ICBME2008
Abstract— In this paper a new method for noise filtering of high resolution stress ECG is presented. The method is based on a perfect reconstruction and linear phase filter banks, which decompose the ECG signal to different frequency subbands. Then different processing algorithms are applied to different frequency subbands to remove various ECG noises such as baseline, power line, EMG noises and motion artifacts.
The proposed processing algorithms for frequency subbands are adaptive. The algorithms first estimate the noise level of the ECG signal and based on the noise level, the parameters of the algorithms are selected. One of the benefits of filter bank is its capability of applying both frequency and time domain processing simultaneously. To apply time domain processes, it is required to extract fiducial points of ECG signal such as Q, R, S and J points. We have also introduced a new multi- resolution curvature scale space based method to extract these points, which can extract these points precisely. The comparison of proposed method with those other methods showed that the proposed methods have promising results.
I. INTRODUCTION
ecording of electrocardiograms (ECG’s) is a useful tool for diagnosing heart disease. Typical cycle of an ECG is shown in Fig. 1. Based on ECG signal shape and distance between fiducial points and other parameters, physicians diagnose heart diseases. Interpretation of the shape of ECG signal and recognition of the fiducial points and calculation of parameters is a tedious task for the physician; 100000 cardiac cycles per patient are recorded in a day and a physician has to interpret this large amount of ECG data to search for only a few abnormal cardiac cycles in the ECG. One solution to this problem is the extraction and interpretation of ECG signal using computer software. However electrocardiographic (ECG) signals may be contaminated by various kinds of noise, which may degrade the interpretation process by the physician or computer software. Therefore, it is necessary to remove various noises of ECG signal before displaying or interpreting it. Fiducial points extraction algorithm should also have less sensitivity to noise.
The noise sources of ECG signals are different and each of them demands a special algorithm to remove. Typical
Alireza Behrad is with Engineering Department of Shahed University, Tehran-Qom Freeway, Tehran, IRAN. (Phone: +98-21 5527-7647; e-mail:
behrad@ shahed.ac.ir).
sources are:
1) Power line interference 2) Motion artifacts
3) Muscle contraction (electromyographic, EMG) 4) Baseline drift and ECG amplitude modulation with
respiration
There are other, less significant noise sources too. A brief description and simulation method of each noise source will be given below. The algorithms for removing different noises will be discussed later.
A. Power line Noise
Power line interference consists of 60 Hz pickup (in the U.S.) and 50 Hz pickup (in IRAN and Europe) and harmonics which can be modeled as sinusoids and combination of sinusoids.
B. Baseline drift and ECG amplitude modulation with respiration
Baseline drift is a low frequency noise which its energy resides mainly in the frequency range of dc to 0.8 Hz.
However in stress ECG its frequency content may be extended up to 1.5Hz. The drift of the baseline with respiration can be represented as a sinusoidal component at the frequency of respiration added to the ECG signal.
C. Motion Artifacts
Motion artifacts are transient (but not step) baseline changes caused by changes in the electrode-skin impedance with electrode motion. As this impedance changes, the ECG amplifier sees a different source impedance, which forms a voltage divider with the amplifier input impedance.
The shape of the baseline disturbance caused by motion artifacts can be assumed to be a biphasic signal resembling one cycle of a sine wave with duration of 100-500ms and amplitude of peak-to-peak ECG amplitude.
A New Method for Noise Filtering of High Resolution Stress ECG
Alireza Behrad
R
Fig. 1. Typical cardiac cycle and fiducial points
ICBME2008
D. Muscle contraction (EMG)
Muscle contractions cause another type of noise in ECG signal. The signals resulting from muscle contraction can be assumed to be transient bursts of zero-mean band-limited Gaussian noise. The frequency content of EMG noise can be in the range of dc to 10000Hz. However the energy of the noise for the frequencies smaller than 10Hz is less significant.
A variety of algorithms have been proposed for the noise filtering of ECG signals. Considering different properties of noise sources, different algorithms are used for filtering of various type of ECG noise.
In this paper we present a new method for noise filtering of high resolution stress ECG. The proposed method is based on filter banks, which are perfect reconstruction and linear phase filter bank. In this method first the signal is decomposed to different frequency subbands. Then different processing algorithms are applied to different frequency subbands. The processing algorithms in the subbands are based on properties of the noise in the subbands and its energy. In other word our method combines different algorithms for filtering of different type of noise in filter bank, where different algorithm are applied to different subbands of the filter bank. In addition to proper results of this method which will be discussed later, the filter bank reduces the computation overhead of the algorithms too. Because the processing algorithms are applied after the downsampling and in the lower sampling frequencies. The processing algorithms are adaptive i.e. we first estimate the noise level of the ECG signal and based on the noise level, the parameters of the algorithms are selected. One of the benefits of filter bank is its capability of applying both frequency and time domain processing simultaneously. To apply time domain processes we need to extract fiducial points of ECG signal such as Q, R, S and J points. We have also introduced a new multi-resolution curvature scale space based method to extract these points.
Following this section, we first review different methods for noise filtering of ECG signal. Then in section III fundamentals of perfect construction and liner phase filter banks are discussed. The proposed algorithms for noise removing of ECG signal and processing methods for different subbands of filter bank appear in section IV. The Section also presents the algorithms for extraction fiducial points. Experimental results are explained in section VI followed by a conclusion in section VII.
II. REVIEW OF PREVIOUS WORKS
In order to suppress the effects of various ECG noise, different algorithms have been presented. Although some of these algorithms may suppress multiple sources of noise, different noise usually demands its own method for noise removing. This is because of different properties of the noise and its frequency content. For example EMG noise has higher frequency content; however baseline wander has
lower frequency content. The amplitude of the various kind of ECG noise is also different. In this section we review the various methods for filtering of various ECG noise. Since our proposed method requires detecting fiducial points of the ECG signal several methods for extraction these points also presented.
A. Suppressing Power Line Noise
Power line interference (either 60 Hz or 50Hz) and its harmonics is one of the noise sources in ECG and biomedical signal recording. One method to eliminate the power line noise is the use of adaptive filters with an external reference [1], [2]. However it is often desirable, or necessary, to reduce power line interference when an external reference is not available. Non adaptive notch filters using IIR or FIR filters [3], [4] are other methods of suppressing power line noise. The main drawback of non adaptive notch filters especially in high resolution ECG is the ringing effect generated by these filters after QRS complex.
Ahlstrom and Tompkins [5] reported on an adaptive power line filter with an internally generated reference. In [4] it is shown that, the adaptive filter with internally generated reference are equivalent to second order non adaptive IIR filter in some sense. However in [6], it is shown that, the most significant distortion of non adaptive notch filters which appears as ringing after the QRS complex are not observed in Ahlstrom and Tompkins adaptive filter. This is because of the incremental adaptation used in Ahlstrom and Tompkins adaptive filter which is not significantly affected by QRS complex. In this reference, it is also shown that Ahlstrom and Tompkins filter has less computation overhead and distortion comparing to non adaptive notch filter.
B. Removing Baseline Wander
Baseline wander has lower frequency content. The suppressing of this type of noise is essential for ECG signal displaying and some algorithms such as ST segment measurement and P wave detection.
Different methods have been proposed for removing baseline wander. Cubic spline [8], [9] is one of the methods used for filtering of the baseline drift. In this method first the fiducial point of the ECG signal is extracted and then using the isoelectric point at ECG signal the baseline drift of the ECG beats are estimated with a cubic spline and subtracted from ECG signal. The main drawback of cubic spline algorithms is that its efficiency is highly dependent to correct fiducial point extraction.
Digital filters [10],[11], including FIR or linear phase IIR filters are other methods of suppressing baseline wander. These methods remove baseline wander by applying a narrowband high pass filter. The -3dB cutoff frequency of these filter may vary from 0.5Hz to 0.8Hz, however in stress ECG the cutoff frequency may also be greater. In [12] the cubic spline and digital filters are
ICBME2008
compared with each other and it is shown that the digital filters have better performance. However it is important to note that when the cutoff frequency of high pass filter is greater than the first main frequency of the ECG signal the distortion of the digital filters is not negligible. This is because of pseudo periodic property of ECG signal. In [13]
a time varying digital filter is used for the filtering of baseline wander, where the baseline of ECG signal is first estimated and the cutoff frequency is then selected base on noise estimation. Morphology based algorithms [15], [16], adaptive filters [17], Bionic wavelet transform [18] and the empirical mode decomposition [19] are other examples of baseline removing filters.
C. Filtering EMG Noise and Motion Artifacts
Similar to other source of noise, digital filters [20], [21]
are useful method for noise filtering of EMG and motion artifacts. Digital low pass filters are usual for removing EMG noise. Digital filters are easy to implement and don't need the extraction of fiducial points. Another group of algorithms use signal overlapping to reduce different kind of noise especially EMG and motion artifact. Mean [22]
and trimmed mean [23], median [24] and ordered average [25] filters are some of noise removing algorithm which are categorized in this group. These algorithms first extract the R points of the ECG signal and divide the signal into beats or epochs. Then by different operations on several consecutive beats, the noise free beats is calculated. This group of algorithm suffers mainly from two problems.
1- The efficiency of the algorithm highly depends on the efficiency of R detection algorithm.
2- Arrhythmias in ECG signal may be considered as outliers and may be replaced with normal beats especially in algorithms with ordered epochs.
Adaptive filters with external reference [26] are other methods for suppressing of EMG. However the reference signal is seldom available and delayed version of ECG signal is sometime used as reference signal. Incremental composite [27] or adaptive filters with R point as reference are also used for filtering of ECG signal; however they highly distort the ECG signal in the case of arrhythmia.
Morphology based filter [28],[29] and Source consistency filter [30] are other methods to suppress EMG
noise. Source consistency filter uses the information of other leads to estimate noise free signal of a lead and utilizes it as a basis for filtering of the ECG signal.
However it is implicitly assumed the other leads are noise free. Filter banks [23], [31] are also used for filtering of EMG and motion artifact. The main advantage of the filter bank is the simultaneously use of time frequency domain information.
D. Fiducial Points Extraction
As mentioned before, some noise removing algorithms need to extract fiducial points of ECG signal. The R point is the most important point of ECG signal which is used for signal overlapping and extraction of other fiducial point.
Variety of algorithms [33]-[35] is used for the detection of this point. When the R point extracted other points such as Q and S points are extracted using slope detection and thresholding [36] or using other method such as neural networks [37],[38] or piecewise derivative dynamic time warping [39].
III. LINEAR PHASE FILTER BANK
Digital Filter banks [40], [41] have been used to decompose a signal into frequency subbands. Then signals in subbands may be processed or coded individually. Such schemes are popular for encoding data from speech and image signals. The process of decomposing and eventual reconstruction is done by what is termed the "analysis- synthesis" filter bank system shown in Fig. 2. In this figure Hi(z) are analysis filters and Fi(z) are synthesis filters. The boxes with
↓ M
denotes decimation or subsampling by M and↑ M
denote expanders, which increase the sampling rate by M. The scheme of Fig. 2 is not proper for software implementation, because the analysis filters is first applied and then data are decimated. However for the data which are not selected by decimator, it is not necessary to apply analysis filters. This is true for synthesis part too. To remedy this problem, the scheme of Fig. 3 is used for implementation, which is termed polyphase implementation of filter banks, where E(z) and R(z) are polyphase matrixes corresponding to analysis and synthesis filters, respectively. These two schemes are equal and can beFig. 2. A M channel uniform filter bank Fig. 3. Polyphase implementation of M channel filter bank
ICBME2008
converted to each other; however the computation overhead of the second scheme is less.
When there are no process in the process boxes in Fig. 2 or Fig. 3, the signals z(n) should be delayed version x(n), which is termed the perfect reconstruction property of the filter bank. One way to do this is to let
R ( z ) = E
−1( z )
, and then choose the matrixE (z )
so that both matrices are FIR. Another approach to design perfect reconstruction system is to choose the matrixE (z )
to be FIR"paraunitary" matrix. A matrix is said to be paraunitary if it satisfies the equation:
I z E z
E ˆ ( ) ( ) =
(1) whereE ˆ ( z ) = E
*( 1 / z *)
. The system can be guaranteed to have perfect reconstruction property by having:R ( z ) = E ˆ ( z )
.Our processing algorithms use both frequency and time domain information by locating fiducial point and considering their location in frequency subbands of filter bank. Therefore we need to design a filter bank which has linear phase property in addition to perfect reconstruction property. Different methods have been proposed for designing linear phase filter banks [40],[41]. The properties of linear phase paraunitary filter banks have been discussed in [41]. The method requires solving an optimization problem. We have implemented a MATLAB program to obtain filter bank coefficient using this method.
IV. PROPOSED NOISE REMOVING ALGORITHM
A. Outline Structure of the Algorithm
Figure 4 shows the outline structure of the algorithms.
As it is shown in this structure, we have considered the suppression of various ECG noise sources. The ECG signal concurrently enters into "R Detection" and "0.5Hz High Pass Filter" boxes. The purpose of "R Detection" box is the extraction of R point which is used for signal overlapping and other fiducial point extraction as well as subbands processing. Since the overlapping of the ECG signal and averaging can not remove the low varying parts of ECG
noise, we have applied a 0.5Hz high pass filter to remove low varying noise of the ECG signal before overlapping it.
The 0.5Hz high pass digital filter doesn't create significant distortion in ECG, because the first main frequency of pseudo periodic ECG signal for the heart rate of greater than 30/min is above 0.5Hz and the heart rate of less than 30/min rarely occurs in stress ECG. Although power line filter can also be included in subbands processing of filter banks, we didn't include the power line filter in filter banks.
This has two reasons: first we have used adaptive filter for this part, which has low computation overhead and the processing of the downsampled version of the ECG signal in filter bank doesn't bring significant speed benefit and second the output of overlapping algorithm would have little noise.
The output of the power line filter and location of R points are used by overlapping algorithm to generate noise free version of the ECG signal. It is important to mention again that ECG overlapping algorithm such as mean or median or ordered average may distort the ECG signal in the case of arrhythmia; however we use the output of the overlapping algorithm for estimation of noise and selection of parameters.
Using the location of R points, the output of power line filter and overlapped ECG signal, we extract the location of other fiducial points such as Q, S and J points. Then two linear phase and perfect reconstruction filter banks are used to decompose the overlapped signal and output of power line filter into subbands. Then using the location of fiducial points and the signals of subbands, different processing algorithms are applied to subbands signal. Finally the output signal is generated using a synthesis filter. Please note that, we need only one synthesis filter bank and the decomposed overlapped ECG signal is not necessary to be reconstructed again. The details of each algorithm are described in the following subsections.
B. 0.5 Hz High Pass Filter
We have used linear phase FIR filter to implement 0.5 Hz high pass filter. The filter is implemented by subtracting the output of a low pass 0.5 Hz filter from the original signal after the compensation for the delay of the filter. The sampling rate of ECG signal is 1116Hz, therefore the direct implementation of the filter increases the computation overhead of the algorithm significantly. To remedy, the original signal is first filtered and down sampled to lower sampling rate and then 0.5Hz low pass filter is applied and the output is up sampled to original sampling rate again.
This method significantly reduces the computation overhead of the algorithm.
C. Power Line Filter
The method of [5] is utilized for power line noise filtering.
The method is an adaptive filter and the output of the filter
Fig. 4. Outline structure of the proposed algorithm.
ICBME2008
is calculated by subtracting the noise estimate e(n) from signal x(n), where the noise estimate is calculated using the following equations:
) 2 ( ) 1 ( ) (
) 60 2 cos(
2
−
−
−
=
=
n e n
Ne n e
T
N π
(2) where T is the sampling period of the ECG signal. The algorithm also corrects the estimation of the e(n) using the value of f(n) which is calculated using the following equation:
[
( ) ( )] [
( 1) ( 1)]
)
(n = x n −e n − x n− −e n−
f (3)
If f(n)<0, the present noise estimate , e(n), is increased by an increment d. If f(n)>0, the present noise estimate , e(n), is decreased by d.
D. ECG Overlapping Algorithm
For noise estimation of ECG signal and adjusting the parameters of the algorithms in subbands processing. We generate a noise free estimation of ECG signal using signal overlapping. As it is shown in Fig. 4, the output of the overlapping algorithm is decomposed using a filter bank and the noise level estimated for each subband separately.
Since the overlapping algorithms are poor in removing low varying ECG noise, we first apply a 0.5Hz high pass filter and then overlapping algorithm is employed. To filter the ECG signal using overlapping algorithm, the ECG signal are divided into epochs or beats using the location of R points. We denote noisy epochs as:
)]
1 ( ),..., 1 ( ), 0 (
[ −
= xk xk xk L
xk (4)
where k is the epoch index and L is the length of the epoch.
Our algorithm for overlapping the ECG signal and creating a noise free estimate of the ECG signal contains the following steps:
• Constitute the set E containing the epoch I, which we are filtering it and the epochs in the neighborhood of it.
E=
{
xI−N/2,xI−N/2+1,...,xI+N/2}
(5)• Resample the epochs in the neighborhood to have the same length as epoch I.
• Sort the epochs in E
E′=sort(E)=
{
x′I−N/2,x′I−N/2+1,...,x′I+N/2}
(6) wherexI′−N/2(n)≤xI′−N/2+1(n)≤...≤xI′+N/2(n) (7)
• Discard bottomαpercent and top αpercent of sorted epochs.
{ }
) ( 2
/
) ( 2
/ ,...,
, 1
N round N
I K
N round N
I J
K J
J
×
− +
=
× +
−
=
′
′
= ′
′′ +
α α x x
x E
(8)
• Calculate the output using Gaussian weighted averaging of epochs inE′′.
∑
=
+ ′
= −
KJ i
i i
k
w x n
J n K
y ( )
1 ) 1
(
(9)where wi are Gaussian coefficients.
E. R Detection and Fiducial Points Extraction
The overlapping algorithm needs to detect R points to divide ECG signal into epochs. We need to extract the location Q, S and J points for subbands processing as well.
We use the method of [36] to detect R points. However we applied some changes to this algorithm to enhance the efficiency of the algorithm in noisy ECG signal. These changes include the adaptively selection of the threshold, proper selection of search region and using QS region width to verify R points.
When the R points are located, other fiducial points are extracted using the location of R points. We search 100ms before the R point for extraction of Q point and 100ms after R point for extraction of S point. The area after S point is also searched for J point. We used curvature scale space method for the extraction of these points. The curvature of two dimensional signals is given by :
2 / 3 2
2 ( , ) )
) , ( (
) , ( ) , ( ) , ( ) , ) (
,
( σ σ
σ σ
σ σ σ
κ X n Y n
n Y n X n
Y n n X
n n
n nn
nn n
+
= − (10)
It can be shown that in the case of one dimensional signal, the Eq. (10) is reduced to the following equation.
2 / 3 2 1) ) , ( (
) , ) (
,
( = +
σ σ σ
κ
n X
n n X
n
nn (11)
) , ( ) ( )) , ( ) ( ( ) ,
( σ x n g nσ x n g nσ
n n
Xn ⊗ = ⊗ n
∂
= ∂ (12)
) , ( ) ( )) , ( ) ( ( ) ,
( 2
2 σ σ
σ x n g n xn g n
n n
Xnn ⊗ = ⊗ nn
∂
= ∂ (13)
where the k(n,σ)denote curvature, g(n,σ)is a Gaussian function of width σ and ⊗ represents one dimensional convolution.
Our algorithms for extraction of different fiducial point are similar o each other and only the search area and the sign of curvature may be different. Our algorithm for extraction of fiducial points consists of the following steps:
• Use the output signal of overlapping algorithm in Fig.
4 and calculate curvature in search area using a high Gaussian width.
• Find the location of extermum in the curvature. In the case of multiple extermums select the one which is the closest to previous estimates of the fiducial point.
• Reduce the Gaussian width step by step and recalculate the curvature. Then correct the location of fiducial point by moving it toward the nearest extermum of the curvature.
• Again use a high Gaussian width and recalculate the curvature for the output signal of the power line filter in Fig. 4.
ICBME2008
• Correct the location of fiducial point using the curvature calculated in the previous step.
• Again reduce the Gaussian width step by step and recalculate the curvature for output of power line filter.
Then correct the location of fiducial point by moving it toward the nearest extermum of the curvature.
F. Subbands Processing
We used 36 channel filter bank (M=36) for processing of ECG signal. The sampling frequency of ECG signal acquired using s stress test ECG equipment is 1116Hz.
Therefore by decomposing the ECG signal using the filter bank, 36 subbands are created in frequency ranges of [0Hz- 15.5Hz], [15.5Hz-31Hz], ...., [542.5Hz-558Hz]. Please note that two filter banks is used to decompose both the ECG signal (output of power line filter) and the output of overlapping filter. We applied different processing algorithms to each subband of ECG signal which are as follows:
• Subband 1([0Hz-15.5Hz])
This subband contains most energy of T and P waves.
Baseline drift is the major noise in this subband, which its frequency content may be extended up to 1.5Hz in stress ECG tests. To remove baseline wander in this subband a time varying filter is used, which employs ten lowpass filters ranging from 0.6Hz to 1.5Hz with the steps of 0.1Hz.
The output of one of these filters is used to remove baseline drift, by subtracting it from the original signal, which called current filter. However as mentioned before, digital filters with high -3dB frequency may remove first main frequency of pseudo periodic ECG signal. To remedy, the overlapped ECG signal in subband 1 is subtracted from the ECG signal of subband 1 and the resultant signal are fed to time varying filter. To determine the current filter in time varying filter, the output of current filter is subtracted form the output of 1.5Hz filter and then the averaging filter is applied to absolute value of the differences. Based on the value obtained the current filter is selected, which higher values implies moving to filters with higher -3dB frequency.
• Subband 2([15.5Hz-31Hz])
QRS region including J point is not modified in this subband and non-QRS regions is attenuated by the factor of
α
, whereα
is calculated using the correlation of ECG signal and overlapped signal in each subband. We calculateα
for each ECG signal individually and using the signal in the window centered on the signal position.• Subband 2 to 6([31Hz-96Hz])
Both QRS and non-QRS regions are attenuated by the factor of
k α
, whereα
is calculated using correlation of the ECG and overlapped ECG signal as before and k is a attenuation factor of subbands. We use the values of k=0.8, 0.6, 0.4, 0.2 for subbands of 3 to 6 respectively.G. Remaining Subbands
Since the ECG signal energy in these subbands are not significant we set the signals in these subbands to zero for stress ECG. However for special purposes and in the case rest ECG proper processing can be employed for these subbands as well.
V. EXPERIMENTAL RESULTS
We used 36 channel filter bank (M=36) for processing of ECG signal. We tested our algorithms with two sets of ECG data including ECG signals of MIT-BIH database and ECG data acquired using a 12 channel stress ECG monitoring hardware designed and developed in Teb Tasvir Medical Eng. Company. The sampling rate of ECG signal is 1116Hz. The MIT-BIH ECG signals also resampled at 1116Hz. We used ten ECG traces with approximate 45 minutes of monitoring time, for the evaluation of proposed algorithm. We achieved the correct R detection ratio of 99.5% compared to 97.6% of original R detection algorithm [36]. We compared our Q and S point detection algorithm with the method of [37]. Tables 1 and 2 show the results of Q and S detection using method of [37] and the proposed method. As it is shown in Table 1 and 2, the proposed algorithm has better result.
We also compared the result of proposed noise filtering algorithm with other methods. To asses the efficiency of different algorithms, simulated noise is added to clean data and after applying the noise removing algorithm the SNR improvement, SNRi, in decibel is calculated according to:
⎟⎟
⎟
⎠
⎞
⎜⎜
⎜
⎝
⎛
= −
X X SNRi X
log ˆ
20 (10)
TABLEI
RESULTS OF QDETECTION ALGORITHMS
Recognition error (ms) Proposed method Method of [37]
0 381 313
2 128 94
4 106 108
6 86 148
8 70 79
More than 8 13 42
Total 784 784
TABLEII
RESULTS OF SDETECTION ALGORITHMS
Recognition error (ms) Proposed method Method of [37]
0 412 341
2 107 102
4 97 105
6 88 123
8 63 84
More than 8 17 29
Total 784 784
ICBME2008
where X is the vector of clean data and Xˆ is the enhanced signal and X denotes the standard deviation of vector X.
Higher SNRi shows the better performance of corresponding algorithm. To asses the distortion resulted from different algorithms, we also calculate the output of different algorithm to clean data and calculate the average root mean square difference as follow:
L i y i x RMSD
L i
2 1
)) ( ) (
( −
=
∑
=(11) where x(i) are clean ECG data with length of L and y(i) are the output of algorithms for clean data. The small value RMSD represents that the corresponding algorithm create little distortion in ECG signal. Table III and IV compare the performance of various noise removing algorithms for different levels of EMG noise. As it is shown in these tables the proposed filter bank method has less distortion (RMSD) and higher SNRi.
Figure 5 depicts the performance of proposed algorithm for canceling of baseline wander.
The efficiency of proposed algorithm for removing power line noise is shown in Fig. 6. As it is obvious from this figure there is no ringing after QRS region and original signal completely has been recovered.
The experimental results showed that the proposed algorithm is robust and can be employed in stress ECG monitoring equipments.
VI. CONCLUSION
In this paper we proposed new methods for noise canceling of ECG signal and fiducial points extraction as well. We used linear phase filter bank for noise suppressing, which different noise removing algorithms are
applied to subbands of filter bank. Noise removing algorithms are adaptive and the parameters of algorithms are calculated by comparing noisy signal with overlapped signal in each subband. One of the advantages of the filter banks is the applying of time and frequency domain processing simultaneously. To use time domain information wee need to extract fiducial points of ECG signal. To extract fiducial points we used a new scale space curvature based algorithm. We applied the proposed algorithm to various ECG signal including MIT-BIH data base. We also employed the proposed algorithm for practical stress ECG monitoring using ECG hardware designed and implemented in Teb Tasvir Medical Eng.
Company and several physicians approved the efficiency of the algorithms. The comparison of the method with those of other methods showed that more proper results could be obtained using the proposed methods.
ACKNOWLEDGMENT
This research was conducted at research and development (R&D) department of Teb Tasvir Medical Engineering Company. The author would like to thank Eng. Saied Yazdani Moghaddam the president of the company for the support of this research.
REFERENCES
[1] B. Widrow, J. R. Glover, J. M. McCool, J. Kaunitz, C. S. Williams, R. H. Hearn, J. R. Zeidler, E. Dong, Jr., and R. C. Goodlin,
"Adaptive noise cancelling: Principles and applications," Proc.
IEEE, vol. 63, Dec. 1975, pp. 1692-1716.
Fig. 6. Power line noise filtering (a) original signal (b) signal after noise addition (c) filtered signal
Fig. 5. Baseline noise filtering (a) noisy signal (b) filtered signal TABLEIII
SIGNAL TO NOISE ENHANCEMENT(SNRI) OF DIFFERENT ALGORITHMS FOR GAUSSIAN NOISE WITH DIFFERENT SNR
Algorithm SNR=10dB SNR=5dB SNR=0dB
Mean 4.681 4.213 4.142
Median 2.827 2.616 1.221
Proposed overlapping filter
26.32 23.374 19.891
Adaptive filter[27] 19.452 18.482 17.531
SCF[30] 13.7474 11.4794 9.4732
Morphology[29] 22.47 20.051 17.306
Proposed Filter bank 32.341 30.85 28.95
TABLEIV
RMSD OF DIFFERENT ALGORITHMS
Algorithm RMSD Mean 0.0244
Median 0.0293
Proposed overlapping filter
0.0058
Adaptive filter[27] 0.019
SCF[30] 0.0078
Morphology[29] 0.0130
Proposed Filter bank 0.0029
ICBME2008
[2] Y. Z. Ider and H. Koymen, "A new technique for line interference monitoring and reduction in biopotential amplifier." IEEE Trans.
Biomed. Eng., vol. 37, 1990, pp. 624-631.
[3] S.-C. Pei and C.-C. Tseng, "Elimination of AC interference in electrocardiogram using IIR notch filter with transient suppression,"
IEEE Trans. Biomed. Eng., vol. 42, no. 7, 1995, pp. 1128-1132.
[4] J. R. Glover, "Comments on digital filters for real-time ECG signal processing using microprocessors," IEEE Trans. Biomed. Eng., vol.
BME-34, 1987, pp. 962.
[5] M. L. Ahlstorm and W. J. Tompkins, "Digital filters for real-time ECG signal processing using microprocessors," IEEE Trans. Biomed.
Eng., vol. BME-32, 1985, pp. 708-713.
[6] P. S. Hamilton, "A comparison of adaptive and nonadaptive filters for reduction of power line interference in the ECG," IEEE Tran.s Biomed. Eng. vol. 43, 1996, pp. 105-109.
[7] A.K. Ziarani AND A. Konrad, "A nonlinear adaptive method of elimination of power line interference in ECG signals," IEEE Tran.s Biomed. Eng. vol. 49, Jun 2002, pp. 540-547.
[8] L. G. Zhou, H. J. Li, and G. Hu, "Real time base-line drift correction and P-wave detection of ECG signal," in Proc. of IEEE International Conference of Engineering in Medicine and Biology Society, Seattle, WA, USA, 1989, pp. 49-50,
[9] F. Gritzali, G. Frangakis and G. Papakonstantinou, “An automatic baseline drift correction method,” in Proc. of Computers in Cardiology conference , Jerusalem, Sep. 1990, pp. 265-267.
[10] S. Canan, Y. Ozbay and B. Karlik, “A method for removing Low varying frequency trend from ECG signal,”in Proc. of 2nd International Conference on Biomedical Engineering Days, Istanbul, Turkey, 1998, pp.144-146.
[11] J. A. van Alsté , W. van Eck and O. E. Herrmenn, "ECG baseline wander reduction using linear phase filters," Computers and Biomedical Research, vol.19 no. 5, Oct. 1986, p.417-427.
[12] J. R. Gradwohl, E. W. Pottala, M .R. Horton and J. J. Bailey,
“Comparison of two method for removing baseline wander in the ECG,” in Proc. of Computers in Cardiology Conference,
Washington, DC, Sep. 1988, pp. 493-496.
[13] L Sornmo, “Time-varying filtering for removal of baseline wander in exercise ECGs,” in Proc. of Computer in Cardiolology, Los Alamitos, CA 1991, pp. 145–148.
[14] S.V. Pandit, “ECG baseline drift removal through STFT,” in Proc.
of 18th Annual International Conference Bridging Disciplines for Biomedicin", Amsterdam, Oct. 1996, pp. 1405-1406.
[15] S. H. Oguz and M. H. Asyali, “A morphology based algorithm for baseline wander elimination in ECG records,” in Proc. International Conf. Biomedical Engineering Days,Turkey, 1992, pp. 157-160.
[16] P. Sun, Q. H. Wu, A. M. Weindling, A. Finkelstein and K. Ibrahim,
"An improved morphological approach to background normalization of ECG signals," IEEE Trans. Biom.Eng, vol. 50, 2003, pp. 117-121.
[17] R. Jane, P. Laguna, N. V. Thakor and P. Caminal, “Adaptive baseline wander removal in the ECG: comparative analysis with cubic spline technique," in Proc. of Computer in Cardiology, 1992,pp.143-146.
[18] O. Sayadi, and M. B. Shamsollahi, Multiadaptive “Bionic Wavelet Transform: Application to ECG Denoising and BaselineWandering Reduction, EURASIP Journal on Advances in Signal Processing, vol. 2007, Article ID 41274, 11 pages, 2007..
[19] B. Weng, M. Blanco-Velasco and K. E. Barner, "Baseline wander correction in ECG by the empirical mode decomposition," in Proc. of the IEEE 32nd Annual Northeast Bioengineering Conference, 2006, pp. 135-136.
[20] M. E. Rahman and M. A. Haque, "Filter based enhancement of QRS complex of electrocardiogram," in Proc IEEE Pacific Rim Conf. on Communications, Computers and signal Processing,2003, pp. 1008- 1011.
[21] X. Hu and V. Nenov, "A single-lead ECG enhancement algorithm using a regularized data-driven filter," IEEE Trans. Biomed. Eng., vol. 53, no. 2, 2006, pp. 347-351.
[22] C. Patil, A. Stephans and A. Warner, "A novel approach to ECG signal averaging in exercise testing" IEE Colloquium on Technological Progress in Cardiology,1990, pp. 8/1-8/3.
[23] V. X. Afonso, W.J. Tompkins, T.Q. Nguyen, K. Mickler and S. Lou,
“Comparing stress ECG enhancement algorithms with an introduction to a filter bank based approach,”IEEE Engineering in medicine and biology magazine, vol. 15, 1996, pp. 37-44.
[24] A. A. Hiasat, M. M. Al-Ibrahim and K. M. Gharaibeh, "Design and implementation of a new efficient median filtering algorithm," in Proc. IEE Vision, Image, and Signal Processing, vol. 146, 1999, pp.
273-285.
[25] J. Ramous, and Y. Pallas-Areny, “ECG noise reduction by ordered averaging,” in Proc. 18th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 1996, pp. 960- 961.
[26] Y. Bensadoun, E. Novakov,abd K. Raoof, ” Multidimensional adaptive method for canceling EMG signal from the ECG signal,” in Proc. 17th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 1995, pp. 173-174.
[27] R. Kepski, J. Cytowski, T. Buchner and L. Malecka, "Adaptive filtering in dynamically changing high resolution ECG," in Proc.
International Conf. Information, Communication and signal processing, Singapore, 1997, pp. 1216-1220.
[28] P. Raphisakl, S. C. Schuckers1 and J. Curry, "An algorithm for EMG noise detection in large ECG data," in Proc. IEEE Computers in Cardiology conference,2004, pp. 369- 372.
[29] C. H. Chu and E. J. Delp, “Impulsive noise suppression and background normalization of electrocardiogram signals using morphological operators,”IEEE Trans. Biomed. Eng., vol. 36, 1989, pp. 262-273.
[30] DW. Morata, “Source consistency filtering –A new tool for ECG noise reduction” in Proc. IEEE Computers in Cardiology conference, 1991, pp. 125-128.
[31] V. X. Afonso, W.J. Tompkins, T.Q. Nguyen, S. Trautmann, and S.
Lou, “Filterbank based processing of the stress ECG,” IEEE-EMBC and CMBEC, vol. 4, 1995, pp. 887-888.
[32] A. Elloum, Z. Oichiri and N. Ellouze, "Denoising ECG contaminated with EMG components baesd on pitch synchronous wavelet analysis," in Proc. IEEE International Conference on Industrial Technology (1CIT), 2004, pp.1660-1663.
[33] G. M. Friesen, T. C. Jannet and M. A. Jadallah, "A comparison of the noise sensitivity of nine QRS detection algorithms," IEEE Trans.
Biomed. Eng., vol. 37, no. 1, 1990, pp. 85-98.
[34] S. Z. Mahmoodabadi, A. Ahmadian, M. D. Abolhasani, M. Eslami1, and J. H. Bidgoli1, “ECG Feature Extraction Based on Multiresolution Wavelet Transform” Proc. IEEE Engineering in Medicine and Biology 27th Annual Conference, China, 2005.
[35] Y. Chen and H. Duan, "A QRS complex detection algorithm based on mathematical morphology and envelope," in Proc. of the 2005 IEEE Engineering in Medicine and Biology, Shanghai, China, Sep.
2005, pp. 4654-4657.
[36] S.J. Weisner, W. J. Tompkins, and B. M. Tompkins, “A compact microprocessor-based ECG ST-segment analyzer for operating room,”IEEE Trans. on Biomed. Eng., vol. BME-29, no 9, 1982, pp.
642-649.
[37] Y. Suzuki, “Self-Organization QRS-Wave Recognition in ECG Using Neural Networks,” IEEE Trans. Neural networks, vol. 6, no.
6, 1995, pp. 1469-1477.
[38] A. Behrad and K. Faez, "New method for QRS-wave recognition in ECG using MART neural network,"in Proc. 7th Australian and New Zealand Intelligent Information Systems Conference, Australia, 2001.
[39] A. Zifan, Sohrab Saberi, M. H. Moradi, and F. Towhidkhah
“Automated ECG Segmentation Using Piecewise Derivative Dynamic Time Warping” Int. J. of Biomed. Science Vol. 1 No. 3 2006
[40] R. L. de Queiroz, T. Q. Nguyen and K. R. Rao, "The GenLOT:
Generalized linear-phase lapped orthogonal transform," IEEE Trans.
Signal Processing, vol. 44, Mar. 1996, pp. 497-507.
[41] A. K. Soman, P. P. Vaidyanathan and T. Q. Nguyen, "Linear phase paraunitary filter banks: Theory, factorizations and designs," IEEE Trans. Signal Processing, vol. 44, no. 12, Dec.1993, pp. 3480-3496.
