You are on page 1of 9

De-noising of Phonocardiograms using the

new modified wavelet threshold method


M. Vishwanath Shervegar Ganesh V Bhat
Assistant Professor, E&C Department Principal
MIT, Kundapura, Udupi, Karnataka CEC, Mangalore, D.K., Karnataka
vishwa@hotmail.co.in ganeshvbhat@yahoo.com

Abstract:
Objective:To improve the effects of de-noising of Phonocardiogram, this work introduces
the basic principles of wavelet threshold de-noising and traditional structures threshold
functions. Meanwhile, it proposes wavelet threshold function and fixed threshold formula
which are both improved here. Methods and materials:First, this work studies the problems
existing in the traditional wavelet threshold functions and introduces the adjustment factors to
construct the new threshold function basis on soft threshold function. Then, it studies the
fixed threshold and introduces the logarithmic function of layer number of wavelet
decomposition to design the new fixed threshold formula. Finally, this work uses hard
threshold, soft threshold, Garrote threshold, and improved threshold function to de-noise
heart sound signals. Results: Experimental results showed that the proposed approach could
improve soft threshold functions with constant deviation and hard threshold function with
discontinuous function problems. The proposed approach could improve the different
decomposition scales that adopt the same threshold value to deal with the noise problems,
also effectively filter the noise in the signals, and improve the SNR and reduce the MSE of
output signals. Conclusion: The new modified threshold method outperforms traditional de-
noising methods in terms of SNR and MSE.
Keywords: SNR,MSE, Phonocardiogram, Wavelet transform, denoising.

1. Introduction
Heart sounds are acoustic waves recorded such as ambient noise, the friction between
from the chest and the backside. These are the chest-piece and skin, as well as lung
transient signals generated by the vibration expiration and inspiration, are usually
of the heart valves as closure and opening unavoidable in the recording of heart
occur, and/or by the vibration of the whole sounds. The accuracy of auscultation
myocardium and associated structures. becomes questionable if the noise and
Heart auscultation is widely used by disturbance are heavy. It is thus desirable
physicians to evaluate heart valve to develop noise and disturbance reduction
functions in medical examinations and algorithms to help physicians and medical
detect the presence of abnormalities. It is students.Literature survey reveals various
not easy to acquire the skill to perform methods that have been proposed to reduce
auscultation, especially in noise and noise in heart sounds. Adaptive noise
disturbance environments. The disturbance cancellation was successfully used to
referred to in this work is the impulsive enhance heart sounds, but an additional
noise or interference characterized by a reference to the noise signal was required
short time duration. Noise and disturbance,
[1], [2]. Meanwhile, the reduced-order Kalman filter was applied to remove noise
From heart sounds [3]. Spectral and there are constant deviations compared
subtraction, which is well known and with the original signals. All the above
widely used in speech enhancement, may mentioned defects cause part of the high
be used to enhance heart sounds [4], [5]. frequency signal information to be
The modified version of spectral missing, which influence the final
subtraction was later used to improve noise processing results[17]. In order to
reduction in [6]. It is found that the noise overcome the defects of hard and soft
reduction of heart sounds was performed threshold function perspective, this work
in the time domain and/or the frequency constructs a new modified wavelet
domain, while a few were done in the threshold function by increasing the
cycle-frequency domain. The performance adjustment factor. In order to overcome
may degrade if noise and disturbance are the problem that the same threshold is used
non-Gaussian, nonstationary, or coloured. to deal with the different decomposition
Robust methods for the noise reduction of scales, this work designs a new variable
heart sounds are therefore needed. Wavelet threshold formula by introducing the
based de-noising methods are also popular logarithmic functions of layer number of
with respect to heart sounds. Wavelet wavelet decomposition.
threshold de-noising method was proposed
by American scholar Donohue. The 2. Methods and Materials
method is simple to calculate and the noise 2.1 The Principle of Wavelet
can be suppressed to a large extent. At the Threshold De-noising
sametime, singular information of the
Noise signals usually appear as high
original signal can be preserved well, so it
frequency signals in signal processing, but
is a simple and effective method [14].
useful signals appear as either low
Various authors [7]-[13] have used
frequency or more smooth signals. The
discrete wavelet transform in various
signals with noise have the features of the
forms. Whether wavelet threshold de-
above.So when the signals are decomposed
noising method is good or bad depends on
by wavelet, the signals with noise in the
two decisive factors. One is threshold 𝜆
high frequency wavelet coefficients,
and the other important factor is the
through threshold quantization threshold
selection of the threshold function [15].As
processing, high frequency wavelet
the most basic threshold function, both
coefficients to reconstruct the signals can
hard and soft threshold functions have
eliminate the noise with the signals. One
their own defects. The breakpoint problem
dimensional signal de-noising process is as
of the hard threshold function makes it
follows: the one dimensional signal is
have no continuity [16]. In the process of
decomposed by wavelet decomposition,
reconstructing heart sound signal,
selecting threshold and threshold function
proneness to volatility and pseudo-Gibbs
to quantify the high frequency coefficients
effect will make the signal distorted while
of wavelet decomposition and reconstruct
the hard threshold function can keep
the one dimensional wavelet. The
singular information of the original signal.
keyfactors affecting the quality of de-
And the constant deviation problem which
noising are de-noising threshold and the
is born with the soft threshold function
selection of threshold function [18].
cannot be overcome by itself, which means
that the reconstructed signal is too smooth
2.2 Wavelet Threshold Function
2.2.1. Classical Threshold expressions such ̅̅̅̅̅
as𝜔 𝑗,𝑘 =
Function. 𝜆2
𝜔𝑗,𝑘 − 𝜔 , |𝜔𝑗,𝑘 | ≥ 𝜆
{ 𝑗,𝑘 } (3)
The hard threshold function sets the 0, |𝜔𝑗,𝑘 | < 𝜆
decomposition coefficient to zero which is
less than the threshold value under 2.2.2 New threshold
different scale spaces and preserves the Function
decomposition coefficient which is greater
than the threshold [19]. This method does The continuity in the soft threshold
not change the local properties of the function is much better, but it has a
signal, but because of the discontinuity, it constant deviation. So, in order to
leads to a certain fluctuation in the overcome its short comings, the soft and
reconstruction of the original signal. hard threshold algorithms are
𝜔𝑗,𝑘, |𝜔𝑗,𝑘 | ≥ 𝜆 compromised. The semisoft threshold
Consider̅̅̅̅̅̅
𝜔𝑗,𝑘 = { } function [22] (4) has been shown in the
0, |𝜔𝑗,𝑘 | < 𝜆
formula
(1)
̅̅̅̅̅
𝜔 𝑗,𝑘 =
The soft threshold function assigns the 𝑠𝑔𝑛(𝜔𝑗,𝑘 ) − (|𝜔𝑗,𝑘 | − 𝑇𝜆), |𝜔𝑗,𝑘 | ≥ 𝜆
specified threshold value of the { }(4)
0, |𝜔𝑗,𝑘 | < 𝜆
decomposition coefficient to zero. After
the algorithm, the decomposition We propose a new threshold method by
coefficient is coherent, but it loses a part of replacing the T parameter by a complex
the high frequency coefficients above the exponential function. Among the soft and
threshold [20].Consider𝜔 ̅̅̅̅̅
𝑗,𝑘 = hard threshold functions, the value of 𝑇 is
𝜆 taken between 0 and 1. Even if the result
𝜔𝑗,𝑘 − 𝜔 , |𝜔𝑗,𝑘 | ≥ 𝜆
{ 𝑗,𝑘 } (2) of semisoft threshold function is between
0, |𝜔𝑗,𝑘 | < 𝜆 them, the value of 𝑇 is fixed. So, there will
still be fixed bias. Consider
Parameter𝜔𝑗,𝑘 is the representation of
𝜔𝑗,𝑘 =
̅̅̅̅̅
estimated wavelet coefficients.Parameter
𝜔𝑗,𝑘 is the representation of the wavelet
coefficients after decomposition.𝜆 is the 𝜆
𝑠𝑔𝑛(𝜔𝑗,𝑘 ) − |𝜔𝑗,𝑘 | − 𝛼 , |𝜔𝑗,𝑘 | ≥ 𝜆
𝛼(|𝜔𝑗,𝑘 |−𝜆)
representation of threshold, and sgn(∗) is [
𝜆
]
the symbolic piecewise function in the ( 𝑒 )
above two formulas[21]. In order to { 0, |𝜔𝑗,𝑘 | < 𝜆 }
overcome the shortcomings of above two
(5)
methods, GaoHongYe proposed Garrote
threshold function in which the de-noising The adjustment factor of the new function
effect is the best compared with above two (5) is different from the semisoft threshold
kinds of traditional threshold functions and function. It consists of a complex
it is better in continuity with the exponential function shown above which
has more adaptability; 𝛼 is the normal
number which can be adjusted freely and
the values of 𝛼 are different with the different signal. When |𝜔𝑗,𝑘 |= ,𝜔
̅̅̅̅̅
𝑗,𝑘 =0
and when |𝜔𝑗,𝑘 |→𝜆 , ̅̅̅̅̅
𝜔𝑗,𝑘 →0.Therefore, 2.3 The Improved Fixed
continuously in place of 𝜆,the new Threshold 𝜆
threshold function has the characteristics
of soft threshold function; when |𝜔𝑗,𝑘 |→∞ , The threshold value 𝜆 is a very important
𝜔𝑗,𝑘 →𝜔𝑗,𝑘 improved threshold function
̅̅̅̅̅ parameter in the wavelet threshold de-
noising algorithm. The performance of
based on ̅̅̅̅̅
𝜔𝑗,𝑘 =𝜔𝑗,𝑘 as the asymptotic line;
wavelet denoising will directly be affected
it can be seen that, with the increase of
by threshold selection. There are some
𝜔𝑗,𝑘 , ̅̅̅̅̅
𝜔𝑗,𝑘 will gradually be close to thresholds, such as rigrsure, heursure,
𝜔𝑗,𝑘 ;when 𝜔𝑗,𝑘 becomes infinite, 𝜔̅̅̅̅̅
𝑗,𝑘 and sqtwolog, and minimaxi [15]. The fixed
𝜔𝑗,𝑘 can be approximated as equal. That threshold isused commonly. Its expression
will not only reduce the discreteness of the is 𝜆 = 𝜎𝑛 √2𝑙𝑛𝑁, 𝜎𝑛 is the noise standard
hard threshold function, but can also avoid variance, and 𝑁 is the length of the signal.
the constant problem in soft threshold When the same threshold value 𝜆 is used
function at the same time. The variables 𝛼 to deal with the different decomposition
in the formula are also really important, scales𝑗, the enhancement effect is not
and the change of the value of 𝛼 affects the ideal. This paper proposes the improved
noise directly. When 𝛼=0, the function will fixed threshold based on the fixed
express soft threshold function and when threshold. Its expression is 𝜆 =
𝛼=∞, the function will express hard 𝜎𝑛 √2𝑙𝑛𝑁(𝑙𝑜𝑔2 𝑗) with the increasing of
threshold function. We now show that soft,
decomposition scale 𝑗, the value of the
hard, Garrote, new threshold function and
improved threshold 𝜆 is gradually reduced.
new modified threshold function in the
And it is consistent with the propagation
form of graphics intuitive. Here, the
characteristics of the noise in different
parameter threshold functionofeachtake 𝜆
scales of the wavelet transform.
= 0.6, 𝛼=10,and−1.5 ≤𝜔𝑗,𝑘 ≤ 1.5. As
shown in Fig 1, the hard threshold function 3. Resultsand discussion
is not continuous at the threshold point, so 3.1 Dataset
that there are fluctuations in the recovery
of the original signal.Compared with the Heart sound present in the Physionet heart
hard threshold function, though there are sound database [23] were taken from
no discontinuity problems in the soft several people around the world, collected
threshold function, there are still some at either a clinical or nonclinical
differences between the original signal and environment, from pathological patients.
the reconstructed signal, because of the The Challenge training set consists of five
constant problems inherent in the soft databases (A through E) containing a total
threshold function. The new threshold of over 3000 heart sound recordings. The
function has continuity and when the duration of heart sound varies from 5
coefficient is reaching 𝜆 it gradually seconds to just over 120 seconds. The
becomes close to zero. Therefore, the new heart sound recordings were obtained from
threshold function overcomes the different locations on the body. Four
deficiencies of discontinuous and constant locations used for acquisition are aortic
deviation. So, de-noisingeffectisbetter. area, pulmonic area, tricuspid area and
mitral area. Abnormal heart sounds were
taken from patients with a confirmed 4, Fig 5 and Fig 6 show the effect of de-
cardiac diagnosis. The patients suffered noising of normal heart sounds using Soft
from a variety of illnesses such as heart threshold, Hard threshold and Garrote
valve defects and coronary artery disease. threshold methods. The de-noising can be
Heart valve defects cover mitral valve visualised by the smoothness observed in
prolapse, mitral regurgitation, aortic the plots. The noisy sound has many
stenosis and valvular surgery. Pathological redundant peaks. Fig 7 shows the de-
patients included both children and adults. noising of heart sound by the new
Each subject/patient have threshold method. Fig 8 shows the plot of
contributedbetween one and six heart abnormal heart sound, while fig 9 shows
sound recordings. All recordings have a the plot of the noisy abnormal heart sound.
sampling rate of 2,000 Hz and are in .wav Fig 10, Fig 11 and Fig 12 show the effect
format. of de-noising of abnormal heart sounds
using Soft threshold, hard threshold and
3.2 De-noising Procedure Garrote threshold methods. Fig 13 shows
There are two important bases for judging the effect of de-noising using the new
the effect of de-noising. One indicator is modified threshold method. The
SNR, and another is MSE [20]. SNR is the comparison between different methods of
proportion of the energy of the signal and de-noisingis shown in Table 1 (Normal
noise inthenoisysignal.MSEandde- HS) and Table 2 (Abnormal HS) in terms
noisingeffecthaveaninversely proportional of average Signal to noise ratio (SNR) and
relationship. It is a means to measure average Mean Square Error (MSE) for all
average errors[21]. The heart sound is sounds. From Table 1 it is clear that the
carefully selected from the Physionet average SNR obtained for the proposed
database and subjected to de-noising method is 1.44 times higher than soft
procedure. The discrete wavelet transform threshold method, 1.12 times higher than
of the heart sound is calculated and the Garrote threshold and 1.03 times higher
signal decomposed into different levels. than Hard threshold method. Also the
We threshold the coefficients using average MSE obtained for proposed
various techniques described above. The method is 3.94 times lower than soft
wavelet family and the number of threshold method, 1.94 times lower than
decomposition levels are selected Garrote threshold and 1.28 times lower
arbitrarily. The inverse discrete wavelet than hard threshold method. From Table 2
transform is calculated on these it is clear that the SNR obtained for the
coefficients. This gives the de-noised proposed method is 1.52 times higher than
signal. Additional noise can be added to soft threshold method, 1.14 times higher
compare the results with the original than Garrote threshold and 1.02 times
signal. The denoising procedure adopted is higher than Hard threshold method. Also
applied to both normal and pathological the MSE obtained for proposed method is
sounds.Fig 1 shows the comparison of the 6.32 times lower than soft threshold
various de-noising methods. Fig 2 shows method, 2.06 times lower than Garrote
the plot of the normal heart sound. Fig 3 threshold and 1.16 times lower than hard
shows the plot of heart sound after it has threshold method.The same results hold
been corrupted with noise. The noise can good to all types of wavelet families and
be colored or white, Gaussian or non- all decomposition levels. These
Gaussian, stationary or non-stationary. Fig
observation suggest that the proposed [2] J. Gnitecki, Z. Moussavi, and H.
method outperforms the other methods. Pasterkamp, “Recursive least squares
adaptive noise cancellation filtering for
4. Conclusion heart sound reduction in lung sounds
This work first introduced the wavelet recordings,” in Proc. 25th Annu. Int. Conf.
threshold de-noising in the traditional IEEE EMBS, Cancun, Mexico, Sep. 17–
threshold function and then proposed the 21, 2003, 2010, pp. 2416–2419.
new wavelet threshold function according [3] S.CharlestionandM.R.Azimi-
to the function curve of the de-noising Sadjadi,“ReducedorderKalmanFiltering for
performance analysis to improve the the enhancement of respiratory sounds,”
function, through the increase in soft IEEE Trans. Biomed. Eng., vol. 43, no. 4,
threshold function adjustment factor. This pp. 421–424, Apr. 1996.
work studied the fixed threshold and
introduced the logarithmic function of [4] S. F. Boll, “Suppression of Acoustic
layer number of wavelet decomposition to noise in speech using spectral subtraction,”
design a new threshold formula. The IEEE Trans. Acoust., Speech Signal
proposed method can make the threshold Process., vol. 27, no. 2, pp. 113–120, Apr.
estimation more accurate and lead to better 1979.
de-noising effects. Experimental results [5] S. H. Nawab, T. F. Quatieri, and J. S.
showed that the proposed method Lim, “Signal reconstruction from short-
improves the SNR of the output signal and time Fourier transform magnitude,” IEEE
reduce the MSE of the output signal. It Trans. Acoust., Speech, Signal Process.,
thus improves the soft threshold function vol. ASSP-31, no. 4, pp. 986–998, Aug.
with constant deviation and the hard 1983.
threshold with discontinuous function
problem. It also improves the different [6] A. S. Paul, E. A. Wan, and A. T.
decomposition scales to adopt the same Nelson, “Noise reduction for heart sounds
threshold value to deal with the noise using a modified minimum-mean squared
problem, and verifies that the proposed error estimator with ECGgating,”inProc.
method is practical, effective, and feasible, 28th IEEE EMBS Annu. Int.
compared with the traditional method of Conf.,NewYorkCity, NY, Aug. 30–Sep. 3,
de-noising threshold function in filtering 2006, pp. 3385–3390.
the noise present in the signal.
[7] Hedayioglu L, Coimbra M, Mattos S.
REFERENCES Heart sound segmentation for digital
stethoscope integration. Master’s thesis.
[1] Y. Bai and C. Lu, “The embedded Porto, Portugal: University of Porto; 2009.
digital stethoscope uses the adaptive noise
cancellation filter and the type I [8] Pedrosa J, Castro A, Vinhoza TT.
Chebyshev IIR bandpass filter to Automatic heart sound segmentation and
reducethenoiseoftheheartsound,”inProc.7th
Int.WorkshopEnterprise Netw. Comput.
Healthc. Ind., Busan, Korea, Jun. 23-25,
2005, pp. 278– 281.
murmur detection in pediatric [16] H.-R. Jia, X.-Y. Zhang, and J. Bai, “A
phonocardiograms. 2014 36th Annual continuous differentiable wavelet
International Conference of the IEEE threshold function for speech
Engineering in Medicine and Biology enhancement”, Journal of Central South
Society (EMBC), Chicago, IL: IEEE; University, vol.20,no.8,pp.2219–
2014. p. 2294–2297. 2225,2013.
[9] Kumar D, Carvallo PD, Antunes M, et [17] L. Su, G. Zhao, and R. Zhang,
al. Wavelet transform and simplicity based “Translation-invariant wavelet de-noising
heart murmur segmentation. Comput method with improved thresholding,” in
Cardiol IEEE. 2006; 33:173–176. Proceedings of the IEEE International
Symposium on Communications and
[10] Quiceno A, Delgado E, Vallverd M,
InformationTechnology(ISCIT’05),vol.1,p
et al. Effective phonocardiogram
p.619–622,October 2005.
segmentation using nonlinear dynamic
analysis and high-frequency [18] A. Phinyomark, C. Limsakul, and P.
decomposition. ComputCardiol. Phukpattaranont,“Optimal wavelet
2008;35:161–164. functions in wavelet denoising for multi-
function myoelectric control,” ECTI
[11] Kouras N, Boutana D, Benidir M.
Transactions on Electrical Engineering,
Wavelet based segmentation and time-
Electronics,andCommunications,vol.8,no.1
frequency caracterisation of some
,pp.43–52,2010.
abnormal heart sound signals. 2012 24th
International Conference on [19] C. He, J. C. Xing, and Q. L. Yang,
Microelectronics (ICM), 2012 Dec 16. “Optimal wavelet basis selection for
IEEE; 2012. p. 1–4. wavelet denoising of structural vibration
signal,”Applied Mechanics and Materials,
[12] Wang P, Kim Y, Ling L, et al. First
vol. 578-579, pp. 1059–1063, 2014.
heart sound detection for phonocardiogram
segmentation. ConfProc IEEE Eng Med [20] J.-Y. Tang, W.-T. Chen, S.-Y. Chen,
Biol Soc. 2006;5:5519–5522. and W. Zhou, “Wavelet based vibration
signal denoising with a new adaptive
[13] Dokur Z, Olmez T. Heart sound
thresholding function,” Journal of
classification using wavelet transform and
Vibration and Shock, vol. 28, no. 7,
incremental self-organizing map. Digit
pp.118–121,2009(Chinese).
Signal Process. 2008;18:951–95.
[21] S.Badiezadegan and R.C.Rose,“ A
[14] D.L. Donoho and I.M.Johnstone
wavelet-basedthresholding approach to
,“Ideal spatial adaptation by wavelet
reconstructing unreliable spectrogram
shrinkage,”Biometrika,vol.81,no.3,pp.425
components,”SpeechCommunication,vol.6
–455,1994.
7,pp.129–142,2015.
[15] C. He, J. Xing, J. Li, Q. Yang, and R.
[22] X. Chen, S. Li, and W. Wang, “New
Wang, “A new wavelet thresholding
de-noising method for speech signal based
function based on hyperbolic tangent
on wavelet entropy and adaptive threshold,
function,” Mathematical Problems in
”Journal of Information and
Engineering, vol. 2015, Article ID
Computational Science, vol.12,no.3,
528656,9pages,2015.
pp.1257–1265,2015.
[23] www.physionet.org/challenge/2016/: accessed date: December 1 2017.

Fig 1 Comparison of Noise Threshold methods Fig 2


Origi
nal
Nor
mal
HS

Fig 3 Noisy Normal HS Fig 4 Soft ThresholdNormal HS

Fig 5 Hard Threshold Normal HS Fig 6 Garrote Threshold Normal HS


Fig 7 New Threshold Normal HS Fig 8 Abnormal HS

Fig 9 Noisy Abnormal HS Fig 10 Soft threshold abnormal HS

Fig 11 Hard Threshold Abnormal HS Fig 12 Garrote Threshold Abnormal HS

Fig 13 New Threshold Abnormal HS

Average Hard Threshold Soft threshold Garrote Threshold New Threshold


SNR 13.3313 9.594 12.2569 13.8125
MSE 0.00018874 0.00058072 0.0002864 0.00014724
Table 1 Results of SNR and MSE after Normal Heart Sound Denoising

Average Hard Threshold Soft threshold Garrote Threshold New Threshold


SNR 20.8484 14.153 18.7712 21.4634
MSE 0.00038803 0.0021157 0.00068919 0.00033528
Table 2 Results of SNR and MSE after Abnormal Heart Sound Denoising

You might also like