Content uploaded by Samit Kumar Ghosh
Author content
All content in this area was uploaded by Samit Kumar Ghosh on Feb 08, 2020
Content may be subject to copyright.
Automated Detection of Heart Valve Diseases using
Chirplet Transform and Multiclass Composite Classifier
with PCG Signals
Samit Kumar Ghosha, R N Ponnalagua, R K Tripathyaand U. Rajendra Acharya b,c,d
aDepartment of Electrical and Electronics Engineering, BITS-Pilani, Hyderabad Campus, Hyderabad 500078, India
bDepartment of Electronics and Computer Engineering, Ngee Ann Polytechnic, Singapore.
cDepartment of Bioinformatics and Medical Engineering, Asia University, Taichung, Taiwan.
dInternational Research Organization for Advanced Science and Technology, Kumamoto University, Kumamoto, Japan.
Abstract
Heart valve diseases (HVDs) are a group of cardiovascular abnormalities, and the causes of HVDs are blood clots, congestive
heart failure, stroke, and sudden cardiac death, if not treated timely. Hence, the detection of HVDs at the initial stage is
very important in cardiovascular engineering to reduce the mortality rate. In this article, we propose a new approach for
the detection of HVDs using phonocardiogram (PCG) signals. The approach uses the Chirplet transform (CT) for the time-
frequency (TF) based analysis of the PCG signal. The local energy (LEN) and local entropy (LENT) features are evaluated
from the TF matrix of the PCG signal. The multiclass composite classi
er formulated based on the sparse representation of the test PCG instance for each class and the distances from the nearest
neighbor PCG instances are used for the classi
cation of HVDs such as mitral regurgitation (MR), mitral stenosis (MS), aortic stenosis (AS), and healthy classes (HC)..
The experimental results show that the proposed approach has sensitivity values of 99.44%, 98.66%, and 96.22% respectively
for AS, MS and MR classes. The classification results of the proposed CT based features is compared with existing approaches
for the automated classification of HVDs. The proposed approach has obtained the highest overall accuracy as compared
to existing methods using the same database. The approach can be considered for the automated detection of HVDs with
Internet of Medical Things (IOMT) applications.
Key words: Heart Valve Diseases (HVDs), PCG, Chirplet Transform, Time-Frequency analysis, Multiclass Composite
Classifier.
Preprint submitted to Journal Name 27 January 2020
1 Introduction
Heart valve diseases (HVDs) have high mortality rates as compared to other cardiovascular diseases (CVDs) [1]
[2]. These diseases occur due to the damage in the heart valves. In the human heart, there are four valves aortic,
pulmonary, mitral and tricuspid valves which help to prevent the backward flow of blood [3]. The mechanical activity
of the heart such as the proper opening and closing of the heart valves are very important for better functioning of the
heart [4]. The medical practitioner uses a stethoscope to listen to the sound produced during the mechanical activity
of the heart. The phonocardiogram (PCG) is a graphical representation of the mechanical activity of the heart which
provides valuable information for the diagnosis of HVDs, congestive heart failure, and anatomical defects. Various
categories of HVDs such as aortic stenosis (AS), mitral regurgitation (MR), mitral stenosis (MS), etc. are diagnosed
using a PCG signal [5] [6]. The MR pathology occurs when the valve between the left atrium and right atrium of the
heart doesn’t close properly causing a backward flow of blood and may result in heart failure [7]. Similarly, the MS
pathology is mainly due to the improper opening of the mitral valve; as a result, the insufficient amount of blood
fows through the heart chambers [5]. The AS pathology restricts the blood flow from the left ventricle to the aorta
and occurs due to narrowing of the aortic valve. The handheld ultrasonic devices (HUDs), echocardiography, cardiac
computed tomography, and the cardiac magnetic resonance (MR) imaging modalities are used for the quantitative
assessment of various types of HVDs [8] [9] [10] [11]. The parameters such as valve area, average transvalvular
gradient, and maximum value of transvalvular velocity have been considered for grading (normal, severe, critical)
of each category of HVD [11]. These parameters are subjectively evaluated from the images of cardiac chambers.
In HUD, proper selection of depth and gain parameters are required to generate high-resolution images of cardiac
chambers for the diagnosis of HVDs [8] [12]. The aforementioned diagnostic modalities are costly, and proper training
is needed to use such devices for accurate detection of HVDs [8] [12]. The PCG is a low-cost diagnostic modality, and
it has been used for the detection of HVDs [13]. The clinical information associated with PCG signal for the diagnosis
of HVDs are systolic and diastolic duration, shapes of S1 and S2 sounds, amplitude of S1 and S2 sound, presence
of abnormal sounds and murmurs [13] [14]. The medical practitioner investigates this information while diagnosing
HVDs. However, in the intensive care unit (ICU), 24 hours continuous recording of PCG signal is considered for the
diagnosis of HVDs [15]. This type of recording produces a huge amount of PCG data, hence it is a cumbersome step
for the medical practitioners for decision making [16]. Therefore, the computer-aided diagnosis (CAD) framework
has been widely used for automated detection of HVDs from the PCG signal [13] [17]. The CAD framework has
mainly three stages: segmentation of cardiac cycles from the PCG signal, the extraction of relevant features, and
classification of HVDs [18] [19]. The extraction of features is very important as it extracts meaningful information
from the cardiac PCG cycle for automated detection of HVDs.
?Corresponding to: R N Ponnalagu (ponnalagu@hyderabad.bits-pilani.ac.in), Department of Electrical and Electronics En-
gineering, BITS-Pilani, Hyderabad Campus, Hyderabad 500078, India
2
In the literature, numerous approaches have been used for the extraction of features for automated detection of
HVDs and a review of these methods is avilable [20]. Machine learning techniques have been used to evaluate the
performance of PCG signal features [20] [21]. In [22], authors have extracted features from PCG signals using Mel
frequency cepstral coefficients (MFCC) and discrete wavelet transform (DWT) methods. Similarly, in [23], wavelet
packet decomposition (WPD) and continuous wavelet transform (CWT) based analysis of PCG signals have been
performed for the detection of HVDs. Moreover, many authors have computed the features in the time-frequency
(TF) domain of the PCG signal for the classification of HVDs [24][25][23][26]. The method proposed by authors in
[27] evaluated various statistical features from the subbands of PCG signal and furthermore they used these features
for the classification of normal and abnormal heart sounds. These features are extracted using a tunable Q wavelet
transform (TQWT) based analysis of PCG signals. Several recent studies have been considered which include the
fast Fourier transform (FFT) [28] [29], short-time Fourier transform (STFT) [30] [31], and TF decomposition (TFD)
[32] techniques for the detection of abnormal heart sounds. These methods have been used for classification models
such as probabilistic neural network (PNN) [29], support vector machine (SVM) [31] classifiers for the classification
of normal and pathological heart sounds. Various deep neural network architechures have also been used [33] [34]
for the detection of abnormal heart sounds. The deep learning approaches require a high volume of training data to
obtain better model parameters for PCG signals. The convolutional neural network (CNN) has yielded slightly lower
performance in the automated catagorization of normal and abnormal heart sounds [35] [33] [34]. Therefore, the
development of a feature-based approach using various advanced signal processing methods is likely to be important
for the accurate classification of HVDs.
In this study, we have considered the chirplet transform (CT) for the analysis of PCG signal and extraction of TF
domain features. It works well for non-stationary signals which have a chirp-like structure [36] [37]. The PCG signal
is a non-stationary signal and it contains multiple sound components such as S1, S2, murmurs, etc. [13]. Therefore,
we can expect that the TF matrix obtained using CT of the PCG signal will effectively assess the information
from this signal for the detection of HVDs. In this work, we have extracted local energy (LEN) and local entropy
(LENT) features from the TF matrix of the PCG signal. The LEN features have shown better performance for the
detection of cardiac ailments using the ECG signal [38]. Motivated from the work reported in [38], we have evaluated
LENT and LEN features from each frequency component of the TF matrix. The composite binary classifier formed
using sparse representation residual and nearest neighbour distances for each class has been used for the detection
of cardiac ailments using the ECG signal [38]. The advantage of the composite classifier is that, it doesn’t require
more instances during training like deep learning models [39]. However, it considers the training instances as a
dictionary for the sparse representation of the test feature vector and the decision-making process of this classifier
is fully distance-based which makes it simple for biomedical applications [38]. Motivated by these advantages, we
have formulated a multiclass composite classifier for the detection of HVDs from the TF domain local features of
the PCG signal. The remainder of this paper is organized as follows. In section 2, the proposed approach for the
3
detection of HVDs is described. In section 3, we present the results and also provide a description of the results. The
conclusions of this paper are provided in section 4.
2 Method
The step-by-step procedure of the proposed approach for the automated detection of HVDs is depicted in Fig.1. The
approach composed of four subsections such as the extraction of cardiac cycles from the PCG signal followed by TF
analysis of each cardiac cycle, extraction of local energy (LEN) and local entropy (LENT) based features from the
TF matrix and finally, the use of a multiclass composite classifier for the classification of HVDs. We have written
detail descriptions of each subsection as follows.
Fig. 1. Step-by-step procedure of the proposed approach for HVDs detection.
2.1 PCG Database and Preprocessing
In this work, we have obtained the PCG signals from a public database 1[22]. There are 1000 PCG recordings
present in the database for different subjects. Out of these 1000 recordings each class contains 200 recordings. There
are five classes of PCG signals given in the database namely the healthy control (HC), AS, MS, MR and mitral valve
collapse (MVP). Each class of PCG recording is sampled at a frequency of 8000 Hz. In this approach, a Butterworth
bandpass filter with cut-off frequencies of 25 Hz and 900 Hz are considered to filter each PCG recordings [40]. We
1https://github.com/yaseen21khan/Classification-of- Heart-Sound- Signal- Using-Multiple- Features-
4
have extracted cardiac PCG cycles from each PCG recording based on the segmentation approach. The segmentation
is done by extraction of the heart sound envelope in PCG recording using Shannon energy. Based on the systolic and
diastolic duration, we have segmented each PCG recording. After the segmentation, we have extracted 2400 cardiac
PCG cycles from 800 PCG recordings of HC, AS, MS and MR classes. The TF analysis of each of these 2400 cardiac
PCG cycles is performed using CT. In the following subsection, we have described the CT for the TF analysis of
each cardiac PCG cycle and feature extraction techniques used in this work.
2.2 Chirplet Transform (CT)
The CT is a TF analysis approach for the analysis of non-stationary signals. For a cardiac PCG cycle, x(n), with
n= 1,2,3........N, and Nis the length of the signal, the CT is defined as follows [36] [37]:
Tα,σ (k, n0) =
N
X
n=1
z(n)e−2πkn
NΨ∗
τ,α,σ (n) (1)
where z(n) is the analytical signal and it is evaluated using Hilbert transform (HT) of the cardiac PCG cycle. The
factor Ψ∗is the complex conjugate of Ψ. The analytical signal is given by z(n) = x(n) + jH [x(n)], where H[x(n)]
is the HT of the cardiac PCG cycle. The window function Ψ∗
n0,α,σ (n) used in CT is given as follows
Ψ∗
n0,α,σ(n) = wσ(n−n0)e−jα
2(n−n0)2(2)
Moreover, the factor wσ(n−n0) is a Gaussian window function and it is given by [36] [37]
ωσ(n) = 1
√2πσ e−n2
2σ2(3)
The TF matrix obtained using the CT of cardiac PCG cycle contains complex values and it can be represented as
Tα,σ (k, n0) = TR
α,σ (k, n0) + j T I
α,σ (k, n0) (4)
The magnitude of TF matrix of cardiac PCG cycle is given as
Tm
α,σ(k , n0) = qTR
α,σ(k , n0)2+TI
α,σ(k , n0)2(5)
In this study, the magnitude of the TF matrix is evaluated for each cardiac cycle of HC, AS, MS and MR classes.
For HC, MS, MR and AS classes, the cardiac PCG cycles are shown in Fig. 2(a)-(d), respectively. It can be observed
that the temporal and spatial characteristics of each type of pathological PCG cycle are different as compared to
the normal PCG signal. The variations are seen in the TF contour plots (as shown in fig.2 (e)-(h)) for HC, AS, MS
and MR classes. Due to pathology, there are clear visual variations in the TF characteristics of the PCG signal.
5
0 2000 4000
Samples
-1
-0.5
0
0.5
1
Amplitude
1000 2000 3000 4000
Samples
0
200
400
frequency
0.01
0.02
0.03
0 2000 4000
samples
-1
-0.5
0
0.5
1
1000 2000 3000 4000
Samples
0
200
400
0.01
0.02
0.03
0.04
0 2000 4000
Samples
-1
-0.5
0
0.5
1
1000 2000 3000 4000
Samples
0
200
400
0.01
0.02
0.03
0 2000 4000
Samples
-0.2
-0.1
0
0.1
0.2
1000 2000 3000 4000
samples
0
200
400
2
4
6
10-3
(a) (b) (c) (d)
(e) (f) (g) (h)
Fig. 2. (a) PCG signal for HC class (b) PCG signal for AS class (c) PCG signal for MS class (d) PCG signal for MR class (e)
TF contour plot for HC class (f) TF contour plot for AS class (g) TF contour plot for MS class (h) TF contour plot for MR
class.
Therefore, the features evaluated using the TF matrix of cardiac PCG cycles can be used for the detection of HVDs.
In this study, we have evaluated local energy and entropy features from the time frequency matrix. The local energy
feature for the kth frequency component is given by
LENk=
N
X
n=1
Tm
α,σ (k, n0)
2(6)
Similarly, the local entropy (LENT) is evaluated using the histogram of kth frequency component and it is denoted as
hb(k) where bis bin number b= 1,2,3.........B [41]. In this work, we have used total bins as 10. Thus, the probability
value for kth frequency component is evaluated as follows [38]:
Pb(k) = hb(k)
PB
b=1 hb(k)(7)
The LENT for kth frequency component is given by
LENTk=−
B
X
b=1
Pb(k)log2[Pb(k)] (8)
In this work, we have extracted LEN features from first 200 frequency components of the TF matrix of PCG
signal. Similarly, the LENT are extracted from first 100 frequency components. Thus, a 300 dimensional feature
vector is formulated for a single cardiac PCG cycle. For all cardiac PCG cycles, the feature matrix is formulated
6
as F∈R2400×300. In the following subsection, the multiclass composite classifier is used for the detection of HVDs
from 3000 dimensional feature vectors of cardiac PCG cycles.
2.3 Multiclass composite Classifier
The multiclass composite classifier is designed using the sparse representation classifier (SRC) for each class and
distances from the nearest neighbours for both normal and pathological classes. The feature matrix is denoted as
F=Rp×z. where pis the number of PCG instances and zcontains the number of TF features. Similarly, the class
label vector for the proposed work is denoted as y∈Rp, with each yi∈ {1,2,3,4}, where class labels such as 1, 2, 3,
4 are assigned for HC, AS, MS and MR classes, respectively. The training and test feature matrices for the proposed
multiclass composite classifier is selected using hold-out and 10-fold cross-validation techniques. The training feature
matrices for HC, AS, MS, MR classes are given as Ftr
l, where l= 1,2,3,4. Similarly, the test feature matrix for
the lth class is denoted as Fte
l. Moreover, the training and test class labels for all four classes are given as ytr
land
yte
l, respectively. For the lth class, the size of training feature matrix is Ftr
l=Rpltr×z. The size of class label for
the lth class is ytr
l∈Rptr
l. In multiclass SRC, the training matrix for each class is considered as a dictionary for the
sparse representation of the test PCG instance. Let’s consider a test instance, fte for the evaluation of the multiclass
composite classifier. This test PCG instance or feature vector is considered as the linear combination of the training
PCG feature vectors for the lth class and it is mathematically written as follows [39]:
fte =γ1lftr
1l+γ2lftr
2l+γ3lftr
3l+......γplftr
pl (9)
if we consider all classes, then fte =γFtr is the representation of the test instance. where, γis termed as the weight
vector and it is given by γ= [γ1γ2γ3γ4]. Where, γ1,γ2,γ3,γ4are the weight vector for the HC, AS, MS, MR
classes. The weight vector γis evaluated based on the optimized problem as follows [38] [42]:
γ= arg min
γkγk0(10)
subject to fte =γFtr . where, kγk0is the L0norm and it measures the non-zero entries in the γvector. The orthogonal
matching pursuit (OMP) is normally used to evaluate the γvector iteratively as the original optimization task is
NP-hard. In OMP algorithm, only those training instances for the lth class is selected which has maximum inner
product with the lth class residual of the test PCG instance. Initially the residual for lth class is selected as the test
PCG instance itself. This algorithm is terminated if the desired sparsity level in the γvector for lth class is achieved.
Therefore the residual of the lth class for the PCG instances fte can be written as Rl=kfte −γlFtr
lk2. where γland
Ftr
lare the weight vector and feature matrix for lth class. For each PCG test instance, four residual vectors such as
R1,R2,R3,R4are evaluated. Similarly, the distance between the test PCG instance and the ith training instance
for the is evaluated as follows [38] [42]:
di
l=
fte −ftri
l
2(11)
7
The average distance value over all the training PCG instance for lth class is evaluated as follows [42]:
Dl=1
ptr
l
ptr
l
X
i=1
di
l(12)
For the composite classifier, the total distance measure is evaluated for the lth class as Tl=Rl+Dl. Hence the final
class level for the test PCG instance is computed as
yte = arg min
l∈1,2,3,4(Tl) (13)
The performance of the multiclass composite classifier is evaluated using different measures for all the test feature
vectors of PCG instances. These measures are precision, sensitivity, specificity, and F-score of each class and the
overall accuracy (OA) [43]. The confusion matrix for the classification of four categories of HVDs is given as follows:
C=
C11 C12 C13 C14
C21 C22 C23 C24
C31 C32 C33 C34
C41 C42 C43 C44
(14)
The precision, the sensitivity, the specificity, the F-score and the accuracy are defined mathematically as follows [44]:
Precisionl(%) = TPl
TPl+ FPl×100 (15)
Sensitivityl(%) = TPl
TPl+ FNl×100 (16)
Specificityl(%) = TNl
FPl+ TNl×100 (17)
F-scorel(%) = 2∗TPl
2∗TPl+ FPl+ FNl×100 (18)
OA (%) = P4
i=1 Cii
P4
i=1 P4
j=1 Cij ×100 (19)
where, TPl, TNl, FPl, FNlare the true positive, true negative, false positive and false negative for the lth class,
respectively.
3 Results and discussion
In this section, the statistical analysis of proposed LEN and LENT features obtained from PCG signals of HC, AS,
MS and MR classes are performed. The performance of the multiclass composite classifier is shown for hold-out and
10-fold cross-validation techniques. For selected LEN and LENT features, we have shown the intra-class variations
8
1234
Classes
0
0.2
0.4
Feat 1
1234
Classes
0
0.2
0.4
0.6
Feat 2
1234
Classes
0
0.5
1
Feat 15
1 2 3 4
Classes
0
0.5
1
1.5
Feat 30
1234
Classes
0
0.01
0.02
Feat 115
1234
Classes
0
0.005
0.01
Feat 145
1234
Classes
0
0.5
1
Feat 175
10-3
1 2 3 4
Classes
0
0.5
1
Feat 195
10-3
1234
Classes
0
1
2
3
Feat 255
1234
Classes
0
1
2
3
Feat 265
1234
Classes
0
1
2
3
Feat 275
1234
Classes
0
1
2
3
Feat 300
(a) (b) (c) (d)
(e) (f) (g) (h)
(i) (j) (k) (l)
Fig. 3. (a) Boxplot for 1st LEN feature (feat 1) for all classes. (b) Boxplot for 2nd LEN feature (feat 2) for all classes. (c)
Boxplot for 15th LEN feature (feat 15) for all classes. (d) Boxplot for 30th LEN feature (feat 30) for all classes. (e) Boxplot
for 115th LENT feature (feat 115) for all classes. (f) Boxplot for 145th LENT feature (feat 145) for all classes. (g) Boxplot
for 175th LEN feature (feat 175) for all classes. (h) Boxplot for 195th LEN feature (feat 195) for all classes. (i) Boxplot for
55th LENT feature (feat 255) for all classes. (j) Boxplot for 65th LENT feature (feat 265) for all classes.(k) Boxplot for 275th
LENT feature (feat 75) for all classes. (l) Boxplot for 100th LENT feature (feat 300) for all classes.
using boxplots in Fig. 3. The boxplots for 1st, 2nd , 15th, 30th , 115th, 145th, 175th , 195th LEN features are depicted in
Fig. 3 (a)-(h). Similarly, in Fig. 3 (i)-(l), we have shown the boxplot for 55th , 65th, 75th , 100th LENT features. The
mean and standard deviation values of selected features for HC, AS, MS and MR classes are shown in Table 1. It is
evident that feat 1, feat 2, feat 15 and feat 115 have lower mean value for the pathological classes as compared to the
HC class. The low pitch sound is present in the PCG signal during AS pathology. Similarly, during MS pathology,
the low pitch (murmur) is present [45]. In the MR class, the diastolic component has less amplitude with multiple
splits. These pathological changes affect TF feature values of the PCG signal [46]. Due to these reasons, the mean
and standard deviation values are different for HC, AS, MS and MR classes. The LENT features measure the non-
linearity in the frequency components of the TF matrix of the PCG signal. The mean and standard deviation values
of LENT features for 55th, 65th , 75th and 100th frequency components are less for the normal class as compared to
the pathological class. The LENT features in the TF domain correctly measure the non-linearity in the PCG signal
during various pathological conditions. Due to this reason, LENT features have different mean values. We have also
shown the significance of LEN and LENT features using the analysis of variance (ANOVA) test in Table 1 [47]. It
can be observed that all 300 features have p-values less than <0.001 and have found to be statistically significant
9
Table 1
Results of statistical analysis for the selected LEN and LENT features for four classes.
Feat
Number
HC
(µ±σ)
AS
(µ±σ)
MS
(µ±σ)
MR
(µ±σ)
p-value
using ANOVA
Feat 1 0.18 ±0.05 0.11 ±0.04 0.06 ±0.08 0.09 ±0.04 <0.001
Feat 2 0.24 ±0.06 0.13 ±0.05 0.06 ±0.09 0.11 ±0.05 <0.001
Feat 15 0.58 ±0.18 0.48 ±0.16 0.21 ±0.22 0.36 ±0.15 <0.001
Feat 30 0.56 ±0.17 0.78 ±0.29 0.32 ±0.20 0.45 ±0.17 <0.001
Feat 115 0.07 ×10−2±0.10 ×10−20.02 ±0.06 0.01 ±0.05 0.05 ×10−2±0.08 ×10−2<0.001
Feat 145 0.03 ×10−2±0.04 ×10−20.44 ×10−2±0.01 0.31 ×10−2±0.01 0.01 ×10−2±0.02 ×10−2<0.001
Feat 175 0.08 ×10−3±0.02 ×10−20.98 ×10−3±0.35 ×10−20.47 ×10−3±0.18 ×10−20.04 ×10−3±0.01 ×10−2<0.001
Feat 195 0.05 ×10−3±0.01 ×10−20.51 ×10−3±0.17 ×10−20.14 ×10−3±0.05 ×10−20.02 ×10−3±0.01 ×10−2<0.001
Feat 255 0.82 ±0.21 2.20 ±0.32 1.99 ±0.37 1.25 ±0.33 <0.001
Feat 265 0.55 ±0.21 1.76 ±0.44 1.73 ±0.38 0.83 ±0.29 <0.001
Feat 275 0.44 ±0.18 1.35 ±0.58 1.46 ±0.40 0.52 ±0.23 <0.001
Feat 300 0.12 ±0.16 0.66 ±0.73 0.84 ±0.50 0.09 ±0.13 <0.001
Table 2
Results of classification using multiclass composite classifier with hold-out cross- validation.
Cross
validation
Measures (%) HC AS MR MS OA (%)
Hold-out
Precision 99.77 ±0.30 97.40 ±1.47 98.41 ±0.25 97.05 ±0.61
98.33%
Sensitivity 98.22 ±0.60 99.44 ±0.39 96.22 ±1.77 98.66 ±0.63
Specificity 99.92 ±0.10 99.09 ±0.52 99.47 ±0.08 98.99 ±0.21
F-score 98.99 ±0.31 98.4±0.75 97.29 ±0.96 97.85 ±0.52
for the catagorization of HVDs.
The multiclass composite classifier performance results for hold-out cross-validation is shown in Table 2. In this study,
the feature vectors for 70% PCG cycles from the feature matrix are considered for training while the remaining 30%
PCG cycles are evaluated during the testing process. It is evident from Table 2 that the sensitivity, specificity,
precision and F-score values are more than 96% for all classes. Similarly, the overall accuracy (OA) value for the
10
Table 3
Results of classification using multiclass composite classifier with ten-fold cross- validation.
HVDs
Measures
(%)
Fold 1Fold 2Fold 3Fold 4Fold 5Fold 6Fold 7Fold 8Fold 9Fold 10 Average
HC
Precision 98.23 100 100 100 100 100 100 100 100 100 99.82 ±0.55
Sensitivity 98.33 98.33 98.33 98.33 95.00 98.33 100 100 98.33 100 98.49 ±1.45
Specificity 100 100 99.44 100 100 100 100 100 100 100 99.94 ±0.17
F-score 100 100 100 97.44 99.16 100 99.16 100 100 97.48 99.32 ±1.04
AS
Precision 98.36 98.36 100 98.36 93.75 95.16 96.77 96.77 96.77 96.77 97.10 ±1.78
Sensitivity 100 100 100 100 98.33 98.33 100 100 100 100 99.66 ±0.70
Specificity 99.44 97.77 98.30 99.44 100 99.44 98.88 100 98.31 98.88 99.04 ±0.75
F-score 98.36 96.67 100 96.00 98.36 97.56 100 99.16 99.17 98.33 98.36 ±1.32
MR
Precision 96.77 98.36 98.33 100 100 100 100 98.28 95.16 100 98.69 ±1.67
Sensitivity 98.33 96.67 98.33 91.67 98.33 96.67 88.33 96.67 100 98.33 96.33 ±3.58
Specificity 99.44 100 99.44 97.77 100 99.44 100 100 98.86 100 99.49 ±0.71
F-score 98.31 95.87 99.16 94.92 98.31 97.44 98.33 100 99.16 97.44 97.99 ±1.60
MS
Precision 98.31 100 95.16 98.36 93.65 95.24 95.24 98.36 100 98.36 97.26 ±2.24
Sensitivity 98.33 100 98.33 96.67 98.33 98.33 98.33 100 100 100 98.83 ±1.12
Specificity 98.88 99.43 98.86 99.44 100 99.44 98.88 98.89 99.43 99.44 99.26 ±0.37
F-score 100 99.16 99.17 96.67 97.52 100 99.17 99.17 100 95.16 98.60 ±1.62
All OA 99.58 97.91 98.75 99.58 97.91 99.16 99.58 99.16 97.50 96.25 98.54 ±1.11
multiclass composite classifier is 98.33%. These high values for all performance measures reveal that the proposed
CT based TF domain features correctly measure the physiological information in the PCG signal for the detection
of HVDs. The robustness of the multiclass composite classifier is also verified using 10-fold cross-validation based
selection of the feature vectors for a PCG signal from the feature matrix. The performance measures such as the
precision, sensitivity, specificity, F-score values for each class and OA value for each fold are shown in Table 3. It is
evident that the specificity and sensitivity values are more than 98% in more than 9 folds for MS and AS classes.
For MR class, the sensitivity values are higher than 98% for five-folds and for the other five-folds, the sensitivity
values lie between 88% and 97%. The OA values of the multiclass composite classifier are more than 97% for nine
folds and only in the 10th fold the OA value is 96.25%. The average OA value is found to be 98.54%. From the above
11
observation, it is evident that the LEN and LENT features efficiently quantify clinical information in the PCG signal
for the automated assessment of HVDs.
Table 4
Comparison of proposed work with existing HVDs detection systems using PCG signals.
Authors Feature extraction method Classifcation method OA (%) Database
used
Yaseen et al.[22] Features evaluated using MFCC of PCG signal Support vector machine
(SVM)
91.60 Github
database
Yaseen et al.[22]
Features evaluated using Discrete wavelet
transform (DWT) of PCG signal
SVM 92.30 Github
database
Yaseen et al.[22] Features evaluated using MFCC and DWT of
PCG signal
SVM 97.90 Github
database
Patidar et al. [27]
Statistical features evaluated using Tunable
Q-Factor Wavelet Transform (TQWT)
Least square SVM 94.01 Public
database
Ari et al. [48] Features evaluated in wavelet domain Least square SVM 91.96 Own database
Li et al. [49] Features based on wavelet packet norm Twin-SVM 85.50 PhysioNet/CinC
Challenge
2016
Ghosh et al. [50]
Magnitude and phase features based on
Synchrosqueezing transform
Random Forest 95.13 Github
database
Safara et al. [51] Wavelet based features SVM with multi-level bias
selection
97.56 Own database
Zheng et al. [52] Energy fraction and sample entropy based fea-
tures
SVM 97.17 Own database
Proposed work LEN and LENT based features using CT of PCG
cycle
Multiclass composite classi-
fier
98.33 Github
database
The objective of this work is to classify various kinds of HVDs from PCG signals using TF based analysis and
multiclass composite classification approach. The LEN and LENT features are evaluated from the TF matrices
to capture the pathological changes in the PCG signals. It has been observed from statistical analysis that, the
proposed LEN and LENT features are discriminative and these features have yielded higher performance for the
detection of HVDs with multiclass composite classifier. Moreover, the proposed work is compared with various
existing techniques for the automated detection of HVDs and the results are shown in Table 4. The method reported
by Yaseen et al. [22] extracted both MFCC and DWT based features from PCG signals and used SVM classification
12
model for the detection of HVDs. From their study, OA values of 91.60% and 92.30% have been reported using
MFCC and DWT features with SVM classifiers. However, combining both MFCC and DWT features have yielded
an OA value of 97.90% in classifying different kinds of HVDs. Similarly, in [50], authors have considered the wavelet
synchrosqueezing transform (WSST) based TF analysis of PCG signals and random forest (RF) classification model
for the classification of different categories of HVDs. They have reported an OA of 95.13% using the same database as
mentioned in [22]. Similarly, Zheng et al. [52] have extracted fraction energy and sample entropy features from PCG
signals and used the SVM model for the classification of normal and abnormal PCG signals. They have obtained an
OA of 97.17% using an SVM classifier. Moreover, other works reported in [51] [48] [27] have computed the wavelet
domain features and used SVM classifiers for the detection of pathological HVDs. However, the proposed method
has yeilded higher performance as compared to these approaches. The advantages of the proposed work are as
follows. We have used novel CT for the TF analysis of PCG signal and non-linear LEN and LENT features for the
classification. A combination of sparse representation based residual and nearest neighbor distances for each class is
used for the classification of HVDs. The proposed classification technique is simple as only distances and residuals
are used. The automated approach developed can be implemented in real-time for the internet-based medical things
(IOMT) applications. The wavelet-based filter-bank approaches can be used for the analysis and classification of
PCG signals [53] [54]. The deep learning methods such as autoencoders [55], convolutional autoencoders [56], and
sequence models [57] can also be used for the evaluation of deep coded features from the TF matrices of PCG
signals for the detection of HVDs. Other diagnostic modalities such as the handheld ultrasonic devices (HUDs),
MRI imaging and echocardiography have been used in the clinical standard for the diagnosis of HVDs [8]. However,
the proposed automated approach uses PCG signals which can be recorded easily using the digital stethoscope.
The proposed automated approach can be implemented in real-time in an embedded platform for the detection and
classification of HVDs.
4 Conclusion
The detection of HVDs based on the TF analysis of the PCG signal has been proposed in this work. The TF matrix
has been evaluated using the CT of the cardiac PCG cycle. The local features such as LEN and LENT have been
computed from the TF matrix. A multiclass composite classifier has been proposed for the detection of HVDs using
LEN and LENT features extracted from PCG signal in the TF domain. This multiclass composite classifier has been
designed using the class-specific residual from SRC and class-specific nearest neighbour distance. The proposed time
frequency-based feature extraction scheme has shown better performance (OA of 98.33%) for the detection of HVDs
using the multiclass composite classifier. In the future, new TF analysis methods can be developed for the extraction
of features from the PCG signal. The deep-learning-based techniques can also be used for the detection of HVDs
from the PCG signal in the TF domain.
13
References
[1] M. D. Combs and K. E. Yutzey, “Heart valve development: regulatory networks in development and disease,” Circulation research,
vol. 105, no. 5, pp. 408–421, 2009.
[2] D. Mozaffarian, E. J. Benjamin, A. S. Go, D. K. Arnett, M. J. Blaha, M. Cushman, S. De Ferranti, J.-P. Despr´es, H. J. Fullerton,
V. J. Howard, et al., “Executive summary: heart disease and stroke statistics—2015 update: a report from the american heart
association,” Circulation, vol. 131, no. 4, pp. 434–441, 2015.
[3] M. Legato, “Gender and the heart: sex-specific differences in normal anatomy and physiology.,” The journal of gender-specific
medicine: JGSM: the official journal of the Partnership for Women’s Health at Columbia, vol. 3, no. 7, pp. 15–18, 2000.
[4] E. Nevo, “Method and apparatus for the assessment and display of variability in mechanical activity of the heart, and enhancement
of ultrasound contrast imaging by variability analysis,” May 1 2001. US Patent 6,224,553.
[5] E. Rapaport, “Natural history of aortic and mitral valve disease,” The American journal of cardiology, vol. 35, no. 2, pp. 221–227,
1975.
[6] S. Sell and R. E. Scully, “Aging changes in the aortic and mitral valves: histologic and histochemical studies, with observations on
the pathogenesis of calcific aortic stenosis and calcification of the mitral annulus,” The American journal of pathology, vol. 46, no. 3,
p. 345, 1965.
[7] M. Enriquez-Sarano, V. T. Nkomo, and H. I. Michelena, “Mitral regurgitation,” in Valvular Heart Disease, pp. 221–246, Springer,
2009.
[8] K. Savino and G. Ambrosio, “Handheld ultrasound and focused cardiovascular echography: Use and information,” Medicina, vol. 55,
no. 8, p. 423, 2019.
[9] K. Maganti, V. H. Rigolin, M. E. Sarano, and R. O. Bonow, “Valvular heart disease: diagnosis and management,” in Mayo Clinic
Proceedings, vol. 85, pp. 483–500, Elsevier, 2010.
[10] L. Chaothawee, “Diagnostic approach to assessment of valvular heart disease using mri—part i: a practical approach for valvular
regurgitation,” Heart Asia, vol. 4, no. 1, pp. 38–43, 2012.
[11] J. F. Glockner, D. L. Johnston, and K. P. McGee, “Evaluation of cardiac valvular disease with mr imaging: qualitative and
quantitative techniques,” Radiographics, vol. 23, no. 1, pp. e9–e9, 2003.
[12] F. M. C´aceres-L´origa, H. P´erez-L´opez, J. Santos-Gracia, and K. Morlans-Hernandez, “Prosthetic heart valve thrombosis:
pathogenesis, diagnosis and management,” International journal of cardiology, vol. 110, no. 1, pp. 1–6, 2006.
[13] V. N. Varghees and K. Ramachandran, “A novel heart sound activity detection framework for automated heart sound analysis,”
Biomedical Signal Processing and Control, vol. 13, pp. 174–188, 2014.
[14] M. S. Manikandan and K. Soman, “Robust heart sound activity detection in noisy environments,” Electronics letters, vol. 46, no. 16,
pp. 1100–1102, 2010.
[15] R. Acharya, S. M. Krishnan, J. A. Spaan, and J. S. Suri, Advances in cardiac signal processing. Springer, 2007.
[16] R. J. Martis, U. R. Acharya, C. M. Lim, K. Mandana, A. K. Ray, and C. Chakraborty, “Application of higher order cumulant
features for cardiac health diagnosis using ecg signals,” International journal of neural systems, vol. 23, no. 04, p. 1350014, 2013.
[17] R. L. Watrous, “Computer-aided auscultation of the heart: From anatomy and physiology to diagnostic decision support,” in 2006
International Conference of the IEEE Engineering in Medicine and Biology Society, pp. 140–143, IEEE, 2006.
14
[18] D. B. Springer, L. Tarassenko, and G. D. Clifford, “Logistic regression-hsmm-based heart sound segmentation,” IEEE Transactions
on Biomedical Engineering, vol. 63, no. 4, pp. 822–832, 2015.
[19] T. R. Reed, N. E. Reed, and P. Fritzson, “Heart sound analysis for symptom detection and computer-aided diagnosis,” Simulation
Modelling Practice and Theory, vol. 12, no. 2, pp. 129–146, 2004.
[20] A. K. Dwivedi, S. A. Imtiaz, and E. Rodriguez-Villegas, “Algorithms for automatic analysis and classification of heart sounds–a
systematic review,” IEEE Access, vol. 7, pp. 8316–8345, 2018.
[21] D. B. Springer, L. Tarassenko, and G. D. Clifford, “Support vector machine hidden semi-markov model-based heart sound
segmentation,” in Computing in Cardiology 2014, pp. 625–628, IEEE, 2014.
[22] G.-Y. Son, S. Kwon, et al., “Classification of heart sound signal using multiple features,” Applied Sciences, vol. 8, no. 12, p. 2344,
2018.
[23] M. Abdollahpur, A. Ghaffari, S. Ghiasi, and M. J. Mollakazemi, “Detection of pathological heart sounds,” Physiological measurement,
vol. 38, no. 8, p. 1616, 2017.
[24] S. Sun, H. Wang, Z. Jiang, Y. Fang, and T. Tao, “Segmentation-based heart sound feature extraction combined with classifier models
for a vsd diagnosis system,” Expert Systems with Applications, vol. 41, no. 4, pp. 1769–1780, 2014.
[25] A. Moukadem, A. Dieterlen, N. Hueber, and C. Brandt, “A robust heart sounds segmentation module based on s-transform,”
Biomedical Signal Processing and Control, vol. 8, no. 3, pp. 273–281, 2013.
[26] I. Maglogiannis, E. Loukis, E. Zafiropoulos, and A. Stasis, “Support vectors machine-based identification of heart valve diseases
using heart sounds,” Computer methods and programs in biomedicine, vol. 95, no. 1, pp. 47–61, 2009.
[27] S. Patidar and R. B. Pachori, “Classification of cardiac sound signals using constrained tunable-q wavelet transform,” Expert Systems
with Applications, vol. 41, no. 16, pp. 7161–7170, 2014.
[28] R. Sara¸cO˘gLu, “Hidden markov model-based classification of heart valve disease with pca for dimension reduction,” Engineering
Applications of Artificial Intelligence, vol. 25, no. 7, pp. 1523–1528, 2012.
[29] F. Beritelli, G. Capizzi, G. L. Sciuto, C. Napoli, and F. Scaglione, “Automatic heart activity diagnosis based on gram polynomials
and probabilistic neural networks,” Biomedical engineering letters, vol. 8, no. 1, pp. 77–85, 2018.
[30] A. Quiceno-Manrique, J. Godino-Llorente, M. Blanco-Velasco, and G. Castellanos-Dominguez, “Selection of dynamic features based
on time–frequency representations for heart murmur detection from phonocardiographic signals,” Annals of biomedical engineering,
vol. 38, no. 1, pp. 118–137, 2010.
[31] S. Kang, R. Doroshow, J. McConnaughey, and R. Shekhar, “Automated identification of innocent still’s murmur in children,” IEEE
Transactions on Biomedical Engineering, vol. 64, no. 6, pp. 1326–1334, 2016.
[32] M. N. Homsi and P. Warrick, “Ensemble methods with outliers for phonocardiogram classification,” Physiological measurement,
vol. 38, no. 8, p. 1631, 2017.
[33] T.-E. Chen, S.-I. Yang, L.-T. Ho, K.-H. Tsai, Y.-H. Chen, Y.-F. Chang, Y.-H. Lai, S.-S. Wang, Y. Tsao, and C.-C. Wu, “S1 and s2
heart sound recognition using deep neural networks,” IEEE Transactions on Biomedical Engineering, vol. 64, no. 2, pp. 372–380,
2016.
[34] V. Maknickas and A. Maknickas, “Recognition of normal–abnormal phonocardiographic signals using deep convolutional neural
networks and mel-frequency spectral coefficients,” Physiological measurement, vol. 38, no. 8, p. 1671, 2017.
[35] O. Faust, Y. Hagiwara, T. J. Hong, O. S. Lih, and U. R. Acharya, “Deep learning for healthcare applications based on physiological
signals: A review,” Computer methods and programs in biomedicine, vol. 161, pp. 1–13, 2018.
15
[36] S. Mann and S. Haykin, “The chirplet transform: Physical considerations,” IEEE Transactions on Signal Processing, vol. 43, no. 11,
pp. 2745–2761, 1995.
[37] Z. Peng, G. Meng, F. Chu, Z. Lang, W. Zhang, and Y. Yang, “Polynomial chirplet transform with application to instantaneous
frequency estimation,” IEEE Transactions on Instrumentation and Measurement, vol. 60, no. 9, pp. 3222–3229, 2011.
[38] R. K. Tripathy, M. R. Paternina, J. G. Arrieta, A. Zamora-Mendez, and G. R. Naik, “Automated detection of congestive heart
failure from electrocardiogram signal using stockwell transform and hybrid classification scheme,” Computer methods and programs
in biomedicine, vol. 173, pp. 53–65, 2019.
[39] M. Cui and S. Prasad, “Class-dependent sparse representation classifier for robust hyperspectral image classification,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 53, no. 5, pp. 2683–2695, 2014.
[40] A. Castro, T. T. Vinhoza, S. S. Mattos, and M. T. Coimbra, “Heart sound segmentation of pediatric auscultations using wavelet
analysis,” in 2013 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC),
pp. 3909–3912, IEEE, 2013.
[41] V. K. Sudarshan, E. Ng, U. R. Acharya, R. S. Tan, S. M. Chou, and D. N. Ghista, “Infarcted left ventricle classification from
cross-sectional echocardiograms using relative wavelet energy and entropy features,” Journal of Mechanics in Medicine and Biology,
vol. 16, no. 01, p. 1640009, 2016.
[42] R. Rosas-Romero, A. D´ıaz-Torres, and G. Etcheverry, “Forecasting of stock return prices with sparse representation of financial time
series over redundant dictionaries,” Expert Systems with Applications, vol. 57, pp. 37–48, 2016.
[43] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern classification. John Wiley & Sons, 2012.
[44] T. Kautz, B. M. Eskofier, and C. F. Pasluosta, “Generic performance measure for multiclass-classifiers,” Pattern Recognition, vol. 68,
pp. 111–125, 2017.
[45] S. Debbal and F. Bereksi-Reguig, “Computerized heart sounds analysis,” Computers in biology and medicine, vol. 38, no. 2, pp. 263–
280, 2008.
[46] V. Nigam and R. Priemer, “Accessing heart dynamics to estimate durations of heart sounds,” Physiological measurement, vol. 26,
no. 6, p. 1005, 2005.
[47] A. Cuevas, M. Febrero, and R. Fraiman, “An anova test for functional data,” Computational statistics & data analysis, vol. 47,
no. 1, pp. 111–122, 2004.
[48] S. Ari, K. Hembram, and G. Saha, “Detection of cardiac abnormality from pcg signal using lms based least square svm classifier,”
Expert Systems with Applications, vol. 37, no. 12, pp. 8019–8026, 2010.
[49] J. Li, L. Ke, and Q. Du, “Classification of heart sounds based on the wavelet fractal and twin support vector machine,” Entropy,
vol. 21, no. 5, p. 472, 2019.
[50] S. K. Ghosh, R. K. Tripathy, R. Ponnalagu, and R. B. Pachori, “Automated detection of heart valve disorders from the pcg signal
using time-frequency magnitude and phase features,” IEEE Sensors Letters, vol. 3, no. 12, pp. 1–4, 2019.
[51] F. Safara, S. Doraisamy, A. Azman, A. Jantan, and A. R. A. Ramaiah, “Multi-level basis selection of wavelet packet decomposition
tree for heart sound classification,” Computers in biology and medicine, vol. 43, no. 10, pp. 1407–1414, 2013.
[52] Y. Zheng, X. Guo, and X. Ding, “A novel hybrid energy fraction and entropy-based approach for systolic heart murmurs
identification,” Expert Systems with Applications, vol. 42, no. 5, pp. 2710–2721, 2015.
[53] M. Sharma, A. A. Bhurane, and U. R. Acharya, “Mmsfl-owfb: A novel class of orthogonal wavelet filters for epileptic seizure
detection,” Knowledge-Based Systems, vol. 160, pp. 265–277, 2018.
16
[54] M. Sharma, P. Achuth, D. Deb, S. D. Puthankattil, and U. R. Acharya, “An automated diagnosis of depression using three-channel
bandwidth-duration localized wavelet filter bank with eeg signals,” Cognitive Systems Research, vol. 52, pp. 508–520, 2018.
[55] U. Raghavendra, A. Gudigar, S. V. Bhandary, T. N. Rao, E. J. Ciaccio, and U. R. Acharya, “A two layer sparse autoencoder for
glaucoma identification with fundus images,” Journal of medical systems, vol. 43, no. 9, p. 299, 2019.
[56] O. Yildirim, R. San Tan, and U. R. Acharya, “An efficient compression of ecg signals using deep convolutional autoencoders,”
Cognitive Systems Research, vol. 52, pp. 198–211, 2018.
[57] O. Yildirim, U. B. Baloglu, R.-S. Tan, E. J. Ciaccio, and U. R. Acharya, “A new approach for arrhythmia classification using deep
coded features and lstm networks,” Computer methods and programs in biomedicine, vol. 176, pp. 121–133, 2019.
17