Seediscussions,stats,andauthorprofilesforthispublicationat:https://www.researchgate.net/publication/23218713

OnlineDigitalFilterandQRSDetectorApplicableinLowResourceECGMonitoringSystems

ARTICLEinANNALSOFBIOMEDICALENGINEERING·SEPTEMBER2008

ImpactFactor:3.23·DOI:10.1007/s10439-008-9553-5·Source:PubMed

CITATIONS

20

READS

193

3AUTHORS:

SerafimDimitrovTabakov

TechnicalUniversityofSofia

7PUBLICATIONS46CITATIONS

SEEPROFILE

IvoIliev

TechnicalUniversityofSofia

27PUBLICATIONS83CITATIONS

SEEPROFILE

VesselaKrasteva

BulgarianAcademyofSciences

76PUBLICATIONS455CITATIONS

SEEPROFILE

Availablefrom:VesselaKrasteva

Retrievedon:14January2016

Online Digital Filter and QRS Detector Applicable in Low Resource

ECG Monitoring Systems

SERAFIM TABAKOV,1 IVO ILIEV,1 and VESSELA KRASTEVA2

1Department of Electronics, Technical University of Sofia, 8 Kl. Ohridski str., 1000 Sofia, Bulgaria; and 2Centre of BiomedicalEngineering ‘Prof. Ivan Daskalov’, Bulgarian Academy of Sciences, Acad. G. Bonchev str., Bl.105, 1113 Sofia, Bulgaria

(Received 9 April 2008; accepted 11 August 2008; published online 28 August 2008)

Abstract—The present work describes fast computationmethods for real-time digital filtration and QRS detection,both applicable in autonomous personal ECG systems forlong-term monitoring. Since such devices work under con-siderable artifacts of intensive body and electrode move-ments, the input filtering should provide high-quality ECGsignals supporting the accurate ECG interpretation. In thisrespect, we propose a combined high-pass and power-lineinterference rejection filter, introducing the simple principleof averaging of samples with a predefined distancebetween them. In our implementation (sampling frequencyof 250 Hz), we applied averaging over 17 samples distancedby 10 samples (Filter10x17), thus realizing a comb filter witha zero at 50 Hz and high-pass cut-off at 1.1 Hz. Filter10x17affords very fast filtering procedure at the price of minimalcomputing resources. Another benefit concerns the smallECG distortions introduced by the filter, providing itspowerful application in the preprocessing module of diag-nostic systems analyzing the ECG morphology. Filter10x17does not attenuate the QRS amplitude, or introduce signif-icant ST-segment elevation/depression. The filter outputproduces a constant error, leading to uniform shifting ofthe entire P-QRS-T segment toward about 5% of the R-peakamplitude. Tests with standardized ECG signals provedthat Filter10x17 is capable to remove very strong baselinewanderings, and to fully suppress 50 Hz interferences. Bychanging the number of the averaged samples and thedistance between them, a filter design with different cut-offand zero frequency could be easily achieved. The real-timeQRS detector is designed with simplified computations oversingle channel, low-resolution ECGs. It relies on simpleevaluations of amplitudes and slopes, including history oftheir mean values estimated over the preceding beats, smartadjustable thresholds, as well as linear logical rules foridentification of the R-peaks in real-time. The performanceof the QRS detector was tested with internationally recog-nized ECG databases (AHA, MIT-BIH, European ST-Tdatabase), showing mean sensitivity of 99.65% and positivepredictive value of 99.57%. The performance of the pre-sented QRS detector can be highly rated, comparable andeven better than other published real-time QRS detectors.

Examples representing some typical unfavorable conditionsin real ECGs, illustrate the common operation of Filter10x17and the QRS detector.

Keywords—Real-time ECG analysis, ECG preprocessing,

High-pass filtering, Baseline wander, Power-line interference

filtering, QRS detection, ECG monitoring of high-risk

cardiac patients.

INTRODUCTION

The long-term electrocardiogram (ECG) monitor-ing of high-risk cardiac patients by computer-assistedbedside or ambulatory systems, demands for fast andreal-time ECG analysis methods, providing high-fidelity of the automated cardiac diagnosis. The scopeof the present article is focused on the two basic pre-processing steps—filtering and the QRS detection,both defining the quality of the input of the ECGinterpretation module.14

The adequate preprocessor filter design must pro-vide maximal artifact rejection with minimal ECGdistortions. However, the spectra of ECG and artifactsin their variety often overlap,30 implying that therewould be a compromise depending on the specificapplication. For example, the baseline wander (BW),frequently induced by motion and respiratory artifactsin long-term ECG recordings, is insufficiently sup-pressed by the 0.05 Hz low-frequency cutoff for diag-nostic electrocardiography,25 recommended mainly topreserve the fidelity of the low-frequency ECG com-ponents during repolarization (ST-segment). BetterBW reduction is achieved in the ECG rhythmmonitors,for which the recommendations are less severe, allow-ing (0.67–1 Hz) low-frequency cutoff at the price ofmarkedly distortion of the repolarization and even theQRS complex amplitude.11 The latter compromise isacceptable for the systems, working under considerableartifacts of intensive body and electrode movements,

Address correspondence to Vessela Krasteva, Centre of Bio-

medical Engineering ‘Prof. Ivan Daskalov’, Bulgarian Academy of

Sciences, Acad. G. Bonchev str., Bl.105, 1113 Sofia, Bulgaria.

Electronic mail: vessika@clbme.bas.bg

Annals of Biomedical Engineering, Vol. 36, No. 11, November 2008 (� 2008) pp. 1805–1815

DOI: 10.1007/s10439-008-9553-5

0090-6964/08/1100-1805/0 � 2008 Biomedical Engineering Society

1805

such as stress testing, personal event/alarm recorders,defibrillators, etc. Such systems typically make thediagnostic interpretation based not on the ST-segmentdeviation, but on proper detection of the QRS com-plexes, rhythm analysis or coherent PQRST templatesalignment. All these methods are improving theiraccuracy when the BW is significantly rejected.

An extensive research is focused on the challenge fordesigning novel digital filters, which can correct BWwhile preserving the fidelity of the QRS and ST-seg-ment levels, exceeding the performance of the analoguefilters. Various approaches have been proposed basedtypically on smart filtering techniques, such as: movingaverage filters repeatedly applied to achieve the pass-band and the stopband specifications10; a combinationof the moving average with a FIR filter designed byusing a Kaiser window4; an offline bi-directional high-pass filter with adaptable cut-off frequency, keepingthe RS and ST changes below a selected threshold7; aquasi real-time procedure for estimation of the BW bylow-pass filtering and consecutive tracking and elimi-nation of the QRS complexes6; the two-step procedurefor baseline estimation by selective ECG filtering andminimization of the residual error27; the nonlinear fil-ter banks allowing low-pass and power-line interfer-ence reduction.17 Adaptive filtering has also beenapplied,30 in particular by cubic spline techniquecombined with adaptive two stage cascade filter13 or byKalman model, accepting the hypothesis that the ECGsignal can be characterized by an autoregressive model,while the BW is estimated as a first order polynomial.22

Another approach has introduced wavelet transformsfor decomposition and reconstruction of the BW.29,31

Although the different filter solutions are attractivebecause of the reasonable frequency characteristics,they have specific disadvantages related either to heavycomputations, or to requirements for certain analysisof the ECG prior to applying the filtering technique,leading to large delay of the ECG signal analysis. Suchfilters are inadmissible for real-time operating systemswith low computation resources.

Different methods for effective rejection of thepower-line interference are proposed, based typicallyon digital notch filters,17,20,24 comb filters19 or adaptivefilters.30 The main drawback is reduction of high-fre-quency components of ECG waves, in particular theQRS complexes. Smart techniques are proposed toovercome this problem, such as the subtraction pro-cedure,18 at the price of auxiliary computations fornon-linear segments detection.

The second main step in the ECG analysis, i.e. theQRS detector, is responsible for the correct localizationof each cardiac contraction that is typically a crucialinput for the ECG interpretation module. Among thehuge amount of QRS detectors, we will not focus on the

offline methods and those, requesting high computationpower, but we will only refer to those methods, whichallow online or quasi-online application by single leadECG analysis. Some of these QRS detection methodsemploy adaptive thresholds continuously updated byeach QRS complex. They rely on evaluation of steepedges and sharp peaks,8 dynamically updated amplitudeand slope thresholds,12 a combination of steep-slopethreshold, integrating threshold and beat expectationthreshold,5 a derivative threshold.15 Other methodscount on ECG decomposition in the frequency domain,such as the hardware filter banks used for joint evalu-ation of the ECG components in different frequencybands,1 or a single matched filter, extracting the mostexpressed QRS frequencies adequate for a dual edgethreshold detector.26 A complicated algorithm usesreconstructed phase portraits to derive mappings in anew dimensional space, defined by the time delay andthe mapping dimension,16 that is additionally combinedby a fill factor, mutual information, and autocorrelationfunction. The large variety of methods for QRS detec-tion and the continuous efforts for their enhancementproves that universally acceptable solution has not beenfound yet.

Fast and simple methods for ECG analysis,including preprocessor filtering and QRS detection arepresented in this work. The two methods are part of areal-time ECG analysis module designed for pocketsizepersonal ECG monitoring devices with low powerconsumption and therefore quite limited computationresources. The system is designed in accordance to allrequirements for adequate ECG signal quality after thepreprocessor filter, as well as high accuracy for detec-tion of the heartbeats. Tests with internationally rec-ognized ECG databases present the performance of theQRS detector allowing comparison to other publishedonline QRS detectors.

METHODS

All signal processing procedures are applied onsingle-lead, small resolution ECG sampled at 250 Hz,8-bit. The presented algorithms are implemented inlow-level programming language for microcontrollersand are executable in real-time.

Combined Digital Filter for High-pass Filteringand Power-line Interference Rejection

The following requirements are set for the design ofthe preprocessor filter:

– to implement both high-pass filtering and power-lineinterference rejection in a single filter;

TABAKOV et al.1806

– to fit the high-pass cut-off of ‘monitor’ type ECG(0.67–1 Hz),11 improving the robustness to BW andreducing the ST-segment distortions;

– to use linear filter equation with integer coefficientsto speed the computations;

We designed a non-recursive averaging filter namedFilterDxN, according to Eq. (1), that embodies twovariables allowing to adjust its amplitude–frequencyresponse: (i) D is the distance between the averagedsamples (as a number of samples); (ii) N is the numberof the averaged samples (an odd number is required toprevent against phase shift between the input and theoutput signals).

SFðiÞ ¼ SðiÞ � 1

N

XðN�1Þ=2

j¼�ðN�1Þ=2Sðiþ jDÞ; ð1Þ

where S is the input signal, SF is the filtered signal, i isthe sample index.

The amplitude–frequency response of the designedFilterDxN (Fig. 1a) resembles a comb-filter. The upperx-axis in Fig. 1a corresponds to the ratio between thesampling frequency (Fs) and the distance D. The zerosof the filter are occurring at integer ratio Fs vs. D. Thenumber of the ripples between the zeros are defined bythe number of the averaged samples N.

By analysis of FilterDxN frequency response, ourchoiceofNandDvalues is conformed toadequatefilteringof ECG signals sampled at Fs = 250 Hz, i.e. D = 10,N = 17 (Filter10x17), following the next considerations:

– N = 17 defines the high-pass cut-off at about1.1 Hz (Fig. 1b—black curve).

– N = 17 presents about 2.7 times steeper slope of thefirst ripple than first-order high-pass filter with cut-off of 1.1 Hz (named Filter 1.1 Hz (HP))—seeFig. 1b. This is important for improving Filter10x17robustness to the baseline drift.

– N = 17 limits the amplitude of the ripples in thepassband at about ±5% (Fig. 1a—black curve).

– D = 10 sets a zero at 50 Hz, thus Filter10x17 isapplicable as a power-line rejecter filter (Fig. 1a—-see the bottom x-axis).

We should note that Filter10x17 works as a com-bined filter (high-pass and power-line interferencerejection), besides, an ordinary low-pass filter is alsoneeded in the ECG system.

Real-time QRS Detector

In our previous work,12 we described a pilot versionof a real-time QRS detector based on the generalconcept for dynamic update of amplitude and slopethresholds, investigated previously by Dotsinsky andStoyanov,8 and Christov.5 In this work, the QRSdetector have been elaborated which considerablyimproved its performance. Detailed description of theQRS detector is provided below.

The QRS detector works with history, storing themean value of the amplitudes of the last four R-peakswith positive or negative polarity, i.e. MAP (meanamplitude of positive peaks) and MAN (mean ampli-tude of negative peaks):

MAP¼ð3MAPþECGiÞ=4; if the R-peak is positive

ð2Þ

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 87.50

0.050.1

0.150.2

0.250.3

0.350.4

0.450.5

0.550.6

0.650.7

0.750.8

0.850.9

0.951

1.051.1

1.151.2

0 6.25 12.5 18.75 25 31.25 37.5 43.75 50 56.25

Mag

nitu

de (

Nor

mal

ized

Uni

ts)

Frequency (Hz)

0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5

FS vs. D

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

1.1

1.2

N=7

N=17

N=11

s

Fs=250 Hz, D=10

Filter 1.1 Hz (HP)

Filter10x17

Frequency (Hz)

(b) (a)

FIGURE 1. (a) The amplitude–frequency response of the designed filter FilterDxN, calculated for different number of the averagedsamples N; The black curve represents Filter10x17. The upper x-axis corresponds to the ratio between the sampling frequency (Fs)and the distance D; the bottom x-axis shows the relevant frequencies calculated for Fs 5 250 Hz and D 5 10. (b) The amplitude–frequency responses of Filter10x17 and first-order high-pass filter with cut-off of 1.1 Hz: zoom in the low frequency range.

Online Digital Filter and QRS Detector Applicable in ECG Systems 1807

MAN¼ð3MANþECGiÞ=4; if the R-peak is negative,

ð3Þwhere ECGi is the amplitude of the ith ECG signalsample, which is recognized as the current R-peak.

History of the R-peak slope is also traced using theslopes of the last four R-peaks. The mean value of theslopes (MSL) is calculated as:

MSL ¼ ð3MSLþ ECGSlopeiÞ=4; ð4Þwhere ECGSlopei is the slope of the current R-peak atsample index i with amplitude ECGi. The slope iscalculated as the sum of the absolute amplitude dif-ferences between the ith sample and ten surrounding itsamples:

ECGSlopei ¼X5

n¼�5

ECGi � ECGiþnj j; ð5Þ

The values of the mean amplitudes and slopes(MAP, MAN, and MSL) are updated after each R-peak, according to (2)–(4). Reset of MAN (MAP) to aminimal value of -100 lV (100 lV) is applied justafter a sequence of at least 10 consecutive positive(negative) R-peaks.

Three dynamically updated thresholds are defined:

– ATP—amplitude threshold for positive peaks. Ondetection of positive R-peak, the ATP value be-comes equal to the positive R-peak amplitude if thelatter is below 1.5 times the MAP, otherwise theATP value becomes equal to the MAP. The ATP

value does not change until the RS deflection is de-tected plus 240 ms after it (to avoid high-amplitudeT-waves). Further, ATP decreases linearly with onebit per sample until the threshold drops down to apreset level (about 100 lV), where it remains con-stant in order to avoid analysis of very low-ampli-tude signals.

– ATN—amplitude threshold for negative peaks. Ondetection of negative R-peak, the ATN value be-comes equal to the negative R-peak amplitude if thelatter is above 1.5 times the MAN, otherwise theATN value becomes equal to the MAN. ATN timecourse changes in analogy to ATP, i.e. after the RSdeflection, the threshold does not change for 240 ms.Further, ATN increases linearly with one bit persample until the threshold rises up to a level of about-100 lV.

– SLT—slope threshold; On R-peak detection, theSLT value becomes equal to the slope of the signalcalculated according to (5), or if the slope is higherthan 1.5MSL, then SLT is set toMSL. In analogy tothe amplitude thresholds, SLT does not changewithin 240 ms after the RS deflection. Further, SLTdecreases with a rate of 1/128 from its current valueper sample, until the minimal value of 1 is reached.

Figure 2 represents an example with positive andnegative R-peaks that is selected for demonstration ofthe thresholds’ time–course. Here, ATP and ATN arepresented by dotted lines in the upper trace (ECG),and SLT is shown in the bottom trace (ECG Slope).The ECG slope is calculated according to (5).

FIGURE 2. File 233D1 from MIT-BIH database (a segment of 20 s) is used to illustrate the operation of the QRS detection algo-rithm. The top trace represents the ECG signal (resolution 20 lV/LSB). The bottom trace is the ECG slope. The circle marks (s)show the positions of the identified maximal amplitude R-peaks (positive and negative). The dotted lines represent the time–courseof the dynamically updated amplitude thresholds (ATP and ATN) and slope threshold (SLT). Markers of the original beat anno-tations (as specified in the MIT-BIH database) are supplied in the top, corresponding to the fiducial points of normal beats (N),ventricular ectopic beats (V) and fusion beats (F).

TABAKOV et al.1808

The R-peak detection criteria are set as follows:

Step 1: Identification of the peak:

signðECGi � ECGi�nÞ � signðECGi � ECGiþnÞ>0; ð6Þwhere n = 1, 2,…,5 is the consecutive number of theanalyzed samples around the peak with index i.

Step 2: Verification of the peak (one of the followingconditions has to be satisfied):

– Condition 1: Verification for steep slope and signifi-cant amplitude of the peak, exceeding the thresholdsfor slope and amplitude:

ðECGSlopei>SLTÞ and

ðECGi>ATPÞ or ðECGi<ATNÞð Þ ð7Þ– Condition 2: the criterion tolerates smaller slopes

above 75% from the threshold SLT, and amplitudeshigher than 20% from the mean amplitude of thelast 4 R-peaks. This criterion is convenient foridentification of typical ventricular extrasystoleswith smoothed QRS waveform, but appearing withhigher R-peak amplitude than the normal sequenceof QRS complexes:

ðECGSlopei>0:75SLTÞ and

ðECGi>1:2MAPÞ or ðECGi<1:2MANÞð Þ ð8Þ– Condition 3: the criterion tolerates R-peaks with

significantly slower slopes—above 30% from SLT,and amplitudes exceeding by 10% the amplitude ofthe last R-peak (considering their absolute values).This condition is valid only if within 240 ms after theR-peak, the QRS waveform changes its polarity andexceeds a threshold of 100 lV in opposite directionthan the detected R-peak. This criterion is createdfor detection of very smoothed, biphasic ventricularextrasystoles with amplitude approaching the one ofthe normal beats:

ðECGSlopei>0:3SLTÞand jECGij>1:1jAmpLastRpeakj andðECGi>0Þ and ðECGiþn<� 100lVÞ 60

n¼1

��� ��

or ðECGi<0Þ and ðECGiþn>100lVÞ 60n¼1

��� �� ð9Þ

Step 3: The R-peak is selected to be either thedetected above peak, or another peak among allsamples within 240 ms after it, with higher amplitude(in absolute value) and slope higher than 1.5 times theslope of the actual detected peak. The detected peak isused for updating of the mean amplitudes MAP, MANand mean slope MSL, as well as the adaptablethresholds ATP, ATN, SLT, following the definitions

given above. An example showing how the thresholdsATP, ATN, SLT are updated dynamically by eachdetected maximal amplitude R-peak (with positive ornegative polarity) is presented in Fig. 2.

RESULTS

Filters’ Performance on Standardized Test ECG Signals

Filters’ output was evaluated by measuring ECGdistortions introduced on the standardized test signalCAL20110. This signal is part of the calibration ECGsfrom the ECG CTS-Test Atlas designed to test theperformance of computerized electrocardiographsaccording to the requirements of the European Con-formance Testing Services.32 CAL20110 simulates amonopolar QRS complex with amplitude of 2 mV, RS-wave, ST amplitude of -200 lV, and heart rate of60 bpm (1 s cardiac cycle)—Fig. 3a. The low frequencysignal component (ST depression) in CAL20110 makestests of the low frequency response relevant. Besides,such large monopolar QRS complexes were reported toproduce a large aftereffect differentiation underdam-ping the initial part of the ST segment.6

The first test compared Filter10x17 vs. a standardfirst-order Filter 1.1 Hz (HP), when CAL20110 is ap-plied at the input. The outputs of both filters are shownin Figs. 3c and 3b, respectively. The error signals arealso shown, estimated as the difference between theoriginal signal CAL20110 and the filters’ output.

The second test compared Filter10x17 vs. an or-dinary filter for power-line interference rejection, whenCAL20110 superimposed with 50 Hz, 1 mV sinusoid isapplied at the input (Fig. 3d). The reference power-linefilter is the moving average comb filter,19 which hadbeen applied for 5 consecutive samples to achieve afirst zero at 50 Hz for sampling rate of 250 Hz. Theoutput of this reference filter, named below Filter50 Hz (Comb), is shown in Fig. 3e. The output ofFilter10x17 is shown in Fig. 3f.

In second supplementary test, CAL20110 superim-posed with 50 Hz, 1 mV sinusoid was applied to theinput of a composite filter arranged by Filter 1.1 Hz(HP) in series to Filter 50 Hz (Comb). This combina-tion of filters is a typical preprocessing solution inECG systems, implementing the same functions as thedesigned Filter10x17. The common performance ofFilter 1.1 Hz (HP) and Filter 50 Hz (Comb) is pre-sented in Fig. 3g. It is compared to the performance ofFilter10x17 shown in Fig. 3f.

The third test assessed the drift reduction of Fil-ter10x17. Sinusoidal drift with 1 mV, 0.5 Hz, and1 Hz, was added to the original ECG signal CAL20110(Figs. 4a and 4c). This relatively high drift frequency

Online Digital Filter and QRS Detector Applicable in ECG Systems 1809

and amplitude can be considered as a heavy experi-mental condition. Moreover, the superposition of aknown drift to the signal permits an accurate assess-ment of the filter output and error (Figs. 4b and 4d).

Performance Assessment of the QRS DetectionAlgorithm

Before reporting the performance of the QRSdetector, it is important to underline that the QRSdetector was tested with filtered ECG signals. Thepreprocessor filter was realized by a single Filter10x17(for high-pass and power-line interference rejection),connected to an ordinary 1-st order low-pass filter withcut-off at 25 Hz (for tremor noise reduction).

ECG Databases

The performance of the QRS detector was esti-mated over a large set of ECG signals taken from

internationally recognized ECG databases, represen-tative for various types of arrhythmia. The ECGsignals included in the study were taken from:

– The Massachusetts Institute of Technology—BethIsrael Hospital (MIT-BIH) arrhythmia database21—all 48 files 9 2 channels (100d1, 100d2,…, 234d1,234d2), each one with duration of 30 min. Since thesampling frequency is 360 Hz, these signals wereadditionally downsampled to 250 Hz.

– The American Heart Association (AHA) Data-base2—70 files 9 2 channels (1001D1, 1001D2,1002D1,…, 7009D1, 7010D2), excluding from thewhole database only 10 files 9 2 channels with ven-tricular fibrillation (8001D1, 8001D2,…, 8010D1,8010D2). Each recording has a 30-min duration. Thesampling frequency is 250 Hz.

– The European Society of Cardiology ST-T Data-base9—all 48 files 9 2 channels freely available fromPhysioNet, each one having a duration of 2 h andsampling frequency of 250 Hz.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-1500

-1000-500

0500

1000150020002500

Am

plitu

de [

uV]

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-750-500-250

0250500750

10001250150017502000

Am

plitu

de [

uV]

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-750-500-250

0250500750

10001250150017502000

Am

plitu

de [

uV]

Time [s]

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-750-500-250

0250500750

10001250150017502000

Am

plitu

de [u

V]

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-750-500-250

0250500750

10001250150017502000

Am

plitu

de [u

V]

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-750-500-250

0250500750

10001250150017502000

Am

plitu

de [u

V]

Time [s]

(a)

(b)

(c)

(d)

(e)

(f)

File: CAL20110 File: CAL20110 + Interference (50 Hz, 1 mV)

Filter 1.1 Hz (HP)

Filter10x17

Filter 50 Hz (Comb)

Filter10x17

error

error

error

error

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-750-500-250

0250500750

10001250150017502000

Am

plitu

de [

uV]

Filter 1.1 Hz (HP) + Filter 50 Hz (Comb) (g)

FIGURE 3. Filter performance: tests for ECG distortions and power-line interference rejection. (a) The test signal CAL20110; (b) theoutput of first-order high-pass filter with cut-off of 1.1 Hz (Filter 1.1 Hz (HP)), when to its input is applied the test ECG signal in (a);(c) Filter10x17 output for the test ECG signal in (a); (d) the signal CAL20110 artificially mixed with interference sinusoid (50 Hz,1 mV); (e) the output of moving average comb filter at 50 Hz (Filter 50 Hz (Comb)) for the test signal in (d); (f) Filter10x17 output forthe test signal in (d); (g) the output of combined filter, implementing Filter 1.1 Hz (HP) and Filter 50 Hz (Comb), when to its input isapplied the test signal in (d); The filters’ output error traces are artificially shifted by 250 lV to improve the reading of the graphs.

TABAKOV et al.1810

The original 12-bit amplitude resolution was re-duced to 8-bit in order to simulate the algorithmoperation with ECG signals sampled by an 8-bit ana-log-to-digital converter. The first channel and the sec-ond channel of each ECG recording were processedindependently. The reference annotations in thedatabases were accepted.

Calculation of Statistical Indices

The QRS detection algorithm was tested by auto-matic checking of the correspondence between thefiducial point and the detection mark for each heart-beat in the databases. Since the fiducial points of theannotations have varying positions in time withinthe onset and the offset of the QRS complexes, weaccepted a valid detection interval, within which thedetection marks of the tested algorithm were consi-dered true positives with respect to the annotations.28

The borders of the interval were set according to therecommendations: 10 ms before and 140 ms after theannotation locations.

Three statistical indices were used to report theperformance of the developed QRS detection algo-rithm: the sensitivity (Se), the positive predictive value(PPV), and the rate of false positives per hour (FPH).3

These are defined as:

Se ¼ TP

TPþ FN� 100 ð%Þ;

PPV ¼ TP

TPþ FP� 100 ð%Þ; FPH ¼ FP

Time;

where TP is the number of correctly detected beats(true positives), FN is the number of undetected beats(false negatives), FP is the number of falsely detectedbeats (false positives), and Time is the duration of theECG recordings.

The calculated statistical indices for the three ECGdatabases are reported in Table 1.

Examples

Two examples are presented to illustrate thecommon performance of the preprocessor Filter10x17

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6-1500-1000-500

0500

100015002000250030003500

Am

plitu

de [u

V]

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6-1500-1000-500

0500

100015002000250030003500

Am

plitu

de [u

V]

Time [s]

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6-1500-1000-500

0500

100015002000250030003500

Am

plitu

de [u

V]

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6-1500-1000-500

0500

100015002000250030003500

Am

plitu

de [u

V]

Time [s]

File: CAL20110 + Drift (0.5 Hz, 1 mV)

Filter10x17

File: CAL20110 + Drift (1 Hz, 1 mV)

Filter10x17

error error

(a)

(b)

(c)

(d)

FIGURE 4. Filter performance: tests for baseline drift reduction. (a) The signal CAL20110 artificially mixed with drift sinusoid(0.5 Hz, 1 mV); (b) Filter10x17 output for the test signal in (a); (c) The signal CAL20110 artificially mixed with drift sinusoid (1 Hz,1 mV); (d) Filter10x17 output for the test signal in (c); The filter’s output error traces are artificially shifted by 500 lV to improve thereading of the graphs.

TABLE 1. Performance statistics for the real-time QRS detection algorithm.

Databases Se (%) PPV (%) TP (no of beats) FN (no of beats) FP (no of beats) FPH (no of beats)

MIT-BIH 99.37 99.51 208701 1315 1033 21.5

AHA 99.32 99.66 322328 2205 1085 15.5

European ST-T 99.85 99.54 836750 1271 3850 20.1

Summary for all databases 99.65 99.57 1367779 4791 5968 19.3

Online Digital Filter and QRS Detector Applicable in ECG Systems 1811

and the QRS detector. These examples are chosenfrom the ECG databases to reproduce some typicalunfavorable conditions, such as large and fast base-line drift and/or arrhythmic changes of the QRSwaveforms. The first example (Fig. 5—top trace)shows a moderate, almost periodical drift, resemblingthe common influence of respiration. The drift issuperimposed on QRS complexes with altering

waveforms (in amplitude and polarity). The secondexample (Fig. 6—top trace) represents a signal withvarying amplitude and type of the QRS complexes,disturbed by extensive drift with fast and slow com-ponents. In all examples, the bottom plot shows therecovery of the ECG at Filter10x17 output, togetherwith marks for the R-peaks found by the QRSdetector.

1751 1753 1755 1757 1759 1761 1763 1765 1767 1769 1771 1773 1775 1777 1779 1781

-50

0

50

N| N| N| N| N| N| N| N| N| N| N| N| N| N| N| N| N| N| N| N| N| N| V| N| N| N| N| N| N| N| N| N| N| V| N| N| N| N| N| N| N| N| N| N| N| N| N|

-50

0

50

N N N N N N N N NNN N N NNN NN N N N NV N NNN N N N N N NV N N N N NN N NN N NNNAnnotation | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |

MIT

-BIH

203

D1

Orig

inal

EC

G

MIT

-BIH

203

D1

Filt

er 1

0x17

Out

put

Time (s)

ADC units

AnnotationADC units

FIGURE 5. The top trace—original 30 s ECG segment taken from file 203D1 (MIT-BIH). It represents a moderate, quasi-periodicaldrift, resembling the common influence of respiration. The drift is superimposed on QRS complexes with varying waveforms (inamplitude and polarity). The original database beat annotations include marks for normal beats (N) and ventricular ectopic beats(V). The bottom trace—recovery of the above signal after Filter10x17. The circle marks (s) show the positions of the identifiedmaximal amplitude R-peaks found by the QRS detector.

1205 1207 1209 1211 1213 1215 1217 1219 1221 1223 1225 1227 1229 1231 1233 123-50

0

50

100

N|

N|

N|

N|

V|

N|

N|

N|

N|

N|

N|

N|

N|

V|

N|

N|

N|

N|

N|

N|

V|

N|

N|

N|

V|

N|

N|

N|

N|

N|

N|

N|

V|

-100

-50

0

50

100

N|

N|

N|

N|

V|

N|

N|

N|

N|

N|

N|

N|

N|

V|

N|

N|

N|

N|

N|

N|

V|

N|

N|

N|

V|

N|

N|

N|

N|

N|

N|

N|

V|

AnnotationADC units

MIT

-BIH

228

D1

Orig

inal

EC

G

MIT

-BIH

228

D1

Filt

er 1

0x17

Out

put

Time (s)

AnnotationADC units

FIGURE 6. The top trace—original 30 s ECG segment taken from file 228D1 (MIT-BIH). It represents a signal with varying qualityand type of the QRS complexes, disturbed by extensive drift with fast and slow components. The original database beat anno-tations include marks for normal beats (N) and ventricular ectopic beats (V). The bottom trace—recovery of the above signal afterFilter10x17. The circle marks (s) show the positions of the identified maximal amplitude R-peaks found by the QRS detector.

TABAKOV et al.1812

DISCUSSION

The present work describes fast computation meth-ods for real-time preprocessing filtration and QRSdetection. The general application of these methodscontributes to the challenge for accurate automaticECG analysis, but actually the most valuable benefit isdirected to autonomous long-term personal ECGmonitors, such as the intelligent holters, the automaticevent/alarm recorders or the personal devices withintermittent telemetric data transfer to a central termi-nal on event-alarm activation. Such systems operateunder unfavorable environmental noise, includingconsiderable artifacts of intensive body and electrodemovements, that request powerful solutions for theirrejection in real-time. On the other hand, the low powerconsumption for long-term continuous operation andtherefore quite limited computation resources, requirefast procedures for real-time ECG signal processing,rather than complicated mathematical transforms.

The absence of baseline wandering and power-lineinterference determines high ECG signal quality, andtherefore improves the automatic ECG interpretation.In this respect, we propose a combined high-passand interference rejection filter Filter10x17, which hasparticular benefits compared to the conventional filters.The first benefit is related to its simple implementationbased on moving averaging of 17 samples distancedeach other by 10 samples. Linear filter equation withinteger coefficients is realized, that increases the speedof computation with integer bit-shift and adds instruc-tions. The second benefit concerns the small ECGdistortions introduced by Filter10x17, providing itspowerful application in the preprocessing module ofdiagnostic systems analyzing the ECG morphology. Inthis respect we compare the performance of Filter10x17vs. conventional filters for ECG monitors using thestandardized test signal CAL20110 (Fig. 3a).32 Fil-ter10x17 presents an error below 100 lV, preserving arelative constant value under the QRS, ST and T-waves(Fig. 3c). Thus no significant reduction of the R-peakamplitude and no distortions of any parts of the ST-segment are observed. Moreover, PQ and ST-deflectionkeep the same amplitude difference as the originalCAL20110 that has a certain diagnostic significance.Filter10x17 presented just the same negligible errorwhen applied for reduction of strong power-line inter-ference (Fig. 3f). In contrast, the test of the referencehigh-pass filter Filter 1.1 Hz (HP) showed an errorreaching values as high as 300 lV during the RS-waveand ST-segment that resulted in about 10% reductionof the R-peak amplitude and underlying of the ST-segment depression with about 50% (Fig. 3b). The testof the reference filter for power-line interference Filter50 Hz (Comb) showed full rejection of the amplitude of

the interference sinusoid, but at the price of abrupt errorof 200 lV just under the QRS complex (Fig. 3e). InFig. 3g we estimated also the error of a combined use ofthe two reference filters—Filter 1.1 Hz (HP) and Filter50 Hz (Comb), which together implement the samefunctions as Filter10x17. The results are not in favor ofthe reference composite filter (compare Figs. 3f to 3e).The abrupt error appearing just under QRS and ST isoriginating from the general filters’ conception to em-ploy a number of close-distance or neighboring samples(recursively or non-recursively), which are part from thesame ECG component. The filter processes all samplestogether and the general result is a partial suppressionof the relevant ECG component. The design of Filter-DxN passes over this conception and involves inputsamples, which are part of different ECG components.For the specific implementation, Filter10x17 uses 17input samples distanced each other by 40 ms, so thatthey belong to different ECG components within aninterval of 680 ms. As a result, the filter output pro-duces a constant error, which leads only to uniformshifting of the whole QRS-T segment in directionopposite to the main R-peak deflection. In our example(Figs. 3c and 3f), the estimated value of this shifting isless than 5% of the QRS amplitude (<100 lV indownward direction). This also explains why Fil-ter10x17 does not produce distortions correlated withthe heart rate, visible from the tests for drift suppression(Figs. 4b and 4d). These tests also demonstrate the verygood performance of the filter in conditions of strongbaseline drift—1 mV, 0.5 Hz is fully suppressed. Theextreme baseline drift of 1 mV, 1 Hz is reduced by 50%.It seems that processing of samples which are infor-mative for a large segment of the artifact (680 ms), isbeneficial for its suppression. Moreover, different cut-off and zero frequencies of FilterDxN could be easilydesigned by various combinations of DxN coefficients,defining different time intervals for integration.

The next aspect of the study concerns an algorithmfor real-time QRS detection. It is again designed withconsiderations of simplified computations applied oversingle channel, low-resolution ECGs. Although theQRS detector has adopted some elements from ourprevious QRS detector,12 it embodies additional smartdetails, leading to improvements of its performancewith about 0.5% (Se from 99.01% becomes 99.65%,PPV from 99.11% becomes 99.57%—Table 2). Thepresent version of the QRS detector, relies ondynamically updated amplitude/slope thresholds, his-tory of mean amplitudes/slopes calculated over thepast beats, and linear logical rules for identification ofthe significant R-peak. Here we will focus on theresults, which have been calculated for the interna-tionally recognized ECG databases—the AHA data-base, the MIT-BIH Arrhythmia database and the

Online Digital Filter and QRS Detector Applicable in ECG Systems 1813

European ST-T database, all being annotated bymedical experts. The statistical assessment on thesepublicly available databases gives us the opportunityto compare our results with already published meth-ods. Under our interest are the methods which applyonline or quasi-online analysis over single lead ECG.All results as reported by the authors, are summarizedin Table 2. The performance of the present methodcould be highly rated, comparable and even better thanother published real-time QRS detectors.

The common application of the QRS detector andFilter10x17 is demonstrated by two examples of ECGsignals with extreme baseline wandering. They havebeen selected to reproduce some typical situationsduring monitoring of cardiac patients, such as consid-erable artifacts of intensive body and electrodemovements (Fig. 6), respiration (Fig. 5), as well asarrhythmic alterations of the QRS waveforms (Figs. 5and 6). Looking at the performance of Filter10x17, weare confident that it is capable to remove very strongbaseline wanderings and thus to fit the output ECGaround the zero line and to enhance the QRS com-plexes. It is important to underline that the P-QRS-Twaves at Filter10x17 output remain almost unaffected,without differentiation or attenuation of their ampli-tudes. The filter does not succeed to restore only theseuseful ECG components, which are merged in the largebaseline wanderings, such as in Fig. 6. Within thesesegments, the QRS detector was unable to recognizeany ventricular contraction. The next look on the QRSdetector reveals its stable work with signals supplied byFilter10x17. We would stress the good resistivity of theQRS detector to the most common problems, avoidingthe false negative detection of small-amplitude QRS justafter QRSwith 3–4 times higher-amplitude (Figs. 5 and6); the false negative detection of consecutive QRSswith altering polarity and amplitude (Figs. 5 and 6); thefalse positive detection of large T-waves (Fig. 6).

CONCLUSIONS

The common application of the two fast proceduresfor Filter10x17 and the designed online QRS detector,provides a powerful tool for long-term ECG moni-toring systems, especially for those with small com-putation resources. We proved the high fidelity foridentification of each cardiac contraction, even underconsiderable artifacts of intensive body, electrodemovements and respiration, as well as arrhythmicalterations of the QRS waveforms. The special benefitof Filter10x17 is that it provides a combined high-passfiltering (1.1 Hz) and power-line interference (50 Hz)rejection, without attenuation of the QRS amplitude orelevation/depression of the ST-segment. The lack ofdistortions correlated with the heart rate, makes Fil-ter10x17 valuable for signal preprocessing in ECGmonitoring systems, which require accurate diagnosticanalysis based on ECG morphology, i.e. identificationof waves, measurement of time intervals and ampli-tudes used for identification of cardiac diseases.

ACKNOWLEDGMENT

This work has been supported by the NationalScience Fund Grant (BY-TH-101)/2005 of the Bul-garian Ministry of Education and Science.

REFERENCES

1Afonso, V., W. Tompkins, T. Nguyen, and S. Luo. ECGbeat detection using filter banks. IEEE Trans. Biomed. Eng.46:192–202, 1999. doi:10.1109/10.740882.2American Heart Association (AHA) ventricular arrhyth-mia ECG database. Emergency Care Research Institute

TABLE 2. Performances of published real-time QRS detectors, as reported by the authors.

Se (%) PPV (%) Method type Testing database

This method 99.65 99.57 R-peak detection using dynamically updated

amplitude/slope thresholds and prehistory of

the mean amplitudes/slopes of the last beats

AHA, MIT-BIH, European ST-T

Iliev et al.12 99.01 99.11 R-peak detection using dynamically updated

amplitude/slope thresholds

AHA, MIT-BIH, European ST-T

Dotsinsky and Stoyanov8 99.04 99.62 Evaluation of steep edges and sharp peaks AHA, MIT-BIH

Christov5 99.69 99.65 Combined adaptive threshold applied on

complex lead synthesized from arbitrary

primary leads

MIT-BIH

Alfonso et al.1 99.59 99.56 Filter-banks based algorithm MIT-BIH

Kunzmann et al.15 98.89 99.87 ECG derivative threshed-based analysis AHA, MIT-BIH, European ST-T

Ruha et al.26 97.8 – Matched filter and dual edge threshold detector 103, 105 files from MIT-BIH

Lee et al.16 99.58 – Time delay-coordinate mapping MIT-BIH

Paoletti and Marchesi23 99.15 – Fast QRS detector designed for noisy applications MIT-BIH

TABAKOV et al.1814

5200 Butler Pike, Plymouth Meeting, PA 19462, USA,1984.3ANSI/AAMI ECE57. Testing and reporting performanceresults of cardiac rhythm and ST segment measurementalgorithms. (AAMI) Recommended Practice/AmericanNational Standard, 1998.4Bai, Y., W. Chu, Ch. Chen, Y. Lee, Y. Tsai, and Ch. Tsai.The combination of Kaiser window and moving averagefor the low-pass filtering of the remote ECG signals. In:Proc. 17th IEEE Symposium on Computer-Based MedicalSystems, 273, 2004.5Christov, I. Real time electrocardiogram QRS detectionusing combined adaptive threshold. BioMed. Eng. OnLine3:28, 2004. http://www.biomedical-engineering-online.com/content/3/1/28. doi:10.1186/1475-925X-3-28.6Christov, I., I. Dotsinsky, and I. Daskalov. High-pass fil-tering of ECG signals using QRS elimination. Med. Biol.Eng. Comp. 30:253–256, 1992. doi:10.1007/BF02446141.7Dotsinsky, I., and T. Stoyanov. Optimization of bi-direc-tional digital filtering for drift suppression in electrocar-diogram signals. J. Med. Eng. Technol. 28:178–180, 2004.doi:10.1080/03091900410001675996.8Dotsinsky, I., and T. Stoyanov. Ventricular beat detectionin single channel electrocardiograms. BioMed. Eng. OnLine3:3, 2004. http://www.biomedical-engineering-online.com/content/3/1/3. doi:10.1186/1475-925X-3-3.9European Society of Cardiology ST-T Database, CNRInstitute of Clinical Physiology, Computer Laboratory, viaTrieste, 41 56100 Pisa, Italy. http://physionet.org/physiobank/database/edb/.

10Faes, Th., H. Govaerts, B. Tenvoorde, and O. Rompelman.Frequency synthesis of digital filters based on repeatedlyapplied unweighed moving average operations. Med. Biol.Eng. Comp. 32:698–701, 1994. doi:10.1007/BF02524254.

11IEC 62D/60601-2-27. Particular requirements for the safetyof electrocardiographic monitoring equipment (equivalentto AAMI EC 13), 1994.

12Iliev, I., V. Krasteva, and S. Tabakov. Real-time detectionof pathological cardiac events in the electrocardiogram.Physiol. Meas. 28:259–276, 2007. doi:10.1088/0967-3334/28/3/003.

13Jane, R., P. Laguna, N. Thakor, and P. Caminal. Adaptivebaseline wander removal in the ECG: comparative analysiswith cubic spline technique. IEEE Comp. Cardiol. 19:143–146, 1992. doi:10.1109/CIC.1992.269426.

14Kligfield, P., L. Gettes, J. Bailey, R. Childers, B. Deal, W.Hancock, G. Herpen, J. Kors, P. Macfarlane, D. Mirvis, O.Pahlm, P. Rautaharju, and G. Wagner. Recommendationsfor the standardization and interpretation of the electro-cardiogram, Part I. The electrocardiogram and itstechnology. J. Am. Coll. Cardiol. 49:1109–1127, 2007.doi:10.1016/j.jacc.2007.01.024.

15Kunzmann, U., G. von Wagner, J. Schochlin, and A. Bolz.Parameter extraction of ECG signals in real-time. Biomed.Tech. 47:875–878, 2002.

16Lee, J., K. Kim, B. Lee, and M. Lee. A real time QRSdetection using delay-coordinate mapping for the micro-controller implementation. Ann. Biomed. Eng. 30:1140–1151, 2002. doi:10.1114/1.1523030.

17Łeskia, J., and N. Henzel. ECG baseline wander andpowerline interference reduction using nonlinear filter

bank. Signal Process 85:781–793, 2005. doi:10.1016/j.sig-pro.2004.12.001.

18Levkov, Ch., G. Mihov, R. Ivanov, I. Daskalov, I. Chris-tov, and I. Dotsinsky. Removal of power-line interferencefrom the ECG: a review of the subtraction procedure,BioMed. Eng. OnLine 4:50, 2005. http://www.biomedical-engineering-online.com/content/4/1/50. doi:10.1186/1475-925X-4-50.

19Lynn, P. Online digital filters for biological signals: somefast designs for a small computer. Med. Biol. Eng. Comp.15:534–540, 1977. doi:10.1007/BF02442281.

20Ma, W. K., Y. T. Zhang, and F. S. Yang. A fast recursive-least-squares adaptive notch filter and it is applications tobiomedical signals. Med. Biol. Eng. Comp. 37:99–103, 1999.doi:10.1007/BF02513273.

21MIT-BIH Arrhythmia Database, http://physionet.ph.biu.ac.il/physiobank/database/mitdb.

22Mneimneh, M., E. Yaz, M. Johnson, and R. Povinelli. Anadaptive Kalman filter for removing baseline wandering inECG signals. IEEE Comp. Cardiol. 33:253–256, 2006.

23Paoletti, M., and C. Marchesi. Discovering dangerouspatterns in long-term ambulatory ECG recordings using afast QRS detection algorithm and explorative data analysis.Comp. Methods Programs Biomed. 82:20–30, 2006.doi:10.1016/j.cmpb.2006.01.005.

24Pei, S., and C. Tseng. Elimination of AC interference inelectrocardiogram using IIR notch filter with transientsuppression. IEEE Trans. Biomed. Eng. 42:1128–1132,1995. doi:10.1109/10.469385.

25Pipberger, H., R. Arzbaecher, A. Berson, et al. Recom-mendations for standardization of leads and of specifica-tions for instruments in electrocardiography andvectorcardiography: report of the Committee on Electro-cardiography. Circulation 52:11–31, 1975.

26Ruha, A., S. Sallinen, and S. Nissila. A real-time micro-processor QRS detector system with a 1-ms timing accu-racy for the measurement of ambulatory HRV. IEEETrans. Biomed. Eng. 44:159–167, 1997. doi:10.1109/10.554762.

27Shusterman, V., S. Shah, A. Beigel, and K. Anderson.Enhancing the precision of ECG baseline correction: selec-tive filtering and removal of residual error. Comp. Biomed.Res. 33:144–160, 2000. doi:10.1006/cbmr.2000.1539.

28Suppappola, S., and Y. Sun. Nonlinear transforms of ECGsignals for digital QRS detection: a quantitative analysis.IEEE Trans. Biomed. Eng. 41:397–400, 1994. doi:10.1109/10.284971.

29Tinati, M., and B. Mozaffary. Wavelet packets approach toelectrocardiograph baseline drift cancellation, Int. J. Bio-med. Imag. Article ID 97157:1–9, 2006.

30Thakor, N., and Y. Zhu. Applications of adaptive filteringto ECG analysis: noise cancellation and arrhythmiadetection. IEEE Trans. Biomed. Eng. 38:785–794, 1991.doi:10.1109/10.83591.

31Xu, L., D. Zhang, and K. Wang. Wavelet-based cascadedadaptive filter for removing baseline drift in pulse wave-forms. IEEE Trans. Biomed. Eng. 52:1973–1975, 2005.doi:10.1109/TBME.2005.856296.

32Zywietz, Chr. CTS-ECG Test Atlas. Hannover: Center forComputer Electrocardiography, Biosignal Processing,Medical School, 1999.

Online Digital Filter and QRS Detector Applicable in ECG Systems 1815