Biomedical Signal Processing and Control 63 (2021) 102212

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

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

Design and implementation of a robust noise removal system in ECG signals

using dual-tree complex wavelet transform
Navdeep Prashar a, *, Meenakshi Sood b, Shruti Jain c
a
Department of Electronics and Communication Engineering, Bahra University, Solan, 173234, Himachal Pradesh, India
b
Department of CDC, NITTTR, Chandigarh, 16009, India
c
Department of Electronics and Communication Engineering, Jaypee University of Information Technology, Solan, 173234, Himachal Pradesh, India

A R T I C L E I N F O A B S T R A C T

Keywords: The key deliverable for any health monitoring system that offers telecardiology services is the recovery of the
Dual-tree complex wavelet transform ECG signal related to cardiac diagnostics. Accurate analysis and diagnosis of heart diseases become difficult due
Threshold functions to noises and artifacts. Therefore, a need for noise removal which is an inverse problem arises that leads to
Threshold values
reliable signal recovery tasks. In this paper, authors have provided a detailed evaluation of the effect of the
Distribution function
Electrocardiogram
choice of the threshold value, threshold algorithm, and distribution function to evaluate the ECG signal de-
noising performance employing Dual tree complex wavelet transform (DTCWT). In this research work, eight
different sets of threshold value selection rules along with six distinct threshold functions are implemented and
evaluated on MIT-BIH arrhythmia database. Authors have proposed an estimator that yields efficient results than
a conventional estimator using Gaussian distribution function and extended the work to Normal distribution. The
results show that the Universal modified threshold level-dependent threshold value selection with Non -Negative
Garrote threshold function delivers higher performance in terms of SNR, lower values of MSE and PRD. The
proposed approach accomplished good performance evaluation results as 58.23 dB, SNR and 9.63e-08, MSE in
comparison with the 48.65 dB, SNR and 8.75e-07, MSE using conventional method for the normal distribution
function. The novelty of this work also lies in an exhaustive empirical comparative analysis with the existing
research work using the same database with the proposed normal distribution technique.

1. Introduction coefficients that are stronger than the noise. The threshold algorithm is
known for its high-speed computation and minimized memory storage.
The automatic electrocardiogram (ECG) analysis algorithm plays an Various pre-processing methods have been proposed prior to evac­
important role in accomplishing quick and accurate Cardiovascular uate distinctive artifacts and to extricate correct morphological char­
(CVD) disease. CVD has been the largest threat to human life for de­ acteristics of the ECG signal [5–9]. FIR based notch filter has been
cades, and millions of people die due to delayed diagnosis and treat­ introduced to remove Power Line Interference (PLI) noise from an ECG
ment. ECG signal shows the graphical portrayal of cardiovascular signal that resulted in a prerequisite of large filter order which assists
movement and its application for the discovery of different heart disease increments the computational complexity [10]. Numerous analysts have
and abnormalities [1,2]. ECG has novel morphological characteristics utilized high order infinite Impulse Response (IIR) filter to take out the
P-QRS-T and on analyzing the variation in these features many cardiac impact of PLI and baseline wander from ECG signal [11]. However, it
diseases can be diagnosed. However, the presence of different noises in has some shortcomings such as large filtering time, a memory that re­
ECG signals has an acute influence on accurate feature extraction [3,4]. sults in inappropriate non- linear filtering. To have faster filtering
The signal denoising comprises expelling the noise from the contami­ response adaptive filtering methods have been used for the elimination
nated signal without changing its characteristics. In this context, the of various noises from ECG signal with smaller residual errors [12].
Thresholding method based on nonlinear multi-resolution analysis is a Authors in [13] proposed a fixed sub-band decomposition approach
viable approach to obtain denoised signal by computing the threshold using a wavelet-based wiener filter that has been introduced to expel
value and comparing each coefficient against a threshold to recover the EMG noise while authors in [14] proposed a Bionic adaptive wavelet

* Corresponding author.
E-mail address: [email protected] (N. Prashar).

https://doi.org/10.1016/j.bspc.2020.102212
Received 23 May 2019; Received in revised form 16 July 2020; Accepted 1 September 2020
Available online 14 September 2020
1746-8094/© 2020 Elsevier Ltd. All rights reserved.
N. Prashar et al. Biomedical Signal Processing and Control 63 (2021) 102212

transform to eliminate high frequency and baseline wander noise.

Wavelet theory-based signal noise residue algorithm for accurate esti­
mation of noise with minimal computation is discussed in [15]. Authors
in [16,17] presented many methods that include temporal average
filtering, Independent Component Analysis (ICA), Linear filtering and
cubic spline filter for noise expulsion from ECG signal. Recently, Discrete
Wavelet Transform (DWT) [18–20] based thresholding has been intro­
duced for ECG signal denoising [21,22]. However, the selection of
appropriate wavelet function [23,24] and thresholding value selection is
the most important parameter for signal denoising [25]. Wavelet
transform-based thresholding introduces aliasing effect that results in
having artifacts in the denoised signal. To overcome this problem Dual
tree complex wavelet transform (DTCWT) has been introduced to get
enhanced reconstructed signal. Implementation of DWT on the ECG
signal imparts some limitations such as lack of shift-invariance property,
aliasing, oscillations, and absence of directionality [26]. To conquer
these constraints and to create an economical tool for signal denoising, a
DTCWT technique is recommended that has an anti-aliasing effect and
shift-invariance property for biomedical signals [27,28]. The proposed
method also overcomes the demerits of existing techniques. Significant
amount of noise in ECG signal may portray the false assessment of heart
condition that results in incorrect analysis of disease. In this manner, it is
exceptionally imperative to expel these noise contents from ECG signal
before it is utilized for further processing. In this research paper, authors
have proposed noise estimators that denoise the ECG signal. The results
Fig. 1. Block diagram of the proposed methodology using DTCWT.
were approved by calculating the performance parameters like SNR,
PRD and MSE. The high SNR and low PRD define the enhancement in the
quality of an ECG signal with proposed technique than other existing inverse DTCWT technique is applied to the modified threshold detailed
techniques. coefficient and the original approximation coefficient to recover the
In this paper, a DTCWT technique based threshold tuning is explored output signal. Lastly, the parameters of the output signal have been
to deliver denoised ECG signal. Threshold tuning is executed by differing evaluated by computing different performance metrics to validate the
the threshold value (th) and its function (fn) at the optimal decomposi­ proposed technique. The proposed method helps in removing the noise
tion level. The contribution of this paper is to propose an estimator that ranging from (0–3 mV).
yields a denoised ECG signal with high SNR value when used with the
optimal combination of different distribution functions. An elaborate 2.1. Pre-processing stage
empirical analysis for selecting the distribution function for obtaining
the best possible threshold function and technique is the prime objective Pre-processing is the essential step in signal processing to eliminate
and novelty of this research work. The DTCWT technique with distinct the noise using various filtering methods by preserving the shape of
threshold tuning parameters has been evaluated based on performance different complexes. Irregularities in beat morphology of the ECG signal
metrics. The simulation results for all the ECG records indicate that the has been decreased by expelling the baseline wander noise through FIR
proposed algorithm is optimal for real-time ECG monitoring in removing High pass filtering of order N = 10 with cut-off frequency 0.5 Hz uti­
various kinds of noises and provides a high-quality clean signal. lizing Blackman windowing procedure having window length N +
This research paper is organized as Section 2 describes a proposed 1 = 11 [3].
methodology employing an estimator technique for optimal distribution
functions using different threshold functions and values; Section 3 ex­
plains the computational results using different distribution functions 2.2. DTCWT signal decomposition
and their comparison with the existing state of the art techniques further
concluded in Section 4. ECG signal is a non-stationary signal with the nonlinear character­
istic that makes it hard to denoise utilizing normal filter. To counter this
restriction, DTCWT has been acquainted to denoise these non-stationary
2. Methodology
signals by virtue of its high performance and qualitative characteristics.
This multilevel signal decomposition process utilizing the DTCWT
The input signal of ECG MIT-BIH Arrhythmia v5 of 100 record has
technique is extensively depicted in [28]. DTCWT technique is
been drawn from the Physio Bank ATM database [29]. This input ECG
composed of two real filter banks (low pass and high pass filter pair).
signal is composed of 1536 samples sampled at 360 Hz frequency with
This technique consists of two DWTs: first DWT offers real & even part
an 11-bit resolution over 10 mV range. This methodology is additionally
and second DWT provides the real & odd or imaginary part of a complex
executed for 101–109 records of the MIT-BIH Arrhythmia database
wavelet. These two real DWTs use two distinct sets of a filter with each
(mitdb) to analyze its credentials for signal denoising. The whole process
satisfies Perfect Reconstruction (PR) conditions [28].
of implementation is shown in Fig. 1.
DTCWT is approximately analytic wavelet expressed by Eq. (1)
In Fig. 1, Input ECG raw signal is fed to the pre-processing block to
remove baseline wander noise. After pre-processing, the DTCWT tech­ Ψ (t) = Ψ h (t) + i Ψ g (t) (1)
nique is applied to perform multilevel decomposition to extract detail
and approximation coefficient. In next step, thresholding process is where Ψ h (t) values are real and even while Ψ g (t) values are real and
performed on detail coefficients by choosing different threshold function odd.
and threshold value selection using conventional and proposed esti­ In this paper, the input ECG signal has been decomposed up to 4
mator technique that resulted to a modified threshold detailed co­ levels of detail and approximation coefficients by employing the DTCWT
efficients (D-1, D-2, D-3, and D-4) at different decomposition level. An technique as illustrated in Fig. 2. Complex detail coefficients have been

2
N. Prashar et al. Biomedical Signal Processing and Control 63 (2021) 102212

extracted at each level of decomposition scale which provides 4 detail Table 1

sub bands (D1, D2, D3, D4) expressed by Eq. (2). Different Threshold Functions [25].
S. Threshold Mathematical Expressions
D1 = D1real + i D1imag (2)
No. Functions

Similarly, complex approximation coefficients are extracted in sub

{ }
1. Hard c(n), for|c(n)| ≥ th
fH=
band (A4) expressed by Eq. (3) 0, otherwise
{ }
2. Soft sign(c(n))(|c(n)| − th), for|c(n)| ≥ th
fs
A4 = A4real +i A4imag (3) =
0, otherwise
3. Non- fnon-
⎧ ⎫
Detail coefficients are high-frequency coefficients and are ordinarily negative
⎨ sign(c(n))(|c(n)| − th ),
⎪ 2
⎪
Garrote for|c(n)| ≥ th ⎬
influenced by noise whereas approximate coefficients are low-frequency neg=
⎪
|c(n)|
⎪
⎩ ⎭
coefficients that are not affected by noise. The thresholding technique is { 0, √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ otherwise
}
4. Hyperbolic 2
applied to the detail coefficients such that the variance of the resulted fhyp = sign(c(n)) c(n) − th 2
, for |c(n)| ≥ th
0, otherwise
threshold coefficient does not increase. ⎧ α α ⎫
⎨ c(n) ∗ |c(n)| − |th| ,
⎪ ⎪
for|c(n)| ≥ th ⎬
ftrim = |c(n)|α
2.3. Thresholding technique 5. Trimmed
⎪
⎩
0, otherwise
⎪
⎭

If α = 1, then it is equivalent to soft thresholding while if α→0

The thresholding technique has been applied to every detail wavelet it is equivalent to hard thresholding.
coefficient to obtain a denoised signal by exploiting the capabilities of fsemi =
⎧ ⎫
signal denoising [30–32]. These detail coefficients are shrunk or scaled ⎪
⎪
⎪
c(n), |c(n)| < th2 ⎪
⎪
⎪
⎪ ⎪
[33,34] based on the selected wavelet function [35]. The performance of
⎨ ⎬
(c(n) − th1 ) ∗ th2
6. Semisoft sign(c(n)) , th2 < |c(n)| ≤ th1
the thresholding technique [36] depends on the threshold value selec­ ⎪
⎪
⎪ th2 − th1 ⎪
⎪
⎪
⎪
⎩ ⎪
⎭
tion and threshold function Thresholding methods are mainly catego­ √̅̅̅̅̅̅̅ 0, |c(n)| < th1
where th1 = th2 for semi-soft thresholding
rized into two sorts- hard thresholding and soft thresholding as proposed
by Donoho and Johnstone [37,38]. Hard thresholding is more sensitive
to signal variations and less steady in comparison with soft thresholding.
All the threshold functions that are used in this research work are
tabulated in Table 1.
Table 2
2.3.1. Threshold functions Different Threshold Value Selection for various Threshold functions [25].
Threshold functions are introduced to confine the input samples to a S. Threshold Value Selection Mathematical Expressions
desired specific range. Implementation of the threshold function No.
impressively evacuates the noise from the signal at each level depending 1. Universal Threshold Thu = σ
√̅̅̅̅̅̅̅̅̅̅̅̅̅
2logN
on the threshold value. A distinct variant of threshold functions is uti­ 2. Universal Threshold level
√̅̅̅̅̅̅̅̅̅̅̅̅̅
Thuld = σj 2lognj
lized to urge a clean ECG signal. In this section, distinctive threshold dependent
functions along with their mathematical expressions have defined as
√̅̅̅̅̅̅̅̅̅̅̅̅̅
3. Universal modified 2lognj
Thumld= σ j
shown in Table 1 [25]. threshold level dependent √̅̅̅̅
nj √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
4. Exponential Threshold level
In Table 1, c(n) represents the wavelet coefficient and th is thresh­ Thexpld = 2(j− J/2) σj 2logn
dependent j
olding value. This threshold (th) is selected from threshold value selec­
5. Minimax threshold Thmini = 0.3936+0.1829*(log nj/log2)
tion to provide the threshold limit at each decomposition level for noise 6. Modified unified threshold
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
2log(N)
elimination. Thmou= σj
log(j + 1)
√̅̅̅̅̅̅
7. Adaptive threshold Selection Thadp= σj wb
2.3.2. Threshold value selection 8. Heursure threshold selection Combination of Universal threshold and
The nature of noise elimination is affected by threshold value se­ Adaptive threshold Selection.
lection. Depending on the selected threshold value, the denoised ECG
signal could have few interferences, some discontinuities/ distortion;

Fig. 2. Multilevel decomposition using DTCWT [26].

3
N. Prashar et al. Biomedical Signal Processing and Control 63 (2021) 102212

contingent upon whether the threshold value was excessively large or 2.6. Performance parameter evaluation
too little. Different types of threshold values selection are tabulated in
Table 2. [25] The performance metrics Signal to Noise Ratio (SNR), Mean Square
These threshold value selections are dependent on the estimation of Error (MSE), and Percent Root Mean Square Difference (PRD) [42] are
noise level ‘σ ’ that assumed to be proportional to the standard deviation computed to examine the efficiency of an output signal. The most
of the coefficients. Conventional estimator ‘σ ’ is expressed by Eq. (4) effective results are achieved once the recovered signal has the highest
( ⃒ ⃒) SNR, smallest PRD and has the lowest MSE.
α = z × median⃒xjk ⃒ (4)
Percent Root Deviation (PRD) is defined by Eq. (6)
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
where xjk represent the detail coefficients at the finest level and z is the √∑N
√ [x(n) − y(n)]2
constant whose value depends on the different distribution function. For PRD = √ n=1 N (6)
√ ∑
the conventional problems, the Gaussian distribution function is x(n)2
considered and the value of z is assumed as 1.4826.
n=1

Mean Square Error (MSE) is expressed in Eq. (7)

2.4. Proposed estimator ‘σ *’ 1 ∑N
MSE = [x(n) − y(n)]2 (7)
N n=1
In 1990, a robust correlation coefficient with a high breakdown point
reliant on the least median of squares (LMdS) regression strategy was where x(n) is the original signal and y(n) is the reconstructed signal.
proposed [39]. The LMdS correlation coefficient tends to provide the Signal -to- Noise Ratio (SNR) is defined by Eq.(8)
very high value of the relationship. As an alternative to overcome this ( )
S
problem, [39] proposed robust correlation coefficient using weighted SNR = 10log (8)
least squares by combining the LMdS estimator with M-estimator During N
2011, a new version of robust correlation coefficient based on the me­
where S represents signal power and N is noise power.
dian using scale estimator median absolute deviation (MAD) was pro­
posed and known as Median-Product (MP) correlation coefficient [40].
3. Results and discussion
They replaced the mean in the classical correlation coefficient into the
median and used MAD in this coefficient calculation. However, MAD
This paper involves processing of ECG signals utilizing pre-
consists of a few drawbacks. Firstly, this estimator has low efficiency,
processing, extraction of detail and approximation coefficient utilizing
which is 37% at the Gaussian distribution and secondly, MAD views a
DTCWT, investigation of different th utilizing diverse fn for different
dispersion of the symmetric distribution. The advantage of the robust
distribution functions and reconstruction of the signal utilizing IDTCWT.
correlation coefficient is that it requires less computing time when
The proposed technique helps in eliminating different ECG noises like
compared. The use of MAD in the equation can be enhanced to another
Baseline wander noise, Power-line interference, Electromyography
robust scale estimator with the goal that this robust correlation coeffi­
noise and Burst noise which is classified as white Gaussian noise (WGN).
cient can perform better.
The frequency range of Burst noise is not well defined and present over
Thus, inspired by their work [41], this investigation intends to
the entire frequency range of ECG signal. The Burst noise is categorized
expand the robust correlation coefficient by proposing an estimator
as electrode pop noise, electrode motion artefact, electro- surgical noise
which yields high SNR. This estimator possesses a high breakdown point
and instrumentation noise. The whole experiment is carried out on MIT-
and more efficient under the normality assumption. We have proposed a
BIH Arrhythmia database (mitdb) without the addition of synthetically
new estimator σ * which yields better SNR for our exploration area of
generated noise. The statistical description and observation of each re­
denoising ECG signal. In this strategy amplitude downscaling of detail
cord of MIT-BIH Arrhythmia database reveals the existence of noise and
coefficients at the finest level is performed by factor 2. As the detailed
artifacts in the Arrhythmia database [29] as shown in Table 3.
coefficients are high-frequency coefficients and are mostly effected by
These noise contents and artifacts are successively removed by the
noise. These noise components are usually present in the amplitude of
proposed methodology adopted in this research work.
detailed coefficients which undergoes a substantial change in the
The raw input ECG signal illustrated in Fig. 3 is initially fed to the
amplitude of an ECG signal. To lessen the effect of such noises, ampli­
pre-processing stage to remove baseline wander noise which is illus­
tude downscaling is performed. After performing amplitude down­
trated in Fig. 4.
scaling, the median of resulted coefficients is taken to minimize the
On observing the Fig. 4, it has been observed that ECG signal is flat at
impact of outliers. The final proposed estimate σ* is given by Eq. (5).
reference line indicating the successive elimination of baseline wander
(⃒⃒ ⃒⃒ )
xjk noise from the raw ECG signal.
α∗ = z × median (5) In the subsequent stage, detail and approximation coefficients are
2

where z is a constant function depending on different distribution

functions. Table 3
Presence of Noise in different records.
2.5. Inverse DTCWT Record Type of noise Present

104 Multiple bursts of muscle noise

In this research paper, the inverse DTCWT technique is implemented 105 High-grade noise and artifact
on the modified threshold detailed coefficients (D-1, D-2, D-3, D-4) and 106 Low frequency noise component
original approximate coefficient (A4) to get the final output signal. The 108 Baseline shifts
inverse DTCWT performed a reverse procedure to reconstruct the signal 111 Short bursts of both baseline shifts and muscle noise
122 Low amplitude high frequency noise throughout the lower channel
where the signal at every level is up-sampled by factor 2, is passed
200 Sporadic bursts of high frequency noise in upper channel and extreme
through the synthesis filter and then added. noise and artifact in lower channel
203 Muscle artifact shifts and baseline shifts in both the channels
214 Two instances of rising artifact amplitude, and one instance of slippage of
tape
222 High-frequency noise / artifact intervals in both channels

4
N. Prashar et al. Biomedical Signal Processing and Control 63 (2021) 102212

3.1. Results of MIT-BIH 100 record data using the conventional estimator
technique

The thresholding technique is applied to all the detail coefficients

using a conventional estimator and different threshold functions and
values. Performance evaluation parameters are calculated in terms of
SNR, MSE, and PRD is tabulated in Tables 4–6 respectively.
From Tables 4–6 it is interpreted that conventional estimator using
Non-negative Garrotte function with universal modified threshold level-
dependent value selection delivers high (45.67 dB) SNR, Low (1.73e-06)
MSE and small (0.005) PRD among all other threshold value selection
Fig. 3. ECG input signal. and threshold functions. The simulation of all threshold functions with
efficient values is shown in Fig. 7.
Fig. 7 shows the output waveform generated with the best combi­
nation of different threshold values and threshold functions to produce
the denoised ECG signal. Among all, the best results are obtained using
Non -negative Garrotte threshold function with universal modified level-
dependent threshold value selection having high SNR and low MSE and
PRD.
All threshold functions shrink detail coefficients which are greater
than the universal threshold towards zero except hard threshold. The
garrote threshold function shrinks to zero fewer number of detail co­
efficients; it accommodates more detail coefficients. This accounts for
garrotte function having the highest gain. The shrinkage in soft
Fig. 4. ECG signal output after Pre-processing stage. threshold is related to the noise level σ (proportional to σ ) whereas the
shrinkage in other functions is related to noise variance σ 2. Furthermore,
the shrinkage in the soft threshold has a big bias due to the shrinkage of
separated up to fourth decay level utilizing the DTCWT procedure as
large coefficients which makes it much less sensitive. This accounts for
shown in Figs. 5 and 6 respectively. In Fig. 5, 768 detail coefficients at
the soft threshold having the least Gain. The Non-negative Garrote
level 1 are acquired which are followed by 384, 192 and 96 coefficients
at sequential levels (2, 3 and 4). The approximation coefficient is a low-
recurrence coefficient achieved at 4 level of signal decomposition. It
comprises of 96 coefficients which are least influenced by commotion as
appeared in Fig. 6. All the simulations were carried out in MATLAB
R2016 a software in Intel (R) Core i5 processor, 64-bit operating
framework.
In this research paper, different threshold value selection (Universal
threshold, Exponential threshold level-dependent, Universal modified
threshold level-dependent, Universal threshold level-dependent, Mini­
max threshold, Modified unified threshold, Adaptive threshold Selec­
tion, Heursure threshold selection) and threshold functions (Hard, Soft,
Non-negative Garrote, Hyperbolic, Trimmed, Semisoft) were applied to
obtain denoised signal. This has been done by using a conventional
estimator (σ) and the proposed estimator (σ*). Extensive empirical
analysis has been performed with different threshold values on varying
the threshold functions for various ECG signals. However, the results of
10 signals are depicted in this research work. Fig. 6. Approximation coefficients using DTCWT at level 4.

Fig. 5. Detailed coefficients using DTCWT at different levels up to the fourth level.

5
N. Prashar et al. Biomedical Signal Processing and Control 63 (2021) 102212

Table 4
SNR evaluation for conventional estimator.
S.No. Threshold Function Thu Thuld Thumld Thexpld Thmini Thmou Thadp Heur

1. Hard 28.77 26.86 43.62 34.96 28.21 30.81 37.45 26.86

2. Soft 26.76 22.55 38.31 28.788 25.38 28.22 31.09 22.55
3. Non-Negative Garrote 28.33 25.90 45.67 42.59 27.49 30.81 34.93 25.90
4. Hyperbolic 14.09 13.91 14.23 14.23 14.15 14.13 14.20 13.91
5. Trimmed 24.31 21.67 26.90 26.27 25.36 24.98 27.48 24.31
6. Semi Soft 10.15 10.78 9.90 10.07 10.30 10.06 10.05 10.78

Table 5
MSE evaluation for conventional estimator.
S.No. Threshold Function Thu Thuld Thumld Thexpld Thmini Thmou Thadp Heur

1. Hard 8.54e-05 7.86e-05 2.78e-06 2.04e-05 9.69e-05 5.32e-05 1.15e-05 1.32e-04

2. Soft 1.35e-04 1.20e-04 1.33e-06 8.48e-05 1.85e-04 9.66e-05 4.98e-05 3.56e-04
3. Non Negative Garrote 9.41e-05 1.64e-04 1.73e-06 3.52e-06 1.14e-04 5.32e-05 2.06e-05 1.64e-04
4. Hyperbolic 0.002 0.0026 0.0024 0.0024 0.0025 0.0025 0.0024 0.0026
5. Trimmed 2.37e-04 4.36e-04 1.30e-04 1.51e-04 1.86e-04 2.03e-04 1.14e-04 2.37e-04
6. Semi Soft 0.006 0.005 0.006 0.006 0.006 0.006 0.006 0.005

Table 6
PRD evaluation for conventional estimator.
S.No. Threshold Function Thu Thuld Thumld Thexpld Thmini Thmou Thadp Heur

1. Hard 0.03 0.04 0.006 0.01 0.03 0.02 0.01 0.04

2. Soft 0.04 0.07 0.012 0.03 0.05 0.03 0.02 0.07
3. Non Negative Garrote 0.03 0.05 0.005 0.007 0.03 0.02 0.01 0.05
4. Hyperbolic 0.19 0.20 0.19 0.19 0.19 0.19 0.19 0.20
5. Trimmed 0.06 0.08 0.04 0.04 0.06 0.05 0.04 0.06
6. Semi Soft 0.31 0.28 0.31 0.31 0.30 0.31 0.31 0.28

estimator has a piecewise linear solution path and is a scaled version of threshold tuning. Moreover, the non-negative garrotte thresholding is
the least square estimate. This estimator is consistent, flexible and eases propelled with characteristics that with probability tending to one, the
of computation. Aliasing in signal processing is an effect that causes solution path having an estimate that correctly identifies the set of prime
diverse signals to become indistinguishable when sampled, that resulted variables and is reliable for the coefficients of the prime variables.
to the distortion in the reconstructed signal from original signal. The Later, comparative table of performance metrics with conventional
proposed technique successively overcomes this effect that make it and proposed estimator techniques using Gaussian distribution function
vigorous to artifacts than other existing techniques. for the efficient results is tabulated in Table 10 which shows a remark­
able performance. The result shows that Non-negative Garrote function
with universal modified threshold level-dependent value selection yields
3.2. Results of MIT-BIH 100 record data with the proposed estimator
more desirable results in terms of SNR, MSE and PRD using the proposed
technique
estimation technique (σ*) than conventional estimator (σ ).
The existing techniques such as DWT lacks in shift-invariance and
Real datasets usually contain a fraction of outliers and other con­
directional selectivity which leads to much less redundancy of transform
taminations. The conventional estimator often gives misleading results
coefficients. These issues have been overcome by utilizing proposed
and is much affected by the outliers. Robust methods are designed to
method confining no oscillations in the transform magnitude near dis­
consider the majority of the data rather than all the data. Therefore,
continuities of a signal. Fig. 8 shows the ECG reconstructed signal that is
robust methods give reasonable results even when data contain a frac­
obtained by Non-Negative Garrote function with universal modified
tion of outliers. To achieve robustness and computational efficiency, a
threshold level-dependent value selection using the proposed estimation
new estimator has been proposed. The conventional estimator uses a
technique.
non-robust estimator as the building blocks. The proposed signal esti­
mator is applied to the detail coefficients. Performance evaluation pa­
rameters are evaluated for the different threshold functions and
3.3. Effect of distribution functions
threshold values. Tables 7–9, results in the SNR, MSE and PRD values
respectively using the proposed estimator technique.
In statistics, a probability distribution gives either the probability of
Computational assessment in terms of SNR, MSE and PRD for pro­
each estimation of an irregular variable (when the variable is discrete),
posed estimator using Gaussian distribution for all combination of
or the probability of the value falling inside a specific interim (when the
threshold functions and threshold value selection are depicted in
variable is continuous). The probability distribution portrays the scope
Tables 7–9 respectively. The result shows that Non-negative Garrotte
of conceivable qualities that an arbitrary variable can accomplish and
function with universal modified threshold level-dependent value se­
the probability that the estimation of the irregular variable is inside any
lection outperforms among all combination of threshold functions and
subset of that range. A probability distribution gives vital information
threshold value selection using proposed estimator delivering high SNR
about the data, how the values are changing, regardless of whether they
(54.98 dB), low MSE (2.03e-07) and small PRD (0.003). This combina­
are grouped together or spread out and whether they are symmetrically
tion of threshold value selection and threshold function using proposed
arranged on the X-axis or not.
estimator encapsulate the noise distributed in the signal by removing the − (x− b)2
redundant wavelet coefficients at every sub-band level to provide fine A Gaussian distribution function is represented as f(x) = ae 2c2 ,

6
N. Prashar et al. Biomedical Signal Processing and Control 63 (2021) 102212

Fig. 7. ECG signal best combinations of threshold values and functions using the Proposed methodology.

Table 7
SNR evaluation for proposed estimator.
S. No Threshold Function Thu Thuld Thumld Thexpld Thmini Thmou Thadp Heur

1. Hard 33.21 28.34 46.27 41.00 30.78 38.62 44.76 28.34

2. Soft 29.08 25.74 44.17 33.50 27.91 31.71 35.97 25.74
3. Non-Negative Garrote 30.54 27.74 54.98 38.04 29.35 34.40 41.01 27.74
4. Hyperbolic 14.16 14.08 14.23 14.21 14.14 14.20 14.22 14.08
5. Trimmed 25.32 23.68 27.42 27.45 24.82 26.13 26.65 23.68
6. Semi Soft 10.02 10.26 9.87 9.97 10.14 9.98 9.89 10.26

Table 8
MSE Evaluation for proposed estimator.
S. No. Threshold Function Thu Thuld Thumld Thexpld Thmini Thmou Thadp Heur

1. Hard 3.06e-05 9.38e-05 1.51e-06 5.09e-06 5.36e-05 8.81e-06 2.14e-06 9.38e-05

2. Soft 7.92e-05 1.71e-04 2.45e-06 2.86e-05 1.03e-04 4.32e-05 1.62e-05 1.71e-04
3. Non Negative Garrote 5.66e-05 1.07e-04 2.03e-07 1.00e-05 7.44e-05 2.33e-05 5.07e-06 1.07e-04
4. Hyperbolic 0.002 0.002 0.002 0.002 0.002 0.002 0.002 0.002
5. Trimmed 1.88e-04 2.74e-04 1.16e-04 1.15e-04 2.11e-04 1.56e-04 1.38e-04 2.74e-04
6. Semi Soft 0.006 0.006 0.006 0.006 0.006 0.006 0.006 0.006

where a is the height of the peak curve, b is the centre of the peak normally distributed random variable with expected value μ = b and
(position) and c is the standard deviation that controls the Gaussian RMS variance σ 2 = c2. Normal Distribution is symmetric about its mean, and
/
width/width of the bell. The parameter a is set equal to 1 σ ( √̅̅̅̅̅ is non-zero over the entire real line.
2π so that
̅)
The normal distribution function for evaluating SNR, MSE and PRD is
the total area under the curve is equal to 1.
used in Tables 11–13 tabulates the SNR, MSE and PRD values respec­
A normalized Gaussian function is the probability density function
tively using different th applied on fn for the proposed estimator.
(pdf) of a Gaussian distribution. The integral of Gaussian function is 1 if
/ From Tables 11–13, it has been observed that the proposed estimator
and only if a = 1 c( √̅̅̅̅̅
2π and in this case the Gaussian is the pdf of a
̅)
using Normal distribution has better performance in terms of SNR, MSE,

7
N. Prashar et al. Biomedical Signal Processing and Control 63 (2021) 102212

Table 9
PRD evaluation for proposed estimator.
S.No. Threshold Function Thu Thuld Thumld Thexpld Thmini Thmou Thadp Heur

1. Hard 0.02 0.03 0.004 0.008 0.02 0.01 0.005 0.03

2. Soft 0.03 0.05 0.006 0.021 0.04 0.02 0.01 0.05
3. Non Negative Garrote 0.02 0.04 0.003 0.001 0.03 0.01 0.008 0.04
4. Hyperbolic 0.19 0.19 0.19 0.19 0.19 0.19 0.19 0.19
5. Trimmed 0.05 0.06 0.04 0.04 0.05 0.04 0.04 0.06
6. Semi Soft 0.31 0.30 0.32 0.006 0.31 0.31 0.31 0.30

Table 10
.Comparative table of performance metrics with the conventional and proposed estimator.
SNR (dB) MSE PRD

Conventional Proposed Conventional Proposed Conventional Proposed

Non-negative Garrote with universal modified threshold level using Gaussian distribution 45.67 54.98 1.73e-6 2.03e-7 0.005 0.003
function

under distributional assumptions in the noise. This study aims to extend

the robust correlation coefficient by applying a proposed estimator. This
estimator possesses a high breakdown point which is 50% and more
efficient under the normality assumption.
The effect of different distribution functions like Gaussian, Normal
and Exponential on the conventional and proposed estimator technique
is tabulated in Table 14. The results tabulated in Table 14 are obtained
by putting the distinct value of constant (z) for each distribution in Eq.
(4) and Eq. (5).
From Table 14 it is observed that the proposed estimator distribution
gives significant results in terms of SNR, MSE, and PRD. The efficient
results were obtained using Normal distribution function as 58.23 dB
SNR and 9.63e-08 MSE for optimal threshold function employing
optimal threshold value.
The SNR and PRD values obtained using different distribution
functions employing the proposed methodology are compared graphi­
cally in Fig. 9. On comparing the output values of SNR for different
distributions with the conventional and proposed estimator, normal
distribution function using the proposed estimator gives the 58.23 dB as
Fig. 8. Original and reconstructed signal using proposed estimator. compared to 54.98 dB with Gaussian distribution function. The value of
PRD is 0.001 as compared to 0.003 from the Gaussian function
The proposed methodology is applied on all records (100–234) of
and PRD as compared with the proposed estimator using Gaussian dis­
MIT-BIH Arrhythmia database for elimination of distinct noises and
tribution. Many problems in engineering practice rely on the Gaussian
artifacts from the ECG signal using different distribution functions as
distribution function of the data, which in many situations is well,
Tabulated in Table 15.
justified. Nominal, however, is useless if the estimators were derived
From Table 15, it has been observed that the proposed methodology

Table 11
SNR evaluation for proposed estimator using Normal distribution.
S.No Threshold Function Thu Thuld Thumld Thexpld Thmini Thmou Thadp Heur

1. Hard 35.61 28.95 50.26 43.74 33.54 47.84 28.95 35.61

2. Soft 30.19 26.55 46.07 35.11 28.88 37.65 26.55 30.19
3. Non-Negative Garrote 32.14 28.24 58.23 40.13 30.54 42.23 28.24 32.14
4. Hyperbolic 14.18 14.10 14.26 14.22 14.16 14.23 14.10 14.18
5. Trimmed 25.67 24.48 27.46 27.09 25.21 26.85 24.14 25.67
6. Semi Soft 10.00 10.19 9.86 9.94 10.08 9.88 10.19 10.00

Table 12
MSE evaluation for proposed estimator using Normal distribution.
S. No. Threshold Function Thu Thuld Thumld Thexpld Thmini Thmou Thadp Heur

1. Hard 1.76e-05 8.17e-05 6.03e-07 2.70e-06 2.83e-05 4.88e-06 1.05e-06 8.17e-05

2. Soft 6.14e-05 1.41e-04 1.58e-06 1.97e-05 8.30e-05 3.09e-05 1.10e-05 1.41e-04
3. Non Negative Garrote 3.92e-05 9.60e-05 9.63e-08 6.22e-06 5.65e-05 1.40e-05 3.04e-06 9.60e-05
4. Hyperbolic 0.002 0.002 0.002 0.002 0.002 0.002 0.002 0.002
5. Trimmed 1.73e-04 2.28e-04 1.15e-04 1.25e-04 1.93e-04 2.43e-04 1.32e-04 2.47e-04
6. Semi Soft 0.006 0.006 0.006 0.006 0.006 0.006 0.006 0.006

8
N. Prashar et al. Biomedical Signal Processing and Control 63 (2021) 102212

Table 13
PRD evaluation for proposed estimator using Normal distribution.
S.No Threshold Function Thu Thuld Thumld Thexpld Thmini Thmou Thadp Heur

1. Hard 0.01 0.03 0.003 0.006 0.02 0.008 0.004 0.03

2. Soft 0.03 0.04 0.005 0.017 0.03 0.02 0.01 0.04
3. Non Negative Garrote 0.002 0.003 0.001 0.009 0.002 0.014 0.006 0.003
4. Hyperbolic 0.19 0.19 0.19 0.19 0.19 0.19 0.19 0.19
5. Trimmed 0.05 0.05 0.04 0.04 0.05 0.06 0.04 0.06
6. Semi Soft 0.31 0.30 0.32 0.31 0.31 0.31 0.32 0.30

Table 14
Comparison Table of performance metrics using different distribution functions.
Non-negative Garrote with a universal modified threshold level dependent SNR MSE PRD

Conventional Proposed Conventional Proposed Conventional Proposed

Gaussian distribution 45.67 54.98 1.73e-6 2.03e-07 0.005 0.003

Normal distribution 48.65 58.23 8.75e-07 9.63e-08 0.003 0.001
Exponential distribution 44.05 53.12 2.52e-06 3.12e-07 0.006 0.004

Fig. 9. SNR and PRD evaluation for different distributions using conventional and proposed estimator.

is efficient for performing denoising of ECG signal using both Normal Electromyogram and White Gaussian noise [43], Wavelet transform
and Gaussian distribution functions. The results further depicts that with Genetic Algorithm (GA + WT) [44], Fourier decomposition method
average value of SNR obtained using proposed estimator is (57.69 dB) (FDM) to separate both BW and PLI simultaneously from the recorded
for Normal distribution and (54.58 dB) for Gaussian distribution while ECG signal and obtain clean ECG data [45], Eigen value decomposition
average value of SNR obtained using conventional estimator is (EVD) to remove Baseline and Power line interference [46], Empirical
(46.23 dB). mode decomposition Wavelet transform based on extended Kalman fil­
ter (EMD-WT-EKF) to filter noisy ECG signal [47]. Powerline interfer­
ence has been removed by using a hybrid approach of empirical mode
3.4. Comparison with existing work decomposition method (EMD) and wavelet analysis [48]. An improved
eigen value decomposition of Hankel matrix and Hilbert transform
The proposed technique is implemented for distinctive database of (IEVDHM-HT)-based method has been applied to detect the presence of
MIT-BIH arrhythmia. The results obtained by our proposed technique coronary artery disease (CAD) using electrocardiogram (ECG) signals
are compared with the existing research techniques for ECG signal [49]. A gradient descent algorithm is used to supervise the training of
denoising as tabulated in Table 16. The result shows a significant in­ the network, which can learn and preserve the small waveform features
crease in SNR using the proposed technique using Gaussian and Normal in the ECG signal [50]. The performance evaluation parameters clearly
distribution functions. The percentage SNR improvement of proposed indicate that our proposed technique outperforms in elimination of
technique using Normal distribution over DT-WT is shown in Table 17 to noise content from the signal as shown in Table 18.
draw a better interpretation. Comparison Table18 further strengthen that our proposed technique
To validate the efficiency of proposed technique, the technique is yields effective results in terms of SNR, PRD, and MSE as compared with
compared with the latest state of art work such as Modified Garrote other state of the art research papers inferable from its high efficiency
Threshold Shrinkage Function for removal of high frequency noise [34], and accuracy.
Mallet tree Decomposition (MTD) using mother wavelets to remove high
frequency noise [42], A hybrid of the -hill climbing metaheuristic al­
gorithm and wavelet transform (WT) to remove Power line noise,

9
N. Prashar et al. Biomedical Signal Processing and Control 63 (2021) 102212

Table 15
Performance Evaluation for MIT-BIH Arrhythmia Database for Normal and Gaussian distribution functions using proposed estimator.
Record Lead Normal Distribution Gaussian Distribution

SNR MSE PRD SNR MSE PRD

100 MLII 58.56 1.78E-07 0.001 55.25 3.83E-07 0.003

100 V5 58.23 9.63E-08 0.001 54.98 2.04E-07 0.003
101 MLII 61.33 3.21E-07 8.58E-04 58.22 1.89E-07 0.001
101 V1 52.60 1.68E-07 0.002 49.09 3.77E-07 0.004
102 V5 59.34 1.02E-07 0.001 55.43 2.00E-07 0.003
102 V2 62.90 7.00E-08 7.16E-04 61.12 1.06E-08 8.79E-04
103 MLII 65.69 8.39E-08 9.23E-04 58.73 4.18E-07 0.001
103 V2 55.70 1.50E-07 0.001 53.03 2.77E-07 0.003
104 V5 56.80 2.48E-07 0.001 53.84 4.91E-07 0.002
104 V2 65.92 7.84E-08 1.03E-05 62.09 9.46E-08 7.16E-04
105 MLII 54.97 4.89E-07 0.001 52.64 8.35E-07 0.002
105 V1 70.02 9.95E-08 0.001 69.77 2.12E-07 0.002
106 MLII 54.27 5.14E-07 0.001 52.23 1.05E-06 0.002
106 V1 44.35 3.70E-07 0.006 40.83 8.32E-07 0.009
107 MLII 57.07 1.39E-06 0.001 54.62 2.45E-06 0.003
107 V1 57.11 8.18E-07 0.001 55.92 1.27E-06 0.003
108 MLII 59.61 1.26E-07 1.00E-03 56.23 2.74E-07 0.002
108 V1 62.85 1.33E-07 7.20E-04 59.70 2.74E-07 0.001
109 MLII 53.05 2.11E-06 0.002 50.71 3.62E-06 0.003
109 V1 56.30 5.87E-07 0.001 53.89 1.12E-06 0.002
111 MLII 55.06 1.80E-07 0.001 52.62 3.15E-07 0.002
111 V1 53.40 3.21E-07 0.002 50.41 6.39E-07 0.003
112 MLII 66.59 1.85E-07 4.68E-04 63.02 4.20E-07 7.07E-04
112 V1 63.97 1.81E-07 6.33E-04 60.69 3.85E-07 9.24E-04
113 MLII 57.38 3.55E-07 0.001 54.87 6.33E-07 0.002
113 V1 66.43 3.66E-08 4.77E-04 63.06 7.95E-08 7.03E-04
114 V5 59.11 6.82E-08 1.10E-03 55.70 1.49E-07 0.001
114 MLII 59.45 6.69E-08 1.10E-03 56.06 1.46E-07 0.001
115 MLII 63.55 1.47E-07 6.64E-04 60.13 3.24E-07 9.85E-04
115 V1 60.78 7.88E-08 9.15E-04 57.05 1.86E-07 0.001
116 MLII 58.02 2.04E-06 1.30E-03 55.40 3.74E-06 0.001
116 V1 59.39 5.05E-07 1.10E-03 55.90 1.13E-06 0.001
117 MLII 65.10 2.31E-07 5.56E-04 61.78 4.95E-07 8.14E-04
117 V2 61.14 2.37E-07 8.77E-04 57.90 5.01E-07 0.001
118 MLII 57.30 2.13E-06 1.40E-03 53.95 4.61E-06 0.002
118 V1 54.86 1.48E-06 1.80E-03 51.91 2.91E-06 2.54E-03
119 MLII 62.85 5.02E-07 7.20E-04 59.99 9.70E-07 1.00E-03
119 V1 61.07 2.16E-07 8.84E-04 57.51 4.92E-07 1.30E-03
121 MLII 68.10 1.01E-07 3.94E-04 64.77 2.16E-07 5.77E-04
121 V1 65.79 1.30E-07 5.13E-04 62.03 3.09E-07 7.91E-04
122 MLII 62.51 4.27E-07 7.49E-04 59.44 8.66E-07 1.10E-03
122 V1 59.30 5.05E-07 1.10E-03 55.97 1.09E-06 1.60E-03
123 MLII 65.12 2.75E-07 5.55E-04 62.20 5.38E-06 7.76E-03
123 V5 60.23 1.71E-07 9.73E-04 56.76 3.82E-06 1.45E-03
124 MLII 70.14 7.37E-09 3.11E-04 66.60 1.66E-07 4.68E-03
124 V4 67.63 8.19E-08 4.15E-04 64.01 1.88E-07 6.30E-04
200 MLII 55.47 5.03E-07 1.70E-03 52.29 1.04E-06 2.40E-03
200 V1 54.03 5.07E-08 2.00E-03 50.60 1.12E-07 3.00E-03
201 MLII 53.03 3.24E-07 2.20E-03 49.75 6.88E-07 3.30E-03
201 V1 50.71 5.51E-07 2.90E-03 47.39 1.18E-06 4.30E-03
202 MLII 58.73 5.66E-08 1.20E-03 55.47 1.20E-07 1.70E-03
202 V1 54.49 9.47E-08 1.90E-03 50.86 2.18E-07 2.86E-03
203 MLII 46.41 5.66E-06 4.80E-03 43.09 1.21E-05 7.00E-03
203 V1 41.01 2.94E-06 8.90E-03 37.56 6.53E-06 1.32E-02
205 MLII 59.93 1.96E-07 1.00E-03 56.45 4.36E-07 1.50E-03
205 V1 65.18 7.90E-08 5.51E-04 61.80 1.72E-07 8.13E-04
207 MLII 63.10 6.68E-08 6.99E-04 59.48 1.54E-07 1.10E-03
207 V1 63.96 7.25E-08 6.34E-04 60.33 1.67E-07 9.63E-04
208 MLII 54.20 1.09E-06 2.00E-03 51.82 1.88E-06 2.60E-03
208 V1 53.16 6.19E-07 2.20E-03 51.40 9.30E-07 2.70E-03
209 MLII 53.81 3.53E-07 2.00E-03 50.34 7.86E-07 3.00E-03
209 V1 39.79 7.21E-07 0.010 36.77 1.45E-06 0.014
210 MLII 58.58 1.09E-07 0.001178 55.49 2.22E-07 1.70E-03
210 V1 46.94 3.09E-07 0.004496 43.50 6.83E-07 6.70E-03
212 MLII 50.03 1.08E-06 0.003152 47.36 2.00E-06 4.30E-03
212 V1 52.21 1.10E-06 0.002453 49.91 1.87E-06 3.20E-03
213 MLII 49.68 5.35E-06 0.003 47.32 9.20E-06 4.30E-03
213 V1 52.55 4.32E-06 0.002 50.08 7.64E-06 0.003
214 MLII 58.88 2.57E-07 0.001 56.28 4.68E-07 0.002
214 V1 55.63 5.63E-07 0.001 53.13 9.99E-07 0.002
215 MLII 50.93 5.54E-07 0.002 47.40 1.25E-06 0.004
215 V1 58.50 1.69E-07 0.001 55.11 3.69E-07 0.002
217 MLII 57.46 5.81E-07 0.001 54.26 1.22E-06 0.002
(continued on next page)

10
N. Prashar et al. Biomedical Signal Processing and Control 63 (2021) 102212

Table 15 (continued )
Record Lead Normal Distribution Gaussian Distribution

SNR MSE PRD SNR MSE PRD

217 V1 59.04 2.44E-07 0.001 55.95 4.97E-07 0.002

219 MLII 59.62 5.91E-07 0.001 56.41 1.24E-06 0.002
219 V1 61.76 3.08E-07 8.17E-04 58.51 6.51E-07 1.20E-03
220 MLII 62.53 2.69E-07 7.48E-04 59.81 5.04E-07 1.00E-03
220 V1 61.59 1.11E-07 8.33E-04 58.01 2.53E-07 1.30E-03
221 MLII 57.52 2.27E-07 1.30E-03 54.42 4.63E-07 1.90E-03
221 V1 55.95 1.76E-07 1.60E-03 52.53 3.88E-07 2.40E-03
222 MLII 53.30 1.32E-07 2.20E-03 50.08 2.77E-07 3.10E-03
222 V1 65.05 4.21E-08 5.59E-04 61.62 9.28E-08 8.30E-04
223 MLII 64.86 1.48E-07 5.71E-04 61.23 3.41E-07 8.68E-04
223 V1 61.89 1.23E-07 8.04E-04 58.88 2.47E-07 1.14E-03
228 MLII 51.10 6.73E-07 2.80E-03 47.93 1.40E-06 4.00E-03
228 V1 55.76 7.08E-07 1.60E-03 52.36 1.55E-06 2.40E-03
230 MLII 57.73 3.39E-07 1.30E-03 54.43 7.26E-07 1.90E-03
230 V1 53.89 4.28E-07 2.00E-03 51.04 8.24E-07 2.80E-03
231 MLII 55.64 2.91E-07 1.70E-03 52.59 5.88E-07 2.35E-03
231 V1 57.99 7.30E-08 1.30E-03 54.47 1.64E-07 1.89E-03
232 MLII 46.86 8.33E-07 4.50E-03 43.44 1.83E-06 6.73E-03
232 V1 52.83 2.24E-07 2.30E-03 49.47 4.86E-07 3.36E-03
233 MLII 57.69 6.42E-07 1.30E-03 55.23 1.13E-06 1.73E-03
233 V1 54.32 7.05E-07 1.90E-03 51.71 1.28E-06 2.60E-03
234 MLII 53.37 3.50E-07 2.10E-03 49.99 7.62E-07 3.17E-03
234 V1 49.48 2.22E-07 3.40E-03 46.05 4.90E-07 5.00E-03

4. Conclusion and future work in the processing of the ECG signals. In future, a novel feature selection
optimization technique will be developed to select the optimal features
Elimination of noise from ECG signals is a challenging problem as form larger feature pool that will be further utilized for assisting the
noise causes serious issues in the inspection of the signal and its analysis. clinicians to detect the heart diseases.
This paper proposes a robust estimator utilizing different optimal dis­
tribution functions for ECG signal denoising by employing DTCWT.
Evaluations were performed on the standard benchmark data of MIT-
BIH arrhythmia database. The proposed approach accomplished good
performance evaluation resulting 58.23 dB, SNR and 9.63e-08, MSE in
comparison with the 48.65 dB, SNR and 8.75e-07, MSE using conven­ Table 17
tional method for the normal distribution function for 100 record. The Percentage improvement in SNR using proposed Technique over DT-WT.
obtained results show an overall improvement in SNR by 19% by using Record Charri et al. Proposed Technique DTCWT Percentage
the proposed estimator. In this paper author has also considered the (2017) DT-WT (Normal distribution) improvement (%)
effect of different distribution functions on proposed technique which 100 45.83 58.23 27.06
shows 54.98 dB, SNR and 2.03e-07, MSE using Gaussian function and 101 49.39 61.33 24.17
53.12 dB, SNR, and 3.12e-07, MSE using an exponential function. The 102 49.87 59.34 18.99
novelty of this work also lies in an exhaustive empirical comparative 103 55.41 65.69 18.55
104 57.51 65.92 14.62
analysis with the existing research work using the same database with 105 66.18 70.02 6.07
the proposed normal distribution technique. The percentage improve­ 106 51.89 54.27 4.59
ment in SNR for 100–109 records employing proposed technique over 107 55.52 57.11 2.86
the existing research technique [25] which shows a considerable 108 48.16 59.61 22.20
109 53.76 56.30 4.72
amount of improvement validating the proposed methodology. The
obtained results show that the proposed estimator is more suitable for
ECG denoising than Donoho’s threshold and can be successfully applied Table 18
Comparison of Performance metrics of various techniques with Proposed
Techniques.
Table 16
S. Technique SNR MSE PRD
Comparison of the SNR (dB) using 100 to 109 records for MIT-BIH arrhythmia.
No. (dB)
Record Charri Charri Proposed Technique Proposed
1. DWT + Garrote [34] 41.79 – –
et al. et al. DTCWT (Gaussian Technique DTCWT
2. MTD [42] 26.53 1.41e- 0.0471
(2017) (2017) DT- distribution) (Normal
04
DWT WT distribution)
3. Hybridizing β-hill climbing + wavelet 16.7
100 42.65 45.83 54.98 58.23 transform [43]
101 48.41 49.39 58.22 61.33 4. GA + WT [44] 19.3
102 45.49 49.87 55.43 59.34 5. FDM [45] 23.3 – –
103 52.20 55.41 58.73 65.69 6. EVD Method [46] 12.42 – –
104 54.13 57.51 62.09 65.92 7. EMD-WT-EKF Method [47] 10.3 – –
105 62.82 66.18 69.77 70.02 8. Deep Factor Analysis [50] 22.62 0.029 –
106 47.45 51.89 52.23 54.27 9. Proposed Technique (Normal) 58.23 9.63e- 0.001
107 54.57 55.52 55.92 57.11 08
108 43.32 48.16 59.70 62.85 10. Proposed Technique (Gaussian) 54.98 2.03e- 0.003
109 51.25 53.76 53.82 56.30 07

11
N. Prashar et al. Biomedical Signal Processing and Control 63 (2021) 102212

CRediT authorship contribution statement [22] S. Rana, S. Jain, J. Virmani, SVM-based characterization of focal kidney lesions
from B-Mode ultrasound images, research journal of pharmaceutical, Biol. Chem.
Sci. 7 (4) (2016) 837–846.
Navdeep Prashar: Conceptualization, Writing - original draft, [23] S. Jain, Classification of mitogen activated protein kinase using different wavelet
Writing - review & editing. Meenakshi Sood: Investigation, Supervi­ transforms (Discrete and gabor), Asian J. Microbiol. Biotechnol. Environ. Sci. 20
sion, Writing - review & editing. Shruti Jain: Investigation, Supervision, (2) (2018) 569–574.
[24] H.S. NiranjanaMurthy, M. Meekashi, ECG Signal Denoising and Ischemic Event
Writing - review & editing. Feature Extraction using Daubechies Wavelets, Int. J. Comput. Appl. 67 (2) (2013)
29–33.
Declaration of Competing Interest [25] O.El. B’charri, R. Latif, K. Elmansouri, A. Abenaou, W. Jenkal, ECG signal
performance de-noising assessment based on threshold tuning of dual-tree wavelet
transform, Biomed. Eng. 16 (2017) 26.
The authors report no declarations of interest. [26] F. Wang, Ji Z, Application of the dual-tree complex wavelet transform in
biomedical signal, Biomed. Mater. Eng. 24 (1) (2014) 109–115.
[27] F.M. Ghombavani, K. Kiani, A powerful novel method for ECG signal de-noising
Appendix A. Supplementary data using different thresholding and dual tree complex wavelet transform, 2nd
International Conference on Knowledge–Based Engineering and Innovation (KBEI)
Supplementary material related to this article can be found, in the (2015) 966–971.
[28] N. Prashar, M. Sood, S. Jain, Semiautomatic detection of cardiac diseases
online version, at https://doi.org/10.1016/j.bspc.2020.102212. employing dual tree complex wavelet transform, Period. Eng. Nat. Sci. 6 (2) (2018)
129–140.
References [29] https://www.physionet.org/physiobank/database/html/mitdbdir/records.htm.
[30] K.D. Priyaa, G.S. Raob, P.S.V. SubbaRaoa, Comparative analysis of wavelet
thresholding techniques with wavelet-Wiener filter on ECG Signal, in: 4th
[1] P. Karthikeyan, M. Murugappan, S. Yacoob, ECG signal denoising using wavelet
International Conference on Recent Trends in Computer Science & Engineering,
threshoding techniques in human stress assessment, Int. J. Electr. Eng. Inform. 4
Procedia Computer Science, 87, 2016, pp. 178–183.
(2) (2012) 306–319.
[31] R. Sharma, R. Kashyap, Design & implementation of denoising of ECG system for
[2] A. Dhiman, A. Singh, S. Dubey, S. Jain, Design of lead II ECG waveform and
medical applications, Int. J. Sci. Eng. Technol. Res. 6 (8) (2017) 1219–1224.
classification performance for morphological features using different classifiers on
[32] S. Mandala, et al., Performance analysis of wavelet-based denoising techniques for
lead II, Res. J. Pharm. Biol. Chem. Sci. 7 (4) (2016) 1226–1231.
ECG Signal, Fifth International Conference on Information and Communication
[3] N. Prashar, J. Dogra, M. Sood, S. Jain, Removal of electromyography noise from
Technology (ICoICT) (2017).
ECG for high performance biomedical systems, Network Biol. 8 (1) (2018) 12–24.
[33] A. Awal, S.S. Mostafa, M. Ahmad, Performance analysis of savitzky-golay
[4] M. Sood, Ensemble framework for epileptic seizure classification from EEG signals,
smoothing filter using ECG signal, IJCIT 1 (2) (2011) 90–95.
In J. Global Trends. Pharm. Sci. 9 (3) (2018) 5838–5847.
[34] A.R. Zubair, Y.K. Ahmed, K.A. Akande, Electromyography noise suppression in
[5] M. Benmalek, A. Charef, F. Abdelliche, Preprocessing of the ECG signals using the
electrocardiogram signal using modified garrote threshold shrinkage function, Afr.
his-purkinje fractal system, IEEE, 7th International Multi-Conference on Systems,
J. Comp. ICT 11 (3) (2018) 85–94.
Signals and Devices (2010).
[35] G.G. Tsaneva, K. Tcheshmedjiev, Denoising of electrocardiogram data with
[6] K.M. Chang, S.H. Liu, Gaussian noise filtering from ECG by wiener filter and
methods of wavelet transform, in: International Conference on Computer Systems
ensemble empirical mode decomposition, J. Signal Process. Syst. 64 (2) (2011)
and Technologies - CompSysTech’13 Ruse, Bulgaria, 2013, pp. 9–16.
249–264.
[36] J. Dogra, M. Sood, S. Jain, N. Prashar, Segmentation of magnetic resonance images
[7] M.P. Chawla, et al., Artifacts and noise removal in electrocardiograms using
of brain using thresholding techniques, 4th IEEE International Conference on
independent component analysis, Int. J. Cardiol. 129 (2008) 278–281.
Signal Processing and Control (2017) 311–315.
[8] M.S. Chavan, R.A. Agarwala, M.D. Uplane, Suppression of noise in the ECG signal
[37] D.L. Donoho, I.M. Johnstone, Ideal Spatial adaption by wavelet shrinkage,
using digital IIR filter, in: Proceedings of the 8th WSEAS International Conference
Biometrika (1994) 425–455.
on Multimedia Systems and Signal Processing, Hangzhou, China, 2008,
[38] D.L. Donoho, De-noising by soft-thresholding, IEEE Trans. Inf. Theory 41 (1995)
pp. 335–343.
613–627.
[9] S.M.M. Martens, et al., An improved adaptive power line interference canceller for
[39] M.B. Abdullah, On a robust correlation coefficient, Statistician 39 (4) (1990)
electrocardiography, IEEE Trans. Biomed. Eng. 53 (2006) 2220–2231.
455–460.
[10] M.S. Chavan, R.A. Agrawala, M.D. Uplane, Design and implementation of digital
[40] A.Z.M. Shafiullah, J.A. Khan, A new robust correlation estimator for bivariate data,
FIR equiripple notch filter on ECG Signal for removal of power line interference, in:
Bangladesh J. Sci. Res. 24 (2) (2011) 97–106.
Wseas International Conference on Signal Processing, 4, 2005, pp. 221–230 (4).
[41] D. Wang, Y.1 Si, W. Yang, G. Zhang, T. Liu, A novel heart rate robust method for
[11] C.B. Mbachu, G.N. Onoh, V.E. Idigo, Ifeagwu Scholastica Nnebe EN : processing
short-term electrocardiogram biometric identification, Appl. Sci. Basel (Basel) 9
ECG signal with kaiser window-based FIR digital filters, Int. J. Eng. Sci. Technol. 3
(2019) 201–221.
(2011) 6775–6783.
[42] N. Prashar, M. Sood, S. Jain, Morphology analysis and time interval measurements
[12] S.G. Tareen, Removal of power line interference and other single frequency tones
using mallat tree decomposition for CVD detection (Eds.). ICAICR 2018, CCIS 955©
from signals. Computer Science and Electronics, Mälardalen University, Sweedan,
Springer Nature Singapore Pte. Ltd., 2019, pp. 71–181.
2008.
[43] Alyasseri, Zaid Abdi Alkareem, et al., Hybridizing β-hill climbing with wavelet
[13] L. Chmelka, J. Kozumplik, Wavelet-based wiener filter for electrocardiogram signal
transform for denoising ECG signals, in: Information Sciences, vol. 429, 2018,
denoising, Comput. Cardiol. 32 (2005) 771–774.
pp. 229–246.
[14] O. Sayadi, M.B. Shamsollahi, ECG denoising with adaptive bionic wavelet
[44] Alan S. Said Ahmad, Majd S. Matti, Omar A.M. ALhabib, Sabri Shaikhow,
transform, in: Proceedings of 28th IEEE,EMBS Annual International Conference,
Denoising of arrhythmia ECG signals, Int. J. Med. Res. Health Sci. 7 (3) (2018)
NewYork, 2006, pp. 6597–6660.
83–93.
[15] M. Khan, F. Aslam, T. Zaidi, S.A. Khan, Wavelet based ECG denoising using signal-
[45] Amit Singhal, et al., An efficient removal of power-line interference and baseline
noise residue method, 5th International Conference on BioinformaTics and
wander from ECG signals by employing Fourier decomposition technique, Biomed.
Biomedical Engineering (2011).
Signal Process. Control (2020).
[16] Taigang He, G. Clifford, L. Tarassenko, Application of independent component
[46] R.R. Sharma, R.B. Pachori, Baseline wander and power line interference removal
analysis in removing artefacts from the electrocardiogram, Neural Comput. Appl.
from ECG signals using eigenvalue decomposition, Biomed. Signal Process. Control
15 (2) (2006) 105–116.
45 (2018) 33–49.
[17] A.El-Dahshan El-Sayed, Genetic algorithm and wavelet hybrid scheme for ECG
[47] R. Sameni, M.B. Shamsollahi, Filtering noisy ECG signals using the extended
signal denoising, Telecommun. Syst. 46 (3) (2010) 209–215.
Kalman filter based on a modified dynamic ECG model, Comput. Cardiol. (2005)
[18] V. Sukanya, C. Saritha, Y.N. Murthy, ECG signal analysis using wavelet transforms,
1017–1020.
Bulg. J. Phys. 35 (2008) 68–77.
[48] S. Agrawal, A. Gupta, Fractal and EMD based removal of baseline wander and
[19] S. Jain, Classification of protein kinase B using discrete wavelet transform, Int. J.
powerline interference from ECG signals, Comput. Biol. Med. 43 (11) (2013)
Inf. Technol. 10 (2) (2018) 211–216.
1889–1899.
[20] T.T. Cai, B.W. Silverman, Incorporating information on neighbouring coefficients
[49] Rishi Raj Sharma, et al., Joint time-frequency domain-based CAD disease sensing
into wavelet estimation, Sankhya The Indian J. Stat. Series B 63 (2) (2001)
system using ECG signal, IEEE Sens. J. 19 (10) (2019).
127–148.
[50] Wang Ge, et al., ECG signal denoising based on deep factor analysis, Biomed. Signal
[21] M. Sood, S. Jain, Ensemble classifier framework for epileptic seizure classification
Process. Control 57 (2019).
of EEG signals, Int. J. Emerg. Technol. Learn. 10 (2) (2019) 23–28.

12