was then transformed by using the formula of the CWT
[39]:
1
C(a, b) =
|a|

y(t)

t b
a

dt

where a and b denote scale and position parameters, respectively,  () denotes a mother wavelet (here it is the Daubechies-2
wavelet), and * is the operator of the complex conjugate. By
appropriately setting the ranges of a and b, a time-frequency representation matrix composed by CWT coefcients was acquired for
each of the 380 murmur segments. Subsequently, the techniques of
singular value decomposition (SVD) and QR decomposition (QRD)
were implemented on these CWT coefcient matrices.
SVD is a valuable tool that was successfully used in signal processing and statistical data analysis [38,40]. An SVD [41] of an m n
matrix X is given by the following:
X = USV T ,
where U is an m n orthonormal matrix (i.e. UT U = I), V is an
n n orthonormal matrix (i.e. VT V = I),and S is an m n diagonal
matrix with nonnegative diagonal elements in decreasing order.
The columns of the matrices U and V are called the left and right
singular vectors, respectively. The diagonal elements in the matrix
S are called singular values, and they represent the importance of
the corresponding left and right singular vectors in the composition
of the matrix X.
QRD [42] is a matrix decomposition that can be used to solve
linear systems of equations. The QRD of a matrix A is of the following
form:
A = QR,
where Q is an orthonormal matrix and R is an upper triangular
matrix.
Because U, V, and Q are orthonormal matrices, their column vectors are orthonormal. Therefore, for each of these column vectors,
their squared elements can be regarded as the components of a
probability mass function. Take the rst left singular vector as an
example. Denote it by U1 = (u11 , u21, . . . , um1 )T . By squaring each of
its elements, a probability mass function is acquired as follows:
fU1 = {u211 , u221 , . . . , u2m1 }

m

u2 = 1. There are two quantitative measures that are

where
i=1 i1
associated with a probability mass function: Shannon entropy and
the Gini index. In this paper, both measures were used to extract
features from murmur segments.
Shannon entropy [43] is a measure of the uncertainty that is
associated with a random variable. It is a key concept in Information
Theory. Assume that a random variable M has a probability mass
function: {p1 , p2 , . . ., pn }. The Shannon entropy of M is dened as
follows:
n


H(M) =

pi log2 pi

i=1

If the values of M are equally likely, then its Shannon entropy is the
largest.
Gini index (or Gini coefcient) is a measure of the statistical dispersion. This index has been widely used in economics for

measuring income inequality. In the area of data mining, the Gini

index is dened as follows [44]:
G(M) = 1

n


p2i

i=1

where the variable M has the probability mass distribution {p1 , p2 ,

, pn }.
In this paper, we took the maximum value of the CWT coefcient matrix and its corresponding scale and position parameters
as the rst three features. The rst 10 largest singular values were
taken as the next 10 features. Regarding the squared elements of
an orthonormal vector as the components of a probability mass
function, the Shannon entropies calculated from the rst 10 left
singular vectors and the rst 10 right singular vectors from SVD of
the CWT coefcient matrix were taken as the next 20 features. The
Gini indices from these 20 left and right singular vectors were taken
as another 20 features. The absolute values of the rst 10 diagonal
elements in the upper triangular matrix R were taken as an additional 10 features. Similarly, Shannon entropy and Gini index were
calculated on each of the rst 10 column vectors of Q. The resulting
10 Shannon entropies and 10 Gini indices were taken as the nal 20
features. In total, 83 features were extracted based on CWT, SVD,
and QRD.
2.4. Feature selection
If all of the extracted features were used, it would require
a considerable amount of computational effort to conduct the
subsequent classication task. To decrease the computational complexity, an appropriate subset of features must be selected. The
sequential forward oating selection (SFFS) algorithm was applied
in this study. SFFS [45] is a feature selection algorithm that rst
adds features and then conditionally removes features. Specically,
in each iteration, SFFS rst adds a new feature into the current
selected feature subset, such that this addition is the best among
all of the possible additions based on some designated criterion;
then, it executes a number of steps of conditional exclusion based
on whether it is true or not that the removal of the least signicant feature improves the criterion value that was obtained from
the previous feature subset that has the same number of features.
If yes, remove the least signicant feature and continue to execute
the steps of conditional exclusion until further removal does not
improve the criterion any more. Then, return to the step of feature
addition. If not, go directly back to the feature addition step. The
SFFS algorithm is terminated when the required number of features
is satised or other terminating conditions are reached. Comparative studies [46,47] have shown that the SFFS outperformed other
prior feature selection methods on many datasets.
2.5. Classication
Because the goal of feature extraction is to distinguish innocent
murmurs from organic murmurs, there are only two classes: the
class that contains innocent murmurs and the class that contains
organic murmurs. We labeled the class of innocent murmurs as
1 and the class of organic murmurs as 0. In most of the existing
literature, articial neural networks (ANN) were adopted for classication. ANNs simulate the activities of human neural networks
and are very useful for modeling complex relationships between
inputs and outputs; therefore, they are suitable for classication.
However, ANNs involve many parameters, such as the number of
hidden layers, the number of neurons in each hidden layer, and
the transfer function in each layer. To obtain the best ANN, it takes
time to nd the most appropriate parameters for its structure. In
this study, we propose to instead use the technique of CART for

Please cite this article in press as: Chen Y, et al. Matrix decomposition based feature extraction for murmur classication. Med Eng Phys (2011),
doi:10.1016/j.medengphy.2011.09.020

G Model
JJBE-1986; No. of Pages 6

ARTICLE IN PRESS
Y. Chen et al. / Medical Engineering & Physics xxx (2011) xxxxxx

Features

Sources

13

The value of the maximum CWT coefcient and the

corresponding scale and position parameters

413

The rst 10 singular values from the SVD of the CWT

coefcient matrix

1423

The Shannon entropies derived from the rst 10 left

singular vectors based on the SVD of the CWT matrix

2433

The Shannon entropies derived from the rst 10 right

singular vectors based on the SVD of the CWT matrix

3443

The Gini indices derived from the rst 10 left singular

vectors based on the SVD of the CWT matrix

4453

The Gini indices derived from the rst 10 right singular

vectors based on the SVD of the CWT matrix

5463

The absolute values of the rst 10 diagonal elements of

R from the QRD of the CWT matrix

6473

The Shannon entropies derived from the rst 10

column vectors of Q based on the QRD of the CWT
matrix

7483

The Gini indices derived from the rst 10 column

vectors of Q based on the QRD of the CWT matrix

classication. CART [44,48] is a binary decision tree algorithm that

can be used for either classication or regression based on the category of the target variable. CART recursively partitions data into
two subsets with similar target values. It grows a tree by producing
exactly two branches for each decision node, based on a splitting
criterion for that node. The splitting criterion is searched among
all of the available variables and all of the possible splitting values,
such that the data in each subset is more pure than the original
data set. A classication tree from CART is related to an impurity
measure, which can be the Gini index or a towing index. The most
commonly used impurity measure is the Gini index. Because of
fewer parameters, CART provides better classication at a faster
speed than ANNs.
3. Results
The extracted features are summarized in Table 1. The rst
column lists feature numbers, and the second column shows the
sources of the corresponding features. In this study, when implementing the step of feature selection, we incorporated it with
the step of classication. Specically, the average classication
accuracy from 10-fold cross-validation of the CART algorithm
was taken as the criterion for the feature selection algorithm of
SFFS. This result occurs because the SFFS is closely related to the
choice of the criterion and our objective is to improve the classication accuracy. The impurity measure adopted for the CART
algorithm was the Gini index. Under this procedure, the following feature subset was selected: {4, 7, 8, 9, 10, 11, 14, 15, 18,
19, 23, 27, 29, 34, 35, 39, 40, 41, 44, 46, 53, 60, 65, 71, 73, 75}.
The rst three features were not selected, which implies that
the maximum CWT coefcient is not of signicance, and we do
not have to consider it. Among the 26 selected features, most
were extracted based on the SVD of the CWT coefcient matrix
while a few were derived from the QRD of the CWT coefcient
matrix.
Fig. 3 illustrates all of the iterations of the SFFS algorithm. We
chose the set of 26 features from iteration 704 because it provided
higher average classication accuracy with a smaller number of features. The average classication accuracy that was obtained from
a 10-fold cross-validation of the CART classication was 90%, with
an average sensitivity of 94% and an average specicity of 83%. The

100

100

50

80

60
0

100

200

300

400

500

Iteration

600

700

800

Number of Features

Table 1
Summary of feature extraction.

Criterion Value (%)

0
900

Fig. 3. The criterion value and the number of features in each SFFS iteration.

sensitivity is dened as the proportion of organic murmurs that

are correctly classied as such, and the specicity is dened as
the proportion of innocent murmurs that are correctly classied
as such. The results obtained are more accurate than the average
auscultation assessment of cardiologists, which is approximately
80% [14,16].
4. Discussion
In this study, we proposed a set of new features for differentiating innocent murmurs from organic murmurs. The
features were extracted from the singular value decomposition and QR decomposition of the continuous wavelet transform
coefcient matrix. The Shannon entropy and Gini index were
then applied on the derivation of orthonormal column vectors
from the matrix decomposition. The sequential forward oating
selection algorithm was used to select features and the classication algorithm of CART was utilized. An average accuracy
of 90%, a sensitivity of 94%, and a specication of 83% were
obtained.
Not only was accuracy much better than the average auscultation accuracy of cardiologists (i.e. 80%) but also the results
were comparable to previous studies. Andrisevic et al. [16] applied
wavelet analysis and articial neural networks to classify normal
and abnormal heart sounds, with a specicity of 70.5% and a sensitivity of 64.7%. Noponen et al. [17] used phono-spectrograms to
analyze heart murmurs, achieving 90% specicity and 91% sensitivity. Ahlstrom et al. [18] examined a large number of features
extracted from a wide range of areas, and the selected feature set
obtained an 86% classication accuracy. Ahlstrom et al. [24] proposed to apply recurrence quantication analysis to distinguish
innocent murmurs from murmurs caused by aortic stenosis, and
acquired a sensitivity of 90% and a specicity of 88%. Strunic et al.
[26] developed a system that can classify normal heart sounds, aortic stenosis, and aortic regurgitation with up to 85 7.4% accuracy
and 95 6.8% sensitivity. Choi and Jiang [30] considered six SVM
modules for murmur classication, with accuracy between 71% and
99.6%. Sengur [31] designed an expert system of heart disease diagnosis with 95.9% sensitivity and 96% specicity. Chauhan et al. [32]
discriminated normal from abnormal heart sounds, with a classication rate found to be 95.7% for continuous murmurs, 96.25%
for systolic murmurs, and 90% for diastolic murmurs. Maglogiannis
et al. [34] classied heart diseases into specic categories. Differentiation between normal and abnormal heart sounds is the rst
stage, with a classication accuracy of 91%. Yuenyong et al. [35]
provided a framework for automatic heart sound analysis without
segmentation. The classication accuracy is 92% for the noisefree case and 90% for both white noise and impulse noise. Chen

Please cite this article in press as: Chen Y, et al. Matrix decomposition based feature extraction for murmur classication. Med Eng Phys (2011),
doi:10.1016/j.medengphy.2011.09.020

G Model
JJBE-1986; No. of Pages 6

ARTICLE IN PRESS
Y. Chen et al. / Medical Engineering & Physics xxx (2011) xxxxxx

et al. [36] investigated childhood murmurs and specied musical

murmurs from other types of murmurs. They achieved 89.02% sensitivity, 84.76% specicity and 87.36% classication accuracy. The
datasets used in the literature were various and the classication
validation processes (e.g. 10-fold cross validation) were different;
therefore, it is not possible to decide which procedure was the best
for such studies in a strict sense. The above comparisons demonstrate that the proposed novel features in this paper perform well
and are complementary to current research for murmur differentiation.
The feature extraction, feature selection, and classication algorithms in this study are novel. Previous feature extraction methods
included time-frequency analysis [21], instantaneous energy and
frequency estimation [22], discrete wavelet transforms [23,35],
recurrence quantication analysis [24], discrete Fourier transforms and Burg autoregressive spectrum analysis [29], normalized
autoregressive power spectral density curves [30], wavelet packet
decomposition and entropy [31,33], and Mel-frequency cepstral
coefcients [32]. The most often used feature selection method
in the literature was principle component analysis [29,31,35].
The existing classication algorithms included articial neural
networks [16,18,20,26,28,29,33,35,36], support vector machines
[30,34,36], articial immune system and fuzzy k-NN [31], hidden
Markov models [32], and discriminant analysis [36]. In this paper,
the matrix decomposition based feature extraction method was
rst introduced. Singular value decomposition, QR decomposition,
Shannon entropy and Gini index were applied. The SFFS algorithm
was used for selecting features, and the CART was used for murmur
classication.
There are some aspects of this research that could be improved
or extended. First, we distinguished only innocent murmurs from
organic murmurs. It would be of further use to specify a certain type
of organic murmur, for example, aortic stenosis (AS), atrial septal defect (ASD), pulmonic stenosis (PS), ventricular septal defect
(VSD), or mitral stenosis (MS). Second, we manually segmented
heart sound recordings in this paper. In future research, we will
investigate the development of an automatic murmur segmentation algorithm, which would be an important building block for an
intelligent murmur differentiation system.
Conict of interest statement
None declared.
References
[1] Heron M, Hoyert D, Murphy S, Xu J, Kochanek K, Tejada-Vera B. Deaths: nal
data for 2006. National vital statistics reports, vol. 57, number 14. Hyattsville,
MD: National Center for Health Statistics; 2009.
[2] Lloyd-Jones D, Adams R, Brown T, Carnethon M, Dai S, De Simone G, et al. Heart
disease and stroke statistics2010 update: a report from the American Heart
Association. Circulation 2010;121(7):e46215.
[3] Myint W, Dillard B. An electronic stethoscope with diagnosis capability. In:
Proceedings of the 33rd southeastern symposium on system theory. 2001. p.
1337.
[4] Tilkian A, Conover M. Understanding heart sounds and murmurs: with an introduction to lung sounds. 4th ed. Philadelphia: Saunders; 2001.
[5] Pelech A. The physiology of cardiac auscultation. Pediatr Clin North Am
2004;51:151535.
[6] Noponen A, Lukkarinen S, Angerla A, Sikio K, Sepponen R. How to recognize the
innocent vibratory murmur? Comput Cardiol 2000;27:5614.
[7] Canent R. Recognizing common innocent heart murmurs in infants and children. Mo Med 1986;83(9):621.
[8] McCrindle B, Shaffer K, Kan J, Zahka K, Rowe S, Kidd I. An evaluation of
parental concerns and misperceptions about heart murmurs. Clin Pediatr
1995;34:2531.
[9] Geggel R, Horowitz L, Brown E, Parsons M, Wang P, Fulton D. Parental anxiety
associated with referral of a child to a pediatric cardiologist for evaluation of a
Stills murmur. J Pediatr 2002;140:74752.
[10] Haney I, Ipp M, Feldmen W, McCrindle B. Accuracy of clinical assessment of
heart murmurs by ofce-based (general practice) paediatricians. Arch Dis Child
1999;81(5):40912.

[11] Gaskin P, Owens S, Talner N, Sanders S, Li J. Clinical auscultation skills in pediatric residents. Pediatrics 2000;105:11847.
[12] Vukanovic-Criley J, Criley S, Warde C, Boker J, Guevara-Matheus L, Churchill W,
et al. Competency in cardiac examination skills in medical students, trainees,
physicians, and faculty: a multicenter study. Arch Intern Med 2006;166:6106.
[13] Pease A. If the heart could speak. Pictures of the Future 2001:601.
[14] Etchells E, Bell C, Robb K. Does this patient have an abnormal systolic murmur?
JAMA 1997;277:56471.
[15] Thompson W, Hayek C, Tuchinda C, Telford J, Lombardo J. Automated cardiac auscultation for detection of pathologic heart murmurs. Pediatr Cardiol
2001;22:3739.
[16] Andrisevic N, Ejaz K, Rios-Gutierrez F, Alba-Flores R, Nordehn G, Burns S. Detection of heart murmurs using wavelet analysis and articial neural networks. J
Biomed Eng 2005;127(6):899904.
[17] Noponen A, Lukkarinen S, Angerla A, Sepponen R. Phono-spectrographic analysis of heart murmur in children. BMC Pediatr 2007;7:23.
[18] Ahlstrom C, Hult P, Rask P, Karlsson J, Nylander E, Dahlstrm U, et al. Feature extraction for systolic heart murmur classication. Ann Biomed Eng
2006;34(11):166677.
[19] Reed T, Reed N, Fritzson P. Heart sound analysis for symptom detection and computer-aided diagnosis. Simul Model Practice Theory 2004;12:
12946.
[20] lmez T, Dokur Z. Classication of heart sounds using an articial neural network. Pattern Recognit Lett 2003;24:61729.
[21] White P, Collis W, Salmon A. Analysing heart murmurs using time-frequency
methods. In: Proceedings of the IEEE-SP international symposium on timefrequency and time-scale analysis. 1996. p. 3858.
[22] Sharif Z, Zainal M, Shaameri A, Salleh S. Analysis and classication of heart
sounds and murmurs based on the instantaneous energy and frequency estimations. In: Proceedings of Tencon, vol. 2. 2000. p. 1304.
[23] Omran S, Tayel M. A heart sound segmentation and feature extraction algorithm
using wavelets. In: Proceedings of the rst international symposium on control,
communications and signal processing. 2004. p. 2358.
[24] Ahlstrom C, Hglund K, Hult P, Hggstrm J, Kvart C, Ask P. Distinguishing
innocent murmurs from murmurs caused by aortic stenosis by recurrence
quantication analysis. Int J Biomed Sci 2006;1(1):2138.
[25] de Vos J, Blanckenberg M. Automated pediatric cardiac auscultation. IEEE Trans
Biomed Eng 2007;54(2):24452.
[26] Strunic S, Rios-Gutierrez F, Alba-Flores R, Nordehn G, Burns S. Detection and
classication of cardiac murmurs using segmentation techniques and articial neural networks. In: Proceedings of IEEE symposium on computational
intelligence and data mining. 2007. p. 397404.
[27] Gupta C, Palaniappan R, Swaninathan S, Krishnan S. Neural network classication of homomorphic segmented heart sounds. Appl Soft Comput
2007;7:28697.
[28] Babaei S, Geranmayeh A. Heart sound reproduction based on neural network
classication of cardiac valve disorders using wavelet transforms of PCG signals.
Comput Biol Med 2009;39(1):815.
[29] Uguz H. A biomedical system based on articial neural network and principal
component analysis for diagnosis of the heart valve diseases. J Med Syst 2010
[Epub ahead of print].
[30] Choi S, Jiang Z. Cardiac sound murmurs classication with autoregressive spectral analysis and multi-support vector machine technique. Comput Biol Med
2010;40(1):820.
[31] Sengur A. An expert system based on principal component analysis, articial immune system and fuzzy k-NN for diagnosis of valvular heart diseases.
Comput Biol Med 2008;38(3):32938.
[32] Chauhan S, Wang P, Sing Lim C, Anantharaman V. A computer-aided
MFCC-based HMM system for automatic auscultation. Comput Biol Med
2008;38(2):22133.
[33] Das R, Turkoglu I, Sengur A. Diagnosis of valvular heart disease through
neural networks ensembles. Comput Methods Programs Biomed 2009;93(2):
18591.
[34] Maglogiannis I, Loukis E, Zaropoulos E, Stasis A. Support vectors machinebased identication of heart valve diseases using heart sounds. Comput
Methods Programs Biomed 2009;95(1):4761.
[35] Yuenyong S, Nishihara A, Kongprawechnon W, Tungpimolrut K. A framework
for automatic heart sound analysis without segmentation. Biomed Eng Online
2011;10:13.
[36] Chen Y, Wang S, Shen C, Choy F. Intelligent specication of childhood musical
murmurs. J Healthcare Eng, in press.
[37] <http://earson.ca/what.html>.
[38] Ari S, Saha G. In search of an optimization technique for articial neural network
to classify abnormal heart sounds. Appl Soft Comput 2009;9:33040.
[39] Vetterli M, Kovacevic J. Wavelets and subband coding. NJ: Prentice Hall; 1995.
[40] Hassanpour H, Mesbah M, Boashash B. Timefrequency feature extraction of
newborn EEG seizure using SVD-based techniques. EURASIP J Appl Signal Processing 2004;16:254454.
[41] Nakos G, Joyner D. Linear algebra with applications. Pacic Grove, CA, USA:
Brooks/Cole Publishing Company; 1998.
[42] Gentle J. Numerical linear algebra for applications in statistics. Berlin: SpringerVerlag; 1998.
[43] Pierce J. An introduction to information theory: symbols, signals and noise. 2nd
ed. Dover Publications; 1980.
[44] Breiman L, Friedman J, Olshen R, Stone C. Classication and regression trees.
Chapman & Hall/CRC Press; 1984.

Please cite this article in press as: Chen Y, et al. Matrix decomposition based feature extraction for murmur classication. Med Eng Phys (2011),
doi:10.1016/j.medengphy.2011.09.020

G Model
JJBE-1986; No. of Pages 6
6

ARTICLE IN PRESS
Y. Chen et al. / Medical Engineering & Physics xxx (2011) xxxxxx

[45] Pudil P, Ferri F, Novovicov J, Kittler J. Floating search methods for feature
selection with nonmonotonic criterion functions. In: Proceedings of the 12th
international conference on pattern recognition. 1994. p. 27983.
[46] Jain A, Zongker D. Feature selection: evaluation, application, and small sample
performance. IEEE Trans Pattern Anal Mach Intell 1997;19:1538.

[47] Kudo M, Sklansky J. Comparison of algorithms that select features for pattern
classiers. Pattern Recognit 2000;33:2541.
[48] Larose D. Discovering knowledge in data: an introduction to data mining. John
Wiley & Sons Inc.; 2005.

Please cite this article in press as: Chen Y, et al. Matrix decomposition based feature extraction for murmur classication. Med Eng Phys (2011),
doi:10.1016/j.medengphy.2011.09.020