ECG Signal Denoising by Empirical Mode Decomposition [602237]

Journal of The Institution of Engineers (India): Series B
ECG Signal Denoising by Empirical Mode Decomposition
–Manuscript Draft–

Manuscript Number: IEIB-D-16-00177
Full Title: ECG Signal Denoising by Empirical Mode Decomposition
Article Type: Original Contribution
Section/Category: Electrical Engineering
Keywords: artifact; baseline wander; ecg; emd; imf
Abstract: The ECG is an important clinical tool to diagnose or to monitor various cardiac
diseases. Like other electrical signals, the ECG signal also corrupted by various kinds
of noise or artifacts which affects diagnosis interpretation and leads to erroneous
results. In order to diagnose a cardiac abnormality an accurate and noiseless ECG
signal is required. In this paper a denoising algorithm based on Empirical Mode
Decomposition (EMD) has been proposed. The performance of present algorithm has
been compared with other established denoising methods. the comparison has been
performed on the basis of statistical tools and morphological study of the signals. The
obtained results shows that the present algorithm performs better than other denoising
techniques.
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation

ECG Signal Denoising by Empirical Mode
Decomposition

Abstract— The ECG is an important clinical tool to diagnose
or to monitor various cardiac diseases. Like other electrical
signals , the ECG signal also corrupted by vari ous kinds of noise
or artifacts which affects diagnosis interpretation and leads to
erroneous results. In order to diagnose a cardiac abnormality an
accurate and noiseless ECG signal is r equired. In this paper a
denoising algorithm based on Empirical Mode Decomp osition
(EMD) has been proposed . The performance of present
algorithm has been compared with other established denoising
methods . the comparison has been performed on the basis of
statistical tools and morphological study of the signals. The
obtained results shows that the present algorithm p erforms
better than other denoising techniques.
Keywords — artifact; baseline wander; ecg; emd ; imf;
I. INTRODUCTION
The electrocardiogram (ECG o r EKG) is a graphical
representation of the time variant voltages produced by
myocardium during the cardiac cycle. The ECG is an
important clinical tool to diagnose various cardiac diseases
and to monitor the conditions of the patients associated with
the heart. It also serves as a timing reference for other
measurements. For clinical applications a noise free and
accurate signal is required in order to diagnose cardiac
diseases accurately. The ECG signal is corrupted by various
kinds of artifacts while rec ording. These artifacts include high
frequency artifacts caused by muscular movement, power
frequency interferences generated by electrical devices present
at recording environments, and low frequency artifacts like
base line wander and baseline drift [1, 2].
Various signal processing techniques have been introduced
and developed during last four decades among which Digital
Signal Processing (DSP) have been preferred over Analog
Signal Processing due to less hardware requirements and
better results [3]. Different researchers have worked on pre –
processing of ECG signals and a number of DSP techniques
have been proposed and developed so far. Some of the
relevant contributions include Finite Impulse Response (FIR)
filters [4, 5], Infinite Impulse Response (I IR) filters [5, 6],
Adaptive filters [7, 8], Principal Component Analysis (PCA)
[9] Independent Component Analysis (ICA), [10] and Wavelet
Transform (WT) [3, 11].
During last decade a new DSP technique, Empirical Mode
Decomposition (EMD) has found broad a pplications in
biomedical signal processing [12]. The EMD was first introduced by N. E. Huang et al. as a Digital Signal Processing
(DSP) technique for nonlinear and non stationary signals
which is best suited for processing ECG signals [13].
Traditional d ata analysis methods, like Fourier and Wavelet –
based methods require some predefined basis functions to
represent a signal whereas, the EMD relies on a fully data
driven mechanism that does not require any priory known
basis. Therefore, it is well suited f or nonlinear and
nonstationary signals, such as biomedical signals [14].
In this paper, an EMD based technique has been presented
for removing various artifacts from ECG signals. The high
frequency artifacts and low frequency artifacts have been
addresse d separately. In order to compare the performance of
present algorithm with traditional denoising techniques, signal
power, Power Spectral Density (PSD) and Signal to Error
Ratio have been calculated.
II. METHODOLOGY
A. Empirical Mode Decomposition (EMD)
Empirical Mode Decomposition (EMD) is a straight
forward, direct and adaptive data decomposition method. The
main idea behind EMD is that every signal is a combination of
some Intrinsic Mode Functions (IMFs) including diverse
oscillations. Therefore, a sig nal can be decomposed into a
number of intrinsic mode functions (IMFs), each of which
must satisfy the following two conditions.
i. In the whole data set, the number of extrema and the
number of zero -crossings must either be equal or
differ at most by one.
ii. At any point, the mean value of the envelope defined
by local maxima and the envelope defined by the
local minima is zero.
The EMD algorithm can be summarized as follows:
i. Identify and connect all local maxima by cubic spline
line to produce the upper envelop e and local minima
to produce the lower envelope
ii. Then take mean of upper and lower envelope (m 1),
and subtract it from the signal x(t) to calculate first
component, h 1, i.e.

11 ()x t m h (1)
iii. Check whether h 1 satisfies the conditions of IMF or
not.
iv. If h 1 satisfies the conditions of IMF then h 1 will be
treated as the first IMF of the signal x(t).
v. If h 1 is not an IMF then treat h 1as the original signal
and repeat steps 1 –4.
vi. In this way the given signal x(t) can be decomposed
as: * On contract Faculty in EED at MMM University of Technology ,
Gorakhpur, email – rohila.ashish6@gmail.com
**Research cum Teaching F ellow in EED at MMM University of
Technology Gorakhpur, email -rajkp007@gmail.co m
*** Professor, EED, MMM University of Technology, Gorakhpur,
email – girivkmmm @gmail.com

Manuscript (excluding authors' names and affiliations)

 ) ( 2
k k x t c t r t
Where, c k(t) is the first IMF and r k(t) is the residual signal
after k sifting.
The decomposition process can be terminated when the
residue r K becomes a monotonic function, a constant or a
function with only one extremum. The original signal x(t) can
be reconstructed from its IMFs as:


1 ( ) ( ) (3)K
kN
kx t c t r t

B. Removing High Frequency Artifacts using EMD
In Empirical Mode Decomposition the hi gh frequency
components of a signal lie within first several IMFs whereas;
low frequency components lie in higher order IMFs. The high
frequency artifacts in ECG signal can be removed by partial
reconstruction of IMFs. The high frequency artifacts from the
signal can be removed by excluding some starting IMFs from
the reconstruction process. However, this method is well
suited for signals in which noise component can be clearly
identified from the signal but in case of ECG signals removal
of lower order IMF s introduce severe distortions in the
signal. This is due to the fact that QRS complex in ECG also
lies within starting IMFs of the signal. In order to preserve
QRS wave, the lower order IMFs must be processed and
included in reconstruction of the denoi sed signal.
In the present work an improved method to denoise ECG
signal has been presented. In this method lower order IMFs
have been summed up and windowed by Tukey window in
order to remove high frequency artifacts and preserve QRS
complex [14].
Following steps have been included in present algorithm.
i. Decompose ECG signal into a number of IMFs with
the procedure described in section II.
ii. Determine the number of IMFs contributing to
noise by performing some statistical tests.
iii. Use proper windowing to remove high frequency
noise and preserve QRS complex.
iv. Reconstruct the signal to filter high frequency noise.
C. Removing Baseline Wander using EMD
Due to the fact that frequency of IMF decreases as the
order of IMF increases, the baseline wander component of
ECG signal exists within last several IMFs [15]. Therefore,
last IMFs should be processed in order to remove baseline
wander. in this work a filter bank of lowpass filters having cut
off frequency 0.5Hz have been drawn and starting from the
last IMF, al l IMFs of the ECG signal up to a stopping
criterion; have been filtered with these low pass filters. The
outputs of these filters are:
( ) ( )* ( ) (4)K
ii
iQb t h t c t


Where, b i(t) is the filtered IMF and h(t) is the impulse
response of lowpass filter. The outpu t bi(t) extract the baseline
wander component in c i(t). the variance of each b i(t) has been
determined as:
21
01var ( ) [ ( ) ] (5)1L
i i bi
rb t b t mL


Where, m bi is the mean value of b i(t). The value of Q can
be selected such that var{b Q+1(t)}>n and var{b Q(t)}<n. The
value of n can be based on prior knowledge or can be experimentally tuned based on baseline wander behaviour.
Then the baseline in the signal can be reconstructed as
( ) (6)K
i
iQBL b t


This baseline wander BL can be subtracted from the
original signa l in order to remove baseline wander from the
ECG signal.
III. RESULTS AND DISCUSSI ONS
In the present work the two types of artifacts i.e. high
frequency artifacts and low frequency artifacts have been
addressed separately. The test results of removing artifacts
from ECG signals have been discussed in following
subsections.
A. Removing High Fre quency Artifacts from ECG Signals
In the present work three signals D_00001, D_00002 and
1L2 from CSE database, one signal 101m from MIT/BIH
Database and two Lab recording signals have been considered
as test signals. These signals are of 10 seconds durati on
sampled with 500Hz sampling frequency. According to method
discussed in II all the signals have been decomposed into a
number of IMFs. The 22 IMFs of signal D_00002 have been
shown in figure 1. From figure 1 it is clear that high frequency
components of signal lie within lower order IMFs and lower
frequency components lie within higher order IMFs.

Fig.1. EMD of noisy signal D_00002. From top to bottom: IMF 1 to IMF 22.
In second step the number of IMFs which contribute most to
noise has been obtained with the statistical test. This test is
called as t -test which is also used in [14]. Since, lower order
IMFs have high frequency component we performed this test
on lower order IMFs to determine if a particular combination
of IMF has zero mean. The number of IMFs which have been
dominated by noise, referred to as noise order. In this test we
have calculated mean of combination of IMFs as:

1( ( )) (7)M
Mi
iH mean c t

Where, c i(t) is the ith IMF and M = 1,2,3… Up to some
significance level the value of M f or which H M is zero has
been referred to as noise order. The role of this noise order can
be considered similar to as cut off frequency in frequency
domain filtering. The noise order indicates the number of
IMFs that should be removed or filtered to remove high
frequency noise. The results of t -test performed on test signals
have been summarized in table 1. The H 5 has insignificant
value as compared with other higher values of M, for the

signal D_00001. Therefore, noise order has been taken as 5 for
this si gnal. Similarly for signal D_00002 H 5 has significant
value as compared with higher values of M, hence its noise
order has been taken as 4.

TABLE : 1 RESULTS OF T-TEST PERFORMED O N TEST SIGNALS
D_00001 AND D_00002 .
M HM for
signal
D_00001 HM for
signal
D_00002 HM for
signal
1L2 HM for
signal
101m HM for
signal
Lab
Record 1 HM for
signal
Lab
Record 2
1 0.0012 0.0442 0.0000 0.0001 0.0001 0.0000
2 0.0532 0.0119 0.0000 0.0001 0.0000 0.0003
3 0.0242 0.0620 0.0000 0.0004 0.0003 0.0002
4 0.4121 0.0868 0.0013 0.0010 0.0005 0.0001
5 0.5568 0.3467 0.0011 0.0017 0.0006 0.0004
6 1.6600 1.0025 0.0000 0.0063 0.0020 0.0010
7 3.5060 2.3822 0.0000 0.0079 0.0025 0.0019
8 5.8747 3.8810 0.0005 0.0089 0.0027 0.0005
9 7.9832 3.5807 0.0024 0.0105 0.0036 0.0017
10 8.0293 3.9233 0.0039 0.0104 0.0019 0.0016
It has been observed that in some cases ECG signal itself
have some zero mean components. In this case the above
procedure results in over smoothing or loss of some useful
information. In order to avoid this situation the noise order N
can be selected as:
N = min (M, 5) (8)
This is due to the fact that IMFs with order greater than
five hardly contains high frequency components. After
obtaining noise order by the procedure discussed above we
have to filter the IMFs that contribute to noise. For windowing
in EMD domain we have employed Tukey window of length
L, and a shape parameter
 . A Tukey window can be
described as:
1
12
21
1
211 cos ,2
( ) 1, (9)
0,tt
w t t
t  

      




And the shape parame ter
 may be defined as:

12
1 (10)2
Where,
1 is the flat region limit and
2 is the transition
region limit. The graphical representation of Tukey Window
has been shown in figure 2.

Fig.2. Tukey Window (tapered cosine) function
In the present work, the value of
 has been taken as 0.25
and window length has been taken as the size of QRS
complex. According to the t -test performed above, first four
IMFs of signal D_00002 contain high frequency noise
components. However, first four IMFs contain high frequency
noise, the QRS complex also present in first few IMFs,
therefore we need to window QRS complex in order to
preserve QRS complex and remove high frequency noise. The windowed QRS component d(t) from first four IMFs has been
shown in figure 3.

Sample NumberAmplitude in µV
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000-400-2000200400Sum of first four IMFs
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000-400-2000200400Windowed QRS Complex

Fig.3. Sum of first four IMFs : c1(t)+c2(t)+c2(t)+c4(t) and windowed QRS
component d(t).
Having removed the high frequency noise from IMFs the
last step in denoising ECG signal is to reconstruct denoised
ECG signal from filtered IMF, d(t) and oth er higher order
IMFs ie. c 5(t) to c 22(t). The filtered signal can be reconstructed
by the formula given below.
1( ) ( ) (11)K
fk
kNECG d t c t


Where, K is the total number of IMF for signal and N is
the noise order obtained from t -test. The original signal
D_00002 and reconstructed denoised signal, and their PSD
curves have been shown in figure 4 and figure 5, respectively.
0 100 200 300 400 500 600 700 800 900 1000-50005001000Original Signal
0 100 200 300 400 500 600 700 800 900 1000-50005001000Denoised SignalSample NumberAmplitude in µV
Fig.4. Original signal D_00002 and denoised D_00002 signal
0 50 100 150 200 250-50050Periodogram Power Spectral Density Estimate
0 50 100 150 200 250-100-50050
Frequency (Hz)Power/Frequency (dB/Hz)(a)
(b)
Fig.5 . PSD Curves (a) original Signal (b) Denoised Signal
The PSD curve given in figure 5 shows that the denoised
signal has all its power concentrated below 1 00Hz unlike the
original signal whose power has been distributed over all
frequencies. This indicates that high frequency noise power
has been removed in the denoised signal.

The signals have been denoised by proposed algorithm and
Power of signals before filtering and after filtering, Percent
Root Difference (PRD) and Signal to Error Ratio (SER) have
been calculated in order to compare it with IIR filtering and
polynomial approximation method.
a) Percent Root Difference (PRD) – It is a measure of
quality of reconstructed ECG signal. Let x[n] and y[n]
be the original and the reconstructed signals,
respectively. Then PRD of the signal may be defined
as.

2
2[ ( ) ( )]100 (12)[ ( )]x t y tPRDxt

b) Signal to Error Ratio (SER) – The Signal to Error
Ratio (SER) has been defined as the ratio of signal
power to the squared error in reconstructed signal. The
SER formula is given as:
2
1
2
1()
(13)
[ ( ) ( )]L
t
L
txt
SER
x t y t




Where, L is the length of the signal
The SER of signals, denoised by different methods have
been given in table 2. It is clear from the table 2 that average
power of the signals has been increased when denoised by
proposed algorithm whereas it decreases when filtered with
Polynomial Approximation and Butterworth IIR filter. The
lower value of PRD and higher values of SER of signals,
denoised with proposed algorithm shows that signal is less
distorted after reconstruction as compared to other two
methods. Also, from morphological study of denoised signal it
is clear that the proposed algorithm gives better results when
compare d to other two methods.
The plots of SER and PRD of signals obtained with
different denoising algorithms have been shown in figure 6 and
figure 7 respectively. From figure 6 it has been cleared that
SER obtained by present algorithm has been observed higher
as compared with other two techniques. The PRD obtained
with present algorithm has been observed lower as compared
with other techniques as shown in figure 7.
Fig.6. SER of Signals Obtained with Different Algorithms

Fig.7. PRD of Signals Obtained with Different Algorithms
B. Removing High Frequency Artifacts from ECG Signals
In the present work five signals from CSE database, five
signals from MIT/BIH database have bee n taken as test
signals. Baseline wander from these test signals has been
removed by present algorithm and Power Spectral Density
(PSD) and Signal power in frequency band (0 -0.5 Hz) has
been calculated. Table 3 shows the average power of original
signals ( Po) and reconstructed signals (P r).
TABLE 3 AVERAGE POWER OF VARIOUS ECG SIGNALS BEFORE
FILTERING AND AFTER FILTERING .
Database Signal
Name Average Power of
Original Signal (P o)
in Frequency Band
(0-0.5Hz) in dB Average Power of
Reconstructed Signal
(Po) in Frequency
Band (0 -0.5Hz) in dB
CSE D_00001 38.43 23.03
CSE D_00002 30.73 21.20
CSE 1L2 28.65 -7.95
CSE 2L2 21.69 -7.75
CSE MA2_022 25.36 1.78
MIT/BIH 100m 59.62 22.68
MIT/BIH 101m 59.59 23.47
MIT/BIH 103m 59.77 25.80
MIT/BIH 105m 59.78 24.75
MIT/BIH 106m 59.89 22.36
TABLE 2: RESULTS OF FILTERING ECG SIGNALS WITH PROPOSED METHOD , POLYNOMIAL APPROXIMATION AND IIR
BUTTERWORTH FILTER .
Original Signal Denoised by Present Algorithm Denoised by Polynomial
Approximation (Polynomial order 9) Denoised by IIR Butterworth filter
(N=2 and f c=100)
Signal Name Average
Power (dB) SER
(dB) PRD
(%) Average
Power (dB) SER
(dB) PRD
(%) Average
Power (dB) SER
(dB) PRD
(%) Average
Power (dB)
D_00001 105.42 19.47 10.63 105.52 9.48 33.56 105.35 15.25 17.27 105.42
D_00002 96.47 19.82 10.20 96.59 8.37 29.11 96.35 12.20 24.54 96.44
1L2 -2.44 36.15 1.56 -2.43 24.49 5.96 -2.44 22.86 2.89 -2.44
101m -20.52 23.50 6.68 -20.43 8.75 36.51 -20.61 14.97 17.83 -20.53
Lab record 1 -57.75 9.99 31.65 -55.10 7.68 41.27 -58.236 9.01 35.40 -57.80
Lab record 2 -43.50 27.35 11.26 -41.58 8.68 36.81 -43.54 10.96 28.29 -43.50

In table 3 average power of different ECG signals in
frequency band (0 -0.5Hz) has been reduced, it shows that
baseline wander has been removed from the reconstructed
signals. Also from morphological study of ECG signals it is
clear that baseline wander has been removed from
reconstructed signal and signal is not distorted unlike the case
of highpass filtering approach where signal distortion is a
major drawback.
In order to examine the performance of present algorithm in
real time situations five ECG signals have been recorded in the
measurement laboratory with intentionally generated baseline
by moving electrode and respiration. The results of experiment
performed have been summarized in table 4.
TABLE 4 AVERAGE POWER OF ECG SIGNALS RECORDED IN
MEASUREMENT LABORATORY BEFORE FILTERING AND AFTER
FILTERING .
Signal
Name Average Power of Original
Signal (P o) in Frequency
Band (0 -0.5Hz) in dB Average Power of
Reconstructed Signal (P o) in
Frequency Band (0 -0.5Hz) in
dB
Lab
Record 1 10.29 -5.45
Lab
Record 2 27.33 -11.63
Lab
Record 3 7.3 -5.57
Lab
Record 4 3.97 -17.00
Lab
Record 5 8.51 -6.43
Table 4 clearly shows that baseline power has been
reduced in reconstructed signal this means the present
algorithm performs better in case of real time situations of
electrode motion and respiration of subject. Original signal
Lab Record 1, reconstructed baseline and reconstructed signal
after removing baseline have been shown in figure 8(a), 8(b)
and 8(c) respectively. The same signal has al so been filtered
with Butterworth IIR filter with cut off Frequency (f c) 0.5 Hz
in order to compare the present algorithm with traditional
method of baseline removal. The ECG signal filtered with
Butterworth IIR filter has been given in figure 9.
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000-2-1012
(b) Reconstructed Baseline
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000-101(c) ECG Signal After Removing Baseline WanderSample NumberAmplitude in mV
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000-2-1012(a) Original ECG Signal
Fig.8.(a) Original signal Lab Record 1, (b) Reconstructed Baseline and (c)
Reconstructed Signal
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000-1012ECG Signal Filtered with Highpass Filter
Fig.9 . ECG Signal Lab Record 1 filtered with Butterworth Highpass Filter The PSD Curve of ECG signal Lab Record 1 has been given
in figure 10. In this figure the distribution of power is reduced
in frequency range of 0 -1Hz it means that lower frequency
components of the signal have been removed.
Fig.10 . PSD Curves of ECG Signal Lab Record 1 (a) Original Signal (b)
Reconstructed Signal.
IV. CONCLUSIONS
In this paper EMD based algorithm for removing artifacts
from ECG signals has been discussed. The test has been
performed over more than 15 ECG signals taken from CSE
database, MIT/BIH database and also on ECG signals
recorded in laboratory with intentionally generated baseline by
electrode motion and respiration in order to examine the
performance of present algorithm in real time situations.
Average power of signal in frequency band 0 -0.5Hz, Pe rcent
Root Difference (PRD) and Signal to Error Ratio (SER) have
been calculated in order to compare the performance of
present algorithm with traditional methods. From quantitative
evaluation and morphological study of original signal and
reconstructed si gnal we can conclude that present algorithm
removes artifacts from ECG signal without introducing
distortions in waveforms.
V. ACKNOWLEDGMENT
This work is supported by University Grant Commission
(UGC), New Delhi, India under Major Research Project (MRP)
scheme. It has been carried out in Electrical Engineering
Department at MMM University of Technology, Gorakhpur,
India.
REFERENCES
[1] Antonio Fasano and Valeria Villani, “ECG Baseline
Wander Removal and Impact on Beat Morphology: A
Comparative Analysis”. Computing in Cardiology, Vol.
40, 2013 , pp. 1167 -1170 .
[2] Hiasat AA, Al -Ibrahim MM, Gharaibeh KM. Design and
implementation of a new effic ient median filtering
algorithm. IEE Proc Vis Image Signal Process, 1999 , pp.
273–278.
[3] Stefan Jurko, Gregor Rozinaj, “High Resolution of the
ECG Signal by Polynomial Approximation”. Radio
Engineering, Vol 15, No. 1, 2006.
[4] Rinky Lakhwani, Shahanaz Ayub, J. P. Saini, “Design
and Comparision of Digital Filters for Removal of
Baseline Wandering from ECG Signal”. Proc. Of IEEE,
5th International Conference on Computational
Intelligence and Communication Networks, 2013 , pp.
186-191.

[5] Ling Zheng, Carolyn Lall, Yu C hen, “Low -Distortion
Baseline Removal Algorithm for Electrocardiogram
Signals”, Co mputing in Cardiology. Vol. 39, 2012 , pp.
769-772.
[6] Piotr Augustyniak, “Continuous Noise Estimation Using
Time -Frequency ECG Representation”, Computing in
Cardiology, Vol. 38, 2011 , pp. 133 -136.
[7] NV Thakor, YS Zhu. “Applications of adaptive filtering
to ECG analysis: Noise cancellation and arrhythmia
detection”. IEEE Trans Biomed Eng, Vol38, No.8, 1991 ,
pp. 785 –794.
[8] Mikhled Alfaouri and Khaled Daqrouq, “ECG Signal
Denoising By Wavelet Transform Thresholding
Mikhled”. American Journal of Applied Sciences Vol. 5
No. 3, 2008 , pp. 276 -281.
[9] Rosaria Silipo and Carlo Marchesi, “Artificial Neural
Networks for Automatic ECG An alysis”. IEEE
Transactions on Signal Processing, Vol. 46, No. 5 , 1998 ,
pp. 1417 -1425 .
[10] Barros, A. Mansour, and N. Ohnishi, “Removing artifacts
from ECG signals using independent components
analysis,” Neurocomputing, 1998 , pp. 173 –186.
[11] Mohammadreza Ravanfar, Leila Azinfar, Riadh Arefin,
Reza Fazel -Rezai, “Electrocardiogram Baseline wander Removal based on Empirical Mode Decomposition”,
Computing in Cardiology. Vol. 41, 2014 , pp.45 -48.
[12] N. E. Hauang, Z. Shen, S.R.Long, M. C. Wu, H. H. Shih,
Q. Zheng, N. C. Yen, C. C. Tung and H. H. Liu, “The
Empirical Mode Decomposition and Hilbert Spectrum for
nonlinear and nonstationary time series analysis”, Proc. R.
Soc. London 454, 1998 , pp. 903 -995.
[13] Manuel Blanco Velasko, Binwei Weng and Kenneth E.
Barner, “ECG signal deno ising and baseline wander
correction based on the empirical mode decomposition”,
Computers in Biology and Medicine, Vol. 38, 2008 , pp.
1-13.
[14] Abdel Ouahab Boudraa, Jean Christophe Cexus," EMD –
based signal filtering", IEEE Transactions on
Instrumentation and Measurement, Vol. 56, No. 6, 2007 ,
pp. 2196 -2202.
[15] Ashish Rohila and V K Giri, "ECG Baseline Wandr
Removal using Empirical Mode Decomposition", Far East
Jouranl of Electronics and Communication Engineering,
Vol. 15, No.2, 2015, pp. 121 -131.