Computer Methods and Programs in Biomedicine 188 (2020) 105304

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine

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

A singular spectrum analysis-based model-free electrocardiogram

denoising technique
Sourav Kumar Mukhopadhyay∗, Sridhar Krishnan
Department of Electrical, Computer, and Biomedical Engineering, Ryerson University, 350 Victoria Street, Toronto, ON M5B 2K3, Canada

a r t i c l e i n f o a b s t r a c t

Article history: Background and objective: An efficient and robust electrocardiogram (ECG) denoising technique caters
Received 20 June 2019 three-fold benefits in the subsequent processing steps: first, it helps improving the accuracy of extracted
Revised 19 December 2019
features. Second, the improved accuracy in the extracted features enhances the performance as well as
Accepted 25 December 2019
the reliability of computerised cardiovascular-disease diagnosis systems, and third, it also makes the in-
terpretation task easier for the clinicians. Albeit a number of ECG denoising techniques are proposed in
Keywords: the literature, but most of these techniques suffer from one or more of the following drawbacks: i) model
ECG distortion or function dependency, ii) sampling-rate dependency, or iii) high time-complexity.
Dynamic window-length
ECG denoising Methods: This paper presents a singular spectrum analysis (SSA)-based ECG denoising technique address-
Mean opinion score ing most of these afore-mentioned shortcomings. First, a trajectory matrix of dimension K × L is formed
Singular spectrum analysis using the original one-dimensional ECG signal of length N. In SSA operation the parameter L, which is
Data-Driven Denoising denoted as the window-length, plays a very important role and is related to the sampling frequency of
the signal. In this research the value of L is calculated dynamically based on the morphological property
of the ECG signal. Then, the matrix is decomposed using singular value decomposition technique, and the
principal components (PC) of the original ECG signal are computed. Next, the reconstructed components
(RC) are calculated from the PCs, and all the RCs are filtered through Butterworth bandpass and notch
filters. An optimum number of filtered RCs are retained based on their significance. Finally, these retained
RCs are summed up to obtain the denoised ECG signal.
Results: Evaluation result shows that the proposed technique outperforms state-of-the-art ECG denoising
methods; in particular, the mean opinion score of the denoised signal falls under the category ‘very good’
as per the gold standard subjective measure.
Conclusions: Both the quantitative and qualitative distortion measure metrics show that the proposed
ECG denoising technique is robust enough to filter various noises present in the signal without jeopardiz-
ing the clinical content. The proposed technique could be adapted for denoising other biomedical signals
exhibiting periodic or quasi-periodic nature such as photoplethysmogram and esophageal pressure signal.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction processing systems. Respiration, patients’ movement, power-line

interference, electromyogram (EMG), electro-surgical and instru-
Electrocardiogram (ECG), which is considered as the prime mentation noises are the common types of noises, which contam-
human physiological signal [1], is the graphical representation inate the ECG signal.
of heart’s electrical activity recorded by means of electrodes Researchers around the globe are continuing to develop robust
from different standardized locations on the body surface. How- and high performance noise elimination techniques from ECG sig-
ever, ECG signals get routinely contaminated during acquisi- nal for over 30 years [2]. In recent times, wavelet transform (WT)
tion by the influence of several high and low frequency noises, [3–11,37], empirical mode decomposition (EMD) and ensemble
which not only exacerbates the task of interpretation for doctors, empirical mode decomposition (EEMD) [7–11]-based ECG denois-
but also degrades the overall performance of computerized ECG ing techniques have drawn significant attention. In [3], WT is
used in combination with β -hill climbing metaheuristic algorithm
∗ for denoising ECG signal, where the original ECG signal is cor-
Corresponding author.
E-mail addresses: [email protected] (S.K. Mukhopadhyay), rupted by mixing with synthetic white Gaussian noise (WGN), and
[email protected] (S. Krishnan). then the optimum filtering parameters such as mother wavelet

https://doi.org/10.1016/j.cmpb.2019.105304
0169-2607/© 2020 Elsevier B.V. All rights reserved.
2 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304

function, number of decomposition levels, thresholding method however empirical, is provided for selecting the window-length
are selected for denoising. In [3] the authors have claimed that L, which is an important parameter to set in SSA operation. In
even for the same ECG signal at a constant sampling rate: (i) the [28], first, synthetic ECG signals are taken and mixed with power-
selection of mother wavelet function varies with varying WGN line interference noise. Then, the noisy signals are processed using
characteristics, and also (ii) the number of decomposition level SSA at different values of L, and the denoised ECGs are shown. In
varies with varying WGN characteristics. But, the reason behind [28], the SSA-based denoising technique is tested on ECG signals
such an interesting fact is left unclarified. Jain et al. have proposed recorded at 250 Hz and 1250 Hz sampling rates with four differ-
an EMD-based ECG denoising method in [7]. In [7], first, the signal ent values of L. The values of L, which are used at 250 Hz sam-
is decomposed into intrinsic mode functions (IMF), and a few pling rate are 10, 13, 15 and 18, and the values of L, which are
high-order IMFs are empirically considered to be corrupted with used at 1250 Hz sampling rate are 50, 65, 75 and 90. The assump-
low frequency noises. Then, these considered IMFs are scaled, tion, based on which the values of L are chosen in [28], is that .
and denoised using a Savitzky Golay filter for the removal of low However, neither the value of L is optimized for different sampling
frequency noises. Rest of the IMFs are filtered using a Riegmann rates, nor the noise removal efficiency at different values of L and
Liouvelle fractional integral filter for high frequency noise removal. (or) at different noise levels are studied.
However, the scaling factors are chosen based on the performance Motivation behind the proposed research work is to develop
of the algorithm only on one ECG segment of 800 samples (2.22 s a high performance and model-free ECG denoising technique. The
of duration). In [8], Rakshit et al. have proposed an EMD, WT proposed technique also addresses most of these above mentioned
and adaptive switching mean filter (ASMF)-based ECG denoising hurdles and shortcomings of other ECG denoising techniques. The
technique, where the noisy signal is decomposed into six IMFs main challenges, which are addressed in this research work are
and one residue signal. Then, each of the first three IMFs are again (1) denoising WGN and random noise-contaminated ECG signals;
decomposed into two levels using WT, and the wavelet coefficients which is the hardest problem faced by the researchers as the band-
are thresholded to discard the high frequency noises. Next, these width of the ECG signal overlaps with the bandwidths of these
three filtered IMFs are added with other three unfiltered ones to two types of noises, (2) efficient enough to expel out other types
get a partially-filtered ECG signal. Then, the R-peaks are detected of noises such as baseline-wander noise, instrumentation noise,
from this partially filtered ECG. Next, the partially-filtered signal power-line interference noise, (3) designing an ECG denoising tech-
is again processed using ASMF. Since the ASMF method highly nique that is adaptive to the sampling rate, and (4) designing an
distorts the QRS complexes regions of the ECG, the previously ECG denoising technique that is data-driven, and whose perfor-
detected R-peak information are used to replace the distorted QRS mance does not depend on any pre-assumed mathematical func-
complexes with their corresponding undistorted ones obtained tion/model unlike the wavelet transform-based ones. In this pro-
before applying ASMF. Use of extended Kalman filter (EKF) is pro- posed research work, first, the original ECG signal of length N is
posed in [13] and [14] for denoising ECG. In [14], first, the noisy arranged in form of a trajectory matrix of dimension K × L, where
ECG signal is preprocessed using variational mode decomposition the window-length L is calculated dynamically. Next, the trajec-
technique along with WT-based method. Then, the R-peaks are tory matrix is decomposed using SVD method. Then, the principal
detected, and these detected ECG-beats are classified into normal components (PC) of the original ECG signal are computed by a lin-
and abnormal with the help of k-nearest neighborhood algorithm. ear combination of the trajectory matrix and the matrix of eigen-
Finally, the signal is denoised using marginalized particle-extended vectors. Next, reconstructed components (RC) are calculated from
Kalman filter. However, the performance of the technique strongly these PCs. All the RCs are filtered through Butterworth bandpass
depends on the accuracy of the classification algorithm. Sparse and notch filters. Thereafter, the dynamic weights of the filtered
decomposition and dictionary learning-based ECG denoising tech- RCs are computed, and an optimum number of filtered RCs are re-
niques are proposed in [15] and [16]. In [15], a moving average tained based on their significance. Finally, these retained RCs are
filter is used to extract the baseline wander noise, and in total summed up to obtain the denoised ECG signal.
eight different threshold values are used. However, no proper The remainder of this paper is organized as follows. The pro-
explanation is given for using such threshold values. posed SSA-based ECG denoising technique is presented in Section
Besides these above mentioned techniques, a number of effi- II. Denoising performance of the proposed technique is presented
cient ECG denoising methods based on spectro-temporal filtering in Section III. Section IV compares the denoising performances of
[12], adaptive Fourier decomposition [17], fractional zero-phase fil- the proposed technique with other state-of-art techniques. Finally,
tering [18,33], non-local means [19], Bayesian framework [20] and discussions and conclusions are drawn in Section V.
adaptive filtering [21] are also proposed in the literature in recent
times. 2. SSA-based ECG denoising technique
Singular spectrum analysis (SSA) is a model-free and data-
driven time-series-decomposition method, which decomposes a The original ECG signal (denoted as E) of length N is mapped
time series into three components: trend, seasonal components into L-lagged vectors of size K, and a trajectory matrix, which is
and noise [22]. These components are generated based on the co- denoted as ETraj , is formed with these lagged vectors. Both the
variance property of the original time series data, and they are in- rows and columns of ETraj are the subseries of E. The first col-
dependent and additive in nature [34,35]. SSA consists of two com- umn of the matrix ETraj contains the entire original-ECG-signal E
plementary stages: decomposition and reconstruction. The decom- (i.e., all the QRS-complexes, P-waves, T-waves, ST-segments and
position stage is divided into two sub-stages: embedding and sin- PR-segments); the second column of the matrix ETraj contains
gular value decomposition (SVD), and the reconstruction stage is the original ECG signal E shifted by 1 samples; the third column
also sub-divided into Eigen-triplet grouping and diagonal averag- contains the original ECG signal E shifted by 2 samples, and
ing. Use of SSA-based methods are proven to be very efficient in likewise, the Lth column contains the original ECG signal E shifted
a wide range of applications including climatology [23], electricity by L-1 number of samples.
consumption forecasting [24], biomedical image processing [25,26], ⎡ ⎤
E1 E2 ... EL
gait parameter estimation [27]. However, the potential of denois-
⎢ E2 E3 ... EL+1 ⎥
ing biosignals; in particular, ECG signal, using SSA-based method ET ra j =⎢ . .. .. ⎥ (1)
is not explored to a good extent to date. Yang et al. have proposed ⎣ .. . ... .
⎦
an SSA-based ECG denoising technique in [28], where a guideline, EK EK+1 ... EL+K−1
 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304 3

in the decreasing order of magnitude, i.e., λ1 ≥ λ2 ≥ … ≥ λL , and

the corresponding eigenvectors of S are denoted as U = [u1 , u2 ,…,
uK ]. The ith collection of (λi , ui , vi ) is call the ith -eigentriple of the
SVD. The energy-contribution of the ith -eigentriple is expressed

L
as λ2i / λ2j , which is also known the singular-spectrum of the
j=1
time series data [39, 40]. If vi = ETTra j ui / λi , then it is possible to
write the trajectory matrix as:
d d
ET ra j = ET ra j_i = λi ui vTi = ET ra j_1 + ET ra j_2 + . . . + ET ra j_d
Fig. 1. . Structures of the l-lagged vectors containing the subseries of E.
i=1 i=1
(3)
where, K = N − L + 1 and K  L.
where d = argmaxi { λi > 0}, ET ra j_i = λi ui vTi , ui is the ith left
From Eq. (1) it is evident that the elements of all the anti-
eigenvector, and vi is the ith right eigenvector [27]. In SVD
diagonals of the trajectory matrix are the same, i.e., Eij = Eji (i =
operation V = [v1 , v2 ,…, vL ] is denoted as the right singular ma-
j). Therefore, the trajectory matrix ETraj is a Henkel matrix.
trix, which contains the information about the most important axis
Selection of the proper value of the window-length L is of
of the data. The first right singular vector v1 reveals the direc-
great importance in SSA operation. The efficiency of SSA in sep-
tion having the most variance, and vL reveals the direction having
arating different signal-components from a composite-signal de-
the least variance. The variance is precisely defined by the singu-
pends on the value of L. An SSA-based method is able to separate
lar values. The strong inter-sample and inter-heartbeat correlation
two signals from a composite signal if their frequency-difference is
of the ECG signal enhances the covariance among the eigenvec-
≥ KL × K1 = 1L Hz. On the other hand, SSA fails to separate two sig-
tor of the matrix U. Since the SVD operation arranges the singular
nals from a composite signal if (1) the amplitudes of both the sig-
values in decreasing order of magnitude, the small singular values
nals are equal, or (2) their frequency-difference is less than 1L Hz.
barely carry any signal-information, and these small singular val-
A detailed mathematical-analysis and theoretical-explanation of
ues could be discarded [36]. The original signal E can be consid-
SSA method can be found in [40]. However, The window-length
ered as the composition of the ECG signal and the noise. If the in-
L should be chosen in such a way that all the ECG-beats (i.e., the
dices I = {i1 ,…ip } (p < d) of the eigenvalues, which hold the most
QRS-complexes) which are present in E, are also to be there in all
of the ECG-signal-information are known, then the matrix corre-
the lagged vectors, i.e., in all the columns of ETraj . A small value of
sponding to the ECG-signal-only can be written as follows [28, 38].
L may not help denoising the signal at all, and the use of a large L
value (i) is redundant, (ii) increases the computational burden, and ⎡ ⎤
E1,1 E1,2 ... E1,K
most importantly (iii) may not guarantee that all the ECG-beats are ip
⎢E2,1 E2,2 ... E2,K ⎥
present in all the columns of ETraj , which leads to a significant loss EECG = ET ra j_i =⎢ . .. .. ⎥ (3a)
of clinical information. In this research work the window-length L ⎣ .. . ... .
⎦
i=i1
is calculated based on the inherent periodicity and morphological EL,1 EL,2 ... EL,K
property of the ECG. Since, the maximum possible heart rate is 240
beats per minute [29], the optimum value of the window-length L However, the indices, which hold the most of the ECG-signal-
can be calculated as below. information are unknown. Next, the PCs are computed by a linear
combination of the trajectory matrix ETraj and the matrix of eigen-
60 × sampling rate (in Hz )
L= (2) vectors V = [v1 , v2 ,…, vL ].
240
PCK×L = ETTra j × VL×L (4a)
Eq. (2) guarantees that (i) all the ECG-beats which are present
in E, are also present in all the columns of ETraj , and (ii) the ECG- The columns of the matrix PCK × L are the principal compo-
beats present in the adjacent columns of the ETraj matrix would nents of E. Then, a matrix of dimension K × L is created by an in-
never overlap with their preceding QRS complexes. verse projection of the PCs and the matrix of eigenvectors V. Finally,
Let us take an example. It is well known that the clinical band- the reconstructed components (RCs) of the original ECG signal are
width of an ECG signal lies between 0.5 to 100 Hz [1]. Now, if an computed by averaging along the anti-diagonals of this matrix [28].
ECG signal is recorded at a sampling rate of 1 kHz, then, follow-
ing Eq. (2), the value of L = 250. Therefore, (1) any high frequency 1
RCi = El,k (4b)
noise above (100 Hz + 1L Hz) 10 0.0 04 kHz, and (2) any low fre- N (Wi )
(l,k )∈Ai
quency noise below (0.5 Hz – 1L Hz) 0.496 Hz could be expelled
where, Wi = {(l, k): l + k = i + 1, 1 ≤ l ≤ L, 1 ≤ k ≤ K}, N(Wi )
out from the ECG using an SSA-based technique.
is the number of elements in Wi . The original ECG signal E can be
Structures of the L-lagged vectors containing the subseries
obtained by summing up all the RCs .
of E are shown in Fig. 1. From Eq. (2) it can be noted that
L ∞ sampling rate of the ECG signal. Therefore, the number of L

columns of the matrix ETraj depends on the sampling rate of the E= RCi (4c)
original ECG signal, and the number of rows of the matrix ETraj i=1

depends on the length of the original ECG signal. Hence, the di- Now, all the RCs are filtered using a zero-phase 4th order But-
mension of the matrix ETraj is dynamic. terworth bandpass filter having lower and upper cut-off frequen-
In the next step, ETraj is decomposed using SVD to represent it cies 0.5 Hz and 100 Hz [1], respectively, and a notch filter to re-
as a sum of rank-one biorthogonal elementary matrices. The SVD move the power-line interference noise. Filtered RCs are denoted
technique factorizes a matrix into the product of another three as FRCs. Fig. 2A shows an example of first and last two RCs and
matrices: an orthogonal matrix (U), a diagonal matrix (S) and the FRCs, and Fig. 2B shows their corresponding frequency spectrums.
transpose of an orthogonal matrix (V). The eigenvalues of the ma- From Fig. 2A it can be noted that the amplitudes of the high-
trix S (= ETTra j × ET ra j ) are denoted as λ = λ1 , λ2 ,…, λL , which are order RCs (RCL − 1 , RCL ) get noticeably suppressed compared to
4 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304

Fig. 2. A: An example of first and last two RCs and FRCs. B: (A) Frequency spectrum of RC1 and FRC1 , (B) Frequency spectrum of RC2 and FRC2 , (C) Frequency spectrum of
RCL − 1 and FRCL − 1 , and (D) Frequency spectrum of RCL and FRCL .

that of the low-order counterparts afterward the bandpass and the maximum and minimum amplitude of each FRC decreases pro-
notch filtering operations, which strongly suggests that the high- gressively, so does their dynamic weights. An example of the dis-
order RCs are noise dominant. Fig. 2B corroborates this notion. tribution of dynamic weights among FRCs are shown in Fig. 3.
Now, to discard the less significant high-order FRCs, a dynamic Next, starting from the first FRC (FRC1 ), an optimum number of
weight calculation-based method is used. Dynamic weights of all FRCs are summed up until their total weight accumulates a pre-
the FRCs are calculated using the following equation. defined threshold value (γ ). This summed up FRCs is considered
K as the denoised ECG, which is denoted as dECG. The FRC-summing
i=1F RCl2 (i ) algorithm is given below.
dW (l ) = L K × 100% (5)
Step 1: set the value of γ , i = 1, temp = 0, ww = 0
i=1 F RCl (i )
2
l=1
Step 2: temp = temp + FRCi
where dW(l) is the dynamic weight of the lth FRC, L denotes the to- Step 3: ww = ww + dW(i)
tal number of FRCs, and FRCl (i) denoted the ith coefficients of the lth Step 4: if ww ≥ γ
FRC. Since the amplitude-bands of FRCs, i.e., the difference between dECG = temp
 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304 5

Table 1
Details of collected ECG data.

Database Sampling rate No. of lead taken No. of samples No. of ECG files
in Hz into consideration in each lead taken

European ST-T 250 2 (ML III and V4) 10 × 105 89

Database (edb)
MIT-BIH 360 1 (ML II) 21.6 × 103 48
arrhythmia
database (mitdb)
MIT-BIH noise 360 2 (ML II and V1) 12
stress test
database (nstdb)
Motion Artifact 500 4 (ECG-1, 2, 3, 4) 4 × 103 27
Contaminated ECG
Database
(macecgdb)
PTB diagnosis ECG 1k 12 (lead I-III, aVR, 60 × 103 519
database (ptbdb) aVL, aVF, V1-V6)

wander as well as high frequency noises above 100 Hz. It is im-

portant to note that the preprocessing is only performed at the
testing phase of the technique, and is not required when denoising
real-time ECG signals. The preprocessed signal is denoted as oECG.
Next, four different types of synthetic noises are added with oECG.
Details of these synthetic noises are given below.

(1) 50 Hz powerline interference noise (PLI).

Fig. 3. Distribution of dynamic weights among FRCs.

P LI = A sin (2π 50t ) (6)
where A = (A1 − A2)/2 where, A1 = maximum amplitde of oECG,
A2 = minimum amplitde of oECG
nopt = i
else (1) White Gaussian noise (WGN) of power 30dB at different stan-
i=i+1 dard deviation values (σ = 0.1 to 5.0) and zero mean (μ = 0).
goto Step 2 (2) Random noise (RN) in the frequency-band 20 Hz – 500 Hz with
endwhere nopt is the optimum number FRCs. different standard deviation values (σ = 0.1 to 5.0) and zero
mean (μ = 0).
3. Experimental results (3) 0.2 Hz baseline wander noise (BWN)
BW N = A sin (2π 0.2t ) (7)
Performance of the proposed SSA-based ECG denoising tech-
nique is tested through both quantitative and qualitative distortion The noisy ECG is denoted as nECG (nECG = oECG + PLI + WGN
measures. Digitized ECG signals are collected from five different + RN + BWN). Now, the nECG signal is denoised using the
databases [30] and are used as the evaluation testbeds of the im- proposed SSA-based denoising technique, which is described in
plemented ECG denoising technique. Details of these collected ECG Section 2. The denoised ECG signal is denoted as dECG.
data are shown in Table 1. In the performance evaluation phase, the metrics: input signal-
In total, the technique is tested on 695 ECG data-files. Out of to-noise ratio (SNRIN ), output signal-to-noise ratio (SNROUT ), im-
695 files, 519 files are taken from PTBDB. Among these 519 files, 17 provement in signal-to-noise ratio (SNRimp ), and percent-root-
files contain bundle branch block ECG, 16 files contain cardiomy- mean-square difference (PRD) are used as quantitative non-
opathy ECG, 15 files contain dysrhythmia ECG, 3 files contain heart diagnostic distortion measures, and wavelet energy-based diagnos-
failure ECG, 7 files contain hypertrophy ECG, 367 files contain my- tic distortion (WEDD) [31] is used as the quantitative diagnostic
ocardial infraction ECG, 4 files contain myocarditis ECG, 80 files distortion measure. SNRIN , SNROUT , SNRimp , and PRD are calculated
contain normal ECG, 1 files contains palpitation ECG, 2 files contain using the following equations.
stable angina ECG, 1 files contains unstable angina ECG, and 6 files N
oECG2i
contain valvular heart disease ECG. All the ECG recordings of nstdb SN RIN = 10log10 Ni=1 (8)
2
are highly corrupted with baseline wander, electromyogram and i=1 inNoisei
electrode motion artifact noises, which are recorded from physi-
where inNoise = PLI + WGN + RN + BWN, and N = length of oECG
cally active volunteers. ECG recordings of macecgdb are highly cor-
rupted with motion artifact noise. ECG signals of these above men- N
dECG2i
tioned databases are recorded using wired ECG acquisition mod- SN ROUT = 10log10 N i=1 (9)
2
ules. No such database is available online to date, where wireless i=1 opNoisei
ECG recordings are freely available. Therefore, the performance of where opNoise = nECG − dECG
the technique is tested only on wired ECG recordings. N
All these digitized ECG signals collected from different i=1 oECG2i − nECG2i
databases, already contain low and high frequency noises, and SN Rimp = 10log10 N (10)
i=1 oECG2i − dECG2i
hence, a little preprocessing is required. Therefore in the testing
phase of the proposed ECG denoising technique, first, all these N
collected ECG signals are preprocessed through a zero-phase But- i=1 oECG2i − dECG2i
P RD(% ) = N × 100 (11)
terworth bandpass filter to remove the low frequency baseline 2
i=1 oECGi
6 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304

Table 2
Performance of the proposed ECG denoising technique on the first lead (ML II) of MIT-BIH arrhythmia database at γ = 99.999%.

PLI noise is generated using Eq. (6), BWN is generated using Eq. (7), and both WGN (30 dB) and RN have σ = 1.5 and μ = 0.
File ID SNRIN SNROUT δ SNRimp PRD (%) WEDD (%)
100 −21.93 −21.95 0.999089 45.38 6.72 3.65
101 −19.70 −19.73 0.998479 43.29 6.62 3.17
102 −18.48 −18.51 0.998379 39.60 8.80 3.74
103 −16.75 −16.75 1 44.56 4.07 2.01
104 −17.56 −17.6 0.997727 39.14 8.34 3.43
105 −19.14 −19.22 0.995838 42.89 6.50 2.84
106 −16.29 −16.35 0.99633 39.83 6.65 2.72
107 −12.20 −12.21 0.999181 44.44 2.44 0.62
108 −21.35 −21.42 0.996732 42.03 9.25 4.38
109 −15.97 −15.97 1 48.44 2.38 0.77
111 −16.69 −16.74 0.997013 38.71 7.93 3.33
112 −14.17 −14.17 1 40.80 4.66 2.51
113 −12.16 −12.18 0.998358 39.55 4.27 1.93
114 −24.12 −24.22 0.995871 42.30 12.34 6.91
115 −15.55 −15.55 1 45.89 3.04 1.95
116 −16.86 −16.89 0.998224 43.23 4.80 1.71
117 −14.67 −14.71 0.997281 42.10 4.25 2.88
118 −15.10 −15.10 1 45.80 2.92 1.36
119 −13.47 −13.54 0.99483 38.57 5.56 2.25
121 −16.03 −16.08 0.996891 39.76 6.51 3.54
122 −11.81 −11.81 1 42.72 2.85 1.24
123 −16.78 −16.79 0.999404 47.00 3.08 2.03
124 −16.20 −16.21 0.999383 48.38 2.46 1.35
200 −16.10 −16.13 0.99814 43.48 4.28 1.77
201 −15.25 −15.28 0.998037 39.01 6.48 3.35
202 −15.29 −15.31 0.998694 41.16 5.09 2.10
203 −16.88 −16.92 0.997636 41.95 5.58 1.82
205 −19.52 −19.52 1 47.37 4.05 2.86
207 −16.64 −16.68 0.997602 41.07 6.00 2.54
208 −14.31 −14.36 0.996518 38.56 6.13 2.14
209 −16.81 −16.83 0.998812 41.02 6.16 3.16
210 −17.10 −17.12 0.998832 42.02 5.67 2.26
212 −15.12 −15.14 0.998679 38.98 6.41 2.64
213 −12.54 −12.55 0.999203 41.57 3.54 1.12
214 −15.35 −15.35 1 45.14 3.24 1.06
215 −16.84 −16.85 0.999407 42.09 5.47 2.29
217 −13.27 −13.30 0.997744 41.17 4.03 1.33
219 −13.63 −13.64 0.999267 44.03 3.02 1.28
220 −16.57 −16.57 1 45.37 3.63 2.07
221 −15.73 −15.75 0.998730 41.64 5.06 2.15
222 −18.28 −18.43 0.991861 35.45 13.86 7.78
223 −17.23 −17.23 1 49.26 2.50 1.31
228 −17.52 −17.65 0.992635 38.90 8.54 4.01
230 −17.69 −17.69 1 48.41 2.91 1.64
231 −16.23 −16.26 0.998155 43.67 4.24 2.74
232 −17.51 −17.52 0.999429 42.07 5.92 3.85
233 −13.77 −13.78 0.999274 46.33 2.35 1.03
234 −15.01 −15.01 1 45.45 3.00 1.71
Average −16.32 −16.35 0.998285 42.70 5.28 2.51

The higher the value of SNRimp , and the lower the value Figs. 4 to 7 show the performance of the proposed ECG de-
of PRD and WEDD, the better the performance. However, except noising technique at different sampling rates and at γ = 99.999%.
PRD and WEDD the other metrics are not standardized yet. As per Fig. 4 shows the oECG, nECG, dECG as well as four different types of
the criteria given in [31], a denoised/reconstructed ECG signal is synthetic noises which are added to oECG. Synthetic PLI noise and
considered as: (1) ‘excellent’, if the WEDD value is <4.517%, (2) BWN, which are shown in Fig. 4 are generated using Eqs. (6) and
‘very good’, if it is within the range 4.517%–6.914%, (3) ‘good’ if (7), respectively, and 30 dB WGN and RN are generated at σ = 1.5
it falls in the range 6.914%−11.125%, (4) ‘not bad’ if it falls in and μ = 0. Exactly the same amount of noises are also added with
the range 11.125%−13.56%, and (5) ‘bad’ beyond that. Also, ac- other signals, which are shown in Figs. 5 to 7, and hence these
cording to the globally considered standard [32], quality of a de- synthetic noises are not shown further in these figures.
noised/reconstructed ECG signal is considered to be (1) ‘very good’ From Table 2, it is noted that the PRD and WEDD values are
if the PRD value lies in between 0 and 2%, (2) ‘good’ if the PRD maximum for the ECG signal contained in the file #222. Fig. 8
value lies in between 2% to 9%, and (3) ‘not good’ if the PRD value shows the original ECG signal contained in file #222 as well as its
is >9%. The ratio of SNRIN to SNROUT is calculated and denoted denoised version. From this figure it can be seen that the original
as δ = SNR IN . The closer the value of δ to unity, the better the per-
SN R
ECG signal itself is very noisy, and the denoising technique extract
OUT
formance. the added synthetic noise and also the other noises which were
All the 48 ECG signals (lead ML II) of MIT-BIH arrhythmia present in the signal beforehand. This is the reason for getting such
database are mixed with PLI noise (generated using Eq. (6)), BWN high values of PRD and WEDD. Moreover, the value of δ is mini-
(generated using Eq. (7)), and both 30 dB WGN and RN at σ = 1.5 mum for this ECG signal; suggesting a high value of SNROUT , which
and μ = 0. Then, these noisy signals are denoised using the pro- indicates that the power of the denoised signal is higher than that
posed technique, and the performance is shown in Table 2. of the power of the extracted noise. Moreover, it can also be seen
 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304 7

Fig. 4. (a) oECG: MIT-BIH 100, sampling rate 360 Hz, (b) PLI noise, (c) 30 dB WGN, μ = 0, σ = 1.5, (d) RN, μ = 0, σ = 1.5, (e) BWN, (f) nECG, (g) oECG and dECG on top of
each other: SNRIN = −21.93dB, SNROUT = −21.95dB, SNRimp = 45.38dB, PRD = 6.73%, and WEDD = 3.65%.

Fig. 5. (a) oECG: edb-e0103, lead-V4, sampling rate 250 Hz, (b) nECG, (c) oECG and dECG on top of each other: SNRIN = −17.94dB, SNROUT = −17.94dB, SNRimp = 48.37dB,
PRD = 3.01% and WEDD = 2.07%.

from Table 2, that in most of the cases, the value of δ is found to BWN noise is generated using Eq. (7), 30 dB WGN and RN with
be very close to unity, which suggests that the denoising technique μ = 0 and σ = 0.1 to 5.0) on MIT-BIH arrhythmia database are
is robust enough to extract most of the synthetically added noises. shown in Fig. 9. Form this figure it can be noted that the variation
Now, the variation in average SNRimp , PRD, and WEDD values at patterns of the average values of SNRimp , PRD, and WEDD with σ
different levels of input noises (PLI noise is generated using Eq. (6), take the shape of polynomials of order 5.

Fig. 6. (a) oECG: macecgdb test01_00 s, signal ECG-3, sampling rate 500 Hz, (b) nECG,(c) oECG and dECG on top of each other: SNRIN = −19.57dB, SNROUT = −19.72dB,
SNRimp = 56.27dB, PRD = 1.46%, and WEDD = 1.46%.
8 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304

Fig. 7. (a) oECG: PTBDB s0431_re (myocardial infraction), lead-I, sampling rate 1 kHz, (b) nECG, (c) oECG and dECG on top of each other: SNRIN = −17.65dB, SNROUT = −17.68dB,
SNRimp = 44.32dB, PRD = 4.64% and WEDD = 1.38%.

Table 3
Variation in the denoising performance of the proposed technique for different
values of γ at a fixed noise level on MIT-BIH arrhythmia database.

PLI noise is generated using Eq. (6), BWN noise is generated using
Eq. (7), 30 dB WGN and RN with μ = 0 and σ = 2.5.
γ in% SNRimp PRD (%) WEDD (%)
90.00 26.95 29.57 16.24
92.00 27.90 26.49 14.03
94.00 29.15 22.94 11.60
96.00 30.71 19.17 9.26
98.00 33.57 13.85 6.36
99.00 35.78 10.80 4.98
99.99 40.84 6.44 3.50
99.995 40.91 6.40 3.50
99.999 41.24 6.19 3.48
Fig. 8. Original ECG signal (file #222, MIT-BIH arrhythmia database) and its de-
100.00 41.24 6.19 3.48
noised version plotted on top of each other.

Now, the variation in the denoising performance of the pro- signal falls under the category ‘not good’ below γ = 99.999%, and
posed technique on mitdb for different values of γ at a fixed noise as per the WEDD criteria the denoised ECG signal falls under the
level is shown in Table 3. category ‘not bad’ below γ = 94%. It important to note that both
Fig. 10 shows the denoising performance of the proposed tech- the values of PRD and WEDD are identical at γ = 99.999% and
nique for different values of γ on a particular ECG signal (file 100%, and the reason is as follows. It is seen in Fig. 2A that the
#200) taken from mitdb. From Table 3 it can be noted that both amplitudes of the high-order FRCs (FRCL − 1 , FRCL etc.) are signif-
the values of PRD and WEDD decreases with increasing the value icantly low (of the order of 10−4 ) compared to that of the low-
of γ , and vice-versa. Also, the values of SNRimp , PRD and WEDD order FRCs (FRC1 , FRC2 etc.). It suggests that most of the signal-
changes rapidly up to a γ value of 99.99%, and then get saturated information is contained in the low-order FRCs. Therefore, addi-
at a γ value of 99.999%. As per the PRD criteria, the denoised ECG tion of those high-order FRCs (FRCL − 1 , FRCL etc.) does not help

Fig. 9. (a) variation in the average values of SNRimp with σ ( SNRimp = 0.0033σ 5 -0.05σ 4 +0.31σ 3 +0.87σ 2 -0.36σ +44), (b) variation of PRD (%) with σ ( PRD(%)= −0.0017σ 5
+0.023σ 4 -0.14σ 3 +0.48σ 2 +0.036σ +4.5), (c) variation of WEDD (%) with σ ( WEDD(%)= −0.0 0 03σ 5 +0.0 043σ 4 -0.026σ 3 +0.085σ 2 +0.83σ +1.1). The σ of WGN and RN is varied
together.
 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304 9

Fig. 10. (a) nECG (PLI noise is generated using Eq. (6), BWN noise is generated using Eq. (7), 30 dB WGN and RN with μ = 0 and σ = 2.5), (b) dECG at γ = 90%, (c) dECG at
γ = 92%, (d) dECG at γ = 94%, (e) dECG at γ = 96%, (f) dECG at γ = 98%, (g) dECG at γ = 99%, (h) dECG at γ = 99.999%, (i) dECG at γ = 100%. From (b) to (i) the oECG and
dECG signals are plotted on top of each other.

enhancing the quality of the reconstructed ECG signal farther, and less at γ = 99.999% compared to that of at 100%; particularly at
therefore, the values of PRD and WEDD do not change noticeably low value σ . Henceforth, a γ value of 99.999% is used in prepar-
above a γ value of 99.999%. However, the value of nopt , i.e., the op- ing all the tables and figures in the rest of this paper. One can
timum number of FRCs required for γ = 99.999% is less than that also chose a γ value of 100%, but it would not help improving the
of for γ = 100%, i.e., (nopt |γ = 99.999% < nopt |γ = 100% ). In our research quality of the reconstructed signal, and also it would increase the
it is observed that nopt γ at a fixed value of σ , and nopt sampling processing time. Now, Table 4 shows the performance of the pro-
frequency of the signal when both σ and γ are constant. Fig. 11 posed denoising technique on ECG signals collected from different
shows the variation of nopt with σ for different values of γ on databases at different noise levels.
two different ECG signals: (i) file#100 taken from mitdb (sampling Striking information, which can be extracted from Table 4 are:
rate 360 Hz) and (ii) file#s0273lre (sampling rate 1 kHz) taken 1) in most of the cases the denoising performance of the technique
from ptbdb. From Fig. 11a, it can be noted that the value of nopt enhances with increasing the sampling rate, 2) in most of the cases
(nopt = 90) for γ = 100% is almost double than that of the value both PRD and WEDD values decrease with increasing sampling rate
of nopt (nopt = 47) for γ = 99.999% at σ = 0.1. It suggests that at a fixed value of σ , and 3) in most of the cases the rate of change
the last-43 FRCs (FRC48 , FRC49 ,…., FRC90 ) do not contain any such of both PRD and WEDD values with σ decreases with increasing
signal-information, which might help improving the quality of the sampling rate. Variation of SNRimp , PRD and WEDD with σ of WGN
reconstructed ECG signal. Similarly, Fig. 11b shows that the value and RN at different sampling rates are shown in Fig. 12.
of nopt (nopt = 250) at γ = 100% is almost five times than that of Now, the performance of the proposed ECG denoising technique
the value of nopt (nopt = 53) for γ = 99.999% at σ = 0.1. It sug- on real-time ECG signals, which are corrupted with real noises (i.e.,
gests that the last-197 FRCs (FRC54 , FRC55 ,…., FRC250 ) do not con- ECG signals, which are not contaminated with synthetic noises)
tain any such signal-information, which could help enhancing the is shown in Fig. 13. This figure demonstrates that the proposed
quality of the reconstructed ECG signal. It is also to be noted from technique is not only efficient in removing synthetic noises, it also
both Fig. 11a and b, that (i) the difference between the values of works proficiently in real-time scenarios.
nopt at γ = 99.999% and 100%, respectively, reduces with increas- Apart from quantitative analysis, qualitative measure of de-
ing the value of σ , and (ii) the values of nopt do not change signifi- noised ECG signal is also very important for true quality assess-
cantly below a γ value of 99%. Therefore, in this research, 99.999% ment. Semi-blind mean opinion score (MOS) test [32] of ten eval-
is found as the optimum value of γ for two reasons: (1) the per- uators (eight researchers who are working in digital signal, bio-
formance measure metrics (SNRimp , PRD, WEDD) provide identical image and bio-signal processing domains and two cardiologists)
results both at γ = 99.999% and 100%, and (2) the value of nopt is has been carried out. In doing so, a web-based survey-form has
been created containing the original and the corresponding de-
noised ECG signals, and the evaluators were invited to take part
in the survey. Six ECG features (QRS complex, T and P waves,
ST segment, PR and QT intervals) are considered for the MOS
test. Eleven ECG signals of MIT-BIH arrhythmia database are taken
(record #100, #106, #112, #117, #123, #202, #208, #214, #220,
#228, 234), and added with synthetic noises (PLI noise is gener-
ated using Eq. (6), BWN noise is generated using Eq. (7), 30 dB
WGN and RN with μ = 0 and σ = 0.1 to 5.0). Then, these noisy
signals, 121 noisy signals in total (each ECG record is added with
WGN and RN at 11 different values of σ along with PLI and
BWN), are denoised using the proposed technique setting γ at
99.999%. In the survey-form the evaluators were asked to com-
pare the original ECGs with their denoised counterparts, and give a
Fig. 11. (a) variation of nopt with σ (0.1 to 5) at different values of γ on the ECG
signal #100 of mitdb (sampling rate 360 Hz, L = 90), (b) variation of nopt with σ score in accordance with their similarities. The quality ratings are
(0.1 to 5) for different values of γ on the ECG signal #s0273lre of ptbdb (sampling 1 (completely different), 2, 3, 4, and 5 (identical). MOS of the de-
rate 10 0 0 Hz, L = 250). noised ECG signals at different values of σ are calculated using the
10 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304

Table 4
Performance of the proposed ECG denoising technique on different ECG databases at different noise levels (γ = 99.999%).

PLI noise is generated using Eq. (6), BWN noise is generated using Eq. (7), 30 dB WGN and RN with μ = 0 and
σ = 0.1 to 5.0.
σ of WGN and SNRimp PRD (%) WEDD (%)
RN
edb, sampling rate 250Hz 0.1 37.68 3.15 0.87
0.5 36.78 4.65 1.99
1.0 35.04 5.52 2.87
1.5 33.33 6.64 3.78
2.0 31.89 7.88 4.75
2.5 30.49 9.25 5.74
3.0 29.30 10.63 6.72
3.5 28.18 12.01 7.73
4.0 27.20 13.58 8.72
4.5 26.32 14.99 9.72
5.0 25.56 16.37 10.70
mitdb, sampling rate 360Hz 0.1 44.33 4.52 1.22
0.5 44.05 4.63 1.57
1.0 43.39 4.92 2.03
1.5 42.70 5.28 2.51
2.0 41.96 5.71 2.99
2.5 41.24 6.19 3.48
3.0 40.50 6.71 3.98
3.5 39.86 7.18 4.47
4.0 39.24 7.72 4.96
4.5 38.68 8.23 5.46
5.0 38.21 8.67 5.94
nstdb, sampling rate 360Hz 0.1 28.13 4.22 1.16
0.5 27.98 4.27 1.41
1.0 27.50 4.46 1.75
1.5 26.80 4.77 2.10
2.0 26.03 5.15 2.45
2.5 25.23 5.59 2.81
3.0 24.44 6.09 3.18
3.5 23.71 6.60 3.54
4.0 22.99 7.16 3.92
4.5 22.33 7.63 4.28
5.0 22.53 7.68 4.30
macecgdb, sampling rate 500Hz 0.1 30.64 3.65 0.98
0.5 30.42 3.73 1.06
1.0 29.52 4.08 1.18
1.5 28.29 4.59 2.15
2.0 27.02 5.24 2.60
2.5 25.77 5.97 3.09
3.0 24.74 6.70 3.63
3.5 23.81 7.48 4.07
4.0 22.98 8.19 4.44
4.5 22.12 9.02 5.07
5.0 21.40 9.74 5.41
ptbdb, sampling rate 1kHz 0.1 25.16 6.18 1.82
0.5 25.19 6.23 1.96
1.0 25.23 6.30 2.07
1.5 25.28 6.32 2.13
2.0 25.30 6.34 2.20
2.5 25.36 6.37 2.21
3.0 25.43 6.39 2.26
3.5 25.48 6.42 2.33
4.0 25.53 6.46 2.40
4.5 25.62 6.62 2.42
5.0 25.78 6.64 2.48

following equation. Gold standard MOS error of each feature and the overall de-
noised ECG signal are calculated using (14) and (15), respectively.
1
Ne Nfe
 
MOS = Qr (e, f e ) (12) MOS ( fe )
Ne N f e M OS f e = 1− × 100% (14)
e=1 f e=1 5

where Ne = Total number of evaluators, Nfe = Total number of fea-

 MOS

M OSe = 1 − × 100% (15)
tures, Qr = Quality rating of the feth feature given by the eth evalu- 5
ator.
Table 5 shows the MOS errors of overall denoised ECG signals
The MOS value of a particular ECG feature is expressed as
at different values of σ on MIT-BIH arrhythmia database.
Ne
From Table 5 it can be noted that the MOS error values increase
MOS ( fe ) =
1
Qr (e, fe ) (13) with increasing values of σ . However, according to the gold stan-
Ne dard MOS error criteria [32], which says that an ECG signal can be
e=1
 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304 11

Fig. 12. Variation of SNRimp , PRD and WEDD with σ of WGN and RN at different sampling rates.

Fig. 13. (a) Performance of the proposed technique on ECG signal #s0 0 02 taken from PTBDB, (b) performance of the technique on ECG signal #105 taken from MIT-BIH
arrhythmia database.

considered as of ‘very good’ quality if the value of MOSe is found

to be ≤15, the denoised ECG signals (overall) fall under the cate-
gory ‘very good’. Values of MOSfe at different values of σ are also
calculated, and are found to be ‘very good’ too as per the criteria.
Run time of an algorithm or technique, which is the function of
the input data file size, is an important figure of merit, and is used
to gauge the time complexity. The proposed ECG denoising tech-
nique is implemented on MATLAB platform with a computer hav-
ing 64-bits Windows 7 Professional operating system, 16GB RAM
and Intel(R) Xeon(R) CPU E3-1225 v3 3.20 GHz. Run times of the
proposed ECG denoising technique as a function of the length of
the signal at different sampling rates are evaluated and shown in
Fig. 14. From this figure it is to be noted that: (i) the time com-
plexity of the proposed technique vary almost linearly with the Fig. 14. Run times of the proposed ECG denoising technique as a function of length
length of the signal, and (ii) rate of change of time-complexity with of the signal at different sampling rates.
the length of the signal is proportional to the sampling rate. Since
the technique is primarily implemented on software platform, the
space complexity and the power requirement are not considered 4. Performance comparisons
in this research. However, the run time of proposed technique can
further be reduced by efficient code-optimization and implemen- Performance of an ECG denoising technique depends on a num-
tation. ber of important parameters such as the considered bandwidth,
sampling frequency at which the signal is acquired, and the added
noise level. Different ECG denoising techniques reported in the lit-
erature are tested on different types of ECG signals and at different
Table 5 noise levels. Therefore, a straightforward comparison of denoising
MOS errors of denoised ECG signals (overall).
efficiency among different techniques is not possible.
σ 0.1 0.5 1.0 1.5 2.0 2.5 It has already been mentioned before that the performance
MOSe (%) 3.33 3.33 4.67 3.33 5.00 4.67 of SSA in denoising ECG signals is studied in [28]. However,
σ 3.0 3.5 4.0 4.5 5.0
MOSe (%) 6.00 11.33 10.67 11.33 13.33
the numerical results of the denoising performance are not pro-
vided in detail in [28], and hence, it is not possible to directly
12 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304

Table 6
Performance comparison of the proposed ECG denoising technique with that of the one proposed in [28].

Database Method Value of L Average SNRimp Average PRD (%) Average WEDD (%)
MITDB (all 48-ECG records) The 14 0.00 682.67 93.53
reproduced 19 0.00 682.67 93.53
algorithm 22 0.00 682.67 93.52
of [28] 26 0.00 682.67 93.52
90 0.00 682.67 93.52
Proposed 90 42.70 5.28 2.51

Table 7
Performance comparison of the proposed ECG denoising technique with that of the EMD-based
method proposed in [7].

MITDB record #115 (800 [7] Proposed

samples) SNRimp MSE SNRimp MSE
Noise types PLI 12.0526 0.00096 31.90 0.000007
WGN 7.53 0.00270 13.33 0.000761
BWN 7.6487 0.00260 29.19 0.000307

compare the performance of the proposed method with that of In [28], the SSA-based denoising technique is tested on ECG sig-
[28]. Based on the spectral analysis of the principal components, nals recorded at 250 Hz and 1250 Hz sampling rates with four dif-
it has been suggested in [28] to choose an L value, which is at ferent values of L. The values of L, which are used at 250 Hz sam-
sampling rate
least twice the factor f requency band o f the signal
. The clinical bandwidth pling rate are 10, 13, 15 and 18, and the values of L, which are used
of an ECG signal lies between 0.5 Hz to 100 Hz [29]. Therefore at 1250 Hz sampling rate are 50, 65, 75 and 90. Following this rule,
rate (in Hz ) rate (in Hz )
, L = C × sampling
100−0.5 Hz = C × sampling
99.5 Hz , where C ≥ 2. In the value of L for an ECG signal recorded at 360 Hz would be 14,
[28], the ECG signals are recorded at 250 Hz and 1250 Hz sam- 19, 22 and 26, and the performance of the reproduced algorithm
pling rates, and then the acquired ECG signals are added with syn- is tested on these values of L, and are shown in Table 6. On the
thetically generated power-line interference noise. Then, the noisy other hand, following Eq. (2), our proposed algorithm calculates
ECG signals are processed with an SSA-based technique for differ- the value of L = 90. The performance of the reproduced algorithm
ent values of L, and the denoising performance at different values is also tested using an L value of 90. Table 6, clearly demonstrate
of L are studied. It has been shown in [28] that the (1) power-line that the algorithm, which is proposed in [28], might able to ex-
interference noise removal efficiency improves with increasing the pel out the powerline-line-interference noise, but its performance
value of L, and (2) the power-line interference noise removal effi- is not satisfactory in presence of other types of noises.
ciency gets particularly better at L values, which are close to the The EMD-based ECG denoising technique proposed in [7] is
integer multiples of the factor sampling 99.5 Hz
rate
. In this proposed SSA- tested only on 800 samples of the ECG data file #115 taken from
based ECG denoising technique, the value of L is calculated us- mitdb. In [7], first, the amplitude of the ECG signal is normalized
ing Eq. (2). For an example, if the sampling rate of the ECG sig- to ±1, and then the signal is added with four different types of
nal is 1250 Hz, then our proposed technique uses a L value of synthetic noises (PLI, WGN, RN and BWN) taking only one type
×1250
313 (= 60 240 ) for denoising the ECG signal, which well-satisfies of noise at-a-time. Next, the noisy ECG signal is denoised, and
the condition for choosing the value of L given in [28]. However, one additional non-diagnostic distortion measure metric, namely,
the denoising efficiency of the technique proposed in [28] is not 
N
mean square error (MSE = (oECG(i ) − dECG(i ))2 /N), is calcu-
tested on ECG signals corrupted with any other types of noises ex- i=1
cept the power-line interference noise. lated along with PRD to show the efficiency of the technique. The
The following steps are performed in order to do a fair com- lower the value of MSE, the better the performance. For the sake
parison of performance between the proposed ECG denoising tech- of a fair comparison, our proposed ECG denoisng technique is also
nique with that of the one, which is proposed in [28]. applied on that particular ECG signal at the same level and type of
The algorithm proposed in [28] is reproduce. noise contamination. The result is shown and compared to [7] in
All the 48 ECG signals (lead ML II) of MIT-BIH arrhythmia Table 7.
database are mixed with PLI noise (generated using Eq. (6)), BWN From Table 7, it can be noted that the performance of the
(generated using Eq. (7)), and both 30 dB WGN and RN at σ = 1.5 proposed ECG denoising technique outperforms the technique re-
and μ = 0 (the same amount of noises, which are used to testify ported in [7] in removing PLI, WGN and BWN. The RN generation
the performance of the proposed algorithm; please see Table 2). technique, which is used in [7] is not mentioned properly, and
The noisy ECG signals are denoised using the reproduced- hence, it is not possible to compare the RN removal efficiency of
algorithm at L values of 14, 19, 22, 26, and 90, and finally, the the proposed technique with that of [7].
denoising-performance of the reproduced-algorithm is compared The EMD and adaptive switching mean filter (ASMF)-based ECG
to that of the proposed one and shown in Table 6. denoising technique proposed in [8] is tested on eight ECG data

Table 8
Performance comparison of the proposed ECG denoising technique with that of the EMD and ASMF-based method proposed in [8].

SNRIN [8] Proposed

in dB SNRimp PRD (%) MSE SNRimp PRD(%) MSE
Noise type: RN 0 9.2980 34.32 0.02022 3.0388 70.22 0.00903
5 9.1351 18.95 0.00655 3.0125 39.55 0.00285
10 8.7879 11.53 0.00232 2.9413 22.06 0.00090
15 8.0516 7.12 0.00088 2.7238 12.80 0.00030
20 5.6733 5.30 0.00050 1.9713 7.83 0.00011
 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304 13

Table 9
Performance comparison of the proposed ECG denoising technique with that of the FZP-based
method proposed in [33].

MITDB record #115 [33] Proposed

SNRimp MSE SNRimp MSE
Noise types PLI 14.2565 0.0128 34.0500 0.000006
WGN 13.4716 0.0153 13.4576 0.000733

Table 10 comparison, the proposed technique is modified so as to obtain a

Performance comparison of the proposed ECG denoising tech-
SNRimp value, which is close to that of [33], and then the perfor-
nique with that of the AFD-based method proposed in [17].
mance is compared and shown in Table 9.
MIT-BIH MSE It can be noted from Table 9, that the PLI removal efficiency of
record # [17] Proposed
the proposed technique outperforms [33]. In removing WGN, our
Noise type: 101 36.0 0.002654
WGN 102 32.6 0.004392 proposed technique achieves a SNRimp value of 13.4576 by setting
103 76.0 0.003356 σ at 0.039, and the obtained SNRimp value is very close to that in
104 55.7 0.003058 [33]. Now, from Table 9, it can be seen that at almost same level
105 73.6 0.004169 of SNRimp , the MSE value obtained using the proposed technique is
106 98.9 0.003828
approximately 21 times better than that of [33].
107 572.6 0.012570
108 24.0 0.002472 In [17], an adaptive Fourier decomposition-based (AFD)-based
109 112.1 0.009362 ECG denoising technique is proposed. In the performance evalua-
201 38.3 0.005129 tion phase of [17], 16 ECG signals are taken from mitdb, and are
202 28.4 0.003801
mixed with WGN at an input SNR level of 10 dB. Then, those
203 321.3 0.007476
205 29.5 0.003213 noisy signals are filtered using the AFD-based technique. Perfor-
207 93.9 0.008841 mance of our proposed ECG denoising technique is also tested on
208 199.2 0.007246 these 16 ECG signals at the same level of noise contamination,
209 60.8 0.005104 and the obtained MSE values are shown and compared to [17] in
Table 10. Table 10 clearly demonstrate that the proposed ECG de-
noising technique outperforms [17].
A dictionary-learning-based ECG denoising technique is pro-
files (#100, #101, #103, #105, #115, #200, #215, #230) taken from posed by Sajita et al. in [15], and the performance of the technique
mitdb. In [8], first, the amplitude of the ECG signal is normalized is evaluated on a number of ECG signals taken from mitdb with
to ±1, and then they are added with synthetic RN in the frequency different combinations of various synthetically generated noises.
band 0.3 Hz – 2 Hz at different levels of input SNR. Finally, the An additional non-diagnostic distortion measure metric, namely,
noisy signals are denoised. For the sake of fair comparison, our maximum absolute amplitude (MAA = max(|oECG − dECG|)), is
proposed denoising technique is also applied on these ECG signals calculated along with SNRimp to show the efficiency of the tech-
after adding synthetic RN at same input SNR levels. RN removal ef- nique. The lower the value of MAA, the better the performance. In
ficiency of the proposed technique is compared to that of [8] in Table 11, the efficiency of the proposed ECG denoising technique is
Table 8. From this Table it is interesting to note that the average compared to that of [15] in removing a combination of BWN and
values of SNRimp and PRD of the method proposed in [8] in remov- PLI. It can be seen from this table that our proposed technique per-
ing RN is about 3 and 2 times, respectively, better than that of the forms better than that of [15] in most of the cases.
proposed technique. However, the average value of MSE obtained
using the proposed technique is about 3 times better than that of 5. Conclusion and discussion
[8].
The fractional zero-phase filtering (FZP)-based ECG denoising A reliable, robust and high performance singular spectrum
technique, which is proposed in [33] is tested only on the ECG analysis-based ECG denoising technique has been proposed in this
record #115 taken from mitdb. In [33], first, the ECG signal is paper. The performance of the technique is tested on a large num-
added with PLI and 15 dB WGN taking one type of noise at-a-time ber of ECG data files taken from five different publicly available
and then the noisy signal is denoised. Frequency and peak-to-peak ECG databases. The reconstructed ECG signal is the noise-reduces
amplitude of the added synthetic PLI noise, which is considered in and smoothed version of the original ECG signal. Since the recon-
[33], are 50 Hz and 0.15 mV, respectively. In [33], the power of the structed components are filtered using zero-phase bi-directional
added WGN (i.e., 15 dB) is mentioned, but the standard deviation bandpass and notch filters, the phase difference, i.e., time-lag be-
value, which is used to generate the WGN is not mentioned. There- tween the original and reconstructed signal is zero. The smoothing
fore, it is difficult to directly compare the WGN removal efficiency and denoising performance of the proposed ECG denoising tech-
of the proposed technique with that of [33]. For the sake of fair nique on real-time ECG signals can be visually examined in Fig. 13.

Table 11
Performance comparison of the proposed ECG denoising technique with that of the dictionary-
learning-based method proposed in [15].

MIT-BIH record # [15] Proposed

SNRimp MAA SNRimp MAA
Noise type: 102 20.90 0.21 21.23 0.2278
BWN+PLI 109 23.75 0.35 33.39 0.2302
111 22.09 0.21 24.61 0.2367
119 23.83 0.30 23.24 0.2242
124 22.43 0.33 37.88 0.2326
203 23.61 0.27 23.95 0.2300
208 23.33 0.34 25.65 0.2301
14 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304

ECG tracings are generally scrutinized for calculating the heart- clearly indicate the QRS complexes regions. Therefore, the pro-
rate and heart-rate-variability, identifying the presence of different posed technique, alongside denoising the ECG, can also be adapted
types of arrhythmias as well as other cardiovascular disorders such for detecting its QRS complexes.
as myocardial infraction, bundle branch block. Therefore, the re- Although a few reliable quantitative ECG distortion measure
constructed, i.e., the denoised ECG signal must be of the same am- metrics, such as percent root mean square difference, are there,
plitude and time scales of the original ECG signal. As the singular nonetheless, the result of subjective assessment test provides bet-
spectrum analysis technique takes the original time-domain ECG ter insight into the quality of the denoised ECG signals as the test
signal as input, and provides the reconstructed components at the is conducted by the cardiologists. The more the ECG signals could
output in the same domain, there happens to be neither time nor be included in the MOS test, the better the performance analysis
amplitude-scale variation with the denoised ECG signal compared would be.
to the original one. In this research the denoising technique of a single-lead ECG
The novelty of the proposed ECG-denoising technique lies in signal using SSA is proposed. The 1 D ECG signal is converted to
its performance. From Table 2 and also from Figs. 4 to 7 it can 2 D through proper embedding operation. The technique can also
be noted that proposed technique is robust enough to bring out be revamped and extended in order to be applicable on tensor data
the ECG signal from a combination of different types of noises so as to denoise multi-lead ECG signals. This will be the topic of
including white Gaussian and random types of noises. Further, our future research work.
Tables 6 to 11 show the competency of the proposed technique
in removing each of the four different types of noises (base- Declaration of Competing Interest
line wander, power-line interference, white Gaussian and random
noises), and also its superiority in comparison with other state-of- The authors report no conflicts of interest. The authors alone
the art methods. Not only the proposed technique performs well are responsible for the content and writing of the paper.
on synthetically-added noises, it is also ace in expelling out the
real noises. Figs. 8 and 13 portray the efficiency of the technique Acknowledgements
in removing real noises.
Both the quantitative and qualitative distortion measure metrics The authors convey their sincere thankfulness to Dr. Vijay
show that the proposed ECG denoising technique is robust enough S Chauhan, MD, FRCPC (Cardiologist, Toronto General Hospital,
to filter various noises present in the signal without jeopardizing Toronto, Canada), Dr. Uma Mandal, M.B.B.S. (College of Medicine
the clinical content. It is also possible that the proposed technique and JNM Hospital, Kolkata, India), and the graduate students of
could be adapted for denoising other biomedical signals exhibit- the Department of Electrical, Computer and Biomedical Engineer-
ing periodic or quasi-periodic nature such as photoplethysmogram, ing, Ryerson University, for their precious time and effort in carry-
esophageal pressure signal. From Table 1 it can be noted that the ing out the subjective quality estimation. We would also thank the
proposed technique is tested on all the available leads of ECG sig- funding provided by the Natural Sciences and Engineering Research
nal of all the five databases, except MIT-BIH arrhythmia database. Council (NSERC) Discovery Grant program of Canada.
There are two leads of ECG signals present in MIT-BIH arrhythmia
References
database: ML II and V5. However, the performance of other ECG
denoising techniques [7,8,15,17 and 33], which are mentioned in [1] A.L. Goldberger, Clinical electrocardiography, a simplified approach, 7th ed.,
Section IV, are reported only on lead ML II of MIT-BIH arrhythmia New York, NY, Elsevier, USA, 2006.
database. Hence, in this research the lead ML II is taken to expedite [2] M.L. Ahlstrom, W.J. Tompkins, Automated high-speed analysis of holter tapes
with microcomputers, IEEE Trans. Biomed. Eng. 30 (10) (1983) 651–657 issOct..
the comparison of performance to other techniques.
[3] Z.A.A. Alyasseri, A.T. Khader, M.A. Al-Betar, M.A. Awadallah, Hybridizing β -hill
It could be said that the (i) optimum value of L, (ii) en- climbing with wavelet transform for denoising ECG signals, Inf. Sci. (Ny) 429
hanced covariance among the eigenvector of the matrix U and also (8) (2018) 229–246 issMarch.
[4] H. Liu, Y. Li, Y. Zhou, X. Jing, T.K. Truong, Joint power line interference suppres-
(iii) carrying out the bandpass and notch filtering operations at
sion and ECG signal recovery in transform domains, Biomed. Signal Process.
the reconstruction-component level are the key reasons for get- Control 44 (6) (2018) 58–66 issJuly.
ting such an excellent ECG denoising performance. The main ad- [5] A. Kumar, R. Komaragiri, M. Kumar, Design of wavelet transform based elec-
vantages of using this proposed technique are: (i) the singular trocardiogram monitoring system, ISA Trans. 80 (9) (2018) 381–398 issSep..
[6] P. Singh, G. Pradhan, S. Shahnawazuddin, Denoising of ECG signal by non-local
spectrum analysis parameters are made adaptive to the sampling estimation of approximation coefficients in DWT, Biocybernet. Biomed. Eng. 37
rate of the ECG signal, and therefore the technique does not re- (3) (2017) 599–610 iss.
quire any manual intervention, and (ii) the proposed technique [7] S. Jain, V. Bajaj, A. Kumar, Riemann Liouvelle fractional integral based empirical
mode decomposition for ECG denoising, IEEE J. Biomed. Health Inform. 22 (4)
is a model-free/function-independent method of ECG denoising. It (2018) 1133–1139 July.
captures the periodicity of the oscillatory modes in ECG signals [8] M. Rakshit, S. Das, An efficient ECG denoising methodology using empirical
better than that of the fixed filtering-based approaches. Also in mode decomposition and adaptive switching mean filter, Biomed. Signal Pro-
cess. Control 40 (2) (2018) 140–148 issFeb..
the case of wavelet transform-based methods the mother wavelet [9] S. Jain, V. Bajaj, A. Kumar, Effective de-noising of ECG by optimised adaptive
function and the number of decomposition levels are to be se- thresholding on noisy modes, IET Sci. Meas. Technol. 12 (5) (2018) 640–644
lected manually, and moreover, the number of decomposition lev- iss.
[10] P. Nguyen, J.M. Kim, Adaptive ECG denoising using genetic algorithm-based
els are also to be adjusted manually for variations in sampling
thresholding and ensemble empirical mode decomposition, Inf. Sci. (Ny) 373
rate. (35) (2016) 499–511 issDec..
The threshold value (γ ) could be set to either 99.999% or [11] S. Kumar, D. Panigrahy, P.K. Sahu, Denoising of electrocardiogram (ECG) signal
by using empirical mode decomposition (EMD) with non-local mean (NLM)
100%, and the quality of the denoised signal is identical for both
technique, Biocybernet. Biomed. Eng. 38 (2) (2018) 297–312 iss.
the threshold values. However, the optimum number of filtered [12] D.P. Tobon, T.H. Falk, Adaptive spectro-temporal filtering for electrocardiogram
reconstruction-components, nopt , required to reconstruct the signal signal enhancement, IEEE J. Biomed. Health Inform. 22 (2) (2018) 421–428
is much higher at γ = 100% than that of at 99.999%. This is the March.
[13] H.D. Hesar, M. Mohebbi, ECG denoising using marginalized particle extended
reason for which γ is set to 99.999% in this research. kalman filter with an automatic particle weighting strategy, IEEE J. Biomed.
Figs. 4 to 7 show that the proposed technique is able to effi- Health Inform. 21 (3) (2018) 635–644 July.
ciently process different types of ECG signals irrespective of their [14] H.D. Hesar, M. Mohebbi, An adaptive particle weighting strategy for ECG de-
noising using marginalized particle extended Kalman filter: an evaluation in
QRS-pattern and clinical morphologies. From Fig. 2, it can be seen arrhythmia contexts, IEEE J. Biomed. Health Inform. 21 (6) (2017) 1581–1592
that the coefficients of certain filtered reconstruction-components Nov.
 S.K. Mukhopadhyay and S. Krishnan / Computer Methods and Programs in Biomedicine 188 (2020) 105304 15

[15] U. Satija, B. Ramkumar, S. Manikandan, Noise-aware dictionary-learning-based [28] B. Yang, Y. Dong, C. Yu, Z. Hou, Singular spectrum analysis window length se-
sparse representation framework for detection and removal of single and com- lection in processing capacitive captured biopotential signals, IEEE Sens. J. 16
bined noises from ECG signal, Healthc. Technol. Lett. 4 (1) (2017) 2–12 issFeb.. (19) (2016) 7183–7193 Oct..
[16] J. Zhu, X. Li, Electrocardiograph signal denoising based on sparse decomposi- [29] S.K. Mukhopadhyay, S. Mitra, M. Mitra, ECG feature extraction using differen-
tion, Healthc. Technol. Lett. 4 (4) (2017) 134–137 issAug.. tiation, hilbert transform, variable threshold and slope reversal approach”, J.
[17] Z. Wang, F. Wan, C.M. Wong, L. Zhang, Adaptive Fourier decomposition based Med. Eng. Technol. 36 (7) (Oct. 2012) 372–386.
ECG denoising, Comput. Biol. Med. 77 (10) (2016) 195–205 issOct.. [30] PhysioNet. [Online]. Available: https://physionet.org/cgi-bin/atm/ATM, accessed
[18] J. Wang, Y. Ye, X. Pan, X. Gao, C. Zhuang, Fractional zero-phase filtering based on: June. 2019.
on the Riemann-Liouville integral, Signal Process. 98 (6) (2014) 150–157 iss- [31] M.S. Manikandan, S. Dandapat, Wavelet energy based diagnostic distortion
May. measure for ECG, Biomed. Signal Process. Control 2 (2) (2007) 80–96 issApril.
[19] Y. Lee, D. Hwang, Periodicity-based nonlocal-means denoising method for elec- [32] Y. Zigel, A. Cohen, A. Katz, The weighted diagnostic distortion (WDD) mea-
trocardiography in low SNR non-white noisy conditions, Biomed. Signal Pro- sure for ECG signal compression, IEEE Trans. Biomed. Eng. 47 (11) (20 0 0)
cess. Control 39 (1) (Jan. 2018) 284–293 iss. 1422–1430 Nov..
[20] S. Cuomo, R. Farina, F. Piccialli, An inverse bayesian scheme for the denoising [33] J. Wang, Y. Ye, X. Pan, X. Gao, Parallel-type fractional zero-phase filtering for
of ECG signals, J. Netw. Comput. Appl. 115 (1) (2018) 48–58 issAug.. ECG signal denoising, Biomed. Signal Process. Control 18 (4) (2015) 36–41 iss-
[21] M. Seo, M. Choi, J.S. Lee, S.W. Kim, Adaptive noise reduction algorithm to im- April.
prove R peak detection in ECG measured by capacitive ECG sensors, Sensors [34] N. Golyandina, A. Zhigljavsky, in: Singular Spectrum Analysis for Time Series,
18 (7) (2018) 1–16 issJuly. Springer-Verlag, Berlin, Germany, 2013, pp. 25–32.
[22] N. Golyandna, A. Zhigljavsky, Singular Spectrum Analysis for Time Series, 1st [35] T.J. Harris, H. Yuan, Filtering and frequency interpretations of singular spec-
ed., New York, NY, Springer, USA, 2013. trum analysis, Physica D Nonlinear Phenomena 239 (20–22) (2010) 1958–1967
[23] L. Zotov, E. Scheplova, MSSA of globally gridded OAM from ECCO, AAM from issOct..
ECMWF, and gravity from GRACE, Proc. 3rd Int. Conf. Digit. Inform. Process. [36] D. Skillicorn, Understanding Complex Datasets: Data Mining with Matrix De-
Data Mining, and Wireless Commun. (DIPDMWC) (2016) 127–132 6-8 July- compositions”, 1st ed., Chapman and Hall/CRC, USA, 2007 New York, NY.
Moscow, Russia. [37] S.W. Chen, H.C. Chen, H.L. Chan, A real-time QRS detection method based
[24] M.L. de Menezes, R.C. Souza, J.F.M. Pessanha, Electricity consumption fore- on moving-averaging incorporating with wavelet denoising, Comput. Methods
casting using singular spectrum analysis, Dyna (Medellin) 82 (190) (2015) Prog. Biomed. 82 (3) (2006) 187–195 issJune.
138–146. [38] O. Haavisto, Detection and analysis of oscillations in a mineral flotation circuit,
[25] S. Wang, et al., Localizing microaneurysms in fundus images through singular Control Eng. Pract. 18 (1) (Jan 2010) 23–30 iss.
spectrum analysis, IEEE Trans. Biomed. Eng. 64 (5) (2017) 990–1002 May. [39] D.G. Manolakis, V.K. Ingle, S.M. Kogon, Statistical and adaptive signal process-
[26] S. Ferdowsi, V. Abolghasemi, Semiblind spectral factorization approach for ing: spectral estimation, signal modeling, Adaptive Filtering and Array Process-
magnetic resonance spectroscopy quantification, IEEE Trans. Biomed. Eng. 65 ing, Artech House, USA, 2005 Norwood, MA.
(8) (2018) 1717–1724 Aug. [40] J. Harmouche, D. Fourer, F. Auger, P. Borgnat, P. Flandrin, “The sliding singu-
[27] D. Jarchi, C. Wong, R.M. Kwasnicki, B. Heller, G.A. Tew, G.Z. Yang, Gait param- lar spectrum analysis: a data-driven nonstationary signal decomposition tool”
eter estimation from a miniaturized ear-worn sensor using singular spectrum IEEE Transactions on Signal Processing, vol. 66, No. 1, Jan 2018.
analysis and longest common subsequence, IEEE Trans. Biomed. Eng. 61 (4)
(2014) 1261–1273 April.