Comparison of Baseline Wander Removal Techniques...

Research ArticleComparison of Baseline Wander Removal Techniquesconsidering the Preservation of ST Changes in the IschemicECG A Simulation Study

Gustavo Lenis1 Nicolas Pilia1 Axel Loewe1 Walther H W Schulze2 and Olaf Doumlssel1

1Karlsruhe Institute of Technology (KIT) Institute of Biomedical Engineering (IBT) Fritz-Haber-Weg 1 76131 Karlsruhe Germany2EP Solutions SA Rue Galilee 7 1400 Yverdon-les-Bains Switzerland

Correspondence should be addressed to Gustavo Lenis gustavoleniskitedu

Received 12 December 2016 Revised 10 February 2017 Accepted 19 February 2017 Published 8 March 2017

Academic Editor Joao Cardoso

Copyright copy 2017 Gustavo Lenis et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

The most important ECG marker for the diagnosis of ischemia or infarction is a change in the ST segment Baseline wander is atypical artifact that corrupts the recorded ECG and can hinder the correct diagnosis of such diseases For the purpose of findingthe best suited filter for the removal of baseline wander the ground truth about the ST change prior to the corrupting artifact andthe subsequent filtering process is needed In order to create the desired reference we used a large simulation study that allowed usto represent the ischemic heart at a multiscale level from the cardiac myocyte to the surface ECGWe also created a realistic modelof baseline wander to evaluate five filtering techniques commonly used in literature In the simulation study we included a totalof 55 million signals coming from 765 electrophysiological setups We found that the best performing method was the wavelet-based baseline cancellation However for medical applications the Butterworth high-pass filter is the better choice because it iscomputationally cheap and almost as accurate Even though all methods modify the ST segment up to some extent they were allproved to be better than leaving baseline wander unfiltered

1 Introduction

Notorious changes in the ST segment (elevation or depres-sion) are the most important ECGmarker when dealing withacute coronary syndrome caused by ischemia or mycardialinfarction [1] Baseline wander is a low frequency artifactin the ECG that arises from breathing electrically chargedelectrodes or subjectmovement and can hinder the detectionof these ST changes because of the varying electrical isoline(Figure 1(a)) The guidelines for diagnosis of myocardialischemia are based on specific small changes in the STsegment Therefore even minor fluctuations in baseline canlead to the decision if a patient is classified as STEMI ornon-STEMI and thus influence the therapeutic approachdramatically [1] This decision is often taken in a preclinicalout-of-hospital scenario with little or no infrastructural noisecontrol A further situation where baseline wander becomescritical for the diagnosis of an ST change is cardiac stresstesting [2] The idea behind this procedure is to measure the

response of the heart to exercise on a treadmill and to studycoronary circulation in comparison to resting conditionsThis test should be terminated immediately if among othercriteria significant ST changes appear in the ECG [3]However the observation of those changes becomes difficultif the ECG baseline is not constant [4]

Hence baseline drift must be removed and as a matterof fact it is a standard signal processing step in many devicesor postprocessing algorithms [5ndash7] The majority of baselinewander removal techniques have in common that they cancelthe low frequency components of the ECG signal By doingso they can also alter the ischemia-induced changes in theST segment and compromise its clinical relevance [8] Anexample of this kind of behavior can be seen in Figure 1(b)A recorded ECG with a clear ST depression is filtered witha high-pass filter diminishing that depression This is thereason why recommendations have been published statingthat high-pass filters with a cut-off frequency not higher than005Hz should be used for these kind of applications [9] Yet

HindawiComputational and Mathematical Methods in MedicineVolume 2017 Article ID 9295029 13 pageshttpsdoiorg10115520179295029

2 Computational and Mathematical Methods in Medicine

minus2minus15minus1

minus050

051

Am

plitu

de (m

V)

2 4 6 8 10 12 140

Time (s)

(a)

V4 unfilteredV4 filtered

0 35 42 32515105Time (s)

minus05

005

115

Am

plitu

de (m

V)

(b)Figure 1 (a) ECG recording corrupted by baseline wander Theremoval of this artifact is necessary when diagnosing a change in theST segment Yet the filtering process can modify the signal as seenin (b) (b) ECG recording with a clear ST depression before (blue)and after (red) high-pass filtering The ST depression is reducedbecause of too strong filtering The signals presented in this figurewere retrieved from the Physionet database [20]

it is also known that baseline wandermay be located at higherfrequencies mandating more intrusive filters with highercut-off frequencies [10 11] Thus finding the filters that arecapable of removing baseline wander without compromisingST changes is a relevant task

Even though some interesting approaches to addressthis question have been reported in literature [12ndash14] itis not trivial to perform such a study using only baselinecorrupted ECG recordings Since the ground truth (artifactfree signal) is not known it is hard to objectively evaluatefilter performance based onmodifications of the ST segmentFurthermore they are prone to detection errors from theautomatic signal processing software because some of themethods require for example a precise annotation of the Rpeak or Pwave [15] On the other hand synthetic signals havealso been used for these applications but they might not berealistic enough to reproduce true changes of the ST segment[16] In order to overcome this limitation in silico studiescan be used The simulation of cardiac electrophysiologyallows us to recreate ischemia at a multiscale level We canadjust the models that govern the action potential (AP) inthe cardiac myocyte to represent ischemia-induced changelet the electrical depolarization and subsequent repolariza-tion propagate in the heart and generate the body surfacepotential map on the chest We can then place electrodeson the torso and extract the simulated 12-lead ECG [1718] With this approach we have the advantage of beingable to locate the most relevant fiducial points (eg QRS

complex or ST segment) in the simulated ECG and are thuscapable of performing an analysis that is free of detectionerrors

By adding baseline wander to the ECG we reproduce thesignals in a controlled environment that would have beenrecorded in real life Since the ground truth is also knownthe performance of the filters can be studied Moreoverthe simulation software allows varying the patient geometrythe location and size of the ischemia in the heart andthe electrical properties of the model This leads to a largenumber of possible ECG signals and ST changes

In this study we address the question of the best suitedtechnique for baseline removal without compromising STchanges For that purpose we use a diverse database ofsimulated ECGs with a state-of-the-art electrophysiologicalmodel add synthetic baseline wander and compare theperformance of five manually tuned filters commonly usedin literature Preliminary results on a smaller study includingfewer filters and less signals were presented at the nationalconference of biomedical engineering in Germany [19]

2 Materials and Methods

21 Generation of Simulated Data The database containingthe 3 times 255 = 765 electrophysiological setups used inthis research project was originally created in our groupwith the aim of studying optimal electrode placement todetect ischemia in the ECG [21 22] The data set wascreated using three different subjects Their geometries wereobtained by segmenting the Visible Man dataset and twoothermagnetic resonance images Distinct tissue classes wereused for the main organs including the ventricles skeletalmuscle fat blood lung liver and spleen Fiber orientationin the ventricles was introduced using a rule-based approach[23]The voxel meshes of the three subjects were created withan isotropic side length no greater than 05mm

A cellular automatonwas chosen to perform the ischemiasimulations [24] This is a computationally cheap methodbased on predefined rules In all simulations the electricaldepolarization wave originates from the Purkinje musclejunctions It then propagates by sequentially activating neigh-boring voxels and triggering an AP in each voxel The repo-larization scheme is thus dependent on theAP correspondingto every region of the heart Transmural heterogeneity andanisotropic tissue conductivities inside the ventricular wallswere included in the model [25]

The AP in each voxel was obtained from a monodomainsimulation carried out using the software acCELLerate andutilizing the Ten Tusscher cell model with a basic cycle lengthof 60 beats per minute [26ndash28] The model parameters wereadjusted to differentiate healthy tissue from ischemic regions[29 30] The latter were placed in all 17 AHA segments in theleft ventricle and varied in size to include subendocardial andtransmural scenarios [31]

From every simulation a QRS complex an ST segmentand a T wave were obtained The QT interval was equal to400ms in all simulations The sampling frequency of thesimulated signals was 500Hz but upsampling to 119891119904 = 512Hzwas carried out to facilitate wavelet-based filtering

Computational and Mathematical Methods in Medicine 3

(a) (b) (c)

minus2

minus1

0

1

2

3

4

Am

plitu

de (m

V)

01 02 030Time (s)

(d)

01 02 030Time (s)

minus2

minus1

0

1

2

Am

plitu

de (m

V)

(e)

minus05

0

05

1

15

2

Am

plitu

de (m

V)

01 02 030Time (s)

(f)Figure 2 Demonstration of the three torsomodels and one example per torso of an ischemic ECG (a) (b) and (c) torsomodel and electrodeplacement of the first second and third geometry used in the study (d) Simulated ECG (Wilson lead V5) with ST depression using the torsomodel displayed in (a)This ECG is the result of a transmural ischemia with a radius of 20mm and located in AHA segment 14 (e) SimulatedECG (Einthoven lead I) with ST elevation using the torsomodel displayed in (b)This ECG is the result of a transmural ischemia with a radiusof 25mm in AHA segment 13 (f) Simulated ECG (Einthoven lead II) with ST depression using the third torso model displayed in (c) ThisECG is the result of a transmural ischemia with a radius of 20mm in AHA segment 5

Finally a quasiperiodical extension was carried out tocreate an ECG with a fixed length of 100 s (51200 samples)In order to make the signal more realistic and recreate heartrate variability (HRV) variable RR intervals were addedRR intervals were modeled with a Gaussian distributionhaving an expected value of 1 s and a standard deviation of50msThese parameters are in accordance with normal shortterm HRV values reported in literature [32] Figure 2 showsthe geometry of all three torso models and the electrodeplacement A simulated signal corresponding to differentECG leads can be seen for each geometry A clear ST changecoming from the ischemia is also present

22 Signal Processing Methods

221 Modeling Baseline Wander We modeled baseline wan-der as a linear combination of sinusoidal functions in thefrequency range from 0 to 05HzThe amplitude and phase of

each of the waveforms were chosen randomly to avoid deter-ministic coherence and to make each randomly generatedsignal unique The upper limit of the frequency band was tentimes larger than what is recommended for a high-pass filterfor ST segment analysis This fact in combination with verylow SNR levels of up tominus10 dBmade this artifact a challengingone and allowed for performance ranking among the removaltechniques [33 34]

Mathematically speaking the model was defined as fol-lows

119887 (119905) = 119862 sdot119870

sum119896=0

119886119896 sdot cos (2120587 sdot 119896 sdot Δ119891 sdot 119905 + 120601119896) (1)

with Δ119891 = 119891119904119873 = 001Hz where 119891119904 = 512Hz was thesampling frequency of the signal and 119873 = 51200 was thetotal number of samples in the signal Additionally 119870 =119891119888119891119904 = 50 was the number of sinusoidal functions that


Am

plitu

de (m

V)

1 2 3 40

Time (s)

minus05

0

05

1

15

(a)

05 10

Frequency (Hz)

0

005

01

015

02

Am

plitu

de (a

u)

(b)

0

005

01

015

02

Am

plitu

de (a

u)

20 400

Frequency (Hz)

(c)

Am

plitu

de (m

V)

1 2 3 40

Time (s)

minus05

0

05

1

15

(d)

0

005

01

015

02

Am

plitu

de (a

u)

05 10

Frequency (Hz)

(e)

0

005

01

015

02

Am

plitu

de (a

u)

20 400

Frequency (Hz)

(f)

Figure 3 (a) Exemplary ECG signal corrupted by an arbitrary realization of the baseline wander model The baseline wander artifact can beseen indicated by the red dashed line An SNR of minus3 dB was chosen for this example (b) Frequency spectrum corresponding to the baselinewander artifact presented in (a) (c) Frequency spectrum corresponding to the signal (ECG plus baseline wander) displayed in (a) (d) Adifferent example of an ECG signal corrupted by another realization of the baseline wander model An SNR of +3 dB was chosen for thisexample (e) Frequency spectrum corresponding to the baseline wander artifact presented in (d) (f) Frequency spectrum corresponding tothe signal (ECG plus baseline wander) displayed in (d)

could be placed within the chosen spectrum of the baselineartifact according to the spectral resolution Δ119891 The constant119886119896 was a random number from the uniform distribution[0 1] and 120601119896 was a random phase in the interval [0 2120587) Foreach ECG signal 119909(119905) we repeated the experiment 100 timesgenerating a new realization of 119887(119905) in each repetition Thescaling factor 119862 was used to increase the total power of theartifact and modify the signal-to-noise ratio (SNR) in theexperimentThe series of different SNRwere chosen from theset [minus10 dB minus3 dB 0 dB +3 dB +10 dB +20 dB]

The baseline is then added to the simulated ECG signal

119910 (119905) = 119909 (119905) + 119887 (119905) (2)

Again 119909(119905) is the reference ECG signal while 119910(119905) isthe corrupted one In this study we created 756 simulations12 ECG channels per simulation 100 realizations of thebaseline wander model per channel and six SNR levelsof each baseline wander realization Thus a total of 5508

million signals were used to evaluate the different baselinewander removal techniques including an extensive variety ofscenarios Figure 3 shows two examples on how the baselinewander corrupted ECG signals and their spectra can look likedepending on the different parameters

222 Signal Processing Workflow 119910(119905) was filtered using fivedifferent baseline wander removal techniques After filteringthe reconstructed signal (119905) was compared with the original119909(119905) The reconstruction was evaluated with respect to threequality criteria and the computation time Figure 4 showsa flow diagram of the signal processing algorithm Thecomponents of the flow diagram will be explained in detailin the next sections Figure 4 shows the processing workflowcreated for this study

223 Baseline Removal Methods We compared five state-of-the-art filtering techniques used regularly in literature


ECG simulation and extension

Baseline wander generation

120601kakC

x(t)

b(t)

y(t)+

(a)

x(t)y(t)

x(t)

Statistical evaluation

Quality measures

Baseline removal technique

(b)

Figure 4 Signal processingworkflowused in this study (a)Creationthe ECG and baseline wander artifact and the superposition tocombine them (b) The simulated and extended ECG is added tothe baseline wander artifact to create the corrupted signal The fivebaseline removal techniques are then applied to the corrupted signalto reconstruct the original one To evaluate filtering performancefour different criteria are applied At the end the results arestatistically analyzed to determine the best filtering method

Butterworth high-pass filter [34] moving median and sub-traction [35] spline approximation and subtraction [36]wavelet-based baseline cancellation [37] and wavelet-basedhigh-pass filtering [38] The filter parameters were chosenproperly to be able to remove the given artifact from the ECGsignal A brief introduction to each method is given in thefollowing sections

224 Butterworth High-Pass Filter The first method has theproperty of being simple easy to implement and applicableinmany scenariosTherefore a classic Butterworth high-passfilter with a total order of four and a cut-off frequency of05Hz was chosen Actually the transfer function of the filterhad an order of 2 but the filtering process was performed inforward and reverse direction creating a zero-phase filteredsignal and a resulting order of four [34]

225 Moving Median and Subtraction The second methodhas the goal of estimating the baseline wander using aconcatenation of two moving median filters and subtractingthat estimate from the corrupted signal The moving medianis based on the same principle as the moving average butinstead of the mean the median within a moving windowof a given length is calculated This filter benefits from theassumption that baseline wander and ECG signal have differ-ent amplitude distributions within the moving windowsThefilter is nonlinear making its behavior more complex [35]

The removal technique started with a window length of400ms corresponding to the QT interval Since windows ofthis short duration can deliver an estimation that is a mixtureof true ECG signal and baseline a second moving medianwith a window of a longer length was applied after the firstestimation We chose the second window to be 2 s long By

doing so the complete frequency band of the artifact wasincluded in the baseline estimation

226 SplineApproximation and Subtraction The idea behindthis method was to detect the center of the PQ intervalin every beat and to interpolate those points to create anestimate of the baseline wander This technique assumesthat the PQ interval corresponds to the isoline of the ECGso that a nonzero signal in this interval must be due tobaseline wander Cubic splines were then used to connectthe PQ center points and reconstruct the artifact Finally theestimated baselinewanderwas subtracted from the corruptedsignal to reconstruct the original ECG [36]

227 Wavelet-Based Baseline Cancellation In this methodthe signal was decomposed using the discrete wavelet trans-form (DWT) and the approximation coefficients at thelowest frequency band were set to zero with the aim offully cancelling baseline wander The filtered ECG was thenreconstructed by synthesizing the modified coefficients Thedecomposition level 119871 for the DWT was chosen such thatthe approximation coefficients in that level correspondedto the frequency band where the artifact was located [3739] Since in every decomposition level the ECG signalwas downsampled by a factor of two its Nyquist frequencywas also halved Thus the necessary decomposition level toachieve a cut-off frequency 119891119888 depending of the samplingfrequency 119891119904 can be given by

119891119888 = 1198911199042119871+1 997904rArr

119871 = log2(119891119904119891119888) minus 1 = 9

(3)

Thewavelet used for this procedurewasDaubechies 8 thathas a compact support of 16 samples and is characterized by8 vanishing moments

228 Wavelet-Based High-Pass Filtering The principle beh-ind this method is very similar to the wavelet-based base-line cancellation but a high-pass filtering is used on theapproximation coefficients instead of setting them to zeroThis is somewhat comparable to a soft threshold on theapproximation coefficients For the wavelet decompositionthe Vaidyanathan-Hoang wavelet was used in accordancewith [38] For the high-pass filtering an infinite impulseresponse filter of order one and a cut-off frequency of 05Hzwas used The reconstructed ECG was synthesized from thefiltered wavelet coefficients

229 Quality Criteria Four performance indexes were cho-sen to evaluate the filters regarding the quality of reconstruc-tion and the clinical applicability in terms of simplicity of thealgorithm The four evaluation criteria will be explained indetail

As all filtered signals undergo a transient oscillation attheir boundaries we removed the first and final second(corresponding approximately to one beat) at the beginningand at the end of each signal from the evaluation analysis


2210 Correlation Coefficient With the aim of quantifyingimpairment in the morphology of the reconstructed sig-nals we used the correlation coefficient It is independentfrom scaling or offsetting the signals and focuses on thematching form of original and reconstructed waveformsMathematically speaking the correlation coefficient betweenthe original signal119909(119905) and the reconstructed one (119905) is givenby

CC 119909 (119905) (119905) = 119864 (119909 (119905) minus 120583119909) ( (119905) minus 120583)120590119909120590 (4)

where 119864sdot denotes the expected value operator 120583119909 is theexpected value of 119909(119905) and 120590119909 is its standard deviation

2211 119897 operator The 119897 operator is a measurement of simi-larity that is actually based on the Euclidian distance betweenthe two signals Mathematically speaking it is defined asfollows

119897 operator 119909 (119905) (119905) = 1 minus 119864 (119909 (119905) minus (119905))2119864 1199092 (119905) + 119864 2 (119905) (5)

The 119897 operator has the advantage of delivering values inthe interval [minus1 +1] It is equal to +1 only if the two signals119909(119905) and (119905) are perfectly equal In contrast to the correlationcoefficient the 119897 operator is sensitive to offsetting and scalingof any of the two signals It was originally introduced by ourgroup to quantify changes in the morphology of the T wave[40]

2212 Deviation of the ST Change For the clinical diagnosisof ST changes thresholds for deviations in the J point ofthe ECG have been recommended [41] Depending on theECG lead patient gender and age the ST changes can bediagnosed if a decrease of as low as 50 120583V is observedThus it is necessary to preserve the original J point value asunchanged as possible after the baseline wander artifact hasbeen removed However it is not always trivial to detect thispoint using automatic signal processing algorithms [42] Inorder to avoid possible detection errors and incorrect resultswe defined a new point inside the ST segment the K point(KP) that can be used in samemannerTheKPwas defined asthe time step for which the envelope (absolute value) amongall ECG leads in a simulation is minimal Mathematicallyspeaking it is given by

KP = min119905119878119879

max119894

1003816100381610038161003816119909119894 (119905ST)1003816100381610038161003816 (6)

In this case the signal 119909119894(119905) denotes the ECG recordedfromone of the 12 standard leads (119894 = 1 2 3 12)The timeinterval 119905ST corresponds to the duration of the ST segmentA complete description of the KP can be found in [21] Thedeviation in KP is then defined as

ΔKP = KPfiltered minus KPoriginal (7)

2213 Computation Time In a clinical environment thecomputation time plays an important role if a fast diagnosisshould be delivered by the physician Faster computationtimes also correlate with simpler algorithms easier to imple-ment for portable or stationary clinical devices Thus thecomputation time needed to process each signal was thefourth performance index The computer used to run thecalculations has a 24GHz Intel Xeon E5645 processorwith 12cores and 64GB of RAM running MacOS and Matlab 2016a

23 Statistical Analysis For the statistical analysis we com-pared the performance of the five baseline removal tech-niques applied on the complete data set with respect to thefour quality criteria mentioned previously In addition wealso quantified all performance indexes for the case that nofilter was applied The idea behind this comparison is to finda ldquoclear winnerrdquo among the filtering techniques that can beapplied in many different scenarios

We hypothesize that the best baseline removal techniquefor a given performance index was the one with the bestmedian This hypothesis was accepted if the statistical distri-bution of that performance index is significantly higher thanall other methods For statistical significance testing we usedthe Wilcoxon signed rank test and a level of significance (119901value) of 5This statistical test is a paired unparametric testcommonly used in the field of signal processing [43ndash45]

When comparing a candidate for best performing filter toall other filters a total of five119901 values are calculatedHoweveronly the lowest 119901 value is considered for further analysis Ifthat 119901 value is below the level of significance the filteringtechnique is leveled as the ldquoclear winnerrdquo

3 Results

31 Performance Evaluation Based on Quality Criteria Wepresent the results of the simulation study in the form ofboxplots and a summarizing table Figures 5(a)ndash5(d) showthe boxplots corresponding to the performance index valuesfor all signals Table 1 contains the median and interquartilerange of each quality criterion and filtering technique The 119901value displayed in the last column of Table 1 is the highest oneamong all five comparisons

According to the chosen evaluation scheme the methodthat best maintained the original ECG morphology (highestcorrelation coefficient) was the wavelet-based baseline can-cellationwith amedian and interquartile range (MED plusmn IQR)of 0993 plusmn 0019 Statistical testing proved that this methodwas indeed a clear winner with a 119901 value lt 10minus6 Secondthe method that restores the signal closest to the original one(highest l operator) was also the wavelet-based cancellationwith a MED plusmn IQR of 0993 plusmn 0018 and a significant 119901 valuelt 10minus6 Third the method for which the ST changes undergothe lowest modifications (KP deviation) was again wavelet-based baseline cancellation with a MED plusmn IQR of 000m119881 plusmn0042m119881 and 119901 value lt 10minus6 Statistical testing proved thatthis method was again a clear winner in this category Lastbut not least the fastest method (computation time) was theButterworth filter with a MED plusmn IQR of 0006 s plusmn 0001 s


03

04

05

06

07

08

09

1

Cor

relat

ion

valu

e

Med

ian

Splin

e

Butte

rwor

th

Wav

elet c

anc

Wav

elet H

P

No

filte

r

(a)

minus02

0

02

04

06

08

1

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

l_op

erat

or v

alue

(b)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

minus08

minus06

minus04

minus02

0

02

04

06

08

K po

int d

evia

tion

(mV

)

(c)

0

05

1

15

2

25

Com

puta

tion

time (

s)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

(d)

Figure 5 Boxplots displaying the results of performance evaluation of the filtering techniques (a)Correlation coefficient between original andfiltered signal (b) l operator between original and filtered signal and (c) KP deviation between original and filtered signal For visualizationpurposes some outliers are not displayed in the figure (d) Computation time needed to filter each signal

Again statistical testing proved this method to be speedierthan all others with a 119901 value lt 10minus6

Finally some particular examples showing how the fil-tered signals compare to the original ones are shown inFigure 6

4 Discussion

In the simulation study we saw that the wavelet-basedbaseline cancellationwas the best performingmethod achiev-ing the highest median and lowest IQR for the correlation


Table 1 Summary of the results obtained for the performance evaluation among the filtersThe values are given as MED plusmn IQRThe 119901 valuesdisplay the statistical significance of the best performing filter (bold numbers) in each category

Filterindexes Correlation l operator KP deviation [mV] Computation time [s]No filter 07794 plusmn 03764 07797 plusmn 04452 00000 plusmn 02802 0 plusmn 0Butterworth 09851 plusmn 00392 09859 plusmn 00372 00010 plusmn 00595 00059 plusmn 00434Median 09904 plusmn 00299 09872 plusmn 00413 minus00033 plusmn 00967 18464 plusmn 00434Spline 09813 plusmn 00486 09716 plusmn 00629 00037 plusmn 00628 00074 plusmn 00434Wavelet cancellation 09928 plusmn 00194 09933 plusmn 00184 00000 plusmn 00419 03892 plusmn 00434Wavelet high-pass 09672 plusmn 00827 09689 plusmn 00816 00000 plusmn 00885 04206 plusmn 00189119901 values lt10minus6 lt10minus6 lt10minus6 lt10minus6

coefficient l operator and KP deviation This result provesthis filter to be not only the most accurate one (MED wasbest) but also the most robust (IQR was lowest) The reasonfor this is probably due to the properties of the chosenwavelet and the relatively high decomposition level Thecombination of these two features matches precisely thetime and frequency domain properties of the artifact In thetime domain the chosen wavelet has 8 vanishing momentsThis means that all polynomials up to a degree of 8 donot correlate with this wavelet and are represented in theapproximation coefficients Since our baseline wander modelis a superposition of sinusoidal functions and they can belocally approximated by a Taylor polynomial of lower degreeit is plausible to believe that the approximation coefficients inthe wavelet transform represent a large portion of the artifactIn addition by a decomposition level of 9 the resultingwavelet filter has a very high order and is able to sharplysplit the spectrum of the corrupted ECG signal at precisely05Hz By setting the resulting approximation coefficients tozero the artifact is cancelled out almost entirely Howeverthis argumentation is only valid if the true ECG signalcannot be locally approximated by a polynomial of order8 or at least not as good as the artifact Furthermore themost relevant spectral components of the pure ECG signalmust be above 05Hz Otherwise the wavelet decompositionwould have combined low frequency ECG components andthe artifact into the approximation coefficients leading to apartial cancellation of the true ECG The results observedhere are also in accordance with what has been reported inliterature [12]

A similar performance in terms of correlation coeffi-cient and l operator was delivered by the median filterHowever the KP deviations had a larger IQR of 967120583Vmaking this method less robust with respect to changesin the ST segment and decrease the certainty of a clinicaldiagnosis [41] In addition the sorting algorithm neededin this technique is computationally intensive leading tolong processing times while filtering the signal This couldbecome amajor drawback in clinical applications In contrastto it the fastest method was the Butterworth filter takingaround 59ms to process the corrupted ECG The speedycomputation can be achieved because the chosen filter orderof two (four in the zero-phase implementation) was relativelysmall The filtering process can be performed very fast as alinear combination of no more than four input and output

signal values for each the forward and the backward filteringprocessMoreover this technique is capable of reconstructingthe original ECG with a median correlation coefficient of0985 and a median l operator of 0986 which ranks thismethod in the third placeMore importantly the Butterworthfilter was the second best performing filter with respect toKP deviations (1 120583119881 plusmn 59 120583119881) This is a surprising result forsuch a simple filter In particular for medical applications thatrequire fast but still accurate signal processing algorithmsthis is the method we would recommend If a zero-phaseimplementation is not required the Butterworth filter canalso be approximated as a finite impulse response filterespecially suited for real time applications

The results demonstrated that even though there weresmall differences among the methods they were all goodperformers in terms of correlation coefficient l operator andKP deviations However such good results are not necessarilyto be expected in real-life applications because a prioriknowledge about the time and frequency domain propertiesof the baseline wander model was used to manually choosefilter parametersThe results are probably overfitted but sinceall filters were adapted to match the baseline wander modelwe assume that the comparability among them should stillbe given Thus the ranking of the filter techniques shouldremain in practice

We also reaffirmed that baseline wander can indeedstrongly affect an ECG The nonfiltered signal had a mediancorrelation coefficient of 0779 and an IQR of the KPdeviation of 2802mV Thus removing the baseline wanderbecomes mandatory to allow any further processing of theECG We also found that even though all methods deliveran improvement the KP deviation after filtering had anIQR of at least 419120583V in every case This means that noneof the filtering techniques is capable of reconstructing theST segment to its exact original shape Hence in clinicalapplications dealing with the diagnosis of ischemia theassessment of the filtered ECG signal has to be carried outvery carefully because the observed ST change might not bethe true one

41 Limitations and Outlook This study ranks popular base-line wander removal techniques using a reference ECG signalthat is free of artifacts but exhibits all the properties ofan ischemic ECG The simulation of realistic ECGs is achallenging task because of the complexity of the underlying


minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

Am

plitu

de (m

V)

02 04 06 080Time (s)

Original signalButterworthMedian

SplineWavelet cancWavelet HP

(a)

02 04 06 080Time (s)

0

05

1

15

2

25

3

35

Am

plitu

de (m

V)



(b)

02 040Time (s)

minus8

minus7

minus6

minus5

minus4

minus3

minus2

minus1

0

1

Am

plitu

de (m

V)



(c)

minus06

minus04

minus02

0

02

04

06

Am

plitu

de (m

V)

01 02 03 04 05 060Time (s)



(d)

Figure 6 Four particular examples showing how the filtered signals compare to the original ones (a) Filtering results for a signal that camefrom the ECG lead I in the first torso model and had an SNR of 0 dB (b) Filtering results for a signal that came from the ECG lead II in thethird torso model and had an SNR of +3 dB (c) Filtering results for a signal that came from the ECG lead V2 in the second torso model andhad an SNR of minus10 dB (d) Filtering results for a signal that came from the ECG lead aVL in the third torso model and had an SNR of minus3 dB


electrophysiological behavior reproduced by the multiscalemodel This model is governed by a large variety of coupleddifferential equations that need to be correctly parametrizedfirst [21] In particular the recreation of the T wave is difficultbecause this wave arises from the heterogeneities in therepolarization of the ventricles and thus the APs have to beadapted to each region of the heart [17] To overcome thisproblem it could be stated that the replication of a referencebeat with ST change from a real recording would be a simplerway of performing the study However this approach has thedrawback that it cannot be assured that the reference beatis completely free of artifacts and should not be used as agolden truth Even if an ECG looks clean to the naked eyelittle low frequency perturbations or a small DC offset isalways present locally Thus it becomes difficult to quantifychanges in morphology because the filter would modify notonly the signal but also the artifacts previously present in itIn addition annotations carried out by a trained physicianwould be necessary to ensure the validity of such a databaseUnder those conditions it is impractical to create a data setwith the same size and diversity as the one presented hereA reduced number of signals would limit the generalizationof the filter ranking A further idea would be the use of anECG synthesizer with adjustable ST change [46 47] Yet thisapproach has the drawback that the user can never be surethat the synthetized ECG is realistic enough to recreate realST changes as they arise from ischemia For these reasons webelieve that the in silico study based on a ischemicmembranemodel delivers a reference signal that is free of artifacts butstill meets the requirements for a realistic ischemic ECG andallows us to generate a large number of scenarios

Even though the total amount of signals used in thisstudy was large there were only three different geometriesand one electrophysiological model (Ten Tusscher) used tosimulate ischemia [26] It would be interesting to use otherelectrophysiological models with other patient geometrieshaving also different fiber orientations in the heart to generatea more heterogeneous data set In addition the atrial activitycharacterized by the P wave can also be included in a futuresimulationThePwavewould affect not only the time domainproperties of the synthesized ECG but also its spectral andstatistical features They pose a further challenge to thebaseline removal methods

A more sophisticated baseline wander model would alsobe of interest It is well-known that respiration cannot onlylead to a floating baseline but it can also modulate the ECGsignal [48 49] Thus a baseline wander model in whichthe artifact affects both the isoline and the amplitude ofthe ECG should be considered in a future research projectAdditionally other spectral properties could also be includedin the artifact Increasing its frequency band up to 3Hz forexample wouldmake the filtering problemmore challengingHowever in those cases the artifact overlaps in the timeand frequency domain with the T wave and the ST segmentThis makes the reconstruction of the original signal verydifficult and a diagnosis is probably no longer possible Otherstudies have investigated how to eliminate baseline wanderwith spectral content up to 08Hz [15 38] In those casesthe spectral content of the artifact starts to overlap with the

dominant frequency of the heart rate in resting conditionsmaking this scenario particularly difficult Thus this kindof removal techniques should only be performed with filtershaving linear phase This was not the case for the majority ofthe filters used here and for that reason we chose our upperfrequency at 05Hz

The filter parameters used in this work were chosen ina heuristic manner with the intention of having a goodperformance for the well-known artifact However they werenot computationally optimized to deliver the best possibleresults For example the order of the Butterworth filterthe length windows of the median filters or the waveletused for the decomposition could all be further optimizedto achieve even better results This optimization processcould also be included in a future work together with othermore less common baseline removal techniques such as theempirical mode decomposition the blind source separationor a Gaussian filter adapted to remove the known spectrumof the artifact [50ndash52]

The use of the KP instead of the J point to evaluate theST change deviations was a strategic decision in order toallow automatized quantification of performance Allowinga trained physician to annotate the J point and perform thestudy with the annotations might deliver different resultsHowever by the large amount of signals simulated a manualannotation becomes unpractical In any case we were consis-tent using the same definition for all filtering techniques andallowing comparability among them

Last but not least we quantified the changes in the STsegment caused by filtering measuring the deviation in KPYet the true clinical impact of the filters on the diagnosisof an ischemia was not studied We did not count howmany ischemia cases would have been missed because of thefiltering processThis question is not easily answered becausethe identification of an ST change can be compromised byother factors besides the filtersThose factors are for examplea large variety of silent ischemia the number of electrodesused in the recording or the placement of the electrodeson the chest of the patient [22] This is the reason whythe sensitivity of the ischemia detection even if a trainedphysician is looking at the ECG can be as low as 45[53] We investigated this limitation in a previous simulationstudy obtaining similar results as the ones observed inclinical studies We also proved that the sensitivity of theischemia diagnosis is strongly dependent on the chosenthreshold defining the ST change [21] In any case it wouldbe interesting for a future project to analyze how these factorsplay together with the baseline removal techniques and howthey all affect the ischemia detection For now it is certainthat the better the filters perform on the corrupted ECG thehigher the chance of a physician making the right decision is

5 Conclusion

In this study we addressed the question of the best suitedtechnique for baseline removal without compromising STchanges in the ischemic ECG For this purpose a largesimulation study with 5508 million signals was carried outThe best performing filter with respect to quality of the


reconstruction turned out to be the wavelet-based baselinecancellation However for medical applications the But-terworth high-pass filter is the better choice because it iscomputationally fast and almost as accurate In addition allthe methods tested proved to be better than leaving baselinewander unfiltered It was also shown that none of methodswas capable of reconstructing the original ECG withoutmodifying the ST segment so the user has to be alwaysvery careful when diagnosing an ST change In future otherbaseline wander models including nonlinear behavior andhigher frequency baseline wander could be used to test themethods in more challenging scenarios

Conflicts of Interest

The authors declare that they have no conflicts of interest

Authorsrsquo Contributions

Gustavo Lenis and Nicolas Pilia contributed equally to thiswork

Acknowledgments

The authors would like to acknowledge the support given bythe Deutsche Forschungsgemeinschaft and the Open AccessPublishing Fund of Karlsruhe Institute of Technology

References

[1] E M Antman D T Anbe P W Armstrong et al ldquoACCAHAguidelines for the management of patients with ST-elevationmyocardial infarctionndashexecutive summary A report of theAmerican College of CardiologyAmerican Heart AssociationTask Force on PracticeGuidelines (WritingCommittee to revisethe 1999 guidelines for the management of patients with acutemyocardial infarction)rdquo Journal of the American College ofCardiology vol 44 no 3 pp 671ndash719 2004

[2] I M Weisman and R J Zeballos Clinical Exercise Testing vol32 KargerMedical and Scientific Publishers Basel Switzerland2002

[3] American College of Sports Medicine ACSMrsquos Guidelinesfor Exercise Testing and Prescription Lippincott Williams ampWilkins 2013

[4] J N Froning M D Olson and V F Froelicher ldquoProblems andlimitations of ECG baseline estimation and removal using acubic spline technique during exercise ECG testing recommen-dations for proper implementationrdquo Journal of Electrocardiol-ogy vol 21 pp S149ndashS157 1988

[5] S Luo andP Johnston ldquoA reviewof electrocardiogramfilteringrdquoJournal of Electrocardiology vol 43 no 6 pp 486ndash496 2010

[6] G Lenis T Baas and O Dossel ldquoEctopic beats and theirinfluence on the morphology of subsequent waves in theelectrocardiogramrdquo Biomedical Engineering vol 58 no 2 pp109ndash119 2013

[7] G Lenis N Pilia T Oesterlein A Luik C Schmitt andODos-sel ldquoP wave detection and delineation in the ECG based on thephase free stationary wavelet transform and using intracardiacatrial electrograms as referencerdquoBiomedical Engineering vol 61no 1 pp 37ndash56 2016

[8] A S Berson and H V Pipberger ldquoErrors caused by inadequacyof low-frequency response of electrocardiographsrdquo in Proceed-ings of the Digest of 6th International Conference on BiomedicalEngineering Society of Med Electr Biomed Eng Ed pp 13ndash14Okomura Printing Co Tokyo Japan August 1965

[9] H V Pipberger R C Arzbaecher A S Berson et al ldquoRecom-mendations for standardization of leads and of specifications forinstruments in electrocardiography and vectorcardiographyrdquoCirculation vol 52 no 2 pp 11ndash31 1975

[10] J A Van Alste and T S Schilder ldquoRemoval of base-line wanderand power-line interference from the ECG by an efficient firfilter with a reduced number of tapsrdquo IEEE Transactions onBiomedical Engineering vol 32 no 12 pp 1052ndash1060 1985

[11] J M Leski and N Henzel ldquoECG baseline wander and power-line interference reduction using nonlinear filter bankrdquo SignalProcessing vol 85 no 4 pp 781ndash793 2005

[12] F A Afsar M S Riaz and M Arif ldquoA comparison of baselineremoval algorithms for electrocardiogram (ECG) based auto-mated diagnosis of coronory heart diseaserdquo in Proceedings of the3rd International Conference on Bioinformatics and BiomedicalEngineering (iCBBE rsquo09) pp 1ndash4 IEEE Beijing China June2009

[13] I Dotsinsky and T Stoyanov ldquoOptimization of bi-directionaldigital filtering for drift suppression in electrocardiogram sig-nalsrdquo Journal ofMedical Engineering and Technology vol 28 no4 pp 178ndash180 2004

[14] J A van Alste W van Eck and O E Herrmann ldquoECG baselinewander reduction using linear phase filtersrdquo Computers andBiomedical Research vol 19 no 5 pp 417ndash427 1986

[15] R Jane P Laguna N V Thakor and P Caminal ldquoAdaptivebaseline wander removal in the ECG comparative analysis withcubic spline techniquerdquo in Proceedings of the in Computers inCardiology 1992 pp 143ndash146 October 1992

[16] I I Christov I A Dotsinsky and I K Daskalov ldquoHigh-passfiltering of ECG signals using QRS eliminationrdquo Medical ampBiological Engineering amp Computing vol 30 no 2 pp 253ndash2561992

[17] D U J Keller D L Weiss O Dossel and G Seemann ldquoInflue-nce of I Ks heterogeneities on the genesis of the T-wave a com-putational evaluationrdquo IEEE Transactions on Biomedical Engi-neering vol 59 no 2 pp 311ndash322 2012

[18] D Potyagaylo E G Cortes W H W Schulze and O DosselldquoBinary optimization for source localization in the inverseproblem of ECGrdquo Medical and Biological Engineering andComputing vol 52 no 9 pp 717ndash728 2014

[19] N Pilia G Lenis A Loewe W Schulze and O Dossel ldquoTheimpact of baseline wander removal techniques on the ST seg-ment in simulated ischemic 12-lead ECGsrdquo Current Directionsin Biomedical Engineering vol 1 no 1 pp 96ndash99 2015

[20] A L Goldberger L A N Amaral L Glass et al ldquoPhysioBankPhysioToolkit and PhysioNet components of a new researchresource for complex physiologic signalsrdquo Circulation vol 101no 23 pp E215ndashE220 2000

[21] A Loewe W H W Schulze Y Jiang et al ldquoECG-baseddetection of early myocardial ischemia in a computationalmodel impact of additional electrodes optimal placement anda new feature for ST deviationrdquo BioMed Research Internationalvol 2015 Article ID 530352 11 pages 2015

[22] A LoeweW HW Schulze Y Jiang MWilhelms and O Dos-sel ldquoDetermination of optimal electrode positions of a wearableECGmonitoring system for detection ofmyocardial ischemia a


simulation studyrdquo in Proceedings of the Computing in Cardiology(CinC rsquo11) pp 741ndash744 IEEE Hangzhou China September2011

[23] D D Streeter ldquoGross morphology and fiber geometry of theheartrdquo in Handbook of Physiology The Cardiovascular Systemvol 59 pp 61ndash290 American Physiology Society BethesdaMdUSA 1979

[24] D Potyagaylo E Gil Cortrsquoes W H W Schulz and O DosselldquoSynthesis and analysis models for sparse signal reconstructionin the inverse problem of ECGrdquo Biomedical Engineering vol 59pp S941ndashS944 2014

[25] D U J Keller F M Weber G Seemann and O Dossel ldquoRan-king the influence of tissue conductivities on forward-calc-ulated ECGsrdquo IEEETransactions onBiomedical Engineering vol57 no 7 pp 1568ndash1576 2010

[26] K H W J Ten Tusscher and A V Panfilov ldquoCell model forefficient simulation of wave propagation in human ventriculartissue under normal and pathological conditionsrdquo Physics inMedicine and Biology vol 51 no 23 pp 6141ndash6156 2006

[27] S A Niederer E Kerfoot A P Benson et al ldquoVerification ofcardiac tissue electrophysiology simulators using an n-versionbenchmarkrdquo Philosophical Transactions of the Royal Society Avol 369 no 1954 pp 4331ndash4351 2011

[28] G Seemann F Sachse M Karl D Weiss V Heuveline and ODossel ldquoFramework for modular exible and efficient solvingthe cardiac bidomain equations using PETScrdquo in Progress inindustrial mathematics at ECMI 2008 pp 363ndash369 SpringerBerlin Germany 2010

[29] KHW J Ten Tusscher andA V Panfilov ldquoAlternans and spiralbreakup in a human ventricular tissuemodelrdquoAmerican Journalof PhysiologymdashHeart and Circulatory Physiology vol 291 no 3pp H1088ndashH1100 2006

[30] M Wilhelms O Dossel and G Seemann ldquoIn silico investiga-tion of electrically silent acute cardiac ischemia in the humanventriclesrdquo IEEE Transactions on Biomedical Engineering vol58 no 10 pp 2961ndash2964 2011

[31] M D Cerqueira N J Weissman V Dilsizian et al ldquoStan-dardized myocardial sementation and nomenclature for tomo-graphic imaging of the heart a statement for healthcare profes-sionals from the cardiac imaging committee of the council onclinical cardiology of the american heart associationrdquo Circula-tion vol 105 no 4 pp 539ndash542 2002

[32] D Nunan G R H Sandercock and D A Brodie ldquoA quan-titative systematic review of normal values for short-termheart rate variability in healthy adultsrdquo Pacing and ClinicalElectrophysiology vol 33 no 11 pp 1407ndash1417 2010

[33] P Kligfield L S Gettes J J Bailey et al ldquoRecommendationsfor the standardization and interpretation of the electrocardio-gram part I the electrocardiogram and its technology a scien-tific statement from the American Heart Association Electro-cardiography andArrhythmias Committee Council onClinicalCardiology the American College of Cardiology Foundationand the Heart Rhythm Society endorsed by the InternationalSociety for Computerized Electrocardiologyrdquo Journal of theAmerican College of Cardiology vol 49 no 10 pp 1109ndash11272007

[34] M S Chavan R A Agarwala and M D Uplane ldquoSuppressionof baseline wander and ower line interference in ECG usingdigital IIR filterrdquo International Journal of Circuits Systems AndSignal Processing vol 2 no 2 pp 356ndash365 2008

[35] P De Chazal C Heneghan E Sheridan R Reilly P Nolan andM OrsquoMalley ldquoAutomated processing of the single-lead electro-cardiogram for the detection of obstructive sleep apnoeardquo IEEETransactions on Biomedical Engineering vol 50 no 6 pp 686ndash696 2003

[36] C R Meyer and H N Keiser ldquoElectrocardiogram baselinenoise estimation and removal using cubic splines and state-space computation techniquesrdquo Computers and Biomedical Re-search vol 10 no 5 pp 459ndash470 1977

[37] A Khawaja Automatic ECG Analysis Using Principal Com-ponent Analysis and Wavelet Transformation vol 3 Univer-sitatsverlag Karlsruhe 2006

[38] K L Park K J Lee and H R Yoon ldquoApplication of a waveletadaptive filter to minimise distortion of the ST-segmentrdquoMedical and Biological Engineering and Computing vol 36 no5 pp 581ndash586 1998

[39] P S Addison ldquoWavelet transforms and the ECG a reviewrdquoPhysiological Measurement vol 26 no 5 pp R155ndashR199 2005

[40] G Lenis Y Lutz G Seeman et al ldquoPost extrasystolic T wavechange in subjects with structural healthy ventricles-measure-ment and simulationrdquo in Proceedings of the 41st Computing inCardiology Conference (CinC rsquo14) pp 1069ndash1072 IEEE Cam-bridge Mass USA September 2014

[41] G S Wagner P Macfarlane H Wellens et al ldquoAHAACCFHRS recommendations for the standardization and inter-pretation of the electrocardiogram part VI acute ischemiainfarction a scientific statement from the American HeartAssociation Electrocardiography and Arrhythmias CommitteeCouncil onClinical Cardiology theAmericanCollege ofCardi-ology Foundation and the Heart Rhythm Society Endorsed bythe International Society for Computerized ElectrocardiologyrdquoJournal of the American College of Cardiology vol 53 no 11 pp1003ndash1011 2009

[42] J Brownfield and M Herbert ldquoECG criteria for fibrinolysiswhats up with the J pointrdquoWestern Journal of EmergencyMedi-cine vol 9 no 1 pp 40ndash42 2008

[43] L G Gamero J Vila and F Palacios ldquoWavelet transformanalysis of heart rate variability during myocardial ischaemiardquoMedical and Biological Engineering and Computing vol 40 no1 pp 72ndash78 2002

[44] L G G Porto and L F Junqueira Jr ldquoComparison of time-domain short-term heart interval variability analysis using awrist-worn heart rate monitor and the conventional electrocar-diogramrdquo Pacing and Clinical Electrophysiology vol 32 no 1pp 43ndash51 2009

[45] K Otsuka G Cornelissen and F Halberg ldquoCircadian rhythmicfractal scaling of heart rate variability in health and coronaryartery diseaserdquo Clinical Cardiology vol 20 no 7 pp 631ndash6381997

[46] D A Bragg-Remschel C M Anderson and R A WinkleldquoFrequency response characteristics of ambulatory ECG moni-toring systems and their implications for ST segment analysisrdquoAmerican Heart Journal vol 103 no 1 pp 20ndash31 1982

[47] P E McSharry G D Clifford L Tarassenko and L A SmithldquoA dynamicalmodel for generating synthetic electrocardiogramsignalsrdquo IEEE Transactions on Biomedical Engineering vol 50no 3 pp 289ndash294 2003

[48] R Pallas-Areny J Colominas-Balague and F J Rosell ldquoEffectof respiration-induced heart movements on the ECGrdquo IEEETransactions on Biomedical Engineering vol 36 no 6 pp 585ndash590 1992


[49] G B Moody R G Mark A Zoccola and S Mantero ldquoDeriva-tion of respiratory signals frommulti-lead ECGsrdquoComputers inCardiology vol 12 no 1985 pp 113ndash116 1985

[50] M Blanco-Velasco B Weng and K E Barner ldquoECG signaldenoising and baseline wander correction based on the empir-ical mode decompositionrdquo Computers in Biology and Medicinevol 38 no 1 pp 1ndash13 2008

[51] Y Luo R H Hargraves A Belle et al ldquoA hierarchical methodfor removal of baseline drift from biomedical signals applica-tion in ECG analysisrdquo The Scientific World Journal vol 2013Article ID 896056 10 pages 2013

[52] G Lenis M Kircher J Lazaro R Bailon E Gil and O DoesselldquoSeparating the effect of respiration on the heart rate variabilityusing Grangerrsquos causality and linear filteringrdquo Biomedical SignalProcessing and Control vol 31 pp 272ndash287 2017

[53] E Tragardh M Claesson G S Wagner S Zhou and OPahlm ldquoDetection of acute myocardial infarction using the 12-lead ECG plus inverted leads versus the 16-lead ECG (withadditional posterior and right-sided chest electrodes)rdquo ClinicalPhysiology and Functional Imaging vol 27 no 6 pp 368ndash3742007

Submit your manuscripts athttpswwwhindawicom

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


MEDIATORSINFLAMMATION

of


Behavioural Neurology

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary and Alternative Medicine

Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom


minus2minus15minus1

minus050

051

Am

plitu

de (m

V)

2 4 6 8 10 12 140

Time (s)

(a)

V4 unfilteredV4 filtered

0 35 42 32515105Time (s)

minus05

005

115

Am

plitu

de (m

V)

(b)Figure 1 (a) ECG recording corrupted by baseline wander Theremoval of this artifact is necessary when diagnosing a change in theST segment Yet the filtering process can modify the signal as seenin (b) (b) ECG recording with a clear ST depression before (blue)and after (red) high-pass filtering The ST depression is reducedbecause of too strong filtering The signals presented in this figurewere retrieved from the Physionet database [20]

it is also known that baseline wandermay be located at higherfrequencies mandating more intrusive filters with highercut-off frequencies [10 11] Thus finding the filters that arecapable of removing baseline wander without compromisingST changes is a relevant task

Even though some interesting approaches to addressthis question have been reported in literature [12ndash14] itis not trivial to perform such a study using only baselinecorrupted ECG recordings Since the ground truth (artifactfree signal) is not known it is hard to objectively evaluatefilter performance based onmodifications of the ST segmentFurthermore they are prone to detection errors from theautomatic signal processing software because some of themethods require for example a precise annotation of the Rpeak or Pwave [15] On the other hand synthetic signals havealso been used for these applications but they might not berealistic enough to reproduce true changes of the ST segment[16] In order to overcome this limitation in silico studiescan be used The simulation of cardiac electrophysiologyallows us to recreate ischemia at a multiscale level We canadjust the models that govern the action potential (AP) inthe cardiac myocyte to represent ischemia-induced changelet the electrical depolarization and subsequent repolariza-tion propagate in the heart and generate the body surfacepotential map on the chest We can then place electrodeson the torso and extract the simulated 12-lead ECG [1718] With this approach we have the advantage of beingable to locate the most relevant fiducial points (eg QRS

complex or ST segment) in the simulated ECG and are thuscapable of performing an analysis that is free of detectionerrors

By adding baseline wander to the ECG we reproduce thesignals in a controlled environment that would have beenrecorded in real life Since the ground truth is also knownthe performance of the filters can be studied Moreoverthe simulation software allows varying the patient geometrythe location and size of the ischemia in the heart andthe electrical properties of the model This leads to a largenumber of possible ECG signals and ST changes

In this study we address the question of the best suitedtechnique for baseline removal without compromising STchanges For that purpose we use a diverse database ofsimulated ECGs with a state-of-the-art electrophysiologicalmodel add synthetic baseline wander and compare theperformance of five manually tuned filters commonly usedin literature Preliminary results on a smaller study includingfewer filters and less signals were presented at the nationalconference of biomedical engineering in Germany [19]

2 Materials and Methods

21 Generation of Simulated Data The database containingthe 3 times 255 = 765 electrophysiological setups used inthis research project was originally created in our groupwith the aim of studying optimal electrode placement todetect ischemia in the ECG [21 22] The data set wascreated using three different subjects Their geometries wereobtained by segmenting the Visible Man dataset and twoothermagnetic resonance images Distinct tissue classes wereused for the main organs including the ventricles skeletalmuscle fat blood lung liver and spleen Fiber orientationin the ventricles was introduced using a rule-based approach[23]The voxel meshes of the three subjects were created withan isotropic side length no greater than 05mm

A cellular automatonwas chosen to perform the ischemiasimulations [24] This is a computationally cheap methodbased on predefined rules In all simulations the electricaldepolarization wave originates from the Purkinje musclejunctions It then propagates by sequentially activating neigh-boring voxels and triggering an AP in each voxel The repo-larization scheme is thus dependent on theAP correspondingto every region of the heart Transmural heterogeneity andanisotropic tissue conductivities inside the ventricular wallswere included in the model [25]

The AP in each voxel was obtained from a monodomainsimulation carried out using the software acCELLerate andutilizing the Ten Tusscher cell model with a basic cycle lengthof 60 beats per minute [26ndash28] The model parameters wereadjusted to differentiate healthy tissue from ischemic regions[29 30] The latter were placed in all 17 AHA segments in theleft ventricle and varied in size to include subendocardial andtransmural scenarios [31]

From every simulation a QRS complex an ST segmentand a T wave were obtained The QT interval was equal to400ms in all simulations The sampling frequency of thesimulated signals was 500Hz but upsampling to 119891119904 = 512Hzwas carried out to facilitate wavelet-based filtering


(a) (b) (c)

minus2

minus1

0

1

2

3

4

Am

plitu

de (m

V)

01 02 030Time (s)

(d)

01 02 030Time (s)

minus2

minus1

0

1

2

Am

plitu

de (m

V)

(e)

minus05

0

05

1

15

2

Am

plitu

de (m

V)

01 02 030Time (s)







119887 (119905) = 119862 sdot119870

sum119896=0




Am

plitu

de (m

V)

1 2 3 40

Time (s)

minus05

0

05

1

15

(a)

05 10

Frequency (Hz)

0

005

01

015

02

Am

plitu

de (a

u)

(b)

0

005

01

015

02

Am

plitu

de (a

u)

20 400

Frequency (Hz)

(c)

Am

plitu

de (m

V)

1 2 3 40

Time (s)

minus05

0

05

1

15

(d)

0

005

01

015

02

Am

plitu

de (a

u)

05 10

Frequency (Hz)

(e)

0

005

01

015

02

Am

plitu

de (a

u)

20 400

Frequency (Hz)

(f)




119910 (119905) = 119909 (119905) + 119887 (119905) (2)








120601kakC

x(t)

b(t)

y(t)+

(a)

x(t)y(t)

x(t)


Quality measures


(b)









119891119888 = 1198911199042119871+1 997904rArr

119871 = log2(119891119904119891119888) minus 1 = 9

(3)







CC 119909 (119905) (119905) = 119864 (119909 (119905) minus 120583119909) ( (119905) minus 120583)120590119909120590 (4)






KP = min119905119878119879

max119894

1003816100381610038161003816119909119894 (119905ST)1003816100381610038161003816 (6)







3 Results




03

04

05

06

07

08

09

1

Cor

relat

ion

valu

e

Med

ian

Splin

e

Butte

rwor

th

Wav

elet c

anc

Wav

elet H

P

No

filte

r

(a)

minus02

0

02

04

06

08

1

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

l_op

erat

or v

alue

(b)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

minus08

minus06

minus04

minus02

0

02

04

06

08

K po

int d

evia

tion

(mV

)

(c)

0

05

1

15

2

25

Com

puta

tion

time (

s)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

(d)




4 Discussion












minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

Am

plitu

de (m

V)

02 04 06 080Time (s)



(a)

02 04 06 080Time (s)

0

05

1

15

2

25

3

35

Am

plitu

de (m

V)



(b)

02 040Time (s)

minus8

minus7

minus6

minus5

minus4

minus3

minus2

minus1

0

1

Am

plitu

de (m

V)



(c)

minus06

minus04

minus02

0

02

04

06

Am

plitu

de (m

V)

01 02 03 04 05 060Time (s)



(d)










5 Conclusion








Acknowledgments


References






























































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b) (c)

minus2

minus1

0

1

2

3

4

Am

plitu

de (m

V)

01 02 030Time (s)

(d)

01 02 030Time (s)

minus2

minus1

0

1

2

Am

plitu

de (m

V)

(e)

minus05

0

05

1

15

2

Am

plitu

de (m

V)

01 02 030Time (s)







119887 (119905) = 119862 sdot119870

sum119896=0




Am

plitu

de (m

V)

1 2 3 40

Time (s)

minus05

0

05

1

15

(a)

05 10

Frequency (Hz)

0

005

01

015

02

Am

plitu

de (a

u)

(b)

0

005

01

015

02

Am

plitu

de (a

u)

20 400

Frequency (Hz)

(c)

Am

plitu

de (m

V)

1 2 3 40

Time (s)

minus05

0

05

1

15

(d)

0

005

01

015

02

Am

plitu

de (a

u)

05 10

Frequency (Hz)

(e)

0

005

01

015

02

Am

plitu

de (a

u)

20 400

Frequency (Hz)

(f)




119910 (119905) = 119909 (119905) + 119887 (119905) (2)








120601kakC

x(t)

b(t)

y(t)+

(a)

x(t)y(t)

x(t)


Quality measures


(b)









119891119888 = 1198911199042119871+1 997904rArr

119871 = log2(119891119904119891119888) minus 1 = 9

(3)







CC 119909 (119905) (119905) = 119864 (119909 (119905) minus 120583119909) ( (119905) minus 120583)120590119909120590 (4)






KP = min119905119878119879

max119894

1003816100381610038161003816119909119894 (119905ST)1003816100381610038161003816 (6)







3 Results




03

04

05

06

07

08

09

1

Cor

relat

ion

valu

e

Med

ian

Splin

e

Butte

rwor

th

Wav

elet c

anc

Wav

elet H

P

No

filte

r

(a)

minus02

0

02

04

06

08

1

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

l_op

erat

or v

alue

(b)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

minus08

minus06

minus04

minus02

0

02

04

06

08

K po

int d

evia

tion

(mV

)

(c)

0

05

1

15

2

25

Com

puta

tion

time (

s)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

(d)




4 Discussion












minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

Am

plitu

de (m

V)

02 04 06 080Time (s)



(a)

02 04 06 080Time (s)

0

05

1

15

2

25

3

35

Am

plitu

de (m

V)



(b)

02 040Time (s)

minus8

minus7

minus6

minus5

minus4

minus3

minus2

minus1

0

1

Am

plitu

de (m

V)



(c)

minus06

minus04

minus02

0

02

04

06

Am

plitu

de (m

V)

01 02 03 04 05 060Time (s)



(d)










5 Conclusion








Acknowledgments


References






























































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















Am

plitu

de (m

V)

1 2 3 40

Time (s)

minus05

0

05

1

15

(a)

05 10

Frequency (Hz)

0

005

01

015

02

Am

plitu

de (a

u)

(b)

0

005

01

015

02

Am

plitu

de (a

u)

20 400

Frequency (Hz)

(c)

Am

plitu

de (m

V)

1 2 3 40

Time (s)

minus05

0

05

1

15

(d)

0

005

01

015

02

Am

plitu

de (a

u)

05 10

Frequency (Hz)

(e)

0

005

01

015

02

Am

plitu

de (a

u)

20 400

Frequency (Hz)

(f)




119910 (119905) = 119909 (119905) + 119887 (119905) (2)








120601kakC

x(t)

b(t)

y(t)+

(a)

x(t)y(t)

x(t)


Quality measures


(b)









119891119888 = 1198911199042119871+1 997904rArr

119871 = log2(119891119904119891119888) minus 1 = 9

(3)







CC 119909 (119905) (119905) = 119864 (119909 (119905) minus 120583119909) ( (119905) minus 120583)120590119909120590 (4)






KP = min119905119878119879

max119894

1003816100381610038161003816119909119894 (119905ST)1003816100381610038161003816 (6)







3 Results




03

04

05

06

07

08

09

1

Cor

relat

ion

valu

e

Med

ian

Splin

e

Butte

rwor

th

Wav

elet c

anc

Wav

elet H

P

No

filte

r

(a)

minus02

0

02

04

06

08

1

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

l_op

erat

or v

alue

(b)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

minus08

minus06

minus04

minus02

0

02

04

06

08

K po

int d

evia

tion

(mV

)

(c)

0

05

1

15

2

25

Com

puta

tion

time (

s)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

(d)




4 Discussion












minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

Am

plitu

de (m

V)

02 04 06 080Time (s)



(a)

02 04 06 080Time (s)

0

05

1

15

2

25

3

35

Am

plitu

de (m

V)



(b)

02 040Time (s)

minus8

minus7

minus6

minus5

minus4

minus3

minus2

minus1

0

1

Am

plitu

de (m

V)



(c)

minus06

minus04

minus02

0

02

04

06

Am

plitu

de (m

V)

01 02 03 04 05 060Time (s)



(d)










5 Conclusion








Acknowledgments


References






























































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of



















120601kakC

x(t)

b(t)

y(t)+

(a)

x(t)y(t)

x(t)


Quality measures


(b)









119891119888 = 1198911199042119871+1 997904rArr

119871 = log2(119891119904119891119888) minus 1 = 9

(3)







CC 119909 (119905) (119905) = 119864 (119909 (119905) minus 120583119909) ( (119905) minus 120583)120590119909120590 (4)






KP = min119905119878119879

max119894

1003816100381610038161003816119909119894 (119905ST)1003816100381610038161003816 (6)







3 Results




03

04

05

06

07

08

09

1

Cor

relat

ion

valu

e

Med

ian

Splin

e

Butte

rwor

th

Wav

elet c

anc

Wav

elet H

P

No

filte

r

(a)

minus02

0

02

04

06

08

1

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

l_op

erat

or v

alue

(b)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

minus08

minus06

minus04

minus02

0

02

04

06

08

K po

int d

evia

tion

(mV

)

(c)

0

05

1

15

2

25

Com

puta

tion

time (

s)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

(d)




4 Discussion












minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

Am

plitu

de (m

V)

02 04 06 080Time (s)



(a)

02 04 06 080Time (s)

0

05

1

15

2

25

3

35

Am

plitu

de (m

V)



(b)

02 040Time (s)

minus8

minus7

minus6

minus5

minus4

minus3

minus2

minus1

0

1

Am

plitu

de (m

V)



(c)

minus06

minus04

minus02

0

02

04

06

Am

plitu

de (m

V)

01 02 03 04 05 060Time (s)



(d)










5 Conclusion








Acknowledgments


References






























































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















CC 119909 (119905) (119905) = 119864 (119909 (119905) minus 120583119909) ( (119905) minus 120583)120590119909120590 (4)






KP = min119905119878119879

max119894

1003816100381610038161003816119909119894 (119905ST)1003816100381610038161003816 (6)







3 Results




03

04

05

06

07

08

09

1

Cor

relat

ion

valu

e

Med

ian

Splin

e

Butte

rwor

th

Wav

elet c

anc

Wav

elet H

P

No

filte

r

(a)

minus02

0

02

04

06

08

1

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

l_op

erat

or v

alue

(b)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

minus08

minus06

minus04

minus02

0

02

04

06

08

K po

int d

evia

tion

(mV

)

(c)

0

05

1

15

2

25

Com

puta

tion

time (

s)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

(d)




4 Discussion












minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

Am

plitu

de (m

V)

02 04 06 080Time (s)



(a)

02 04 06 080Time (s)

0

05

1

15

2

25

3

35

Am

plitu

de (m

V)



(b)

02 040Time (s)

minus8

minus7

minus6

minus5

minus4

minus3

minus2

minus1

0

1

Am

plitu

de (m

V)



(c)

minus06

minus04

minus02

0

02

04

06

Am

plitu

de (m

V)

01 02 03 04 05 060Time (s)



(d)










5 Conclusion








Acknowledgments


References






























































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















03

04

05

06

07

08

09

1

Cor

relat

ion

valu

e

Med

ian

Splin

e

Butte

rwor

th

Wav

elet c

anc

Wav

elet H

P

No

filte

r

(a)

minus02

0

02

04

06

08

1

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

l_op

erat

or v

alue

(b)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

minus08

minus06

minus04

minus02

0

02

04

06

08

K po

int d

evia

tion

(mV

)

(c)

0

05

1

15

2

25

Com

puta

tion

time (

s)

Splin

e

Med

ian

No

filte

r

Wav

elet H

P

Butte

rwor

th

Wav

elet c

anc

(d)




4 Discussion












minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

Am

plitu

de (m

V)

02 04 06 080Time (s)



(a)

02 04 06 080Time (s)

0

05

1

15

2

25

3

35

Am

plitu

de (m

V)



(b)

02 040Time (s)

minus8

minus7

minus6

minus5

minus4

minus3

minus2

minus1

0

1

Am

plitu

de (m

V)



(c)

minus06

minus04

minus02

0

02

04

06

Am

plitu

de (m

V)

01 02 03 04 05 060Time (s)



(d)










5 Conclusion








Acknowledgments


References






























































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


























minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

Am

plitu

de (m

V)

02 04 06 080Time (s)



(a)

02 04 06 080Time (s)

0

05

1

15

2

25

3

35

Am

plitu

de (m

V)



(b)

02 040Time (s)

minus8

minus7

minus6

minus5

minus4

minus3

minus2

minus1

0

1

Am

plitu

de (m

V)



(c)

minus06

minus04

minus02

0

02

04

06

Am

plitu

de (m

V)

01 02 03 04 05 060Time (s)



(d)










5 Conclusion








Acknowledgments


References






























































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
























5 Conclusion








Acknowledgments


References






























































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of






















Acknowledgments


References






























































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of























































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















Comparison of Baseline Wander Removal Techniques...

Documents

Transcript of Comparison of Baseline Wander Removal Techniques...