This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2944027, IEEE
Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

Removal of Power Line Interference from ECG

signals using Adaptive Notch Filters of Sharp
Resolution
Binqiang Chen1, Weifang Sun2, and Wangpeng He3 (IEEE MEMBER)
1.
School of Aerospace Engineering, Xiamen University, Xiamen 361005, Fujian Province, P. R. China
2.
College of Mechanical and Electrical Engineering, Wenzhou University, Wenzhou 325035, Zhejiang Province, P. R. China
3.
School of Aerospace Science and Technology, Xidian University, Xi’an 710054, P. R. China
Corresponding author: Weifang Sun (e-mail: [email protected]).
This research is supported financially by National Natural Science Foundation of China (No. 51605403), Natural Science Foundation of Guangdong
Province, China (No. 2015A030310010), Natural Science Foundation of Fujian Province, China (No. 2016J01012).

ABSTRACT The noise cancellation in electrocardiogram (ECG) signal is very influential to distinguish the
essential signal features masked by noises. The power line interference (PLI) is the main source of noise in
most of bio-electric signals. Digital notch filters can be used to suppress the PLI in ECG signals. However,
the problems of transient interferences and the ringing effect occur, especially when the digitization of PLI
does not meet the condition of full period sampling. In this paper, to obtain a better cancellation of the PLI,
a designing approach, generating adaptive notch filter (ANF) of sharp resolution, is proposed. The proposed
method is concise in algorithm and achieves a more comprehensive reduction of the PLI. It only requires on
one fast Fourier transform on the input signal. The spectrum correction method, based on the information
from the FFT spectrum of the corrupted signal, is utilized to estimate the harmonic parameters of the PLI.
The information of a few main lobe spectral bins in the FFT spectrum is merged such that a compensation
signal can be synthesized. By subtracting the compensational signal from the original measurement, the PLI
within the investigated signal can substantially reduced. A distinguished advantage of the proposed ANF
lies in the fact that no parameters are required to be specified, making the algorithm easier to be
implemented. The proposed ANF outperforms conventional notch filters because it not only alleviates the
undesirable effects but also better preserves the QRS-complex features in the filtered signal.

INDEX TERMS Electrocardiogram (ECG); Power line interference; Spectrum correction; Adaptive Notch
Filter;

I. INTRODUCTION methodologies, based on the classical theory of Fourier

The measurement of the electrical activity of the heart is analysis, have found many applications. Either finite
referred to as an electrocardiogram (ECG) [1,2]. The impulse response filters or infinite impulse response filters
reliable and accurate feature measurement of ECG signals can be designed for this purpose. Piskorowski utilized
plays an important role for the effective diagnosis of multiple notch filtering methods to suppress harmonic
cardiovascular disease [3-5]. However, during the process power line interferences in ECG signals [12]. Vallejo
of the ECG measurement, power line interference (PLI), proposed a spectro-temporal filtering technique for
which is of 50Hz in frequency, emerges as a major source electrocardiogram enhancement and found a gain in signal-
of interference that significantly deteriorates signal quality to-noise ratio improvement [13]. Wang put forward a novel
if it is left untreated [6,7]. design method to realize adaptive notch filters with infinite-
To retrieve the actual ECG signal and to extract more impulse response such that harmonics in the signal can be
accurate information for further analysis, the cancellation of removed [14].
PLI is necessary and thus attracts extensive attentions [8- Time-scale analysis (TSA) is also utilized in PLI
10]. Many advanced signal processing techniques, which cancellation and noise reduction of biological signals
can remove the PLI from noisy ECG signals, have been [15,16]. TSA methods can be regarded as filterbank based
reported in the literature [11]. Notch filter based techniques, which consist of a set of digital band-pass

VOLUME XX, 2017 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2944027, IEEE
Access

filters [17-19]. There are studies reporting the usage of the Usually in the digitization of a physical process, a time
wavelet transform and the empirical mode decomposition series of even spaced samples is recorded. In this section,
for suppressing PLI. Liu investigated the time, second-order we focus on explaining how the energy leakage problem
difference, and wavelet domain as sparse domains for the occurs in spectrum based techniques. We will show that the
ECG signal and the frequency domain as the sparse domain condition of fully period sampling (FPS), which is
for PLI [20]. Suchetha conducted a comparative analysis of determined by sampling parameters, has a substantial
EMD based filtering methods for 50 Hz noise cancellation impact on the spectrum analysis of signals.
in ECG signal [21]. Yadav proposed a novel non-local
A. Fundamentals of spectrum analysis of digital
wavelet transform method for ECG signal denoising by signals
exploiting the local and non-local redundancy present in the
signal [22]. Sayed introduced an effective hybrid scheme Modern spectrum analysis relies heavily on the theory of
for the denoising of ECG signals corrupted by non- discrete time Fourier Transform (DTFT). Cooley and
stationary noises using genetic algorithm and wavelet Tukey developed the celebrated fast algorithm of fast
transform [23]. Fourier transform (FFT), which decomposes a signal into
Besides the above methodologies, other advanced the sum of a few sinusoidal waves. For a signal x (t )
methods, such as least mean square based adaptive filtering sampled at the frequency fs and the length N , the
[24], adaptive Fourier transform [25], Kalman Filter [26], associated FFT spectrum xˆ(k ) is a complex valued series
nonlinear Bayesian filtering [27], eigen-value spaced at a uniform spatial interval of fs / N in the
decomposition [28], have also been investigated for the frequency domain.
purpose of ECG denoising. Recently, artificial intelligence N -1 -j
2 pnk
based techniques are also introduced to enable intelligent xˆ(k ) = å x (n )e N
, (1)
analysis of biomedical signals with noises [29-31]. Wang n =0

proposed an adversarial m ethod for ECG signals de-noising where the constant j indicates the imaginary number and
and achieved up to about 62% improvement on the SNR of k = 0,1, , N - 1 . The relationship between the sampling of
de-noised signals on average compared with state-of-the-art x (n ) and that of xˆ(k ) is shown in Figure 1.
technologies [32]. Huang investigated the classification of
t  1 / f s
ECG signals using the combination of short time Fourier
transform and convolutional neural network [33]. Wang
Amplitude

proposed a fast and accurate ECG classification system

based on deep learning [34].
The problem of PLI removal in ECG signal analysis can
be interpreted as the reduction of sinusoidal waves at some
specific frequencies. We proposed a novel adaptive notch Time
filter for PLI cancellation. The core of the proposed (a)
algorithm is the ratio based spectrum correction (RBSC). In
Amplitude

the procedure, only one fast Fourier transform is required to

f  f s / N
be performed on the original ECG signal. A pair of spectral
bins located within the main lobe of the Fourier spectrum
can be used to estimate the amplitude, the frequency and
the initial phase of the PLI such that a compensation signal
can be constructed. By subtracting the compensation signal Frequency
from the original ECG signal, the PLI can be effectively (b)
suppressed.
Two performances of two window based RBSC are FIGURE 1. Relationship between (a) the digitized signal and (b) the
compared and the Hanning window is chosen to be used in associated FFT spectrum.

the algorithm. Numerical results show that the proposed B. Condition of full period sampling for simple
ANF is more effective in suppressing undesirable harmonic wave
phenomenon of transient interference and the ringing effect,
which widely exist in notch filter based ECG de-noising A simple harmonic wave (SHW) is referred to as a dynamic
techniques. ECG measurements from the dataset released process in which a sinusoidal wave of a specific frequency
by MIT-BIH are used to test the proposed method in actual is included (Figure 2). The mathematical definition of a
ECG analysis. The results show that the ANF not only can SHW can be defined as
suppress the undesirable phenomena, but also produces less shw(t ) = A ⋅ cos(2p fct + f) , (2)
residual QRS-complex features, which outperforms the
performances of two comparison notch filters. where A denotes the amplitude; fc denotes the frequency
of the SHW; and f denotes the initial phase of the process.
II. ENERGY LEAKAGE OF SINUSOIDAL WAVE IN
DIGITAL SAMPLING

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2944027, IEEE
Access

It can be inferred that an SHW, by nature, is a band limited C. Negative effect caused by non-FPS condition
signal of extremely narrow bandwidth (Figure 3(a)). In the case of non-FPS condition, the energy leakage occurs
An SHW, of finite evenly spaced samples, is described to across the entire frequency axis. A notch filter is designed
meet the demand of FPS if the following condition is in the frequency domain. As such, no matter what kind of
satisfied. notch filter is used, the FIR type or the FIR type, only a
fc portion of the leakage components of PLI can be reduced.
N⋅ Î + , (3)
fs Owing to the incomplete reduction, an undesirable
phenomenon of ringing effect is inevitable. The
where  + denotes the set of positive integers. The Fourier
phenomenon is reflected as highly oscillated fluctuations of
spectrum of an SHW satisfying the FPS condition is a
waveforms in the time domain. Although the amplitudes of
discrete Dirac sequence in the frequency domain (Figure
the oscillations are not large, it still complicates subsequent
3(b), denoted as processes of ECG analysis.
ìï1, f = fc

shw( f ) = d( f - fc ) = ïí , (4)
ïï0, f ¹ fc III. RATIO BASED SPECTRUM CORRECTION METHODS
ïî
BASED ON RECTANGLE WINDOW AND HANNING
WINDOW
A window function is often used in spectrum analysis
Amplitude

because it is beneficial to suppress the energy leakage

problem. RBSC can be interpreted as a preprocessing of the
FFT spectra. More accurate harmonic information of Ac ,
fc and fc can be estimated using a few main lobe spectral
bins (MLSBs) located within the main lobe of a window
Time
function in the frequency domain.
FIGURE 2. The waveform of an SHW in the time domain.
A. Spectrum correction using rectangle window
Otherwise it will become a discretized version of a
broadband signal whose shape is determined by the FFT of The mathematical definition and the frequency response of
the windows function (Figure 3(c)). Two most commonly a rectangle window is shown in Equation (5,6).
used types of window are the rectangular window and the
Wr (n ) = 1 for n = 1,2, , N , (5)
Hanning window.

 ( f  fc ) (a) sin(Lw / 2) j (N -1)w / 2

Wˆr (w ) = e , (6)
sin(w / 2)
Amplitude

1.2
1 Main lobe (a)
0.8
Side lobes
Frequency 0.6
0.4
 ( f  fc ) (b) 0.2
0
Amplitude

−3 −2 −1 0 1 2 3
Index of spectral bin
1.2 Actual spectral bin
T (k) (b)
1 T (k  1)
0.8
Frequency 0.6
(c) 0.4 T (k)
k
Amplitude

0.2
Broad band spectrum 0
k 1
k
Spectral bin number

FIGURE 4. (a) The frequency response of a rectangle window; and (b)

the MLSBs for spectrum correction.
Frequency
The window has the same length as that of the input signal.
FIGURE 3. (a) The actual spectrum of an SHW; (b) the narrow band The modulus function of the frequency response is
spectrum of an SHW when it satisfy the FPS condition; and (c) the
broad band spectrum of an SHW when it does not satisfy the FPS
condition.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2944027, IEEE
Access

sin(pk ) L sin(pk ) ìï 
T (k ) = » , (7) ïïA = pDk ⋅ T (k ) ⋅ 2[1 - (Dk )2 ]
sin(pk / L) pk ïï c
sin(pDk )
ïï
ïï fs
and the shape of T (k ) is plotted in Figure 4(a). In the main í fc = (k + Dk ) , (14)
ïï L
lobe of the function, two MLSBs are utilized (Figure 4(b)). ïï Im(k ¢)
They are denoted as T (k) and T (k + 1) . ïïfc = arctan - pDk
ïï Re(k ¢)
Let the normalized error Dk = k ¢ - k , where k ¢ is the ïî
index of the window vertex associated with the actual 1
spectral bin. This error can be computed as (a)
1 T (k + 1)
Dk = - =- , (8) 0.5
1+v T (k) + T (k + 1)

where v = T (k) / T (k + 1) . Finally, the harmonic information 0

−1 −0.5 0 0.5 1
of an SHW can be obtained as below. Normalized angular frequency/
1.2
ì
ï pDk ⋅ T (k ) T ( k ) k  Actual spectral bin
ï
ïAc = 1
ï
ï sin(pDk ) 0.8 T ( k  1)
ï
ï
ï
ï fs 0.6
í fc = (k + Dk ) , (9) T ( k  1) ( b)
ï
ï L 0.4
ï T ( k  2)
ï Im(k ¢) 0.2
ïfc = arctan
ï - pDk 0
ï
ï Re(k ¢) k  1 k k  1 k  2
ï
î
Spectral bin number

B. Spectrum correction using Hanning window FIGURE 5. (a) The frequency response of a Hanning window; and (b)
the MLSBs for spectrum correction.
The mathematical definition and the frequency response of
a Hanning window is shown in Equation (10,11). IV. DESIGN OF ADAPTIVE NOTCH FILTER OF SHARP
RESOLUTION BASED ON RBSC
1 1 æ 2pn ö÷
Wh (n ) = - cos ççç ÷ for n = 0,1,2 , L - 1 , (10) If the frequency of the PLI is exactly 50Hz, it is feasible to
2 2 çè L ø÷÷
avoid energy leakage by properly selecting the sampling
ìï 1 1é æ 2p ö æ 2p öù üï
parameters. However, the value of the actual frequency is
Wˆh (w ) = ïí Wˆr (w ) + êêWˆr çççw - ÷÷÷ +Wˆr çççw + ÷÷÷úú ïýe - jNw /2 , (11) just approximately 50. An error of the power frequency in
ïïîï 2 4 êë èç L ø÷ èç L ø÷úû ïïþï
the range of [-2, 2]Hz can be found according to power
The modulus function of the frequency response is quality control regulations, for example in China. RBSC
based techniques can be employed to design adaptive notch
sin(pk ) 1 filter of extremely sharp resolution. The procedure of the
T (k ) = ⋅ , (12)
2pk 1-k2 proposed methodology is stated as below.
Step 1). Let the analyzed signal be ecg(t ) of sampling
and the shape of T (k ) is plotted in Figure 5(a). When the frequency fs and length N . Perform the fast Fourier
FPS condition is not satisfied, there are four MLSBs existed transform on the signal ecg(t ) .
in the main lobe. However, only the two with larger (f )
ecg(t ) ¾¾¾
FFT
ecg
amplitudes are selected (Figure 5(b)). The error between k ¢ ,
the index of the actual spectral bin, and k , is calculated Step 2). Search two spectral bins of large amplitudes
used the following equation within the main lobe of the PLI in the frequency domain.
Estimate the harmonic information ( APLI , fPLI and fPLI ) of
v -2
Dk = , (13) the PLI by ratio based spectrum correction methods.
v +1
( f ) ¾¾¾¾
ecg RBSC
 APLI , fPLI , fPLI
where the v = T (k) / T (k + 1) . Finally the harmonic
information of an SHW can be estimated as Step 3). Construct a compensation signal compPLI (t )
based on the extracted harmonic information.
compPLI (t ) = APLI ⋅ cos(2p fPLIt + fPLI )

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2944027, IEEE
Access

Step 4) Subtract compPLI (t ) from the corrupted ECG The algorithm belongs to the type of parameter free. A
signal. flow chart of the proposed algorithm is displayed as in
Figure 6.
ecg(t ) = ecg(t ) - compPLI (t )


ECG signal corrupted by PLI

Compensation signal
compPLI (t ) = APLI ⋅ cos(2p fPLIt + fPLI )

Search MLSBs in the main lobe Estimate harmonic

FFT on the ECG signal
Information of PLI
1.2 Actual spectral bin
T ( k)
1 T (k  1)
0.8
0.6
0.4 T (k)
0.2 k

0
k 1
k
Spectral bin number

Filtered signal

FIGURE 6. Flow chart of the algorithm of the adaptive notch filter.

2
V. COMPARISON OF PERFORMANCES OF TWO errrec = yrec (t ) - x (t ) , (18)
2
SPECTRUM CORRECTION METHOS
Either of the two windows can be employed to estimate the Where || ⋅ ||2 means to compute the 2-norm of input series.
harmonic frequency of PLI in ECG signals. In this section,
we test the estimation performance of the RBSCs with the Let the frequency error be ranged in the interval of [0,
two types of windows. 0.5]Hz, the reconstruction errors of the two RBSC methods
are shown in Figure 7. It can be seen that when
A. Comparison of performance for noise-free SHW Df Î [0, 0.25]Hz , the rectangular window based RBSC has
In this subsection, a digital signal x (t ) containing one SHW better performance. While when Df Î [0.25, 0.5]Hz , the
is simulated. Hanning window based RBSC has better performance. No
matter what kind of window is used, the reconstruction
x c (t ) = Ac cos(2p fct + fc ) , (15)
error is small and acceptable for high precision estimation.
where the amplitude Ac is chosen as 1 without loss of
generality. The sampling length of celebrated datasets of
ECG measurement are mainly 200~400Hz. Therefore, we
set the sampling frequency of x (t ) as 1, and 250 digital
samples are utilized in the following comparison.
Mean value of the error

To simulate the error of power line frequency, the

parameter fc is set as
fc = 50 + Df , (16)

Supposing that yrec (t ) and yhan (t ) are synthesized signals

reconstructed using the rectangular window based RBSC
and the Hanning window based RBSC, respectively, the
reconstruction errors between x (t ) and the synthesized
signals using the following formulae.

errrec = yrec (t ) - x (t ) ,
2
(17) f [Hz]
2

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2944027, IEEE
Access

FIGURE 7. (a) The frequency response of a Hanning window; and (b)

the MLSBs for spectrum correction.

A. Comparison of performance for SHW corrupted by

Amplitude
Gaussian noises
In this subsection, we create a noisy version of the signal
x c (t ) as

xc (t ) = x c (t ) + wgn(t ) . (19)

Time [s]
The same sampling parameters are adopted as those in xc (t ) . FIGURE 9. The waveform of the simulated simple harmonic wave.
The signal-to-noise ratio of the noisy signal is set as 20.
The resulting reconstruction errors using the two types of Transient interference
windows are shown in Figure 8. For all values of Df , the

Amplitude
curves of the reconstruction errors based on Hanning
window RBSC is stable. While for the case of rectangle
window, large errors are produced when Df Î [0.01, 0.13] . Ringing effect
This phenomenon is caused by one MLSB small in
amplitude. Time [s]

FIGURE 10. The waveform of the simulated simple harmonic wave.

The Fourier spectra of the simulated signal and the

filtered signal are shown in Figure 11. The scale of Y-axis
this figure is set as in the mode of logarithm. The leakage
Mean value of the error

components, whose frequencies are closed to 50Hz, are

effectively reduced; however spectral bins with larger
distances from the central frequency of the notch filter are
preserved. This leads to the undesirable phenomena of
filtering based on the IIR filter.

f [Hz]
Amplitude

FIGURE 8. (a) The frequency response of a Hanning window; and (b)

the MLSBs for spectrum correction.

C. Result analysis
The actual ECG measurement is not free of noise. As
shown in the comparisons, the Hanning windows based
RBSC is more robust in correction harmonic information of
SHW not satisfying FPS condition. Therefore, we choose
the type of window, which is utilized in the algorithm of Frequency [Hz]
Section IV, as the Hanning window.
FIGURE 11. The Fourier spectra of the simulated signal and the filtered
VI. COMPARISON OF THE PROPOSED METHOD WITH signal by the IIR notch filter.
CONVENTIONAL NOTCH FITLER IN REDUCTION OF Figure 12 shows the filtering result using the proposed
SHW
method. The maximal absolute value of the signal is
Following the definition in Equation (15), an SHW with
smaller than 4 ´ 10-3 . Although the ringing effect still exists,
frequency is 50.1Hz is simulated and the waveform is
it can be neglected due to extremely small amplitudes. The
shown in Figure 9.
An IIR notch filter is constructed to remove the SHW. root-mean-square values of the ringing parts of the two
The central frequency and the bandwidth of the notch filter filtered signals in Figure 10 and Figure 12 are 0.024 and
are set as 50Hz and 0.4Hz. The filter result is shown in 0.016 respectively, which demonstrates the enhancement of
Figure 10. Undesirable phenomena of the transient the proposed method in the suppression of the ringing effect.
interference at the beginning samples and the ringing effect
can be easily recognized in Figure 11.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2944027, IEEE
Access

(a)

Amplitude
Amplitude

Time [s]
Time [s]
FIGURE 12. The waveform of the filtered signal using the proposed (b)

Amplitude
method.

Figure 13 shows the Fourier spectra of the simulated

signal and the filtered result by the proposed method. The
amplitudes of the leakage components across the entire
frequency axis are substantially reduced.
Time [s]

1000
(c)

Amplitude
500

-500
Amplitude

0 0.5 1 1.5 2 2.5 3 3.5 4

Time [s]
FIGURE 14. (a) The waveform of the original ECG signal; (b) the
simulated PLI; and (c) the simulated ECG signal with PLI.

Frequency [Hz]
Amplitude

FIGURE 13. The Fourier spectra of the simulated signal and the filtered
signal by the proposed method.

VII. NUMERICAL ANALYSIS OF THE PROPOSED

METHOD IN ECG SIGNALS
To evaluate the performance of the proposed method in
processing ECG signals corrupted by PLI, measurements of
ECG form the dataset of MIT-BIH polysomnographic
database (http://www.physionet.org/cgi-bin/atm/ATM) are
utilized. In this section, the original ECG signal to be
analyzed is denoted as d (n ) . The sampling frequency and Frequency [Hz]
the sampling length of the signal is 250 and 1024. FIGURE 15. The Fourier spectra of the original ECG signal and the
An SHW is simulated to be the PLI of the original ECG simulated signal corrupted by PLI.
signal. The harmonic information of the SHW are Based on the proposed ADF method, the filter signal and
ì
ï reconstruction errors are shown in Figure 16. The
ïA = 250
ï c
ï
construction error is pretty small. The largest value of the
í fc = 50.1Hz . (20) errors is smaller than 1.5.
ï
ï
ï
ïf = 15
ï
î c
The waveform of the simulated PLI is shown in Figure
14(b). By summing the original ECG signal and the SHW,
the simulated signal with PLi is shown in Figure 14(c). The
SNR of the simulated signal is 1dB. The Fourier spectra of
the original ECG signal and the simulated signal are shown
in Figure 15.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2944027, IEEE
Access

(a) (a)

Amplitude
Amplitude

Time [s] Time [s]

(b) (b)

Amplitude
Amplitude

Time [s] Time [s]

FIGURE 16. (a) The filtered signal using the proposed method; and (b) FIGURE 18. (a) The filtered signal using the conventional IIR filter; and
the construction errors between the filtered signal and the original ECG (b) the construction errors between the filtered signal and the original
signal. ECG signal.

The comparison between the Fourier spectrum of the The other comparison method is an improved IIR with
original ECG signal and that of the filtered signal is shown transient interference suppression ability [35]. The filtered
in Figure 17. It can be seen, the two signals are very result and the construction error are shown in Figure 19. In
similar except for a few spectral bins round the Figure 19(b), the transient interferences at the beginning
neighborhood of the actual frequency of the PLI. samples and the residual QRS-complex features are much
smaller than those in Figure 18(b).

(a)
Amplitude
Amplitude

Time [s]
(b)
Amplitude

Frequency [Hz]
FIGURE 17. Comparison between the original ECG signal and filtered Time [s]
signal.
FIGURE 19. (a) The filtered signal using the improved IIR filter; and (b)
We utilized two types of IIR notch filter to process the the construction errors between the filtered signal and the original ECG
signal. One is the conventional IIR filter employed in signal.

Section VI. The filtered signal and its Fourier spectrum are Comparisons between the waveform of the original ECG
shown in Figure 18. Although the waveform of the filtered signal and those of the filtered signal generated by the two
signal is very similar to the original ECG signal. The curve comparison IIR notch filters are shown in Figure 20. In
of errors between the two signals reveals that phase shift is Figure 20(a), relatively large errors are found in the high
produced in the IIR filtering process and causes very large frequency range of the proposed method. To make a more
errors in the neighboring areas of the impulsive QRS- quantitative analysis, the maximal value and the RMS value
complex features. The maximal error can be as high as 300. of the construction errors are computed. The related
On the other hand, transient interferences are also detected information is shown in Table 1. The construction error by
at some beginning samples of the filtered signal. the proposed method is the smallest in the indicators of
maximal error and RMS of errors. These comparisons have
validated that the proposed ANF outperforms traditional
IIR notch filter based techniques.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2944027, IEEE
Access

103 conventional IIR notch filter. Results show that the

The origianl ECG signal
(a) The filtered signal proposed ANF of sharp resolution can better suppress the
ringing effect and causes no transient interferences at the
102
beginning samples of the filtered signal.
(3) The ECG measurements form the dataset released by
MIT-BIH are used for test the performance of the proposed
Amplitude

101
method. An additive 50Hz PLI is added to the original ECG
signal to simulate ECG signals corrupted by PLI. The
100 proposed method is compared with two IIR notch filters. It
is found that the maximal value and the RMS value of
construction errors, produced by the proposed ANF, are
10-1
smallest among the results of three methods, indicating an
improved SNR in the filtered signal.
10-2
0 20 40 60 80 100 120
REFERENCES
Frequency [Hz] [1] E. N. Bruce, Biomedical Signal Processing and Signal Modeling,
Wiley, New York, 2001.
[2] P. E. McSharry, G. D. Clifford, L. Tarassenko, L. A. Smith, A
(b) dynamical model for generating synthetic electrocardiogram signals,
IEEE Trans. Biomed. Eng., 2003, 50(3): 289–294.
[3] K. Friganovic, D. Kukolja, A. Jovic, M. Cifrek, G. Krstacic,
Amplitude

Optimizing the Detection of Characteristic Waves in ECG Based on

Processing Methods Combinations, IEEE Access, 2018, 6: 50609-
50626.
[4] N. Vemishetty, P. Patra, P. K. Jha, and et al, Low Power Personalized
ECG Based System Design Methodology for Remote Cardiac Health
Monitoring, IEEE Access, 2016, 4: 8407-8417.
[5] O. Yildirim, R. S. Tan, U. R. Acharya, An efficient compression of
ECG signals using deep convolutional autoencoders, Cognitive
Systems Research, 2018, 52:198-211.
[6] H. Q. Liu, Y. Li, Y. Zhou, and et al. Joint power line interference
suppression and ECG signal recovery in transform domains.
Biomedical Signal Processing and Control, 2018, 44: 58-66.
[7] Y. D. Lin, Y. H. Hu, Power-Line Interference Detection and
Frequency [Hz] Suppression in ECG Signal Processing, IEEE Transactions on
Biomedical Engineering, 2008, 55(1): 354-357.
FIGURE 20. (a) The filtered signal using the conventional IIR filter; and [8] C. Chandrakar, M.K. Kowar, Denoising ECG signals using adaptive
(b) the construction errors between the filtered signal and the original filter algorithm, Int. J. Soft Comput. Eng., 2012, 2(1): 120-123.
ECG signal. [9] J.R. Glover, Adaptive noise canceling applied to sinusoidal
interferences, IEEE Trans. Acoust. Speech Signal Process., 1977, 25:
TABLE I 484-491.
COMPARISONS BETWEEN THE PROPOSED METHOD AND TWO TYPES OF IIR [10] S. C. Pei, C. C. Tseng, Elimination of AC interference in
NOTCH FILTERS electrocardiogram using IIR notch filter with transient suppression,
The conventional IIR IEEE Trans. Biomed. Eng., 1995, 42(11): 1128-1132.
The The improved IIR filter
filter [11] R. R. Sharma, R. B. Pachori, Baseline wander and power line
proposed
method Entire Ringing Entire Ringing interference removal from ECG signals using eigenvalue
signal part signal part decomposition, Biomedical Signal Processing and Control, 2018, 45:
Maximal 33-49.
1.1601 9.0063 6.0187 370.58 370.58
error [12] Jacek Piskorowski. Suppressing harmonic powerline interference
RMS of using multiple-notch filtering methods with improved transient
0.4436 1.8412 1.0015 60.7558 60.9910
errors
behavior. Measurement, 2012, 45(6): 1350-1361.
[13] D.T. Vallejo, T. Falk, Adaptive spectro-temporal filtering for
VII. CONCLUSIONS electrocardiogram signal enhancement, IEEE J. Biomed. Health
Inform., 2018, 22: 421-428.
The findings the paper are summarized as below. [14] Q. S. Wang, X. L. Gua, J. Y. Lin, Adaptive notch filter design under
(1) The PLI emerges in the form of simple harmonic multiple identical bandwidths, AEU - International Journal of
wave. The phenomenon of energy leakage occurs when the Electronics and Communications, 2017, 82: 202-210.
[15] M. Suchetha, N. Kumaravel. Empirical mode decomposition based
sampling of the PLI does not meet the condition of FPS. In filtering techniques for power line interference reduction in
such situations, conventional notch filters can not suppress electrocardiogram using various adaptive structures and subtraction
all spectral bins related to energy leakage, which causes methods. Biomedical Signal Processing and Control, 2013, 8(6): 575-
undesirable effects in the filtered signal. 585.
[16] S. Poornachandra. Wavelet-based denoising using subband dependent
(2) A novel adaptive notch filter based on RBSC is threshold for ECG signals. Journal of Digital Signal Processing, 2008,
proposed in this paper. In noisy measurements, the Hanning 18: 49-55.
window based RBSC is more robust in acquiring high [17] Naveed ur Rehman, D.P. Mandic. Filter bank property of multivariate
precision harmonic information. In simulations of SHW empirical mode decomposition. IEEE Transactions on Signal
Processing, 2011, 59(5): 2421-2426.
reduction, the proposed method was compared with a

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2944027, IEEE
Access

[18] B. Q. Chen, Z. S. Zhang, Y. Y. Zi, Z. J. He, C. Sun. Detecting of

transient vibration signatures using an improved fast spatial–spectral
ensemble kurtosis kurtogram and its applications to mechanical
signature analysis of short duration data from rotating machinery.
Mechanical Systems and Signal Processing, 2013, 40(1): 1-37.
[19] Y. Liu, Y. Li, H. Lin, H. Ma. An amplitude-preserved time-frequency
peak filtering based on empirical mode decomposition for seismic
random noise reduction. IEEE Geosci. Remote Sens. Lett., 2014,
11(5): 896-900.
[20] H. Q. Liu, Y. Li, Y. Zhou, X. R. Jing, T. K. Truong. Joint power line
interference suppression and ECG signal recovery in transform
domains. Biomedical Signal Processing and Control, 2018, 44: 58-66.
[21] M.Suchetha, N. Kumaravel, M. Jagannath, Saravana Kumar
Jaganathan. A comparative analysis of EMD based filtering methods
for 50 Hz noise cancellation in ECG signal. Informatics in Medicine
Unlocked, 2017, 8: 54-59.
[22] S. K. Yadav, R. Sinha, P. K. Bora. Electrocardiogram signal
denoising using non-local wavelet transform domain filtering. IET
Signal Process., 2015, 9: 88-96.
[23] El-Sayed A. El-Dahshan. Genetic algorithm and wavelet hybrid
scheme for ECG signal denoising. Telecommunication Systems, 2011,
46(3):209–215.
[24] F. Xiong, D. Y. Chen, Z. H. Chen, S. M. Dai. Cancellation of motion
artifacts in ambulatory ECG signals using TD-LMS adaptive filtering
techniques. Journal of Visual Communication and Image
Representation, 2019, 58: 606-618.
[25] Z. Wang, F, Wan, C. M. Wong, L. M. Zhang. Adaptive Fourier
decomposition based ECG denoising. Computers in Biology and
Medicine, 2016, 77: 195-205.
[26] N. Zeng, Z. Wang, Y. Li, M. Du, X. Liu, A hybrid EKF and
switching PSO algorithm for joint state and parameter estimation of
lateral flow immunoassay models, IEEE/ACM Transactions on
Computational Biology and Bioinformatics, ,2012, 9(2): 321-329.
[27] R. Sameni, M.B. Shamsollahi, C. Jutten, G.D. Clifford. A nonlinear
Bayesian filtering framework for ECG denoising. IEEE Trans.
Biomed. Eng., 2007, 54(12): 2172-2185.
[28] R. R. Sharma, R. B. Pachori. Baseline wander and power line
interference removal from ECG signals using eigenvalue
decomposition. Biomedical Signal Processing and Control, 2018, 45:
33-49.
[29] Y. Peng, Z. Wu, J. Jiang. A novel feature selection approach for
biomedical data classification. Journal of Biomedical Informatics.
2010, 43(1):15-23.
[30] Y. Peng , X. Zhang. Integrative data mining in systems biology: from
text to network mining. Artificial Intelligence in Medicine, 2007,
41(2):83-86.
[31] N. Zeng, Z. Wang, B. Zineddin, Y. Li, M. Du, L. Xiao, X. Liu and T.
Young, Image-based quantitative analysis of gold
immunochromatographic strip via cellular neural network approach,
IEEE Transactions on Medical Imaging, vol. 33, no. 5, pp. 1129-
1136, 2014.
[32] J. L. Wang, R. F. Li, R. Li, K. Q. Li, H. B. Zeng, G. Q. Xie, L. Liu.
Adversarial de-noising of electrocardiogram. Neurocomputing. 2019,
349: 212-224.
[33] J. S. Huang, B. Q. Chen, B. Yao, W. P. He. ECG Arrhythmia
Classification Using STFT-Based Spectrogram and Convolutional
Neural Network. IEEE Access, 2019, 7: 92871 - 92880.
[34] E. K. Wang, X. Zhagn, L. Y. Pan. Automatic classification of CAD
ECG signals with SDAE and bidirectional long short-term network.
IEEE Access, 2019, 7, 1-1.
[35] S. C. Pei, and C. C. Tseng. Elimination of AC interference in
electrocardiogram using IIR notch filter with transient suppression.
Biomedical Engineering, IEEE Transactions on 42.11 (1995): 1128-
1132.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.