spie proceedings [spie spie defense, security, and sensing - orlando, florida (monday 25 april...

Empirical Mode Decomposition of the ECG signal for NoiseRemoval

Jesmin Khana, Sharif Bhuiyana, Gregory Murphya and Mohammad Alamb

aTuskegee University, Department of Electrical Engineering, 307 Luther H. Foster Hall,Tuskegee, AL 36088, USA

bUniversity of South Alabama, Department of Electrical and Computer Engineering, 307University Blvd., Mobile, AL 36688, USA

ABSTRACT

Electrocardiography is a diagnostic procedure for the detection and diagnosis of heart abnormalities. The elec-trocardiogram (ECG) signal contains important information that is utilized by physicians for the diagnosis andanalysis of heart diseases. So good quality ECG signal plays a vital role for the interpretation and identificationof pathological, anatomical and physiological aspects of the whole cardiac muscle. However, the ECG signalsare corrupted by noise which severely limit the utility of the recorded ECG signal for medical evaluation. Themost common noise presents in the ECG signal is the high frequency noise caused by the forces acting on theelectrodes. In this paper, we propose a new ECG denoising method based on the empirical mode decomposi-tion (EMD). The proposed method is able to enhance the ECG signal upon removing the noise with minimumsignal distortion. Simulation is done on the MIT-BIH database to verify the efficacy of the proposed algorithm.Experiments show that the presented method offers very good results to remove noise from the ECG signal.

Keywords: Electrocardiogram (ECG), ECG denoising, Empirical mode decomposition (EMD), muscle artifact(ma), and electrode motion artifact (em)

1. INTRODUCTION

Electrocardiography (ECG) measures the heart’s electrical activity to help evaluate its function and identify anyproblems that might exist. The ECG can help determine the rate and regularity of heartbeats, the size andposition of the heart’s chambers, and whether there is any damage present. The information in the recordedECG signal is interpreted by a machine and drawn as a graph. The graph consists of multiple waves, whichreflect the activity of the heart. The height, length, and frequency of the waves are read in the following way:

The number of waves per minute on the graph is the heart rate.

The distances between these waves is the heart rhythm.

The shapes of the waves show how well the heart’s electrical impulses are working, the size of the heart, andhow well the individual components of the heart are working together.

The consistency of the waves provides relatively specific information about any heart damage present.

A person’s heartbeat should be consistent and even. ECG signal provides information about abnormallyslow and fast heart rates, abnormal rhythm patterns, conduction blocks (short-circuits of the heart’s electricalimpulses that cause rhythm inconsistencies between the upper and lower chambers). In the past, the ECG wasrecorded on a machine that drew on long strips of paper, with records from each electrode presented in a standardsequence. Now the ECG tracings are stored as computer files that can be preprocessed before the analysis and

Further author information: (Send correspondence to Jesmin Khan)Jesmin Khan: E-mail: [email protected], Telephone: 1 334 727 8987Sharif Bhuiyan: E-mail: [email protected], Telephone: 1 334 727 8989Gregory Murphy: E-mail: [email protected], Telephone: 1 334 727 8995Mohammad Alam: E-mail: [email protected], Telephone: 1 251 460 6117

Invited Paper

Optical Pattern Recognition XXII, edited by David P. Casasent, Tien-Hsin Chao, Proc. of SPIE Vol. 8055805504 · © 2011 SPIE · CCC code: 0277-786X/11/$18 · doi: 10.1117/12.884744

Proc. of SPIE Vol. 8055 805504-1

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 09/17/2013 Terms of Use: http://spiedl.org/terms

the visual interpretation. The analysis and interpretation of the ECG signals evolved from the simple visualinspection to completely automated diagnosis systems1 during the last five decades.

It is very important to be able to extract the parameters of the ECG signal correctly without noise as themorphology of ECG signal has been used for recognizing the heart activity.2 In order to acquire a completepicture and detailed information about the electrophysiology of the heart diseases and the ischemic changes thatmay occur like the myocardial infarction, conduction defects and arrhythmia; the ECG signal must be clearlyrepresented and filtered, to remove out all noises and artifacts from the signal. The main sources of the noiseand the artifacts are: the high-frequency noise such as the electromyography (EMG) noise caused by the muscleactivity and the artifacts caused by the motion of the patient or the leads. In this paper, the goal is to separatethe legitimate ECG signal from the undesired artifacts using EMD, to allow an easy interpretation and diagnosis.

The ECG signal is one of the biomedical signals that is considered as a non-stationary signal and needs acautious processing for the noise removal.3,4 A recent efficient technique for the processing of a non-stationarysignal is the empirical mode decomposition (EMD). The EMD5 is used for the decomposition of a non-linear andnon-stationary signals in the time-frequency domain. It is a preprocessing stage for the efficient computation ofthe instantaneous frequency through the Hilbert transform.6 The EMD is considered as an alternative methodto the wavelet analysis,7 the WignerVille distribution, and the short-time Fourier transform. It is especially wellsuited for the analysis of nonlinear and nonstationary signals, such as biomedical signals. This motivates us touse the EMD for the noise removal from the ECG signal.

The contribution of this work is the use of the EMD to enhance the ECG signal after removing the noise. Theperformance of the proposed algorithm is verified through different experiments performed over the MITBIHarrhythmia database.8 Experiments are carried out both for synthetic and real noise cases. The experimentalresults show that the proposed method is a good tool for ECG signal noise removal. The remainder of thepaper is organized as follows. A brief literature review is presented in Section 2. The EMD technique is brieflydescribed in Section 3. The proposed algorithm for the noise removal is explained in Section 4. Experimentalresults are demonstrated in Section 5. Concluding remarks and recommendations for future improvement aregiven in Section 6.

2. RELATED WORK

In the literature, several methods have been presented to enhance the ECG signal. The most widely used isthe least mean square adaptive algorithm (LMS), such as the adaptive impulse correlated filter (AICF),9,10 thetime sequence adaptive filter (TSAF)11,12 and the signal-input adaptive filter (SIF).13 Additionally, there aresome recent contributions on the ECG signal enhancement using a wide range of different techniques, such as theperfect reconstruction maximally decimated filter banks,14 the nonlinear filter banks,15 advanced averaging,16,17

the wavelet transform,18–22 singular value decomposition,23 and independent component analysis.24

As an effective signal processing tool, the EMD has also been used extensively as reported in the literature.For example, Flandrin et al.25 have applied the EMD for the detrending and denoising of the signal based on thefact that EMD behaves as a wavelet-like dyadic filter bank for fractional Gaussian noise. As an application inthe biomedical engineering, Huang et al. used the EMD for the analysis of the blood-pressure.26 Liang et al.27

presented a work on artifact reduction in electrogastrogram signals based on EMD. The EMD is utilized in28 toextract the lower esophageal sphincter pressure in the gastroesophageal reflux disease. For the analysis of theheart rate variability (HRV), the EMD has been exploited in.29,30 Specifically for the ECG signal processing, theEMD is applied in31 for the investigation of the chaotic nature of the ECG signal and in32 for the enhancementof the ECG signal.

3. EMD OVERVIEW

The empirical mode decomposition (EMD) decomposes a signal into a finite number of (zero mean) frequencyand amplitude modulated signals called intrinsic mode functions (IMFs). The basic idea embodied in the EMDanalysis, as introduced by Huang et al.,26 is to allow for an adaptive and unsupervised representation of theintrinsic components of linear and non-linear signals, based purely on the properties observed in the data withoutappealing to the concept of stationarity.



The aim of the EMD is to decompose the signal into a sum of intrinsic mode functions (IMFs). An IMFis defined as a function with equal number of extrema and zero crossings (or at most differed by one) withits envelopes, as defined by all the local maxima and minima, being symmetric with respect to zero. An IMFrepresents a simple oscillatory mode as a counterpart to the simple harmonic function used in Fourier analysis.The first IMF contains the highest local frequencies of oscillation while the final IMF, or the residue, contains asingle extremum, a monotonic trend, or simply a constant.

As an example, an original 1-D signal/data is being denoted as d(t), which will be decomposed by employingthe EMD process. From the decomposition, consider the ith IMF as fi(t) and the residue as r(t). For the signald(t) there are m data points such that x ∈ 1, 2, . . . ,m. Multiple iterations are required to obtain an IMF, fi(t)in a classical and/or other standard EMD processes. The intermediate state of an IMF in the jth iteration canbe denoted as fi,j(t). In this process, the IMF fi(t) is obtained from its source signal/data si(t). With the abovedefinitions, the basic steps of a classical EMD method can be summarized as

(i) Start with i = 1, and si(t) = d(t). If si(t) (i.e., d(t)) is not the only component, (that means if si(t) is nothaving the residue properties), then go to step (ii).

(ii) Make j = 1, and fi,j(t) = si(t).

(iii) Extract the local maxima points of fi,j(t), which is known as the maxima map and denoted as pi,j(t).Similarly, extract the local minima points of fi,j(t), which is known as the minima map and denoted asqi,j(t). Note that the data points in the maxima and minima maps corresponding to the local maxima andminima points in the intermediate state of an IMF, fi,j(t), will be non-zero; while the other points in themaxima and minima maps will be zero.

(iv) Construct the upper envelope, ui,j(t), and the lower envelope, li,j(t), for the intermediate state of an IMF,fi,j(t), from the maxima points in pi,j(t) and minima points in qi,j(t), respectively.

(v) Find the mean envelope as mi,j(t) =ui,j(t)+li,j(t)

2 .

(vi) Calculate fi,j+1(t) as fi,j+1(t) = fi,j(t)−mi,j(t).

(vii) Check whether fi,j+1(t) follows the properties of an IMF. Although there can be various methodologies toverify the properties of IMF, the most used method is to calculate a deviation error, denoted as δi:j,j+1.

26,33

This deviation error, which is similar to a mean squared error (MSE), can be defined as given below.26,33,34

δi:j,j+1 =

m∑

x=1

|fi,j+1(t)− fi,j(t)|2|fi,j+1(t)|2 , (1)

If the calculated deviation error is within the deviation error threshold, δT , then the corresponding IMF isassumed to have the required IMF properties. Normally, a low deviation error threshold, δT (e.g., below0.5) is chosen to ensure nearly zero envelope mean of the IMF.

(viii) If fi,j+1(t) satisfies the IMF properties as per step (vii), then take fi(t) = fi,j+1(t); set si+1(t) = si(t)−fi(t),and i = i+ 1; and go to step (ix). Otherwise, set j = j + 1, go to step (iii) and continue up to step (viii).

(ix) Find out the number of extrema points (maxima and minima together), denoted as εi, in si(t). If εi is lessthan the extrema threshold, εT , the residue, r(t) = si(t); and the decomposition is complete. Otherwise,go to step (ii) and continue up to step (ix). The most appropriate value for the extrema threshold for 1-Dsignal, εT , is 2.

The above process of extracting the IMFs one by one is also called sifting. It may be mentioned that thesecond and subsequent source signal/data are also known as intermediate residues, ri(t).

26,35 Accordingly, ifthere are K IMFs in a decomposition, then si+1(t) = ri(t) for i ∈ 1 : K and rK(t) = r(t). For the convenienceof discussion, the first IMF, f1(t), is called IMF-1; the second IMF, f2(t), is called IMF-2, and so on. If no



0 200 400 600 800 1000 1200 1400 1600 1800 2000

f 1(t)

f 2(t)

f 3(t)

f 4(t)

f 5(t)

f 6(t)

f 7(t)

f 8(t)

f 9(t)

f 10(t)

f 11(t)

f 12(t)

f 13(t)

Figure 1. The EMD of a clean ECG Signal. Top plot is for original signal. From 2nd at the top to the bottom are theIMFs.

processing is done on the empirical mode functions, the summation of all the empirical mode functions and theresidue returns the original data/signal back as given by

K∑

i=1

fi(t) + r(t) = d̃(t), (2)

where, ideally, d̃(t) = d(t), but practically, these two quantities will differ only by the round-off and truncationerrors introduced by the manipulating software and/or computer. The EMD separates the intrinsic signalcomponents gradually from the highest to the lowest local time variation/oscillation. If the EMD is interpretedas a time-scale analysis method, lower-order IMFs and higher-order IMFs correspond to the fine and coarsescales, respectively. The residue itself can also be regarded as an IMF, which brings some convenience in themathematical representation of the EMD.

4. NOISE REMOVAL FROM ECG

The EMD decomposes a signal into IMFs with decreasing frequency content. By the convention, the lower-order IMFs capture fast oscillation modes while higher-order IMFs typically represent slow oscillation modes.Following the procedure of the EMD as described in Section 3, the EMD of clean and noisy ECG signals areillustrated in the following two examples, thus revealing specific patterns associated with the QRS complex andnoise in the EMD domain. At first the EMD of a clean ECG signal (first lead of record 103) from the MITBIHarrhythmia database is shown in Fig 1. In this figure, the original signal is at the top plot, and the remainingplots show all the IMFs from low to high orders. It can be seen from Fig 1 that the frequency content of eachIMF decreases as the order of IMF increases. Also the first three IMFs have the oscillatory patterns similar tothe QRS complex, which has strong high-frequency components.

Next the EMD of a noisy ECG signal is shown in Fig 2. In this figure, the noisy signal is at the top plot,which is obtained by adding Gaussian noise to the clean signal in Fig 2. The IMFs of the noisy signal are



0 200 400 600 800 1000 1200 1400 1600 1800 2000

f 1(t)

f 2(t)

f 3(t)

f 4(t)

f 5(t)

f 6(t)

f 7(t)

f 8(t)

f 9(t)

f 10(t)

f 11(t)

Figure 2. The EMD of a noisy ECG Signal. Top plot is for original noisy signal. From 2nd at the top to the bottom arethe IMFs.

shown in the remaining plots of Fig 2. Compared to the clean signal case, the first IMF of the noisy signalcontains strong noise components. The second and the third IMFs have the oscillatory patterns similar to theQRS complex. The above analysis of EMD on clean and noisy ECG indicates that it is possible to removethe noise and at the same time preserve the important signal information by processing the IMFs before thereconstruction.

Considering the noise as high-frequency signal the noise removal from the ECG signal is carried out basedon the facts that the noise components lie in the first several IMFs, the most ECG signal power is concentratedin lower frequencies (higher order IMFs), the QRS complex spreads across the mid- to high-frequency bands.However, these facts, as well, complicates ECG signal denoising. Since, simple low-pass filtering or simplyremoving lower order IMFs in the reconstruction of the original signal will introduce severe signal distortion (forexample R-wave amplitude attenuation and QRS complex distortion) and thus loss of important informationfrom the signal; we have to take special care in processing the IMFs to remove noise. As most of the noisecomponents are contained in the lower order IMFs, we can remove the noise from those lower order IMFs beforesumming up all the IMFs in the reconstruction of the signal. Therefore, it is essential to find the order of theIMF up to which most of the noise components are contained in. Knowing the order, the IMFs correspondingto the noise are processed first and then the reconstruction of the original signal is obtained by summing up theclean IMFs and the remaining IMFs.

Now before the application of any noise removal algorithm, we must know what essential information mustbe preserved in the reconstructed ECG signal. As shown two pulses from an ECG signal in Fig 3, the followinginformation are considered as important by the doctors:

1. The fiducial points: P, Q, R, S, T as shown in Fig 3.

2. The R-R interval: R1R2, R2R3, . . . in Fig 3.

3. The QRS complex duration: Q1S1, Q2S2, . . . in Fig 3.



Figure 3. A synthetic ECG signal showing the fiducial points, QRS complex and R-R interval.

0 200 400 600 800 1000 1200 1400 1600 1800 2000

0 200 400 600 800 1000 1200 1400 1600 1800 2000

Figure 4. R-points in the clean and noisy ECG signal as shown by the dash-dotted lines.

So in order to remove noise and at the same time to preserve the important information, we follow thefollowing procedures in our proposed algorithm:

Step 1: Find the points R1, R2, . . . . Finding those points are not difficult even for the noisy ECG signal asthe R-waves are very dominant in the ECG signal. In Fig 4, we have shown the R-points in the clean and noisyECG signals. Once we find those points, we know the R-R intervals and the mid points of the R-R intervals asR2−R1

2 = T1,R3−R2

2 = T2, . . . .

Step 2: Find the duration of the QRS complex Q1S1, Q2S2, . . . . From the experimentation on several cleanand noisy ECG signals, we observe that the first three IMFs contain the most part of the QRS complex oscillation.But we also observe that the first IMF contains most of the noise as well. So, based on the experiments, we havechosen the 2nd and 3rd IMFs to be used for the calculation of the QRS complex duration.

The QRS complex and the oscillatory patterns in the 2nd and 3rd IMFs are illustrated in Fig 5 for bothclean and noisy ECG signals. In these two figures, the ECG signal is plotted in a thick solid line and the thin lineis the sum of the 2nd and 3rd IMFs: d2(t) = f2(t) + f3(t). From this figure, we can see that the QRS complexis bounded by the second zero-crossing points of d2(t) at the left (the Q-points: Q1, Q2, . . . ) and at the right(the S-points: S1, S2, . . . ) from the fiducial point, R; as shown in Fig 5. Even in the noisy case (Fig 5 (b)),this revelation holds, which validates the usage of the 2nd and 3rd IMFs for the determination of QRS complexduration. Fig 6 shows the determined QRS complexes for the whole ECG signal for both noise-free and noisycases.

Step 3: Apply appropriate window functions in some low order IMFs such that those will preserve the QRScomplex, P wave and T wave; and remove the noise. In this step, we also need to find the number of low orderIMFs, where the window function will be applied. In order to find this number, we consider that the windowfunction should be applied to those IMFs that are dominated by noise.

For ECG signals, the contaminating noise is usually zero mean while the signal is nonzero mean. Sincelower-order IMFs contain the noise, we perform a test on the IMFs to determine if a particular IMF has a meangreater than a threshold value, if the m-th IMF is found to have a mean greater than the threshold such that allthe IMFs of order less than m has a mean less than the threshold, then the window function is applied to thefirst (m− 1) IMFs.



400 450 500 550 600 650 700

d2(t)

400 450 500 550 600 650 700

d2(t)

Figure 5. QRS duration determination from d2(t). a) Clean ECG and b) Noisy ECG.

0 200 400 600 800 1000 1200 1400 1600 1800 2000

D(t)

0 200 400 600 800 1000 1200 1400 1600 1800 2000

Dn(t)

Figure 6. Determined QRS duration for the ECG signal. a) Clean ECG and b) Noisy ECG.

In Fig 7, it has been demonstrated how the window functions are chosen with variable widths for the ECGsignal. As the first IMF contains the most noise and the QRS complex oscillation, only one window functionis applied at the QRS complex to preserve it. Fig 7 (a) shows the window used for the first IMF, which isa Rectangular window. The reason for choosing this window is that it is flat over the duration of the QRScomplex to preserve it. Since the window size is equals to the QRS duration determined in the previous step,these window functions adjust their sizes from one pulse to the next for a given ECG signal. The flat region ofRectangular window is τ1 = S1 −Q1.

For the 2nd, 3rd and up to (m − 1)-th IMFs, along with the Rectangular window at the QRS complex, wealso apply two more Nuttall windows at the P- and T-waves. Fig 7 (b) displays the windows used for 2nd, 3rdand up to (m − 1)-th IMFs. The reason for choosing Nuttall window is that it has the similar shape of the P-and T-waves. The right cutoff for the P-wave Nuttall window is Q2 − τ2, where τ2 = S2−Q2

3 ; and the left cutoff

is Q2 − τ2 − τ5, where τ5 is the width for this window and τ5 = (Q2 − S2−Q2

3 )− (Q2−T1

2 ). The left cutoff for theT-wave Nuttall window is S1 − τ3, where τ3 = τ1; the right cutoff is S1 − τ3 + τ4, where τ4 is the width for theT-wave Nuttall window and τ4 = T1 − [S1 + (S1 −Q1)].

5. EXPERIMENTAL RESULTS

In this section we demonstrate the simulation results of the proposed algorithm to remove noise from ECG signal.A noisy segnal Dn(t) is generated by adding noise to the clean ECG signal D(t). Dn(t) = D(t) + n(t); wheren(t) is the noise added to the clean signal. Synthetic noise is added to the signal from the first lead of record 103from the MITBIH arrhythmia database. This signal is chosen because it captures normal sinus rhythms and isreasonably free of noise. Gaussian noise is added to the original clean signal to yieldDn(t). The IMFs of the noisy



�2

�1 �3

R1S1Q1R2S2Q2

T1T1+

2

12 TQ �

�4 �5

�1

R1S1Q1 R2S2Q2T1

Figure 7. a) Window function applied to the 1st IMF and b) window function applied to the 2nd, 3rd . . . (m-1)-th orderIMF.

signal are obtained by applying the EMD. The QRS complex duration is determined from d2(t) = f2(t) + f3(t)as shown in Fig 5 (b). In order to determine the number of lower order IMFs to be multiplied by the windowfunction, the threshold mean is set to be 0.5. In this test, the first six IMFs have been found to have a meanless than the threshold mean. So m− 1 = 6 and we will apply the window function to these first six IMFs.

In applying the window function, the first IMF has a window function W1(t), which is a series of Rectangularwindows centered at the QRS complexes. For the 2nd, 3rd, . . . (m-1)-th IMFs, the window function W2(t) is aseries of Rectangular and Nuttall windows. Finally, the reconstructed signal can be expressed mathematicallyas,

Dr(t) = f1(t)W1(t) +

m−1∑

k=2

fk(t)W2(t) +

K∑

k=m

fk(t) (3)

where, fk(t) is the k-th IMF, K is the total number of IMFs for Dn(t) and Dr(t) is the reconstructed ECG signal.Finally, the reconstructed signal obtained by the above equation is plotted in Fig 8. In this figure, the top plotis the noisy signal, the middle plot is the reconstructed signal. The bottom plot shows both the original signal(solid line) and the reconstructed signal (dash-dotted line). From the result it can be seen that the proposedmethod can remove noise from ECG signal effectively.

Next we do experiment with ECG signal corrupted by real noise. Two real noise signals known as electrodemotion artifact (in the MIT-BIH noise stress test database the record ’em’) and muscle (EMG) artifact (in theMIT-BIH noise stress test database the record ’ma’), are added with the clean ECG signal. Fig 9 shows theclean ECG signal at the top, then the ’em’ signal and the third plot from the top is for ’ma’ noise. The bottomplot shows the total noise (the summation of ’em’ and ’ma’ noises) added to the clean ECG signal. The secondplot from the top of Fig 10 shows the noisy ECG signal. Between the two types of noise, the electrode motion,



0 200 400 600 800 1000 1200 1400 1600 1800 2000800

1000

1200

1400

1600

0 200 400 600 800 1000 1200 1400 1600 1800 2000800

1000

1200

1400

1600

0 200 400 600 800 1000 1200 1400 1600 1800 2000800

1000

1200

1400

1600

Figure 8. ECG denoising for Gaussian noise. From top to bottom: a) The noisy signal. b) The reconstructed signal. c)The reconstructed and the clean ECG signal on the same plot.

0 200 400 600 800 1000 1200 1400 1600 1800 2000500

1000

1500

0 200 400 600 800 1000 1200 1400 1600 1800 2000-500

0

500

0 200 400 600 800 1000 1200 1400 1600 1800 2000-100

0

100

0 200 400 600 800 1000 1200 1400 1600 1800 2000-500

0

500

Figure 9. ECG signal corrupted by real noise. a) The clean ECG signal. b) The real ’em’ noise signal. c) The real ’ma’noise signal. d) The total noise added to corrupt the clean ECG signal.

’em’ artifact is generally considered the most troublesome, since it can mimic the appearance of ectopic beatsand cannot be removed easily by simple filters, as can noise of other types.

Now for this real noise removal, we apply the same procedure described above. Fig 10 shows the performanceof the proposed method for real noise removal. The top plot shows the original signal, the second plot is for thenoisy signal, and the third plot displays the reconstructed signals from the proposed EMD-based method. Thebottom plot shows both the original signal (solid line) and the reconstructed signal (dash-dotted line) in thesame plot. The figure shows that the significant noise components are eliminated by the proposed method.

6. CONCLUSION

A method to remove noise from ECG signal using EMD is proposed in this paper. The effectiveness of theproposed algorithm is shown through an experiment on both synthetic and real noisy ECG signal. Resultsindicate that the EMD is an effective enhancement tool for the ECG signal. The techniques used here can beapplied in practical stress ECG tests and long-term Holter monitoring as in these cases strong noise is presentin the recorded ECG.

REFERENCES

[1] Merri, M., Alberto, M., and Moss, A., “Dynamic analysis of ventricular representation duration from 24-hourrecording,” IEEE Transactions on biomedical engineering 40(12) (1993).



0 200 400 600 800 1000 1200 1400 1600 1800 2000800

1000

1200

1400

0 200 400 600 800 1000 1200 1400 1600 1800 2000800

1000

1200

1400

1600

0 200 400 600 800 1000 1200 1400 1600 1800 2000800

1000

1200

1400

1600

0 200 400 600 800 1000 1200 1400 1600 1800 2000800

1000

1200

1400

1600

Figure 10. ECG denoising for real noise. From top to bottom: a) the clean ECG signal. b) The noisy signal. c) Thereconstructed signal. d) The reconstructed and the clean ECG signal on the same plot

Proc. of SPIE Vol. 8055 805504-10


[2] Mainadri, L., Bianchi, A., Baselli, S., and Cerutti, “Pole-tracking algorithms for the extraction of time-variant heart rate variability spectral parametrs,” IEEE Transactions on biomedical engineering 42(3),20–31 (1995).

[3] Shrouf, A., The linear prediction methods analysis and compression, PhD thesis, Slask Technical Universityin Gliwice (1994).

[4] Donoho, D. L., “De-noising by soft thresholding,” IEEE Transaction on Information Theory 41, 613–627(May 1995).

[5] Huang, W., Shen, Z., Huang, N., and Fung, Y., “Engineering analysis of biological variables: an example ofblood pressure over 1 day,” in [Proc. Nat. Acad. Sci. ], 48164821 (1998).

[6] Huang, N. and Attoh-Okine, N., [The HilbertHuang Transform in Engineering ], CRC Press in N. Huang,N. Attoh-Okine (Eds.), Boca Raton, FL (2005).

[7] Flandrin, P., Rilling, G., and Goncalves, P., “Empirical mode decomposition as a filter bank,” IEEE Trans-actions on Signal Processing 11(2), 112–114 (2004).

[8] Goldberger, A., Amaral, L., Glass, L., Hausdorff, J., Ivanov, P., Mark, R., Mietus, J., Moody, G., Peng,C., and Stanley, H., “Physiobank, physiotoolkit, and physionet: components of a new research resource forcomplex physiologic signals,” Circulation 101 23, 215–220 (2000).

[9] Laguna, P., Jane, R., and Meste, O., “Adaptive filter for event-related bioelectric signals using an impulsecorrelated reference input: Comparison with signal averaging techniques,” IEEE Transaction on BiomedicalEngineering 39, 1032–1044 (October 1992).

[10] Slock, D., “On the convergence behavior of the lms and the normalized lms,” IEEE Transaction on SignalProcessing 41, 2811–2825 (September 1993).

[11] Ferrara, E. and Widrow, B., “The time-sequenced adaptive filter,” IEEE Transaction on Acoustics, Speechand Signal Processing 29, 679–683 (June 1981).

[12] Ferrara, E. and Widrow, B., “Fetal electro-cardiogram enhancement by time-sequenced adaptive filtering,”IEEE Transaction on Biomedical Engineering 29, 458–460 (June 1982).

[13] Almenar, V. and Albiol, A., “A new adaptive scheme for ecg enhancement,” Signal Processing 75, 253–265(1999).

[14] Afonso, V., Tompkins, W., Nguyen, T., Michler, K., and Luo, S., “Comparing stress ecg enhancementalgorithms,” IEEE Eng. Med. Biol. Magazine 15(3), 37–44 (1996).

[15] Leski, J. and Henzel, N., “Ecg baseline wander and powerline interference reduction using nonlinear filterbank,” Signal Processing 35(4), 781793 (2004).

[16] Iravanian, S. and Tung, L., “A novel algorithm for cardiac biosignal filtering based on filtered residuemethod,” IEEE Trans. Biomed. Eng. 49(11), 13101317 (2002).

[17] Leski, J., “Robust weighted averaging,” IEEE Trans. Biomed. Eng. 49(8), 796804 (2002).

[18] Tikkanen, P., “Nonlinear wavelet and wavelet packet denoising of electrocardiogram signal,” Biol. Cy-bern. 80(4), 259267 (1999).

[19] Ho, C., Ling, B., Wong, T., Chan, A., and Tam, P., “Fuzzy multiwavelet denoising on ecg signal,” Electron.Lett. 39(16), 11631164 (2003).

[20] Ercelebi, E., “Electrocardiogram signals de-noising using lifting-based discrete wavelet transform,” Comput.Biol. Med. 34(6), 479493 (2004).

[21] Poornachandra, S. and Kumaravel, N., “Hyper-trim shrinkage for denoising of ecg signal,” Digital SignalProcessing 15(3), 317327 (2005).

[22] Alfaouri, M. and Daqrouq, K., “Ecg signal denoising by wavelet transform thresholding,” American Journalof Applied Sciences 5(3), 276–281 (2008).

[23] Paul, J., Reddy, M., and Kumar, J., “Data processing of stress ecg using discrete cosine transform,” Comput.Biol. Med. 28(6), 639658 (1998).

[24] Barros, A., Mansour, A., and Ohnishi, N., “Removing artifacts from electrocardiographic signals usingindependent components analysis,” Neurocomputing 22, 173186 (1998).

[25] Flandrin, P., Goncalves, P., and Rilling, G., “Detrending and denoising with empirical mode decomposition,”in [Proceedings of XII EUSIPCO 2004 ], (September 2004).

Proc. of SPIE Vol. 8055 805504-11


[26] Huang, N. E., Shen, Z., Long, S. R., Wu, M. C., Shih, H. H., Zheng, Q., Yen, N., Tung, C. C., and Liu,H. H., “The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary timeseries analysis,” in [Proceedings of Mathematical, Physical and Engineering Sciences, Royal Society London ],903–995 (1998).

[27] Liang, H., Lin, Z., and McCallum, R., “Artifact reduction in electrogastrogram based on empirical modedecomposition method,” Med. Biol. Eng. Comput. 38(1), 3541 (2000).

[28] Liang, H., Lin, Q., and Chen, J., “Application of the empirical mode decomposition to the analysisof esophageal manometric data in gastroesophageal reflux disease,” IEEE Trans. Biomed. Eng. 52(10),16921701 (2005).

[29] Echevarria, J., Crowe, J., Woolfson, M., and Hayes-Gill, B., “Application of empirical mode decompositionto heart rate variability analysis,” Med. Biol. Eng. Comput. 39(4), 471479 (2001).

[30] Balocchi, R., Menicucci, D., Santarcangelo, E., Sebastiani, L., Gemignani, A., Ghelarducci, B., and Varanini,M., “Deriving the respiratory sinus arrhythmia from the heartbeat time series using empirical mode decom-position,” Chaos Solitons Fractals 20(1), 171177 (2004).

[31] Salisbury, J. and Sun, Y., “Assessment of chaotic parameters in nonstationary electrocardiograms by use ofempirical mode decomposition,” Ann. Biomed. Eng. 32(10), 13481354 (2004).

[32] Blanco-Velasco, M., Weng, B., and Barner, K., “Ecg signal denoising and baselinewander correction basedon the empirical mode decomposition,” Computers in Biology and Medicine 38, 1–13 (2008).

[33] Kizhner, S., Blank, K., Flatley, T., Huang, N. E., Petrick, D., and Hestnes, P., “On certain theoretical de-velopments underlying the Hilbert-Huang transform,” in [Proceedings of 2006 IEEE Aerospace Conference ],4738, 1–14 (March 2006).

[34] Bhuiyan, S. M. A., Attoh-Okine, N. O., Barner, K. E., Ayenu-Prah, A. Y., and Adhami, R. R., “Bidimen-sional empirical mode decomposition using various interpolation techniques,” Advances in Adaptive DataAnalysis (2009).

[35] Huang, N. E., Shen, Z., and Long, S. R., “A new view of nonlinear water waves: The Hilbert spectrum,”Annual Rev. Fluid Mech. 31, 417–457 (1999).

Proc. of SPIE Vol. 8055 805504-12


spie proceedings [spie spie defense, security, and sensing - orlando, florida (monday 25 april...

Documents