Biomedical Signal Processing and Control 8 (2013) 575–585

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

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

Empirical mode decomposition based filtering techniques for power

line interference reduction in electrocardiogram using various
adaptive structures and subtraction methods
M. Suchetha a,∗ , N. Kumaravel b
a
Velammal Engineering College, Anna University, Chennai, India
b
Anna University, Chennai, India

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 vital sign monitoring measurement of the cardiac activity. One of the main
Received 12 December 2012 problems in biomedical signals like electrocardiogram is the separation of the desired signal from noises
Received in revised form 4 April 2013 caused by power line interference, muscle artifacts, baseline wandering and electrode artifacts. Different
Accepted 3 May 2013
types of digital filters are used to separate signal components from unwanted frequency ranges. Adaptive
Available online 28 June 2013
filter is one of the primary methods to filter, because it does not need the signal statistic characteristics. In
contrast with Fourier analysis and wavelet methods, a new technique called EMD, a fully data-driven tech-
Keywords:
nique is used. It is an adaptive method well suited to analyze biomedical signals. This paper foregrounds
ECG denoising
Power line interference
an empirical mode decomposition based two-weight adaptive filter structure to eliminate the power line
Noise cancelation interference in ECG signals. This paper proposes four possible methods and each have less computational
Adaptive filter complexity compared to other methods. These methods of filtering are fully a signal-dependent approach
with adaptive nature, and hence it is best suited for denoising applications. Compared to other proposed
methods, EMD based direct subtraction method gives better SNR irrespective of the level of noises.
© 2013 Elsevier Ltd. All rights reserved.

1. Introduction artifacts, so as to present an ECG that facilitates easy and accurate

The electrocardiogram (ECG) is a graphical representation of
the functionality of human heart and is an important tool used for
diagnosis of cardiac abnormalities. In clinical environment during
acquisition, the ECG signal gets added to various types of artifacts.
The predominant artifacts present in the ECG include: Power-line
Interference and Baseline Wander. This artifact strongly affects
the ST segment, degrades the signal quality, frequency resolution,
and produces large amplitude signals in ECG that can resemble
P–Q–R–S–T waveforms. Consequently, these factors mask small
features that are important for clinical monitoring and diagno-
sis of heart failure. Annulment of these artifacts in ECG signals is
an important factor for better diagnosis. The extraction of high-
resolution ECG signals from recordings which are perturbed by
noise is an important issue. The goal of ECG signal denoising is
to separate the wanted signal components from the unwanted
∗ Corresponding author at: Department of Electronics and Communication,
Velammal Engineering College, Anna University, Chennai 600066, India.
Tel.: +91 9443396450.
E-mail addresses: suchethashivani@gmail.com, suchetha mt@yahoo.com
(M. Suchetha).
interpretation.
Traditional noise reduction method is based on standard filter-
ing, either by low pass filter or high pass filter. A low pass filter
removes high frequency noise, while a high pass filter removes
low frequencies, such as baseline wander and respiration interfer-
ence. However, these ideal filters will remove the desired signal
component within the band of noise since both signal and noise
display an overlapping spectrum. Another conventional method
to remove power line interference is to use a notch filter, which
is tuned to the frequency of interference. The problem with finite
impulse response (FIR) notch filter is that the notch has a relatively
large bandwidth, which attenuates the needed signal components
within the bandwidth [1,2].
The advanced signal processing methods applied on the studies
of ECG noise reduction is nonlinear filter bank [3], indepen-
dent component analysis [4], adaptive filtering [5,6] and wavelet
transforms [7]. Adaptive filtering is a powerful signal detection
technique. The least mean square (LMS) algorithm introduced by
Widrow and Hoff (1959) is one of the most widely used algorithms
in adaptive filtering due to its simplicity and ease of computa-
tion [8]. The concept of wavelet transform was proposed by the
geophysicist Morlet in analysis and processing geophysical data
in France in 1984. An advantage of wavelet transform is that the
windows vary, and it has an infinite set of possible basis functions.

1746-8094/$ – see front matter © 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.bspc.2013.05.001
576 M. Suchetha, N. Kumaravel / Biomedical Signal Processing and Control 8 (2013) 575–585
Donoho and Johnstone (1994) have developed hard and soft shrink-
age functions [9,10]. In the case of soft shrinkage the insignificant
empirical wavelet coefficients are set to zero if it lies between the
ranges of thresholds and shrink the coefficients to zero in a lin-
ear fashion. In case of hard shrinkage the insignificant empirical
wavelet coefficients are vanished if it lies between the thresh-
old ranges and the remaining coefficients are used to represent
the signal. Poornachandra and Kumaravel (2005) developed a sub-
band adaptive shrinkage function for denoising of ECG signals
[11,12]. The shrinkage function is incorporated at the vicinity of
power line frequency by selecting the proper subband level. But
the concept of wavelet thresholding relies on the assumption that
signal magnitudes dominate the magnitudes of the noise in a
wavelet representation so that wavelet coefficients can be set to
zero if their magnitudes are less than a predetermined thresh-
old [12]. Another limitation of wavelet approach is that the basis
functions are fixed and, thus, do not necessarily match all real
signals.
Empirical mode decomposition (EMD) is a recently introduced
technique and it is also used for processing non-linear and non-
stationary signals in addition to stationary signals [13]. EMD has
the property of adaptive and signal-dependency, making this tech-
nique well suited for biomedical signal analysis. It is proposed as a
preprocessing stage to efficiently compute the instantaneous fre-
quency through the Hilbert transforms [14,15], however, it can be
applied independently also. It is discussed in the literature that
EMD behaves as a “wavelet-like” dyadic filter bank for Gaussian
noise [16,17]. It is an iterative algorithm that computes the maxi-
mum and minimum extreme. The major advantage of the EMD is
that the basis functions are derived from the signal itself. Thus the
analysis is adaptive in contrast to the traditional methods where
the basis functions are fixed. The EMD is based on the sequential
extraction of energy associated with various intrinsic time scales
of the signal, starting from finer temporal scales (high-frequency
modes) to coarser ones (low-frequency modes). Thus according
to this decomposition any signal can be represented as the sum
of intrinsic mode functions (IMFs) and a residue [18,19]. An IMF
is defined as a function with equal number of extreme and zero
crossings with its envelopes, as defined by all the local maxima and
minima. The filtering of noises in arrhythmia signals using ensem-
ble empirical mode decomposition algorithm is proposed in the
paper [20]. The principle of the EEMD is to add white noise into the
signal with many trials. The noise in each trial is different and the
added noise can be canceled out on average. Thus, as more and more
trials are added to the ensemble, the residual part is the signal. But
these types of filtering further disturb the morphology of ECG sig-
nal due to different trials in addition of white noise. A robust power
line suppression system based on an extended version of kalman
filter and an improved version of EMD is used to attenuate the QRS
complex of ECG signal [21].
This work discusses various possible EMD based adaptive fil-
tering techniques and subtraction techniques for denoising the
contaminated ECG signal. The decomposed IMF levels will sepa-
rate the frequency components present in the noisy ECG signal.
The 50 Hz complex sinusoidal interference is separated in the first
IMF level, so a two-weight adaptive filter is used in the first level to
remove the interference. The different possible ways of using the
adaptive filter and subtraction techniques are discussed in detail in
Section 3. This approach does not need the thresholding of signal
levels like wavelet thresholding. It is also not necessary that the
signal amplitude be higher than the noise amplitude. This makes
the techniques applicable to any kind of noisy nature of the signal.
The present approach of filtering is fully a data-dependent and an
adaptive method.
This paper is organized as follows: Section 2 discusses basics
of empirical mode decomposition and adaptive filtering; Section 3
studies the theme of the paper and the proposed model; and Section
4 examines the results of the MATLAB simulation.
2. Introduction to empirical mode decomposition and
adaptive filtering
2.1. Empirical mode decomposition
The EMD was recently proposed by Huang et al. [13] as a tool to
adaptively decompose a signal into a collection of AM–FM compo-
nents. Other data analysis methods, like Fourier and wavelet-based
methods require pre-defined basis functions to represent a signal.
The EMD is fully a data based mechanism that does not require
any a priori known basis. It is especially well suited for analysis of
biomedical signals. The goal of empirical mode decomposition is
to represent a signal as an expansion of adaptively defined basis
functions with well-defined frequency localization levels. It will
decompose the input signal into a sum of IMFs. To decompose a
signal x(t), the EMD algorithm works as follows:
(1) Identify all maxima and minima of the signal x(t).
(2) Connect all the local maxima of the signal x(t) by a cubic spline
to form the upper envelope, represented as eu (t). Similarly, con-
nect all the local minima to form the lower envelope, el (t).
(3) Compute the mean, m1 (t) of the two envelopes.
m1 (t) = [eu (t) + el (t)]/2.
(4) Compute the detail, h(t), by subtracting the mean from the sig-
nal, h(t) = x(t) − m1 (t).
(5) Repeat the iteration on the residual m1 (t), steps (3) and (4).
Continue until the residual is such that no IMF can be extracted
and represents a monotonic function.
The above procedure to extract the IMF is referred to as the
sifting process. Finally, the EMD of the original signal can be repre-
sented as the sum of IMFs and a residue
N
x(t) = ci (t) + r(t) (1)
i=1
where ci (t) is the sum of IMF levels and r(t) is the resultant residue.
An IMF is a function that satisfies the two following conditions:
(a) the number of extreme and the number of zero crossings must
either equal or differ at most by one, and (b) the mean value of the
envelope defined by the local maxima and the envelope defined by
the local minima is zero at every point. The detailed procedure for
empirical mode decomposition is shown in Fig. 1.
2.2. Two-weight adaptive filter
An adaptive filter is the natural choice for removal of power
line interference which can adjust its coefficients according to cer-
tain algorithm. Widrow and Hoff et al. (1959) introduced the least
mean square (LMS) algorithm and are widely used in adaptive fil-
tering algorithms. It is more powerful and simple in computation.
The main aim of an adaptive filter is to calculate the difference
between the desired signal and the adaptive filter output, e(n). This
error signal is feedback into the adaptive filter and its coefficients
are changed algorithmically in order to minimize a function of this
difference, known as the cost function [8]. When the adaptive filter
output is equal to the desired signal, the error signal goes to zero.
Thus with each iteration of the LMS algorithm, the filter tap weights
of the adaptive filter are updated according to the weight update
equation:
w(n + 1) = w(n) + e(n)x(n) (2)
 M. Suchetha, N. Kumaravel / Biomedical Signal Processing and Control 8 (2013) 575–585 577

Fig. 1. Flow chart for the EMD procedure.

x(n) = [x(n)x(n − 1)x(n − 2) . . . x(n − N + 1)]T denotes the input signal vector

w(n) = [w0 (n)w1 (n)w2 (n) . . . wN−1 (n)]T represents the coefficients of the adaptive FIR filter tap weight vector at time n

The parameter  is known as the step size parameter. Selection

of a suitable value for  is important for the performance of the
LMS algorithm, if the value is too small the time the adaptive fil- This study employs the adaptive noise canceller with two
ter takes to converge on the optimal solution will be too long; if adaptive weights. The adaptive filter has two adjustable weights,
 is too large the adaptive filter becomes unstable and its output w1 (n + 1) and w2 (n + 1). One weight, w1 (n + 1), is fed directly from
diverges. the reference point. The second weight, w2 (n + 1), is fed from a 90◦ -
phase-shifted version of the reference input. The sum of the two
weighted versions of the reference signal is then subtracted from
2.3. Implementation of the LMS algorithm
the ECG output to produce an error signal. This error signal together
with the weighted inputs is applied to the LMS algorithm, which
The each iteration of the LMS algorithm requires three distinct
controls the adjustments applied to the two weights. In this case,
steps in this order:
the adaptive noise canceller acts as a variable “notch filter”. The
structure of two-weight adaptive filter is as shown in Fig. 2.
1. The output of the FIR filter, y(n) is calculated
N−1
y1(n) = i=0
w1i (n)x(n − i) = wT1 (n)x(n), where N is the filter lengh,
N−1 (3)
y2(n) = i=0
w2i (n)x90 (n − i) = wT2 (n)x90 (n)
The frequency of the sinusoidal interference in the ECG output
is the same as that of the sinusoidal reference signal. However, the
For the two-weight adaptive filter output,
amplitude and phase of the sinusoidal interference in the ECG out-
y(n) = y1(n) + y2(n) (4) put are unknown. The two weights w1 (n + 1) and w2 (n + 1) provide
the two degrees of freedom required to control the amplitude and
phase of the sinusoidal reference signal so as to cancel the 50 Hz
2. The value of the error estimation is calculated
interference contained in the ECG output.
e(n) = d(n) − y(n) (5)
3. Power line interference reduction using EMD based
where d(n) is the primary signal. filtering
3. The tap weights of the FIR filter are updated in preparation for
the next iteration 3.1. Rationale

The objective of this paper is the reduction of power line

w1 (n + 1) = w1 (n) + x(n)e(n)
frequency noise hidden in ECG signals. In the presence of such
interferences, the recovered signal, d(kTs )is expressed as
w2 (n + 1) = w2 (n) + x90 (n)e(n)
d(kTS ) = x(kTS ) + ˛i cos(ωi kTS + ϕi ) (6)
578 M. Suchetha, N. Kumaravel / Biomedical Signal Processing and Control 8 (2013) 575–585
Fig. 2. Structure of two-weight adaptive filter.
where x(kTs ) is the ECG signal, Ts is the sampling period, ωi , i = 1,
2, . . ., M are the known radian frequencies while ˛i and ϕi rep-
resent the unknown amplitudes and phases, respectively, of the
interfering sinusoids. It is assumed that ˛i > 0 and ϕi ∈ [0, 2] for all
i.
The main aim is to extract x(kTs ) from the corrupted signal d(kTs )
without any distortion. The recovered signal ˆ x(kTs ) is of the form
∧ below.
x(kTS ) = d(kTS ) − ∧ ˛i cos(ωi kTS + ∧ ϕi )
where ˆ˛i , ˆϕi , and ˆx(kTs ) are the estimates of ˛i , ϕi and x(kTs ),
respectively.
The original ECG signal corrupted by power line interference is
decomposed into a series of IMFs which are extracted from local
high frequency to local low frequency. Fig. 3 shows the EMD for
original ECG signal with all the IMFs from low to high orders. From
the plot it is interpreted that the frequency content of each indi-
vidual IMF decreases as the order of IMF increases. Note that the
oscillatory patterns observed in the first three IMFs are mainly due
to the QRS complex, which has strong high-frequency components.
EMD can act as a dyadic filter bank with adaptive bandwidth which
can separate the signal components in the temporal domain [20].
Thus the low level IMF acts as high pass filter, high level has the
effect of low pass filtering and the mid-band IMF levels acts as band
pass filter.
The proposed approach utilizes the two-weight adaptive fil-
tering in various possible ways by combining with EMD–IMFs for
the elimination of power line interference. These approaches give
smooth reconstructed signal and better visual presentation.
3.2. Proposed techniques
3.2.1. Empirical mode decomposition based adaptive filtering
without reference (Technique A)
As we can see from Fig. 3, the frequency content of each individ-
ual IMF decreases as the order of IMF increases. The spectral plot
of each IMF level in Fig. 4 shows the presence of frequency com-
ponent present in the signal. The EMD is based on the sequential
extraction of energy associated with various intrinsic time scales
of the signal, starting from finer temporal scales. Thus irrespective
of the noise level, the high frequency components get separated
in the first IMFs. The noise encountered in ECG applications is
usually located in the high-frequency band. Although most ECG sig-
nal power is concentrated in lower frequencies, the QRS complex
spreads across the mid- to high-frequency bands [22,23]. The EMD
is applied on the noisy signal and various levels of intrinsic mode
functions are obtained. The 50 Hz power line interference gets sep-
arated in the first IMF level (IMF 1) and the remaining levels are
free from the interference component. So a two-weighed adaptive
filter is applied in IMF level 1, thereby it controls the amplitude and
phase of the signal. The adaptive filtering architecture is as shown
in Fig. 2. Primary signal is taken as a noisy ECG signal. IMF 1 signal
is applied as the reference signal and the adaptation is performed.
The detailed steps involved in the filtering process are explained
(7)
1. Apply EMD on the noisy ECG signal and the intrinsic mode func-
tions are obtained.
2. Frequency spectrum for all levels of IMFs is plotted and it shows
the frequency separation at various levels.
3. The 50 Hz interference components get separated in the first IMF
level. So adaptive filtering is done in the first IMF level.
4. A two-weight adaptive filter is designed, with contaminated ECG
signal, d(n) = x(n) + v(n) as primary signal and reference signal
as first level IMF. Where x(n) is the ECG signal and v(n) is the
sinusoidal interference.
5. A 90◦ phase shifted version of IMF 1 is applied as another input
to the LMS block and the weight update is carried out.
6. The error signal, e(n) is estimated as the difference between the
primary signal and the output signal.
7. Estimated signal is again fed back as input and the iteration pro-
cess is repeated.
The adaptive filter essentially minimizes the mean-squared
error between a primary input and the reference input [24]. The
reconstructed signal shows the clean ECG signal. Compared to
adaptive filter with reference method, in this method the part of
the noisy signal is given as reference and subsequently the adap-
tation is carried out. It has the advantage that no source signal is
needed for adaptation.
3.2.2. Empirical mode decomposition based adaptive filtering
using subtraction techniques
3.2.2.1. Direct subtraction method (Technique B). The main draw-
back in conventional method of adaptive filtering is that one cannot
perform the subtraction of noise from the contaminated signal
directly as both the primary signal and reference signal are of out
of phase [25]. In the proposed method IMF level 1(IMF 1) is taken
as the reference signal and noisy ECG signal as primary signal. The
IMF 1 can be directly subtracted from the primary signal as both the
 M. Suchetha, N. Kumaravel / Biomedical Signal Processing and Control 8 (2013) 575–585 579

IMF 1

Amplitude(mV)
1
0
-1
0 100 200 300 400 500 600 700 800 900 1000
Samples
IMF 2
Amplitude(mV)

5
0
-5
0 100 200 300 400 500 600 700 800 900 1000
Samples
IMF 3
Amplitude(mV)

5
0
-5
0 100 200 300 400 500 600 700 800 900 1000
Samples
IMF 4
Amplitude(mV)

0
-5
0 100 200 300 400 500 600 700 800 900 1000
Samples
IMF 5
Amplitude(mV)

0.5

0
-0.5
0 100 200 300 400 500 600 700 800 900 1000
Samples

IMF 6
9
1
Amplitude(mV)

0
-1
0 100 200 300 400 500 600 700 800 900 1000
Samples
IMF 7
2
Amplitude(mV)

0
-2
0 100 200 300 400 500 600 700 800 900 1000
Samples
IMF 9
2
Amplitude(mV)

0
-2
0 100 200 300 400 500 600 700 800 900 1000
Samples
IMF 10
1
Amplitude(mV)

0
-1
0 100 200 300 400 500 600 700 800 900 1000
Samples
residue
2
Amplitude(mV)

0
-2
0 100 200 300 400 500 600 700 800 900 1000
Samples

Fig. 3. EMD of a noisy ECG and its IMF 1-11.

reference and the primary signal have the same phase. This method
of subtraction is simple and has fewer computations.
3.2.2.2. Indirect subtraction method (Technique C). In this proposed
approach, the decomposed IMF level 1 of EMD is first band pass
filtered and the 50 Hz interference band is separated. The pass band
frequency range for the band pass filter is chosen as 48–51 Hz. The
resultant filtered signal is subtracted directly from the noisy ECG
signal. The filter structure used in indirect subtraction method is
shown in Fig. 5.
The reconstructed signal is obtained without destroying the
neighboring signal components. Therefore there is no possibility of
noise component overlapping with harmonic components of ECG
signal.
3.2.3. Empirical mode decomposition based adaptive filtering by
extracting the 50 Hz interference band using band pass filter
(Technique D)
The decomposed intrinsic mode functions are mainly the
AM–FM components of the noisy signal. So the IMF 1 will be the
amplitude modulated version of the contaminated signal. To avoid
the destruction of ECG signal, the frequency range from 48 Hz to
51 Hz is band pass filtered. Thus the range of BPF is considered as
the range of unwanted interference in the original signal. The adap-
tive filter structure is modified as shown in Fig. 6. IMF 1 is taken as
the reference signal and it is passed to a band pass filter. The fil-
tered band of signal is applied as input to the adaptive structure.
The phase shifted version of the filtered signal is given as another
input to the adaptive filter. Adaptation is performed by the feedback
580 M. Suchetha, N. Kumaravel / Biomedical Signal Processing and Control 8 (2013) 575–585

Fig. 4. Spectrum for IMF 1-4.

Fig. 5. Method for indirect subtraction.

of the estimated error signal, e(n). By properly adjusting the filter
coefficients, the reference waveform will be changed in magnitude
and phase to model the IMF 1.
The estimated source signal is given by subtraction of the filtered
output from the corrupted waveform. In this new adaptive interfer-
ence cancelation system, which is characterized by the amplitude
and phase parameters, the frequencies of the interfering signal are
IMF 1. The main advantage of this technique is that no other ECG
components get removed rather than only the 50 Hz interference
part.
4. Simulations and results discussion
The current study focuses on implementing the proposed tech-
niques on ECG waveforms with a sampling rate of 256 samples per
second. The simulations are carried out in MATLAB environment.
In the simulation, synthetic ECG signals are generated by using
the software available on physionet with a sampling frequency of
256 Hz. The ECG signal is added with 50 Hz frequency sinusoidal
noise r(n) = s(n) + v(n), where v(n) is the sinusoidal interference,
v(n) = Asin(2 × pi × 50 × t) for the simulation purpose as illustrated

Fig. 6. EMD based adaptive filter structure using peak filter.

 M. Suchetha, N. Kumaravel / Biomedical Signal Processing and Control 8 (2013) 575–585 581

in Fig. 7(b). The spectral output of noise embedded ECG shows a
spike of 50 Hz interference. Two types of input samples are consid-
ered for various studies. They are: simulated ECG with sinusoidal
interference; ECG signal from MIT-BIH arrhythmia database. The
practical ECG with 1000 samples is downloaded from the Phy-
sioBank of MIT-BIH arrhythmia database whose sampling rate is
360 Hz. The study is conducted on 8 patients ECG signals in normal
and 12 ECG signals in abnormal conditions. This database contains
48 records, each containing two-channel ECG signals for 30 min
duration selected from 24-h recordings of 47 individuals. There are
116,137 numbers of QRS complexes in the database. Usually for the
adaptive filters reference signal is taken from the external source
but in the proposed methods the reference signal in the form of IMF
1 is derived from noisy signal and it is given as reference to adap-
tive filter. Since the noise is taken from the primary signal, the exact
noise present in the noisy ECG signal can be obtained. EMD based
methods performs well on both simulated and actual noisy data.
The denoised waveforms obtained on simulated signal are shown
in Fig. 7.
The time and spectral differences between the original and the
filtered ECGs confirm that this method of filtering does not affect
the signal and spectral components of ECG as shown in Fig. 11.
The interference frequency is selected from the range 48–51 Hz and
added to the pure ECG, and the performance of the EMD based noise

Fig. 7. (a) Original ECG signal and its spectrum, (b) power line interference embedded ECG and its spectral response, (c) recovered ECG signal using notch filter and its
spectral response, (d) recovered ECG signal using wavelet transform and its spectral response, (e) recovered ECG signal using EMD based adaptive filtering (Technique A)
and its spectral response, (f) recovered ECG signal using EMD direct subtraction (Technique B) and its spectral response, (g) recovered ECG signal using EMD based indirect
subtraction (Technique C) and its spectral response, (h) recovered ECG signal using EMD based adaptive filtering using band pass filter(Technique D) and its spectral response.
582 M. Suchetha, N. Kumaravel / Biomedical Signal Processing and Control 8 (2013) 575–585

Fig. 7. (Continued ).

cancelation methods are studied. The comparisons are made at 5%, SNR values are high for all the signals irrespective of the abnor-
10%, 15%, 25%, 30% noise level. The IIR notch filter is designed with malities. The SNR and MSE are defined as
notch frequency as 50 Hz .The main drawback of using notch filter is  x[n]

that it would also filter the ECG signal component present at 50 Hz. SNR(dB) = 20 log (8)
x[n] − d(n)
Also the notch should be too narrow for the 50 Hz rejection.
Adaptive filters can adjust to unknown environment, and even
track signal or system characteristics varying over time. Figs. 8 and 9 1 N 2
MSE = (x[n] − x̂[n]) (9)
show the SNR versus noise level in percentage for the proposed N n=1
methods using practical ECG signals from MIT-BIH database and
the simulated ECG signal. The practical ECG signals are downloaded

N 2
from the database records with patients having left bundle branch n=1
(x[n] − x̂[n])
PRD =  N
× 100 (10)
block (109), right bundle branch block (118), atrial premature beat x2 [n]
i=1
(119), premature ventricular contraction and normal (209). The
 M. Suchetha, N. Kumaravel / Biomedical Signal Processing and Control 8 (2013) 575–585 583

SNR vs Noise level plot record 103 MSE plot Tech. A

56 1
record 109 Tech. B
record 118 0.9 Tech. C
54 record 119 Tech. D
record 209 0.8 wavelet
52 Notch
0.7
50
SNR

0.6

MSE
48 0.5
0.4
46
0.3
44
0.2

42 0.1
5 10 15 20 25 30
Noise level (%) 0
5 10 15 20 25 30
Noise level (%)
Fig. 8. Plot of SNR against %noise level using Technique C for practical ECG records
from MIT-BIH database.
Fig. 10. Plot of MSE against %noise level.

SNR vs Noise Level Tech. A

65 Tech. B
60 Tech. C From Fig. 7 it can be observed that EMD based techniques
Tech. D
55 wavelet are fully capable of suppressing the power line interference from
Notch
50 ECG. The recovered signal is uncontaminated and flawless. The fre-
SNR (dB)

45 quency spectrum shown in Fig. 7(b) depicts the existence of power

40
35
30
25
20
15
5 10 15 20 25
Noise level (%)
Fig. 9. Plot of SNR against %noise level.
where x(n) is the original signal, x̂(n) is the recovered signal and
d(n) is the noisy ECG signal.
For better denoising perfromance, signal to noise ratio (SNR)
should be high and mean square error (MSE), percentage root mean
square error difference (PRD) should be low.
All the proposed techniques are compared with wavelet trans-
form method of denoising. Fig. 9 shows that in the case of wavelet
transform, the SNR values are not improving much depending
upon the different noise levels. It is known from the literature that
symlets family of wavelets give details more accurately than that
of other wavelets like coiflet, haar, Daubechies, Biorthogonal and
Meyer. Also, these wavelets show similarity with the QRS com-
plexes and as in the case of ECG signal, their energy spectra are
concentrated around low frequencies. Therefore, in this paper, sym-
let 4 is selected as the mother wavelet while performing wavelet
analysis. Two-level wavelet decomposition with symlet 4 bases
function is exploited for the implementation of the thresholding
in the DWT domain.
line interference. Its removal is evidenced in Fig. 7(c)–(g) when
using the proposed methods.
The results obtained indicate that the proposed techniques can
eradicate the noise component distributed around the 50 Hz fre-
quency without disturbing the signal component. On interpreting
30 the parameters like SNR and RMS error it is apparent that EMD
based indirect subtraction (Technique C) gives a better perfor-
mance irrespective of all noise levels. It also maintains the stability
in all noise levels compared to the other three proposed meth-
ods. The RMS error is almost nearing zero as shown in Fig. 10. In
Tables 1 and 2 tabulates different parameters values obtained at
different noise levels for the four methods.
Direct subtraction also maintains a similar performance to indi-
rect subtraction but as the noise level increases there is a slight
variation in RMS error. SNR values of all the denoising methods
are consistent for noise level less than 10% and it decreases gradu-
ally as noise increases. Visual representation of proposed methods
exhibit a superior presentation and gives better performance. The
spectrum of recovered signal shows a clear elimination of 50 Hz
interference. The properties of signal dependency and adaptive
nature comply well with denoising the power line interference,
thus making it potentially suitable for real time implementation.
As in the case in wavelet based methods, the EMD methods do
not have any fixed basis functions and it is derived from the sig-
nal itself. Adaptive filters are often realized as a set of program
instructions running on an arithmetical processing device such as a
microprocessor or DSP chips. The complexity in implementing the
conventional adaptive filter was very high as the minimum MSE
of biological signals was achieved with several iterations, which

Fig. 11. (a) Plot of spectral subtraction between original and reconstructed signal using Technique A and (b) plot of spectral subtraction between IMF 1 and reconstructed
signal using Technique C.
584 M. Suchetha, N. Kumaravel / Biomedical Signal Processing and Control 8 (2013) 575–585

Table 1
SNR and MSE comparison of proposed techniques for different noise levels.

Noise SNR (dB) MSE

level (%)
Notch Wavelet Technique A Technique B Technique C Technique D Notch Wavelet Technique A Technique B Technique C Technique D

5 32.98 20.68 44.72 50.97 60.31 34.496 0.03 0.64 0.0013 0.0012 4.13E+05 0.0609
10 30.45 20.83 44.01 49.69 58.74 34.008 0.05 0.64 0.0018 0.0017 2.08E+05 0.0643
15 28.05 20.55 41.58 48.87 57.04 32.684 0.1 0.65 0.0127 0.0017 7E−05 0.104
20 25.52 20.81 41.29 46.18 55.28 31.681 0.15 0.611 0.0153 0.0041 9.4E−05 0.1102
25 23.12 20.31 39.1 45.63 54.17 30.996 0.22 0.7 0.025 0.0045 0.00016 0.1385
30 22.35 19.47 38.51 43.4 53.55 30.513 0.33 0.915 0.0301 0.0052 0.00018 0.154

Table 2
RMSE and PRD comparison of proposed techniques for different noise levels.

Noise RMSE PRD (%)

level (%)
Notch Wavelet Technique A Technique B Technique C Technique D Notch Wavelet Technique A Technique B Technique C Technique D

5 0.1635 0.8 0.037 0.0344 0.0064 0.247 0.03 0.033 0.048 0.042 0.0000253 0.042
10 0.2296 0.8 0.043 0.0409 0.0046 0.254 0.03 0.033 0.046 0.04 0.0000296 0.04
15 0.3082 0.8 0.011 0.0412 0.0084 0.323 0.04 0.031 0.057 0.054 0.0000339 0.054
20 0.3926 0.751 0.123 0.0639 0.0097 0.332 0.03 0.029 0.097 0.075 0.0000875 0.075
25 0.4827 0.83 0.159 0.0634 0.0127 0.372 0.04 0.027 0.079 0.082 0.0000910 0.082
30 0.5725 0.95 0.173 0.0719 0.0134 0.393 0.04 0.035 0.1303 0.097 0.0001580 0.097

resulted in delay in denoising process. This was the main constraint
of real-time implementation. EMD based method simply remove
the components, which are not contributed in the signal charac-
teristics. The real-time implementations of these techniques are
simpler and are unlikely to introduce delay problems in the denois-
ing process. Real-time EMD can be possible to implement by using
a sliding window. This is done by block wise extraction of ECG using
sliding windows [26].
5. Conclusion
This paper has proposed novel techniques based on various EMD
based adaptive filter structures. The results shows that EMD based
noise cancelation methods are found to perform relatively bet-
ter than some of the state-of-art algorithms like notch filtering,
wavelet transforms, and this algorithm can be invoked as and when
power line interference is sensed on the iso-electric potential line.
The proposed power line frequency elimination techniques have
potential application in data acquisition systems, which are gener-
ally encountered by power line frequency noise problems. From the
analysis presented in this paper, it can be concluded that the EMD
based noise annulment method is adaptive to power line interfer-
ence in ECGs and can offer SNR improved performance. Out of the
four techniques proposed EMD based direct subtraction method
gives better results irrespective of the noise levels and also it is
computationally efficient.
References
[1] P. Strobach, Single section least square adaptive notch filter, IEEE Transactions
on Signal Processing 43 (8) (1995) 2007–2010.
[2] M. Ferdjallah, R.E. Barr, Adaptive digital notch filter design on the unit circle
for the removal of power line noise from biomedical signals, IEEE Transactions
on Biomedical Engineering 41 (6) (1994) 529–536.
[3] J.M. Leski, N. Henzel, ECG baseline wander and power line interference reduc-
tion using nonlinear filter bank, Signal Process 35 (4) (2004) 781–793.
[4] A.K. Barros, A. Mansour, N. Ohnishi, Removing artifacts from electrocardio-
graphic signals using independent components analysis, Neurocomputing 22
(1998) 173–186.
[5] P.S. Hamilton, A comparison of adaptive and non adaptive filters for reduc-
tion of power line interference in the ECG, IEEE Transactions on Biomedical
Engineering 43 (1) (1996) 105–109.
[6] A.K. Ziarani, A. Konrad, A nonlinear adaptive method of elimination of power
line interference in ECG signals, IEEE Transactions on Biomedical Engineering
49 (6) (2002) 540–547.
[7] D.L. Donoho, I.M. Johnstone, Adapt to unknown smoothness via wavelet
shrinkage, Journal of the American Statistical Association 90 (1995)
1200–1224.
[8] S. Haykin, Adaptive Filter Theory, third ed., Prentice-Hall Inc., Englewood Cliffs,
NJ, 1996.
[9] S. Poornachandra, Wavelet-based denoising using subband dependent
threshold for ECG signals, Journal of Digital Signal Processing 18 (2008)
49–55.
[10] D.L. Donoho, De-noising by soft thresholding, IEEE Transactions on Information
Theory 41 (3) (1995) 613–627.
[11] S. Poornachandra, N. Kumaravel, A novel method for the elimination of power
line frequency in ECG signal using hyper shrinkage function, Journal of Digital
Signal Processing 18 (2008) 116–126.
[12] S. Poornachandra, N. Kumaravel, Hyper-trim shrinkage for denoising of ECG
signal, Digital Signal Process 15 (3) (2005) 317–327.
[13] N.E. Huang, et al., The empirical mode composition and the Hilbert spectrum
for nonlinear and non-stationary time series analysis, Proceedings of the Royal
Society London A 454 (1998) 903–995.
[14] Naveed ur Rehman, D.P. Mandic, Filter bank property of multivariate empiri-
cal mode decomposition, IEEE Transactions on Signal Processing 59 (5) (2011)
2421–2426.
[15] A.-O. Boudraa, J.-C. Cexus, EMD-based signal filtering, IEEE Transac-
tions on Instrumentation and Measurement 56 (December (6)) (2007)
2196–2202.
[16] Z. Wu, N.E. Huang, A study of the characteristics of white noise using the empir-
ical mode decomposition method, Proceedings of the Royal Society London A
460 (June (2046)) (2004) 1597–1611.
[17] J.C. Echevarría, J.A. Crowe, M.S. Woolfson, B.R. Hayes-Gill, Application of
empirical mode decomposition to heart rate variability analysis, Medical and
Biological Engineering and Computing 39 (4) (2001) 471–479.
[18] V. Nitish, Thakor, Y.-S. Zhu, Applications of adaptive filtering to ECG analysis:
noise cancellation and arrhythmia detection, IEEE Transactions on Biomedical
Engineering 38 (August (8)) (1991) 785–794.
[19] P. Flandrin, G. Rilling, P. Goncalves, Empirical mode decomposition as a filter
bank, IEEE Signal Processing Letters 11 (2) (2004) 112–114.
[20] K.-M. Chang, Arrhythmia ECG noise reduction by ensemble empir-
ical mode decomposition, Journal of Sensors 10 (2010) 6063–6080,
http://dx.doi.org/10.3390/s100606063.
[21] G. Li, X. Zeng, X. Zhou, et al., Robust suppression of nonstationary power-
line interference in electrocardiogram signals, Physiological Measurement
(England) 33 (July (7)) (2012) 1151–1169.
[22] M. Blanco-Velasco, B.W. Weng, ECG signal denoising and baseline wander cor-
rection based on the empirical mode decomposition, Journal of Computers in
Biology and Medicine 38 (2008) 1–13.
[23] H.L. Liang, et al., Artifact reduction in electro gastrogram based on the empir-
ical model decomposition method, Medical and Biological Engineering and
Computing 38 (1) (2000) 35–41.
[24] W. Jie-Sheng, Z. Yong, Z. Ping, Research on denoising algorithm for ECG sig-
nals, in: Proceedings of the 29th Chinese Control Conference, China, July 29–31,
2010.
[25] C. Levkov, G. Mihov, Removal of power-line interference from the ECG: a
review of the subtraction procedure, Biomedical Engineering Online 4 (2005)
50.
[26] G. Rilling, P. Flandrin, P. Goņcalv‘es, On empirical mode decomposition and
its algorithms, in: Proceeding of Nonlinear Signal and Image Processing,
2003.
 M. Suchetha, N. Kumaravel / Biomedical Signal Processing and Control 8 (2013) 575–585
M. Suchetha received the B.E. degree from Noorul Islam
college of Engineering, Manonmaniam Sundaranar Uni-
versity in the year 1999, the master degree from Mepco
Schlenk Engineering College, Madurai Kamaraj Univer-
sity in the year 2001, and she is currently pursuing
doctoral program in Anna University, Chennai in bio-
signal processing. She is working as Assistant Professor in
Department of Electronics and Communication Engineer-
ing, Velammal Engineering College, Chennai from the year
2007 onwards. She has presented a few papers in inter-
national journals and conferences. Her areas of interest
are bio-signal processing using wavelet transform, EMD,
adaptive filter, and analysis of various filtering methods.
She is a life member of technical bodies like ISTE.
585
N. Kumaravel received the Ph.D. degree from Anna Uni-
versity in bio-signal processing. Currently, he is working
as Professor, Department of Electronics and Communica-
tion Engineering, Anna University, Chennai, India. He has
published several papers in various national and inter-
national journals and conferences. His areas of interest
are noise cancelation techniques using genetic algorithm,
neural network, wavelet-based adaptive filtering, and
medical image processing. He is a member of technical
bodies like IEEE, IE, IETE, ISTE, and Biomedical Society of
India.