Biomedical Signal Processing and Control 25 (2016) 178187

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

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

A comprehensive performance analysis of EEMD-BLMS and DWT-NN

hybrid algorithms for ECG denoising
Kevin Krgaard, Sren Hjllund Jensen, Sadasivan Puthusserypady
Department of Electrical Engineering, Technical University of Denmark, 2800 Kgs, Lyngby, Denmark

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

Article history: Electrocardiogram (ECG) is a widely used non-invasive method to study the rhythmic activity of the
Received 24 July 2015 heart. These signals, however, are often obscured by artifacts/noises from various sources and mini-
Received in revised form mization of these artifacts is of paramount importance for detecting anomalies. This paper presents a
19 November 2015
thorough analysis of the performance of two hybrid signal processing schemes ((i) Ensemble Empirical
Accepted 30 November 2015
Mode Decomposition (EEMD) based method in conjunction with the Block Least Mean Square (BLMS)
Available online 22 December 2015
adaptive algorithm (EEMD-BLMS), and (ii) Discrete Wavelet Transform (DWT) combined with the Neu-
ral Network (NN), named the Wavelet NN (WNN)) for denoising the ECG signals. These methods are
Keywords:
Electrocardiogram (ECG)
compared to the conventional EMD (C-EMD), C-EEMD, EEMD-LMS as well as the DWT thresholding
Denoising (DWT-Th) based methods through extensive simulation studies on real as well as noise corrupted ECG
Ensemble empirical mode decomposition signals. Results clearly show the superiority of the proposed methods.
(EEMD) 2015 Elsevier Ltd. All rights reserved.
Block least mean square (BLMS)
Discrete Wavelet Transform (DWT)
Neural Networks (NN)

1. Introduction Methods of decomposing the signals into sub-components for

Electrocardiogram (ECG) is a measure of the electrical activity
of the heart, and is obtained by surface electrodes at standardized
locations on the subjects chest. During acquisition, various arti-
facts/noises such as the baseline wander, power-line interference,
muscle contraction and electrode movements obscure the ECG. It
is important that these artifacts are minimized for the clinicians to
make better diagnoses on heart problems.
Conventional lters such as the nite impulse response (FIR)
[1,2], innite impulse response (IIR) [3,4], lter banks [5], poly-
nomial lter [6] and Wiener lter [7] have been proposed in the
literature to minimize artifacts. Other approaches for ECG denois-
ing include adaptive lters, namely the least mean square (LMS),
recursive least square (RLS) and their variants such as the block LMS
(BLMS), normalized sign-sign LMS (NLMS) etc., [820]. Adaptive
Kalman lter and extended Kalman lter were also suggested by
some researchers [2124]. Promising performances were obtained
by nonlinear methods such as the Bayesian ltering and nonlinear
projective ltering [25,26].
Corresponding author. Tel.: +45 45253652.
E-mail address: spu@elektro.dtu.dk (S. Puthusserypady).
noise reduction have become popular and were proposed for
denoising the ECG signals. They include the independent compo-
nent analysis (ICA), singular value decomposition (SVD), empirical
mode decomposition (EMD), and ensemble EMD (EEMD) [2731].
Soft, hard and adaptive thresholding methods have also been pro-
posed on EMD and EEMD schemes [32]. Wavelet transformation
(WT) has been shown to be a powerful tool for denoising signals
in the frequency domain [33,34] and has been proposed for ECG
denoising [3541]. It has conventionally been used by applying
soft or hard thresholds on the obtained discrete WT (DWT) coef-
cients [42,43]. A combination of DWT with Wiener ltering has
been proposed by Kesler et al [44]. Recently, hybrid schemes have
been proposed to improve the denoising performance. For exam-
ple, in [9], the EMD and EEMD methods were used to provide the
reference inputs to the BLMS adaptive lter. In another scheme,
the DWT and Neural Networks (NN) are combined to minimize the
noise in ECG and have shown to provide better performances than
conventional wavelet methods [4547].
In this paper, a comprehensive analysis of the performances of
two hybrid schemes for denoising ECG signals is presented. They
are: (i) the EEMD based method in conjunction with the BLMS
adaptive lter, namely the EEMD-BLMS approach, and (ii) the DWT
based method combined with the NN, named as the Wavelet NN

http://dx.doi.org/10.1016/j.bspc.2015.11.012
1746-8094/ 2015 Elsevier Ltd. All rights reserved.
 K. Krgaard et al. / Biomedical Signal Processing and Control 25 (2016) 178187 179

(WNN) approach. These methods are compared to the conventional

EMD (C-EMD), conventional EEMD (C-EEMD), EEMD-LMS, and Dis-
crete Wavelet Transformation thresholding (DWT-Th) methods
[42] by performing extensive simulation studies on real as well
as simulated noise corrupted ECG signals. These six methods are
interesting as they are able to adapt to the time varying nature of
ECG and minimize the noises with a minimum signal distortion
[31]. Furthermore, these methods are able to separate the noise
components from the recorded ECG signals automatically, and can
therefore be used in situations where ECG data from only one lead
is available, like in an ambulatory ECG.
The remainder of the paper is organized as follows: Section 2
describes the six ECG denoising approaches proposed in this study
followed by a detailed discussion of the results on simulated and
real data in Section 3. The paper concludes with some nal remarks Fig. 1. Block diagram of the EEMD-LMS/BLMS adaptive lter.
in Section 4.
2.2. EEMD-BLMS and EEMD-LMS
2. Materials and methods
Adaptive ltering schemes require the primary (desired signal +
The proposed ECG denoising approaches are briey described noise) and reference (noise which is correlated to the noise in the
in this section. First of all, the observed ECG signal is modelled as, primary input) inputs. In this work, only the primary input (x(n))
x(n) = d(n) + v(n), (1) is available. The IMFs derived from x(n) using the C-EEMD method
are used as reference inputs to the LMS and BLMS lters.
where d(n) represents the desired ECG signal, and v(n) is the noise The framework for both the LMS and BLMS adaptive ltering
which corrupts the desired signal. is illustrated in Fig. 1. Here, IMFk (n) ; k = 1, 2, . . ., K is used as the
kth reference input, yk (n) is the corresponding noise estimate from
2.1. C-EMD and C-EEMD the lter, and ek (n) = ek1 (n) yk (n) is the error signal with eK (n) =
d(n) = x(n) v(n). Here the estimate of the noise component v(n)
C-EMD is an iterative process by which a signal is decomposed K
in the primary input is given by, v(n) = y (n).
k=1 k
into subcomponents, namely the intrinsic mode functions (IMFs)
In the BLMS lters, the lter coefcients are updated block-wise,
[48]. From the observed signal x(n), the lower and upper envelopes
unlike in the conventional LMS lters where the coefcients are
(envl (n) and envu (n)) are found by interpolating the local maxima
updated in a sample-by-sample fashion. The computational steps
and minima with cubic splines. The average of these envelopes pro-
for the BLMS lter (say the kth lter in Fig. 1) is described here.
vides an approximation of the lowest frequencies present in the
First, the signal inputs to the BLMS lters are partitioned into non-
signal x(n), which is subtracted from the signal to produce,
overlapping blocks of length L. These blocks are then ltered by a
1   FIR lter of length P. The lter coefcients are kept xed for each
x(n) = x(n) envl (n) + envu (n) . (2)
2 block of data, and the adaptation of the coefcients are performed
This process is repeated on x(n) and the subsequent signals until the block-wise [49]. Let m = 0, 1, . . . denote the block index and the lter
mean value found by the envelopes are close to zero, and the num- coefcient vector for mth block of the kth lter be,
ber of maxima and minima are equal or differ utmost by one. The T
wk (m) = [w0k (m), . . ., wP1
k
(m)] , m = 0, 1, . . . (5)
resulting signal at this point is the rst IMF (IMF1 (n), representing
the highest frequencies). The second IMF is found by performing With the original sample index, n = mL + i, i = 0, 1, . . ., P 1, and m,
the same process on the residue signal, r1 (n) = x(n) IMF1 (n). For the input vector to the kth adaptive lter at sample index n can be
each repetition of the process, a new IMF and residue are obtained, written as,
and the process is stopped when the residue becomes monotonic.
At this point, a certain number of (say K) IMFs have been obtained, uk (n) = [IMF k (n), IMF k (n 1), . . ., IMF k (n P + 1)]T (6)
and the original signal could then be synthesized as,
The corresponding lter output yk (n) = wTk (m)uk (n) is the noise

K estimate and the error signal is, ek (n) = ek1 yk (n) with
x(n) = rK (n) + IMF i (n), (3) e0 (n) = x(n). Filter weights wk (m) are updated as [49]:
i=1

L1
where rK (n) is the residue after the extraction of IMFK . wk (m + 1) = wk (m) + blms
k
uk (mP + i)e(mP + i). (7)
In the literature, C-EEMD is described as a noise assisted data i=0
analysis method, wherein the mode mixing problem associated
with EMD is reduced through averaging multiple EMD results [29]. Here blms
k
is the step-size parameter and is chosen as,
It is achieved by simulating several white noise variants and adding
2
them to the signal x(n) to generate xv (n) (v = 300 1000) as shown 0 < blms
k
P , (8)
in Eq. (4) and performing EMD on each of them. P 
i=1 i

xv (n) = x(n) + wv (n), (4) where i s are the eigenvalues of the input covariance matrix. If the
lter length P is chosen to be equal to L, the update rule can be
where controls the amplitude of the added white noise. An aver-
written as,
age of these 3001000 sets of IMFs of xv (n)s produce the nal
IMFs. In both the C-EMD and C-EEMD approaches, the denoised 
L1

ECG signals are obtained by subtracting the IMFs corresponding to wk (m + 1) = wk (m) + blms
k
uk (mL + i)ek (mL + i). (9)
the noise components from the observed ECG signals [2931]. i=0
180 K. Krgaard et al. / Biomedical Signal Processing and Control 25 (2016) 178187

individual coefcients [43]. The outputs of hard and soft threshold-

ing are described by Eqs. (14) and (15), respectively.

W (j, q) if W (j, q) > Tj
zj (q) = (14)
0 if W (j, q) < Tj .

W (j, q) Tj if W (j, q) > Tj
zj (q) = (15)
0 if W (j, q) < Tj .

Tj s in the equations are the desired thresholds [50]. We use

Daubechies-8 (Db8) in this study to nd the DWT of x(n) [51]
and the thresholds are found manually from the rst set of x(n)
(observed ECG signal) by nding a constant Q proportional to the
coefcients power, such that:

Tj = Qj Cj , j = 0, 1, . . ., 11, (16)

where Cj is the power of the detail coefcient at the jth scale level.
The threshold is then applied to all ECG signals as per Eq. (15) and
subsequently the inverse DWT (IDWT) is performed to reconstruct
the denoised ECG signal [42].
In the WNN approach, wavelet decomposition is performed on
x(n), and the resulting coefcients (detail and approximation) are
fed to the NN for the estimation of d(n). The signal x(n), n = 0, 1,
. . ., N 1, of length N samples is rst zero padded (256 zeros in
the beginning and in the end) to generate a new signal (x (n))
of length N + 512. To mitigate the varying effect of zero-padding,
Fig. 2. Flowchart of the EEMD-BLMS algorithm. a sliding window approach for extracting wavelet coefcients is
adopted. From x (n), using a sliding window (with an overlap of
The conventional LMS lter can be seen as a special case of the 512 samples), N smaller data segments of length 513 samples are
BLMS lter with a block length L = 1. Hence, the block index m can extracted. Each of these smaller segments are subjected to DWT
be replaced by the sample index n and thereby the update rule for which results in 513 sets of 14 wavelet coefcients (13 detail
the kth LMS lter is given by Eq. (10): (W ) and 1 approximation(W )) which are arranged in a matrix
of dimension 14 513. From these wavelet coefcients matrices,
wk (n + 1) = wk (n) + lms uk (n)ek (n). (10)
k the 14 coefcients corresponding to specic smaller data seg-
The step-size lms for the LMS lters are chosen as, ments (for example, x(n)) is extracted as the 257th column of the
k
respective matrices. The corresponding input vector xDWT (n) (of
2 length 154 = 14 11) to the NN is obtained by concatenating 11
0 < lms
k
, (11)
max columns (252:262) of the respective wavelet coefcient matrix.
where max is the maximum eigenvalue of the input covariance The incorporation of wavelet coefcients of the ten neighbor-
matrix. A owchart of the EEMD-BLMS algorithm can be seen in ing samples of x(n) into xDWT (n) provides the NN with temporal
Fig. 2. information and thereby improves its performance. A multilayer
perceptron (MLP(154:38:38:1)), trained using the Bayesian regu-
2.3. DWT-Th and WNN larization backpropagation (BRBP) algorithm, is used in this study
[52,53]. A owchart describing the steps in the DWT-NN algorithm
DWT is used to obtain the frequency information over time for can be seen in Fig. 3.
the given signal x(n). It has the inherent property of increasing the
temporal resolution for high frequencies, and increasing frequency 3. Results and discussion
resolution for low frequencies making it a suitable tool for analyzing
biological signals. Given x(n), its DWT can be derived as follows: The proposed methods are tested on simulated (noise cor-
Let = [ (0), (1), . . ., (N 1)]T and = [(0), (1), . . ., rupted ECG) as well as on real ECG signals and the performances
(N 1)] be two orthogonal basis functions and let j0 = 0, N = 2j ,
T
are compared quantitatively by the signal-to-noise ratio (SNR)
j = 0, 1, . . ., J 1 and q = 0, 1, . . ., 2j 1. Here, j represents different improvement.
scale levels, and q represents the translation in each scale level. The  N1

wavelet coefcients (W approximation and W detail) corre- n=0

d2 (n)
SNR = 10log10 N1  2 , (17)
sponding to x(n) are given by,
n=0
d(n) d(n)
1 
N1
W (j0 , q) = x(n)j0 ,q (n), q, (12) where, d(n) is the estimated ECG signal, d(n) is the desired ECG
N
n=0 signal and N is the length of the data. While calculating the SNR
values, the rst 2000 samples (corresponding to four seconds) and
1 
N1
last 500 samples (one second) are omitted in all methods to account
W (j, q) = x(n) j,q (n), q, j > j0 . (13)
N for the convergence of LMS and BLMS algorithms as well as distor-
n=0
tion of the wavelet estimation due to zero-padding. Paired t-test
Most wavelet denoising approaches rely on wavelet threshold- is performed on these SNR values to analyze similarities in the
ing, by imposing a soft or hard threshold either globally or on the performances.
 K. Krgaard et al. / Biomedical Signal Processing and Control 25 (2016) 178187 181

Fig. 4. Power spectrum of desired (blue) and 0dB (red) Type-I ECG. (For interpreta-
tion of the references to color in this gure legend, the reader is referred to the web
version of this article.)

Fig. 3. Flowchart of the DWT-NN algorithm.

In the adaptive lters, the lter order (P) is chosen by empirical

analysis (Section 3.2.2; Fig. 7). The block length (L) of the BLMS lter
is set equal to P in order to eliminate the redundant operations and
waste of lter tap weights [49]. The step-sizes are chosen according
to Eqs. (8) and (11).
3.1. Data details
The real ECG data (ECG1ECG10) from 10 individuals used in this
study are obtained from the MIT/BIH ECG-ID database [54]. The data
is sampled at 500 Hz and contains both noisy recordings (denoted
as x(n)) as well as the corresponding cleaned (e.g., wavelet drift
correction, band-pass ltering and smoothing [55]) signals denoted
as d(n). In this study, ECG6ECG10 are used to train the NN and
the remaining 5 data sets (ECG1ECG5) are used for testing the
proposed denoising strategies.
3.2. Results on simulated data
Fig. 5. Power spectrum of the IMFs for the 0dB Type I signal (ECG1). (For interpre-
tation of the references to color in this gure legend, the reader is referred to the
web version of this article.)
(blue curve) is shown in Fig. 4. The two added frequency compo-
nents (0.1 Hz and 50 Hz) can clearly be seen in the red curve.
3.2.1. C-EMD and C-EEMD
In the C-EMD, the noise components are identied from the
resulting IMFs. Fig. 5 shows an example of the extracted IMFs from
Type-I signal at 0 dB. It can be seen that IMF10 and IMF11 are chosen
as the baseline references, and IMF1 as the power-line noise com-
ponent. Adjusting the amplitude of the added white noise in Eq. (4)
is indeed important in order to minimize the mode mixing, while
avoiding redundant IMFs [57]. Fig. 5 is analyzed at each SNR level
to select an appropriate white noise amplitude.
Tables 1 and 2 show the results obtained using the C-EMD and
C-EEMD methods on Type-I and Type-II simulated signals for the 5
ECG signals at 5 different input SNR values. In the case of C-EEMD

The efciencies of the proposed methods are evaluated by per- Table 1

forming studies on two types (Type-I and Type-II) of simulated Comparison of output SNRs on Type-I signals.
noise contaminated ECG signals. Type-I signals are generated by
Output SNR (dB)
adding low (0.1 Hz) and high (50 Hz) frequency sinusoids to the
desired ECG signals (d(n)s) and Type-II signals are generated by Input SNR (dB) ECG1 ECG2 ECG3 ECG4 ECG5 Avg.
adding three sinusoids (0.1 Hz, 7 Hz and 50 Hz) to d(n)s. The 0.1Hz C-EMD
sinusoid is chosen to account for the possible baseline wander 10 6.19 5.67 5.96 4.19 6.42 5.69
caused by the respiration and the 7 Hz sinusoid is chosen to cre- 5 7.63 0.11 0.68 0.13 1.21 1.14
0 3.84 3.80 4.86 3.10 3.35 3.79
ate a noise within the ECG bandwidth [56] and the 7 Hz sinusoid
5 6.14 8.16 6.69 7.16 7.87 7.20
is chosen over the colored noise to maintain a clear identication 10 9.83 7.70 7.89 8.87 11.36 9.13
of the IMF containing the noise as the noise power decreases. The
C-EEMD
amplitudes of all noise components are scaled evenly so that they 10 7.27 10.11 11.72 11.38 12.69 10.63
contribute equally to the desired input SNR values. Five sets of 5 15.55 15.43 16.62 15.04 18.25 16.18
Type-I and Type-II signals are generated corresponding to 5 ECG 0 17.93 17.48 18.14 16.37 19.54 17.89
signals (ECG1ECG5) at 5 different input SNR values. 5 18.55 16.56 18.34 15.85 12.49 16.36
10 18.31 16.69 15.64 15.81 21.13 17.52
The Fourier spectrum of the simulated Type-I signal at 0dB (red)
along with the spectrum of the corresponding desired ECG signal Note: Bold values indicate the average output SNRs.
182 K. Krgaard et al. / Biomedical Signal Processing and Control 25 (2016) 178187

Table 2
Comparison of output SNRs on Type-II signals.

Output SNR (dB)

Input SNR ECG1 ECG2 ECG3 ECG4 ECG5 Avg.

C-EMD
10 11.56 10.66 11.84 11.39 -11.93 11.48
5 6.54 5.70 5.70 5.25 6.74 5.99
0 1.08 1.52 1.64 1.14 0.72 1.22
5 2.28 2.14 0.98 1.11 2.62 1.83
10 3.83 4.62 3.31 4.86 5.37 4.40

C-EEMD
10 0.05 0.37 0.06 0.42 0.035 0.17
5 4.56 2.57 2.57 3.08 3.86 3.33
0 6.52 6.25 6.17 4.77 3.36 5.41
5 8.86 8.64 8.47 7.33 7.25 8.11
10 10.01 9.30 9.82 8.03 10.44 9.52 Fig. 7. SNR values for various lter lengths of the low frequency (blue) and high
frequency lter (red). (For interpretation of the references to color in this gure
Note: Bold values indicate the average output SNRs. legend, the reader is referred to the web version of this article.)

3.2.2. EEMD-LMS and EEMD-BLMS

approach, the IMFs estimate are obtained by averaging 300 sets of
Performance comparison of the EEMD-LMS and EEMD-BLMS
IMFs as described in Section 2.1.
methods for denoising ECG signals are discussed in this section.
It is clear from Table 1 that C-EEMD outperforms C-EMD in esti-
In all simulations, the number of EMD iterations for each EEMD
mating the ECG noises. As can be seen, C-EEMD on average has
is set to 300. From the obtained IMFs, the ones corresponding to
an SNR improvement of 7.5220.63 dB for the 5 input SNR val-
the noise frequencies are identied and used as reference inputs to
ues. The C-EMD results on the other hand show a signicantly
both the LMS and BLMS based schemes.
lower improvement, and even a decline in the average output SNR
Before performing the simulations, an analysis to determine the
(9.13 dB) for the case with an input SNR of 10 dB. This decline may
optimum lter lengths (P) for the LMS and BLMS schemes to lter
have been caused by the mode mixing in C-EMD, where the ECG
out the low (0.1 Hz) and power-line frequency noises (individually)
frequency components leaked into the noise estimates. Results
is performed on Type-I ECG1 signal at 0 dB SNR. In this, the lter
in Table 2 clearly show the superiority of the C-EEMD method
length of one of the lters is kept constant (10), while it is varied
compared to the C-EMD. Nevertheless, in both cases, the obtained
through a range of predened values for the other lter. The prede-
output SNRs are much lower than that for the Type-I signals. Inter-
ned numbers are chosen to accommodate for P = L as mentioned in
estingly, the average output SNRs of the C-EMD method is lower
the beginning of this section. Fig. 7 shows the output SNRs as a func-
than the input SNRs in all cases. The reason for this decline in perfor-
tion of these lter lengths. The blue curve shows the output SNR val-
mance can be attributed to the fact that the subtraction of the IMF
ues when the powerline lter length is kept constant (10) and the
corresponding to the 7Hz (sinusoid) noise component may be caus-
baseline lter length is varied from 1 to 200. Likewise, the red curve
ing the rejection of the in-band ECG frequency component. Fig. 5
shows the output SNR values when the baseline lter length is kept
clearly shows the overlap of the spectrums of IMFs inside the ECG
constant (10), and the powerline lter length is varied from 1 to 200.
bandwidth, which causes the noise reference for the 7 Hz sinusoid
The step-size parameter for this experiment is chosen according to
to contain some of the desired ECG signal.
Eq. (11) in all cases. It is clear from the graph that the output SNR
The time plots of the denoised ECG signals (ECG1) using C-EMD
decreases beyond P = 10 for the lter to reduce the low frequency
and C-EEMD on a Type-I signal (0dB) can be seen in Fig. 6. The signal
noise and for the power-line noise, the SNR value saturates beyond
output of the C-EEMD method clearly shows that it closely matches
P = 100. Accordingly, P = 10 is chosen for the baseline lter and 100
the desired ECG signal, whereas the result from the C-EMD method
for the powerline lter. The gure indicates that the powerline lter
seems to retain some of the low frequency (0.1 Hz sinusoid) noise
length could be reduced to 40 without signicantly compromising
components. Results from these two methods have been examined
the SNR improvement but with increased computational gain.
using paired t-test with 5% signicance level, revealing a difference
Performance of both EEMD-LMS and EEMD-BLMS schemes for
in performance between the two methods for all (ve) SNR levels
denoising ECG signals are compared on Type-I and Type-II signals
on both Type-I and Type-II signals.
and the results are tabulated in Tables 3 and 4. For Type-I signals,

Table 3
Comparison of output SNRs on Type-I signals.

Output SNR (dB)

Input SNR ECG1 ECG2 ECG3 ECG4 ECG5 Avg.

EEMD-LMS
10 17.64 18.71 16.55 18.25 20.81 18.39
5 19.00 21.79 20.98 21.72 20.69 20.84
0 19.88 23.21 21.77 23.06 22.50 22.08
5 20.53 22.63 22.24 21.87 22.03 21.86
10 20.42 21.78 19.68 20.65 25.12 21.53

EEMD-BLMS
10 17.76 18.75 16.59 18.29 20.85 18.45
5 19.07 21.76 21.04 21.85 20.83 20.91
Fig. 6. Time plots of the original and denoised Type-I 0dB signal (ECG1) using C- 0 20.00 23.27 21.85 23.18 22.56 22.17
EMD and C-EEMD. (a) Desired (red) and noisy signal (blue); (b) Desired (red) and 5 20.51 22.71 22.30 21.95 22.10 21.91
results obtained from C-EEMD (green) and C-EMD (black). (For interpretation of the 10 20.48 21.88 19.85 20.89 25.45 21.71
references to color in this gure legend, the reader is referred to the web version of
this article.) Note: Bold values indicate the average output SNRs.
 K. Krgaard et al. / Biomedical Signal Processing and Control 25 (2016) 178187 183

Table 4
Comparison of output SNRs on Type-II signals.

Output SNR (dB)

Input SNR ECG1 ECG2 ECG3 ECG4 ECG5 Avg.

EEMD-LMS
10 4.55 4.05 3.91 3.25 5.74 4.30
5 5.65 3.63 3.42 3.59 6.46 4.55
0 5.73 4.97 4.21 4.19 6.60 5.14
5 6.31 6.39 6.48 5.13 6.73 6.21
10 6.85 6.68 6.51 5.59 7.23 6.57

EEMD-BLMS
10 4.72 4.27 4.04 3.48 6.34 4.57
5 5.71 3.76 3.61 3.77 6.61 4.69 Fig. 9. Learning curves of the LMS (black) and BLMS (red) adaptive lters. (For inter-
0 5.87 5.03 4.29 4.20 6.68 5.21 pretation of the references to color in this gure legend, the reader is referred to the
5 6.49 6.51 6.70 5.21 6.81 6.34 web version of this article.)
10 7.08 6.88 6.84 5.66 7.45 6.78

Note: Bold values indicate the average output SNRs.
it is interesting to note that there is only a marginal difference
between the LMS and BLMS based schemes, as the average output
SNR values and hence the SNR improvements (11.5328.39 dB for
LMS based scheme and 11.7128.45 dB for the BLMS based scheme)
for all the cases are more or less even, though the BLMS based
scheme performs slightly better. For Type-II signals (Table 4), both
methods struggle to improve SNR values when the input SNR is
above 5 dB. The LMS still performs more or less similar to BLMS,
however the BLMS (SNR improvement: 3.22 to 14.57dB) performs
slightly better than the LMS scheme (SNR improvement: 3.43 to
14.30dB). The poor performance of these methods at high input
SNRs may again be attributed to the fact that the 7 Hz noise ref-
erence contain some real ECG information (Fig. 5). A comparison
of the LMS and BLMS based schemes to the C-EMD and C-EEMD
schemes clearly show that the hybrid methods are signicantly
better, be it for Type-I or Type-II signals. Paired t-test with a 5%
signicance level indicate a different performance for both meth-
ods on all SNR levels and signal types, except Type-I with 5dB SNR.
Here the p-value (0.07) is just above the signicance level, support-
ing the statement that the two methods have different performance
groups.
The time plots of the denoised ECG signals from the EEMD-LMS
and EEMD-BLMS ltering methods on 0 dB Type-I signal for ECG1
is shown in Fig. 8. The initial transients before convergence can be
clearly seen in the gure due to the adaptive algorithms. Both meth-
ods give good estimates of the desired ECG signal (this is reected
in the output SNRs as well: 19.88 dB (LMS) and 20.00 dB (BLMS)).
Analysis of the convergence of the algorithms is important when
we deal with adaptive algorithms. Fig. 9 shows the learning curves
(samples versus the mean square error (MSE) curve) for both the
LMS and BLMS based schemes obtained as the average of 300
Monte-Carlo simulations. The analysis is performed on a 0dB Type-
I signal (ECG1) with lter lengths 10 and 100. The corresponding
step sizes are, for LMS: 5 104 lms 4 lms
max and 8.5 10 max , respec-
tively and for BLMS, they are 5 103 blms
max and 8.5 103 blms
max ,
blms lms
respectively. Here, max and max are found from Eqs. (8) and (11),
respectively. Both schemes can be seen to converge more or less
identically although the LMS based scheme converges a little faster.
Both schemes reach an MSE value <0.001 in about 2000 samples.
The MSE values for the BLMS scheme are slightly lower compared
to the LMS scheme (see the zoom in Fig. 9) explaining why the BLMS
based scheme works slightly better than the LMS based scheme in
denoising the ECG signals.
3.2.3. DWT-Th and WNN
Here, the denoising performances of the DWT-Th and WNN
methods on simulated noise are compared. Both these methods are
set up as per the discussions in Section 2.3. In the WNN method,
the NN is trained on ECG6-ECG10 and tested on ECG1-ECG5
(details provided in Section 3.1). For the DWT-Th, the thresholds
described in Eq. (16) are obtained from the 0dB Type-I ECG1 signal.
These parameters are used for the rest of the simulation studies in
this section.
In Tables 5 and 6, the output SNR values are listed for both
the DWT-Th and WNN methods when running the simulations on
Type-I and Type-II signals for 5 different input SNR values. It can
be seen that for both types of noise, the WNN method provides
signicantly better performance (average SNR improvement for
Type-I: 8.6624.69 dB and for Type-II: 3.4719.65 dB) compared
to the DWT-Th method (Type-I: 4.2114.05 dB and for Type-II:
2.204.70 dB). It is interesting to note that these methods unlike
the other methods (C-EMD, C-EEMD, EEMD-LMS and EEMD-BLMS),
do not produce a negative SNR improvement (average). Another
observation is that though the performance of these methods (on

Table 5
Comparison of output SNRs on Type-I signals.

Output SNR (dB)

Input SNR ECG1 ECG2 ECG3 ECG4 ECG5 Avg.

DWT-Th
10 4.40 3.96 3.37 3.83 4.71 4.05
5 8.01 7.07 8.17 6.78 9.33 7.87
0 11.52 10.15 11.85 9.45 13.63 11.32
5 13.80 12.27 14.28 11.08 15.86 13.46
10 14.69 13.33 15.21 11.77 16.11 14.21

WNN
10 11.32 19.41 11.24 16.34 15.13 14.69
5 12.22 19.58 12.11 17.60 17.38 15.78
Fig. 8. Time plots of the original and denoised Type-I 0dB signal (ECG1) using EEMD- 0 15.60 23.71 17.04 20.46 18.35 19.03
LMS and EEMD-BLMS. (a) Desired (red) and noisy signal (blue); (b) Desired (red), 5 15.73 23.03 17.11 20.21 18.03 18.82
results obtained from EEMD-BLMS ltering (black) and EEMD-LMS ltering (green). 10 16.00 22.56 17.39 19.36 17.98 18.66
(For interpretation of the references to color in this gure legend, the reader is
referred to the web version of this article.) Note: Bold values indicate the average output SNRs.
184 K. Krgaard et al. / Biomedical Signal Processing and Control 25 (2016) 178187

Table 6
Comparison of output SNRs on Type-II signals.

Output SNR (dB)

Input SNR ECG1 ECG2 ECG3 ECG4 ECG5 Avg.

DWT-Th
10 5.08 5.26 5.13 5.57 5.42 5.30
5 0.15 0.40 0.28 0.79 0.41 0.40
0 4.69 4.26 4.53 3.74 4.38 4.32
5 8.89 8.52 8.87 7.51 9.15 8.59
10 12.88 11.86 12.74 10.30 13.21 12.20

WNN
10 9.17 11.12 8.26 12.49 7.19 9.65
5 9.92 12.54 9.93 11.14 8.76 10.46 Fig. 11. Training curve for 0dB type I signals.
0 11.15 13.29 11.19 12.68 10.11 11.68
5 12.25 13.84 12.94 12.76 10.60 12.48 of the proposed denoising schemes, they have been extended to
10 13.28 14.14 13.54 13.46 12.93 13.47 ve sets of real, noise contaminated ECG signals (ECG1 to ECG5),
Note: Bold values indicate the average output SNRs. which are downloaded from the MIT/BIH ECG-ID database [54].
The database also contains the desired (noise minimized) ECG sig-
Type-I signals) is signicantly lower than the LMS and BLMS based nals which we use to compare the performance of the proposed
schemes (Table 3), they are are much better than the C-EMD and algorithms for denoising the signals.
C-EEMD based methods. Paired t-test with a 5% signicance level The parameter settings for different methods used for the sim-
also indicate a different performance for these methods except for ulation studies have been used in the real data analysis as well.
10 dB Type-I and Type-II with p-values of 0.055 and 0.11, respec- Since the desired ECG signals corresponding to ECG1 to ECG5 are
tively. This can be seen from Tables 5 and 6 where the difference in available, the output SNR values for the denoised signals using the
mean performance becomes less compared to variation within the proposed methods are calculated (using Eq. (17)) and tabulated in
10 dB signals. Table 7. Since we observe a degradation in performance when using
The time plots for the denoised ECG signals corresponding to IMFs inside the bandwidth of the ECG signals in the simulation stud-
the DWT-Th and WNN methods when run on 0 dB Type-I signal ies, only IMFs outside the ECG bandwidth are considered for the
(ECG1) are shown in Fig. 10. Both methods provide good denois- EMD and EEMD based methods.
ing performances however the DWTTh is shown to cause minor From Table 7, it can be seen that the C-EEMD is once again
issues at the QRS complexes. The WNN method on the other hand, shown to perform signicantly better than the C-EMD method
is shown to retain some low-frequency component in the denoised for all ECG signals considered. The use of adaptive lters (LMS
ECG signal. and BLMS) further improves the performance signicantly. Surpris-
It is important to study the learning efciency (training curve) ingly, the EEMD-LMS based method provides an SNR improvement
of the MLP network. For this, the MLP is trained on 5 ECG data sets of 1.05 dB more than EEMD-BLMS. When it comes to the wavelet
(ECG6-ECG10) of 5000 samples. For each sample (x(n)), the 154 based methods, the WNN method is once again shown to be supe-
DWT coefcients are determined and used as the inputs to the MLP. rior to the DWT-Th and is comparable to both the LMS and BLMS
The corresponding desired signals (d(n)s) are used as the target. based methods. However, the DWT-Th performs the worst among
This huge set of (5 5000) input-target pairs form one epoch. The all the methods with an output (average) SNR of 1.61 dB.
MLP is trained up to 10 epochs, the mean square error (MSE) is The time plots and the corresponding noise spectra for all six
determined and the learning curve is plotted in Fig. 11. It can be proposed approaches in denoising ECG5 are discussed in detail
observed that the MSE decreased with an increase in the number here. As can be seen from Table 7, the inherent SNR of real ECG5
of epochs, until achieving the lowest value at about 7 epochs and is estimated to be 16.01 dB. The noisy ECG signal (green) and the
remained more or less unchanged with further training. corresponding desired signal (orange) are shown in Fig. 12(a). It is
clear that the real ECG5 has a signicant amount of low frequency
3.3. Results on real data as well as high frequency noise components. Subtracting the
desired signal from the original signal provides the noise that
After the extensive studies on ECG signals contaminated with is present in the real signal and the corresponding noise power
simulated noises (Type-I and Type-II) to assess the performance spectrum (black) is plotted in Fig. 12(c) along with the noise
power spectrum after EMD subtraction (red). Here, the presence
of low frequency noise is still evident in the signal after the C-EMD
processing. This can be attributed to mode mixing in the IMFs
containing the lowest frequencies.
For all EEMD based methods, the same IMFs are identied as
noise components. This is done to alleviate the inherent problem

Table 7
Input and output SNRs for the real ECG data.

Inherent SNR ECG1 ECG2 ECG3 ECG4 ECG5 Avg.

5.97 7.72 9.96 5.23 16.01 6.59

C-EMD 10.13 9.77 3.07 9.26 8.36 6.89

C-EEMD 13.33 12.19 9.36 9.78 9.02 10.74
EEMD-LMS 13.36 14.51 10.66 13.51 13.91 13.19
Fig. 10. Time plots of the original and denoised Type I 0dB signal (ECG1) using EEMD-BLMS 13.37 12.45 9.94 11.65 13.29 12.14
DWT-Th and WNN. (a) Desired (red) and noisy signal (blue); (b)Desired (red), and DWT-Th 9.28 1.29 2.32 7.08 4.70 1.61
results obtained from DWT-Th (black) and WNN (green). (For interpretation of the WNN 11.68 16.34 8.26 11.41 11.92 11.92
references to color in this gure legend, the reader is referred to the web version of
this article.) Note: Bold values indicate the average output SNRs.
 K. Krgaard et al. / Biomedical Signal Processing and Control 25 (2016) 178187
Fig. 12. Time plots of the original (noisy), desired and denoised ECG5, as well as the
power spectrum of the original and residual noise. (a) Desired (orange) and original
signal (green);(b) desired (orange) result obtained through C-EMD (blue); (c) noise
spectrum before (black) and after C-EMD (red). (For interpretation of the references
to color in this gure legend, the reader is referred to the web version of this article.)
in EEMD where the same number of IMFs is not produced for dif-
ferent trials. In Fig. 13, results from the C-EEMD denoising on ECG5
is shown. Here, EEMD produced 12 IMFs and from these, IMFs 1
and 2 have been identied as the power-line reference, and the
three lowest IMFs (10, 11, and 12) are identied for the baseline ref-
erence. The corresponding noise power spectra have been shown
in Fig. 13(b); the red curve here shows the residual noise power
spectrum. In Fig. 13(a), the denoised ECG signal (blue) still contains
some high frequency components even after ltering. This is visi-
ble in the residual noise power spectrum (Fig. 13(b)) where a peak
at 50 Hz is seen. In this case, the 50 Hz noise appeared between
the lter banks of the rst and second IMF, dividing the noise
power between them. However, including the second IMF in the
high frequency noise component would result in subtraction of
ECG components and thereby further attenuation of the desired
ECG signal.
Results from the EEMD-LMS method when applied to ECG5 are
shown in Fig. 14. The desired signal (orange) and the corresponding
denoised signal (blue) are shown in Fig. 14(a). It can be clearly seen
that after about 4 s, the denoised ECG signal closely follows the
desired ECG signal, showing the effectiveness of the EEMD-LMS
method in denoising. The initial transients can be attributed to the
convergence time of the adaptive algorithm, which is about 4 s,
according to the learning curve shown in Fig. 9. Accordingly, in the
noise power spectrum estimates as well as the SNR calculations,
Fig. 13. Time plots of the desired and denoised ECG5, as well as the power spectrum
of the original and residual noise. (a) Desired (orange) and ECG5 obtained through
C-EEMD (blue); (b) noise frequencies before (black) and after C-EEMD (red). (For
interpretation of the references to color in this gure legend, the reader is referred
to the web version of this article.)
185
Fig. 14. Time plots of the desired and denoised ECG5, as well as the power spectrum
of the original and residual noise. (a) Desired (orange) and ECG5 obtained through
EEMD-LMS (blue); (b) noise frequencies before (black) and after EEMD-LMS (red).
(For interpretation of the references to color in this gure legend, the reader is
referred to the web version of this article.)
the rst 4 s of the noise data have been excluded. Fig. 14(b)
however, reveals the presence of a very small amount of low
frequencies, which can be neglected. The output SNR performance
(29.92 dB improvement (Table 7)) of this method is superior to all
other methods.
Fig. 15 illustrates the corresponding plots of the denoised ECG
with the EEMD-BLMS approach. Like in the LMS approach, the
BLMS scheme also needs about 4 s to converge to the optimum
lter coefcients. As can be seen in Fig. 15(a), the denoised ECG
is able to follow the desired ECG quite closely after 4 s. The noise
power spectra shown in Fig. 15(b) also shows a more or less simi-
lar performance, as with the EEMD-LMS approach. The output SNR
performance is very impressive here, although slightly inferior to
the EEMD-LMS approach. This shows the superiority of these meth-
ods in minimizing noise components from the ECG signal. Although
according to the convergence analysis (Fig. 9), the BLMS method is
supposed to provide a better performance compared to the LMS,
the SNR values (Table 7) show little inferior performance here in
the real data case. This discrepancy of the BLMS method in the sim-
ulated and the real data case could be caused by the addition of
white noise, which the BLMS struggles to remove. This is in line with
[17] which reports that BLMS lters are not optimal for minimiz-
ing the white noise. Fig. 15(b) shows a power spectrum similar to
the one obtained from EEMD-LMS. Here one can see that the back-
ground noise level is slightly higher than in the LMS case (Fig. 14),
supporting its little inferior performance.
Time and power spectrum plots of the signal using the DWT-Th
method are shown in Fig. 16. As can be seen from the time-plots, the
denoised ECG signal (blue) still contains a large amount of baseline
wander (in the range of 0.10.3 Hz), which can be seen in the noise
power spectrum as well (Fig. 16(b)). Fig. 16(b) also shows some
Fig. 15. Time plots of the desired and denoised ECG5, as well as the power spectrum
of the original and residual noise. (a) Desired (orange) and ECG5 obtained through
EEMD-BLMS (blue); (b) noise frequencies before (black) and after EEMD-BLMS (red).
(For interpretation of the references to color in this gure legend, the reader is
referred to the web version of this article.)
186 K. Krgaard et al. / Biomedical Signal Processing and Control 25 (2016) 178187
Fig. 16. Time plots of the desired and denoised ECG5, as well as the power spectrum
of the original and residual noise. (a) Desired (orange) and ECG5 obtained through
DWT-Th (blue); (b) noise frequencies before (black) and after DWT-Th (red). (For
interpretation of the references to color in this gure legend, the reader is referred
to the web version of this article.)
Fig. 17. Time plots of the desired and denoised ECG5, as well as the power spectrum
of the original and residual noise. (a) Desired (orange) and ECG5 obtained through
WNN (blue); (b) noise frequencies before (black) and after WNN (red). (For inter-
pretation of the references to color in this gure legend, the reader is referred to the
web version of this article.)
residual high frequency components (around 50 Hz), although not
as prominent as the baseline wander. Due to the presence of these
residual noise components, the SNR improvement is not signicant
as seen in Table 7, where the output SNR after denoising is 4.70 dB,
giving an SNR improvement of 11.11 dB.
In Fig. 17, the results of the WNN denoising method is illustrated.
Although Table 7 shows that the SNR improvement of WNN is about
2 dB lower than the EEMD-LMS (27.92 dB), Fig. 17(b) reveals high
attenuation of both low and high frequency noises. The time-plots
(Fig. 17(a)) also clearly show clear matching of the denoised signal
to the desired ECG signal. The lower SNR performance of the WNN
method may be attributed to the fact that there is a small amount
of low frequency noise component still present in the denoised ECG
(according to residual noise power spectrum in Fig. 17(b)).
4. Conclusion
ECG is one of the most widely used non-invasive method to
understand the rhythmic activities of the heart. These signals how-
ever are obscured by unwanted noises/artifacts which are to be
minimized before the clinicians could make a proper diagnosis.
In this article, a comprehensive analysis of the performance of
six different ECG denoising algorithms based on both the EMD
and wavelets is performed. They are the C-EMD, C-EEMD, EEMD-
LMS, EEMD-BLMS, DWT-Th and WNN methods. Results from the
simulated and real data show that the EEMD-LMS and EEMD-
BLMS hybrid methods provide the best performances, followed by
another hybrid scheme, the WNN method. This method however,
produced signicantly better performance than the DWT-Th. The
C-EEMD method is shown to outperform the C-EMD method in
estimating the noises present in the ECG but inferior to the hybrid
schemes in terms of denoising the ECG signals.
References
[1] J.A. Van Alste, T.S. Schilder, Removal of base-line wander and power line inter-
ference from the ECG by an efcient FIR lter with a reduced number of
taps, in: Proceedings of the 2nd International Conference on vol. 32, 1985, pp.
10521060.
[2] T. Slonim, M.A. Slonim, E. Ovsyscher, The use of simple FIR lters for ltering
of ECG signals and a new method for post-lter signal reconstruction, in: Pro-
ceedings of Computers in Cardiology Conference, London, 1993, pp. 871873.
[3] M.S. Chavan, R.A. Agarwala, M.D. Uplane, Suppression of baseline wander and
power line interference in ECG using digital IIR lter, Int. J. Circuit Syst. Signal
Process. 2 (2) (2008) 132142.
[4] S. Cuomo, G. De Pietro, R. Farina, A. Galletti, G. Sannino, A novel O(n) numerical
scheme for ECG signal denoising, Proc. Comput. Sci. 51 (2015) 775784.
[5] V. Alfonso, W. Thomkins, T. Nguyen, S. Trautmann, S. Luo, Filter bank-based
processing of the stress ECG, in: Proceedings of the 17th IEEE Engineering in
Medicine and Biology Society Conference vol. 2, 1995, pp. 887888.
[6] S. Hargittai, Savitzky-Golay least-squares polynomial lters in ECG signal
processing, in: Proceedings of the Computers in Cardiology, Lyon, 2005, pp.
763766.
[7] K.M. Chang, S.H. Liu, Gaussian noise ltering from ECG by wiener lter
and ensemble empirical mode decompostion, J. Siganl Proc. Syst. 64 (2011)
249264.
[8] N.V. Thakor, Y.-S. Zhu, Applications of adaptive ltering to ECG analysis: noise
cancellation and arrhythmia detection, IEEE Trans. Biomed. Eng. 38 (8) (1991
Aug) 785794.
[9] J. Jenitta, A. Rajeswari Dr., Denoising of ECG signal based on improved adaptive
lter with EMD and EEMD, in: Conference on Information and Communication
Technology, 2013, pp. 957962.
[10] G.M.S. Sajjad, H. Rahman, A.K. Dey, A.M. Biswas, Z. Islam, A.K.M.J. Hoque, Per-
formance comparison of modied LMS and RLS algorithms in denoising of ECG
signals, Int. J. Eng. Technol. 2 (3) (2012) 466468.
[11] C. Chandrakar, M.K. Kowar, Denoising ECG signals using adaptive lter algo-
rithm, Int. J. Soft Comput. Eng. 2 (1) (2012) 120123.
[12] M. Rahman, R. Shaik, D. Reddy, Adaptive noise removal in the ECG using the
Block LMS algorithm, in: Proceedings of the 2nd International Conference on
Adaptive Science and Technology (ICAST), 2009, pp. 380383.
[13] Y. Weiting, Z. Runjing, An improved self-adaptive lter based on LMS algorithm
for ltering 50 Hz interference in ECG signals, in: Proceedings of the 8th Interna-
tional Conference on Electronic Measurement and Instruments ICEMI07, 2007,
pp. 3-8743-878.
[14] M. Rahman, R. Shaik, D. Reddy, Noise cancellation in ECG signals using nor-
malized sign-sign LMS algorithm, in: Proceedings of the IEEE International
Symposium on Signal Processing and Information Technology, (ISSPIT), 2009,
pp. 288292.
[15] S. Olmos, L. Srnmo, P. Laguna, Block adaptive lters with deterministic ref-
erence inputs for event-related signals: BLMS and BRLS, IEEE Trans. Signal
Process. 50 (5) (2002) 11021112.
[16] S.A. Rehman, R.R. Kumar, Performance comparison of adaptive lter algorithms
for ECG signal enhancement, Int. J. Adv. Res. Comput. Commun. Eng. 1 (2)
(2012).
[17] S. Malik, N.A. Saqib, Adaptive ltering for noisy and mismatched conditions
in speaker recognition, in: Proceedings of the IEE International Conference on
Signal and Image Processing Applications, 2011, pp. 162167.
[18] M.Z.U. Rahman, R.A. Shaik, D.V.R.K. Reddy, Adaptive noise removal in the ECG
using the block LMS algorithm, in: Proceedings of the 2nd International Con-
ference on Adaptive Science & Technology, 2009, pp. 380383.
[19] N.K. Muhsin, Noise removal of ECG signal using recursive least square algo-
rithms, Al-Khwarizmi Engg. Jl. 7 (1) (2011) 1321.
[20] S. Olmos, L. Srnmo, P. Laguna, Block adaptive lters with deterministic ref-
erence inputs for event-related signals: BLMS and BRLS, IEEE Trans. Signal
Process. 50 (5) (2002).
[21] M. Mneimneh, E. Yaz, M. Johnson, R. Povinelli, An adaptive Kalman lter for
removing baseline wandering in ECG signals, in: Proceedings of the 33rd Annual
International Conference Computers and Cardiology, 2006, pp. 253256.
[22] R. Sameni, M.B. Shamsollahi, C. Jutten, Filtering electrocardiogram signals using
the extended Kalman lter, in: Proceedings of the 27th Annual International
Conference IEEE Engineering in Medicine and Biology Society (EMBS), 14
September, Shanghai, China, 2005, pp. 56395642.
[23] R. Sameni, M.B. Shamsollahi, C. Jutten, M. Babaie-Zadeh, Filtering noisy ECG
signals using the extended Kalman lter based on a modied dynamic ECG
model, in: Proceedings of the 32nd Annual International Conference Computers
in Cardiology, 2528 September, Lyon, France, 2005, pp. 10171020.
[24] O. Sayadi, M.B. Shamsollahi, ECG denoising and compression using a modi-
ed extended Kalman lter structure, IEEE Tans. Biomed. Eng. 55 (9) (2008)
22402248.
[25] R. Sameni, M.B. Shamsollahi, C. Jutten, G.D. Clifford, A nonlinear Bayesian l-
tering framework for ECG denoising, IEEE Trans. Biomed. Eng. 54 (12) (2007)
21722185.
[26] T. Schreiber, D.T. Kaplan, Nonlinear noise reduction for electrocardiograms,
Chaos 6 (1) (1996) 8792.
 K. Krgaard et al. / Biomedical Signal Processing and Control 25 (2016) 178187
[27] A.K. Barros, A. Mansour, N. Ohnishi, Removing artifacts from electrocardio-
graphic signals using independent components analysis, Neurocomputing 22
(1-3) (1998) 173186.
[28] J.S. Paul, M.R.S. Reddy, J. Kumar, Data processing of stress ECG using discrete
cosine transform, Comput. Biol. Med. 28 (6) (1998) 639658.
[29] W. Zhaohua, H. Norden E, Ensemble empirical mode decomposition: a noise
assisted data analysis method, in: Advances in Adaptive Data Analysis, 2009,
pp. 141.
[30] N. Pan, V. Mang I, M. Peng Un Pun Sio Hang, Accurate removal of baseline
wander in ECG using empirical mode decomposition, in: Proceedings of the
Joint Meeting of the 6th International Symposium on Noninvasive Functional
Source Imaging of the Brain and Heart and the International Conference on
Functional Biomedical Imaging, 2007, pp. 177180.
[31] C. Kang-Ming, Arrhythmia ECG noise reduction by ensemble empirical mode
decomposition, Sensors 10 (6) (2010) 60636080.
[32] G. Singh, G. Kaur, V. Kumar, ECG denoising using adaptive selection of IMFs
through EMD and EEMD, in: Proceedings of the International Conference on
Data Science & Engineering (ICDSE), 2014, pp. 228231.
[33] V. Prasad, T.S. Latha, M. Suresh, Denoising of biological signals using wavelets,
Int. J. Curr. Eng. Technol. 3 (3) (2013) 863866.
[34] R. Cohen, Signal Denoising Using Wavelets, 2012, Project Report
(http://tx.technion.ac.il/rc).
[35] K. Daqrouq, ECG baseline wander reduction using discrete wavelet transform,
Asian J. Inf. Technol. 4 (11) (2005) 989995.
[36] B. Arvinti, M. Costache, D. Toader, M. Oltean, A. Isar, ECG statistical denoising in
the wavelet domain, in: Proceedings of the 9th IEEE International Symposium
on Electronics and Telecommunications, 2010, pp. 307310.
[37] K. Daqrouq, A.-R. Al-Qawasmi, ECG enhancement using wavelet transform,
WSEAS Trans. Biol. Biomed. 7 (2) (2010) 6272.
[38] G. Georgieva-Tsaneva, K. Tcheshmedjiev, Denoising of electrocardiogram data
with methods of wavelet transform, in: Proceedings of the International Con-
ference on Computer Systems and Technologies, 2013, pp. 916.
[39] M.F.M. Kania, Wavelet denoising for multi-lead high resolution ECG signals,
Meas. Sci. Rev. 7 (2) (2007).
[40] P.S. Gokhale, ECG signal de-noising using discrete wavelet transform for
removal of 50Hz PLI noise, Int. J. Emerg. Technol. Adv. Eng. 2 (5) (2012)
8185.
[41] O. Sayadi, M.B. Shamsollahi, Multiadaptive bionic wavelet transform: applica-
tion to ECG denoising and baseline wandering reduction, EURASIP J. Adv. Signal
Process. 2007 (2007), http://dx.doi.org/10.1155/2007/41274, Article ID 41274,
11 p.
187
[42] M. Alfaouri, K. Daqrouq, ECG signal denoising by wavelet transform threshold-
ing, Am. J. Appl. Sci. 5 (3) (2008) 276281.
[43] G. Georgieva-Tsaneva, K. Tcheshmedjiev, Denoising of electrocardiogram data
with methods of wavelet transform, in: Proceedings International Conference
on Computer Systems and Technologies Compsystech13, 2013, pp. 916.
[44] H.A. Kestler, M. Haschka, W. Kratz, et al., De-noising of high-resolution ECG
signals by combining the discrete wavelet transform with the Wiener lter,
in: Proceedings Computers in Cardiology, September, Cleveland, OH, 1998, pp.
233236.
[45] Q. Zhang, A. Benveniste, Wavelet Networks, IEEE Trans. Neural Netw. 3 (6) (1992
November) 889898.
[46] S. Poungponsri, X. Yu, Electrocardigram (ECG) signal modeling and noise reduc-
tion using wavelet neural networks, in: Proceedings of the IEEE International
Conference on Aut. and Logis, 2009, pp. 394398.
[47] S. Poungponsri, X. Yu, An adaptive ltering approach for electrocardiogram
(ECG) signal noise reduction using neural networks, Neucomputing 117 (2013)
206213.
[48] N.E. Huang, Review of empirical mode decomposition, Proceedings of the SPIE
4391 (2001) 7180.
[49] S. Haykin, Adaptive Filter Theory, 3rd ed., Prentice-Hall, Inc. A S, New Jersey,
1996, pp. 449451, ch. 10, sec. 1.
[50] S. Mallet, A Wavelet Tour of Signal Processing, 2nd ed., A Harcourt Sci. and
Technol. Comp, London, 2001, ch. 4 & 7, sec 3, pp. 8991 & 255259.
[51] B.N. Singh, A.K. Tiwari, Optimal selection of wavelet basis function applied to
ECG signal denoising, Dig. Signal Process. 16 (3) (2006) 275287.
[52] X. Pan, B. Lee, C. Zhang, A comparison of neural network backpropagation algo-
rithms for electricity load forecasting, in: Proceedings of the IEEE International
Workshop on Intelligent Energy Systems, 2013 November, pp. 2227.
[53] F.D. Foresee, M.T. Hagan, Gauss-Newton Approximation to Bayesian Regular-
ization, in: Proceedings of the 1997 International Joint Conference on Neural
Networks, 1997, pp. 19301935.
[54] The MIT-BIH ECG-ID Database [Online]. Available at: http://www.physionet
.org/cgi-bin/atm/ATM.
[55] Biometric Human Identication based on ECG [Online]. Available at:
http://physionet.org/physiobank/database/ecgiddb/biometric.shtml.
[56] V.K. Murthy, T.M. Grove, G.A. Harvey, L.J. Haywood, Clicinal usefullness of ECG
frequency spectrum analysis, in: Proceedings of Annual Symposium on Com-
puter Applications in Medical Care, 1978, pp. 610612.
[57] J. Zhang, R. Yan, R.X. Gao a, Z. Feng, Performance enhancement of ensem-
ble empirical mode decomposition, Mech. Syst. Signal Process. 24 (7) (2010)
21042123.