9 qrs detectionalgorithms

8/3/2019 9 QRS detectionalgorithms

1/14

I E E E T R A N S A C T I O N S O N B I O M E D I C A L E N G I N E E R I N G . VOL . 37. N O . I. J A N U A R Y 1990 85

A Comparison of the Noise Sensitivity of Nine QRSDetection AlgorithmsGAR Y M. FR IESEN, THOMAS C . JANNETT, M E M B E R , I E E E , MANAL AFIFY JADALLAH,STANFOR D L. YATES, STEPHEN R . QUINT, M E M B E R , IEEE,

AN D H . TROY NAGLE, FELLOW, I E E E

Abstract-The noise sensitivities for nine different Q R S detection al-gorithms were measured for a normal, single-channel lead 11, synthe-sized EC G corrupted with five different types of synthesized no ise. Thenoise types were electromyographic interference, 60 Hz powerline in-terference, baseline drift due to respiration, abrupt baseline shift, anda composite noise constructed from all of the other noise types. Thepercentage of QR S complexes detected, the number of false positives,and the detection delay were measured.

None of the algorithms were ab le to detect all QR S complexes with-out any false positives for all of the noise types at the highest noiselevel. Algorithms based on amplitude and slope had the highest per-formance for EMG-corrupted EC G. An algorithm using a digital filterhad the best performance for the composite noise corrupted data.

I . INTRODUCTIONN recent years the trend toward automated analysis ofI lectrocardiograms has gained momentum. Many sys-tems have been implemented in order to perform suchtasks as 12-lead off-line electrocardiogram analysis, Hol-ter tape analysis, and real-time patient monitoring . Recentliterature dealing with design consideration of cardiacpacem akers suggests that the latest generation of these de-vices employs an ECG analysis capability. All of theseapplications require the accurate detection of QR S com-plexes in the presence of noise. M any of the existing ECGanalysis programs req uire a relatively noise-free digitizedECG. Data corrupted with noise must either be filtered ordiscarded. ECG quality assurance not only requires hu-man or software noise detection schemes, but can alsoresult in the loss of clinically significant data. Filteringcan alter the signal and may require substantial compu-tational overhead. These issues are important design con-sideration for applications in real-time heart monitoring.Manuscript received November 16, 1987; revised September 2, 1988.This work was supported by the Engineering E xperiment Station at AuburnUniversity and the North Carolina Biotechnology Center.G. M. Friesen is with Auburn University, Auburn, AL 36830.T . C. Jannet t is with the Department of Electrical Engineering, Univer-M. A. Jadallah is with Alcatel Network Systems, Research TriangleS. L. Yates is with Bell Northern Research, Research Triangle Park.S.R. Q uint is with the Biomedical E ngineering Curriculum . UniversityH. T. Nagle is with the Department of Electrical and Co mpu ter Engi-IEEE Log Number 893 154 .

sity of Alabama, Birmingham, AL, 35233.Park, NC.NC .of North Carol ina, Chapel Hil l , NC 275 14.neering North Carolina State University, Raleigh, NC 27695.

The purpose of this paper is to quantify the relativenoise susceptibility of nine different QR S detectionschemes. The database consists of a synthesized normalECG (used as a gold standard) corrupted with four levelsof each of four types of noise. Th ese four noise types werealso combined to form a fifth composite noise source. Th eability of each algorithm to detect QR S complexes and tolocate the onset of each complex was measured on eachnoise-corrupted ECG as well as on the noise-free ECG.The n ine QR S detection algorithms were chosen from aliterature survey. Each algorithm was programmed inFortran from its published description. T he algorithms arenonadaptive. In programming each one, we attempted toimplement the essential features described by its authors.When possible, we tuned each algorithm for the noisesources we are applying by adjusting its parameters(thresholds, weighting constants, etc.). In other words,we attempted to make each of the nine QR S algorithmsperform at its best for our five noise types. The input sig-nal to all nine algorithms is a synthesized, ideal, invariantECG. Consequently. the exact location of the QR S com-plex is known before the noise sources are added.The results of this study will help in the developmentof a more robust clinical instrument by making the front-end signal processing more effective. A procedure to befollowed might be:1) Evaluate a QR S detection algorithm by using a goldstandard ecg waveform;2) evaluate a QR S detection algorithm using a set stan-dard ECG waveforms such as the MIT/BIH database;3 ) implement the algorithm in a VLSI chip with built-in self test features; and4) include the chip in a prototype clinical instrument.This paper addresses the techniques in step 1).In the next section we discuss the properties of noiseartifacts in ECG signals. Then we select certain featuresof the artifacts for our noise simulation. Next we describethe details of our implementation of the nine QR S algo-rithms. Finally we present the simulation results and com-pare the algorithms performance.

11. NOISEARTIFACTSN ECGsElectrocardiographic (ECG) signals may be corruptedby various kinds of noise. Typical examples are:

0018-9294/90/0100-0085$01 .O O O 1990 IEEE


2/14

86 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, V O L . 37, N O. I . JANUARY 1990

1) power line interference2) electrode contact noise3) motion artifacts4) muscle contraction (electromyographic, EMG)5 ) baseline drift and ECG amplitude modulation with6) instrumentation noise generated by electronic de-7) electrosurgical noise,

respirationvices used in signal processing, andand other, less significant noise sources [l]. A brief de-scription of each noise signal listed will be given below,and methods for approximating these signals will be dis-cusse d. Identification of the pertinent c harac teristic s ofeach noise signal will be given.A . Power Line Interference

Power line interference consists of 60 H z pickup (in theU . S . ) and harmonics which can be modeled as sinusoidsand combination of sinusoids [ 2 ] .See Fig . l (a) . Charac-teristics which might need to be varied in a model ofpower line noise include the amplitude and frequencycontent of the signa l. These characteristics are generallyconsistent for a given measurement situation and, onceset, will not change during a detector evaluation.Typical parameters:Frequency content-60 H z (fundamental) with har-Amplitude-up to 50 perce nt of peak-to-peak EC Gmonicsamplitude

B . Electrode Contact NoiseElectrode contact noise is transient interference causedby loss of contact between the electrode and skin, whicheffectively disconnects the me asurement system from thesubject. See Fig. I(b). The loss of contact can be per-manent, or can be intermittent, as would be the case whena loose electrode is brought in and out of contact with the

skin as a result of movem ents and vibration. This switch-ing action at the measurement system input can result inlarge artifacts since the ECG signal is usually capacitivelycoupled to the system. With the amplifier input discon-nected, 60 H z interference may be significant.Electrode contact noise can be modeled as a random ly-occurring rapid baseline transition (step) which decaysexponentially to the baseline value and has a superim-posed 60 Hz component. This transition may occur onlyonce or may rapidly occur several times in succession.Characteristics of this noise signal inc lude the amplitudeof the initial transition, the am plitude of the 6 0 Hz com-ponent, and the time constant of the decay.Typical parameters:Duration-1 sAmplitude-maximum recorder outputFrequency-60 Hz Time constant-about 1s

C. Motion ArtifactsMotion artifacts are transient (but not step) baselinechanges caused by changes in the electrode-skin imped-

(d )Fig. 1. Noisy e lectrocardiogram s: (a) powerline and motion art i fact ; (b)loose contacts; (c) respirat ion and EMG; (d) Instrumentation saturation.

ance with electrode motion. A typical example is shownin Fig . l (a) . As th is impedance changes, the ECG ampli-fier sees a different sourc e impedan ce, w hich form s a volt-age divider with the amplifier input impedance. There-fore, the amplifier input voltage depends on the sourceimpedance, which changes as the electrode positionchanges. The usual cause of motion artifacts will be as-sumed to be vibrations or movement of the subject. Theshape of the baseline disturbance caused by motion arti-facts can be assumed to be a biphasic signal resemblingone cycle of a sine wave. The peak amplitude and dura-tion of the artifact are variables.Typical parameters:Duration-100-500 ms


3/14

FRIESEN er a l . : NOISE SENSITIVITY OF QR S ALGORITHMS 87

Amplitude-500 percent of peak-to-peak ECG am-plitudeD. Muscle Contractions (EMG)

Muscle contractions cause artifactual millivolt-levelpotentials to be generated. T he baseline electromyogramis usually in the microvolt range and therefore is usuallyinsignificant. As shown in Fig. l(c), the signals resultingfrom muscle contraction can be assumed to be transientbursts of zero-mean band-limited Gaussian noise. Thevariance of the distribution may be estimated from thevariance and duration of the bursts.Typical parameters:

amplitudeStandard Deviation- 10 percent of peak-to-peak ECGDuration-50 msFrequency Content-dc to 10 000 H z [ l ]E. Baseline Drift and ECG A mplitude Modulation w ithRespiration

The drift of the baseline with respiration can be repre-sented as a sinusoidal component at the frequency of res-piration added to the ECG signal. See Fig. l(c) . The am-plitude and frequency of the sinusoidal component shouldbe variables. The amplitude of the ECG signal also variesby about 15 percent with respiration. The variation couldbe reproduced by amplitude modulation of the ECG bythe sinusoidal component which is added to the baseline.

Amplitude variation- 15 percent of peak-to-peakBaseline variation-15 percent of p-p EC G amplitude

Typical parameters:(p-p) ECG amplitudevariation at 0.15 t o 0 . 3 H z

F. Noise Generated by Electronic Devices Used inSignal ProcessingArtifacts generated by electronic devices in the instru-mentation system as shown in Fig. l(d) cannot be cor-rected by a QR S detection algorithm. The input amplifierhas saturated and no information about the EC G can reachthe detector. In this case an alarm must sound to allert theECG technician to take corrective action.

G. Electrosurgical NoiseElectrosurgical noise completely destroys the ECG andcan be represented by a large amplitud e sinusoid with fre-quencies approximately between 100 kHz and 1 MHz.Since the sampling rate of an ECG signal is 250 to 1000Hz, an aliased version of this signal would be added tothe ECG signal. The amplitude, duration, and possibly

the aliased frequency should be variable.Typical parameters:Amplitude-200 percent of peak-to-peak ECG am -Frequency Content-Aliased 100 kHz to 1 MHz Du-plituderation-1-10 s

111. TH E ECG DAT ABA SEA. The Uncorrupted Signal

The uncorrupted signal was obtained by recording theLead I1 ECG of a human volunteer on a Beckman analogEC G strip chart recorder inside a shielded instrumentationcage. This recording was manually digitized. The digi-tized single cycle of the ECG was copied and appendedto itself repetitively in ord er to form 37 cycles of the ECGfor a total of 32 s of data. T he digitized EC G was plottedon a graphics terminal and subsequently edited until it ap-peared consistent with the analog recording. T he digitizedECG was then transferred to a VAX 11/780. The digi-tized ECG has an equivalent sampling rate of 250 Hz , andis stored in a Fortran data file as a linear array of 8192single precision floating point numbers. The heart rate isa constant 69 beats per min, the QR S width is 88 ms, andthe R-wave amplitude is 1.08 mV. These characteristicsfall within the normal range according to Ganong [3]. Aplot of a segment of the normal, uncorrupted ECG ap-pears in Fig . 2 .B. The Simulated Noise

Since the purpose of this study was to ev aluate the noiserejection properties of nine QR S detection algorithms, weselected four different representative noise sources forsimulation:1) electromyographic interference because of its ran-dom properties and high frequency content,2) powerline interference because it is ubiquitous,3) baseline drift due to respiration because of its low4) abrupt shifts in the baseline d ue to its large first de-5 ) a composite of all of the above.Although we did not specifically simulate electrosurg-ical and instrumentation noise, they behave similarly tothe random model w e use for EM G. Using the same rea-soning, motion artifact is much like baseline drift in res-piration so it also was not specifically modeled.Now let us describe our noise models more precisely.In what follows, the dynamic characteristics of the fivenoise types are specified. In each case, the maximum am-plitude was selected well beyond the values of the typicalparameters given earlier to challenge the algorithms whileallowing reasonable results at intermediate noise levels.Each of the five types of noise is added to an uncorruptedECG at four different levels: 25, 50, 75, and 100 percentof the maximum am plitude.1 ) Electromyographic Inte$erence: This type of noise

is simulated by adding random noise to the ECG. Themaximum noise level is formed by adding random singleprecision numbers of +50 percent of the ECG maximumamplitude to the uncorrupted ECG. The reduced noiselevels are formed by scaling the random numbers by theappropriate amount. The random numbers are generatedusing the VAX-11 Fortran random number generator. A

frequency properties,rivative, and


4/14

88 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 37. NO. 1. JANUARY 1990

1:L- 1 ! - 1 . 1 . 1 i0 1 2 3 4Fig . 2 . Uncorrupted ECG.

. .

asa8YasU)

0-

cnfU-

ol5w)N

zU0)

cnfwY)-

cn5U0-

Fig

3 -2 -1 -0 -

.i- 1 I I

0 1 24 1J

2 -

- 1 ! 10 1 2

- 0 I0 1

. .1 !. .0

A -

I

L I0 1

3 . ECG corrupted with electromy ographic noise.

4 1J3 -2 -1 -0 -

I ,

0

23 I1

h1

I ,0

413 1 n I2

- .0 11

2

210

- 1- I

0Fig . 4. ECG corrupted with 60 Hz powerline interference.

plot of the ECG corrupted by electromyographic noise isgiven in Fig. 3.2) Powerline Interference: Sixty Hertz noise is gen-erated using the Fortran sine function generator. Themaximum noise level corresponds to a peak-to-peak am-plitude of 0.333 m V . A plot of the ECG corrupted by thistype of noise is given in Fig. 4.Harmonic powerline linefrequencies were not modeled since the 60 Hz componentwill be dominant.


5/14

FRIESEN et a l . : NOISE SENSITIVITY OF QR S ALGORITHMS

4 -3 -2 -1 -0

- 1 -

4 1

-b

mmLm0-

I

4

0In

--Kwc

c+--aUm

wLo

c40

mUIm

w0n

--

c+-caU0

wLo

E+--aUmw00

6 -4 -2 -0 -

-2 -1

J3 -2 -1 -o -A

3 -2 -1 -

0 -A

- 2 1 . I . I . I . i0 1 2 3 4

4 -2 -

0 -

-20 1 2 3 4

- 4 1 . I . I . I . i0 1 2 3 46:

-2 -- 4 : . I . I . , . i

0 1 2 3 4Fig. 5 . ECG corrupted with baseline drift due to respiration.

3 ) Base line Dri p Due to Respiration: This type of in-terference is simulated by adding a low frequency sinu-soid to the uncorrupted E CG. The frequency is 0 . 3 3 3 H zand the maximum amplitude is 1.0 mV peak. It is gen-erated in the same manner as the powerline interference,except that the frequency and the maximum amplitude aredifferent. Amplitude effects due to respiration were notmodeled. A plot of the ECG corrupted by baseline driftdue to respiration is given in Fig. 5 .

~

mmUmUI

0-

4

r)

w--

4

c)

wIn

-

4

r)w0n

-

4

c)-wLo-

+r)w0

--

89

4 13 -2 -1 -0 -4

4 13 -2 - t 1 I I1 -o -A P

4 1

0 1 2 3 4

4 ,

- 2 1 . I . I 1 I . 10 1 2 3 4

3 -2 -1 -0 -

- I -4 ,

- 1-2.

0 1 2 3 4Fig. 6. ECG corrupted with abrupt baseline shift

4 ) Abrupt Shift in Baseline: This type of interferencerepresents an abrupt shift in baseline due to movement ofthe patient while the ECG is being recorded. It is simu-lated by adding a dc bias for a given segm ent of the ECG.The m aximum noise level consists of six alternating base-line shifts of +0.5 or -0.5 mV. This resulted in fivebaseline shifts of + 1 O o r - 1 O mV and one of +0.5 mV.The reduced noise levels are scaled by the appropriateamount. A plot of the ECG corrupted by abrupt shifts in


6/14

90 IEEE TRANSACTION S ON BIOMEDICAL ENGINEERING, VOL. 37. NO. I, JANUARY 199G4 -

.a 3 -k 2 --) 1 -= 0 -00- 1

noise types to 50 percent of maximum and then summingthem. The reduced noise levels of the composite areformed by scaling the composite noise after i t has beenof the ECG s corrupted by the composite is given in Fig.formed and then adding it to the uncorrupted ECG. A plot

8 I 7 .

UEr 3 -- 2 -0 1 -w

-5

0- 1 ! I 1

0 1 2

I

0 - A

.Um 3 -- 2 -U 1 -wLn 0 -FI

-f- I I

0 1 2

4 10Er 3- 2U 1w 0m

-5

- 1 4 I 10 1 2

4. a 3- 2U 1wm Op.

U-I

- 1 ! I I0 1 2

5 1Ur 4- 3f0 2w 100 0c

0 1 2Fig. 7. EC G corrupted with composite noise.

baseline appears in Fig. 6. In reality, shifts in ECG base-lines are less abrupt than our model; hence we are per-forming a worst case simulation for the derivative-basedQR S algorithms.5) Composi te Noise: A com posite noise is constructedby combining all of the noise types described above. T hemaximum noise level is constructed by reducing the max-imum noise levels for each of the previously described

The software fo r this experiment is written in VAX-11Fortran and executed on a VAX-11/780 computer . Thesoftware package consists of five major components. Th ecollection of algorithms an d digitized E CG data files makeup two of the components. A scoring routine and a routineto tabulate the results account for two additional b locks.The fifth module is a main program which integrates thefour modules l isted above. T he algorithms are called se-quentially. Each of the data files illustrated in Figs. 2-7are submitted to the algorithm for QR S detection. Whenan algorithm detects a Q RS candidate, the scoring routineis called by the algorithm. Following the scoring proce-dure, the algorithm resumes its search for the next QR Sstarting at the next da ta point until the next QR S candidateis detected or until the d ata is exhausted. This procedurecontinues until all of the algorithms have been run o n eachof the ECG data files. The results are then tabulated.B. The Scoring Cr iteria

The scoring routine compares the onset of the QRS can-didate to a key file containing the locations of all of thevalid QR S onsets. If the candidate onset falls between theactual onset and the end of the QR S complex (2 2 samplepoints or 88 ms following the actual onset), it is scoredas a QRS detection. If the candidate onset occurs outsideof these boundaries, i t is counted as a false positive. Th epercentage of QR S complexes correctly detected is cal-culated at the end of each run by dividing the number ofQRS complexes correctly detected by the total number ofactual QR S complexes.An indication of the average time delay required fordetection is also calculated after each run . If the detectionoccurs after the actual onset but before the end of the QR Scomplex, it is classified as a late detection. The numberof sample points between the onset and the detection aresummed fo r all of the detections of that run. This numberis divided by the total number of detections to give theaverage time delay (in 4ms sample points).Since the search for the next QR S complex resumes atthe next sample point following the candidate onset, it ispossible that the same Q RS complex will be detected morethan on ce. O nly the first correct detection will be includedin the scoring of QRS-complexes found or the m easure ofdelay. S ubseq uent detections which fall within the bound-aries of the first detected QR S should b e ignored.In some systems, the search for the next Q RS does notresume until a minimum delay has passed. T his delay, orwindow, corresponds to the minimum distance to the nextQRS complex based on the m aximum possible heart rate.


7/14

FRIESEN rt al . : NOISE SENSITIVITY OF QR S ALGORITHMS 91

This method was not used since false positives could beskipped over, or a valid QR S complex might be over-looked if the QR S candidate w as actually a false positivelocated just before the real QR S complex. In our study,we did not use the windowing approach per se since itwould enhance the performance of the nine algorithms.We did adjust our scoring algorithm so that multiple de-tections of the same QR S would not result in a penalty.v . DESCRIPTIONF TH E ALGORITHMS

A. Algorithm Selection CriteriaA large number of QR S detection schemes are describedin the literature [2], [4]-[37]. It would be impractical tocompare all of them. Several considerations were used tolimit the number of QR S detections schemes to a reason-able cross section of the different basic techniques de-scribed in the literature.The two basic criteria used in selection were complex-ity and performance. Only relatively simple algorithmswere used. Any of the algorithms used in this study couldeasily execute in real-time on a common 8-bit micropro-

cessor with enough reserve processing capacity for otherrelated functions such as preprocessing o r EC G analysiscapabilities. Certain schemes, such as Holter tape analy-sis, require faster than real-time capability which placeseven more importance on this restriction.The performance criterion was the basis for rejectingQR S detection algorithms which were highly noise sen-sitive and gave large numbers of false positives at lownoise levels. All of the algorithms should be insensitiveto low noise levels if they are to be realistic candidatesfor incorporation in clinical systems.Each algorithm in this study is based on a specificscheme presented in the literature. However, they are notcopies and should be considered as a generic adaptationof the fundamental concept. A great deal of experimen-tation was used to determine which of the available QRSdetection algorithms were to be used and to determine theexact form of the adaptation. There are four basic typesof algorithms included in this study with at least twovariants of each type. The basic type is designated by atwo letter prefix AF for algorithms based on both am-plitude and first derivative, FD for algorithms basedon first derivative only, FS algorithms utilize both firstand second derivative, and the last category was desig-nated DF which refers to digital QRS pass filters.B. Algorithm Parameter Determination

Each of the algorithms used in this study employed oneor more preset constants, either as multipliers or asthresholds. In some cases these constants were not givenin the literature, w hile in others they were not com patiblewith the data format that we used. A tuning procedure w ascarried out in order to determine the value for these con-stants which would give the best results for the compositenoise corrupted data. The approximate values of theseconstants were determined by observing intermediate

stages of the algorithms when the normal uncorruptedECG was applied. The precise value of these constantswas determined by varying each of the constants in analgorithm independently and recording the combinationsof constants which gave the best results when the 75 or100percent composite noise corrupted ECG was applied,preference given to the highest noise level that allowedreasonable results. The rough constants were varied fromapproximately 75 to 125 percent of the preliminary valueusing at least ten different increments. If any of the finalvalues of the constants fell on or near the end points, theprocess was repeated using the final value from the pre-vious run as the starting point for the next run. The scor-ing criteria for the selection of the constants were basedon obtaining the highest possible value for the differencebetween the number of QR S complexes correctly d etectedand the number of false positives. If a range of values ofa particular constant gave equally good results, the pro-cess was carried out on the 75 and 100 percent levels ofthe other types of noise. The results of these runs sug-gested the values of the constants. Finally, the algorithmwas executed on all of the noise-corrupted data and theresults were tabulated. Th e results were checked to insurethat no decrease in performance had occurred followingthe final selection of the constants compared to the bestresults obtained during the tuning procedure.C. Algorithms Based on Amplitude and First Derivative

I ) A F l : The concept for this QRS complex detectorwas derived from the algorithm developed by Moriet-Ma-houdeaux [4].L et X ( n ) = X ( O ), X ( 1 ) , * X ( 8 1 9 1 )represent a one-dimensional array of sample points of thesynthesized digitized EC G. An amplitude threshold is cal-culated as a fraction of the largest positive valued elementelement of that arrayAmplitude threshold= 0 .3 max [ X ( n ) ] 0 < n < 8191

The first derivative Y ( n ) s calculated at each point ofX (n ) such tha tY ( n ) = X (n + 1 ) - X (n - 1) 1 < n < 8190.

A QR S candidate occurs when three con secutive points inthe first derivative array exceed a p ositive slope thresholdand are followed within the next 100 ms by two consec-utive points which exceed the negative (descending slope)threshold. All data points in the ECG between the onsetof the rising slope and before the end of the descendingslope must meet or exceed the amplitude threshold.Y ( i ) , Y ( i + l ) , Y ( i + 2 ) > 0.5an d

Y ( j ) , Y ( j + 1 ) < - 0 . 3where

( i + 2 ) < j < ( i + 2 5 )


8/14

92 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 37 . N O . I . JANUARY 1990

an dX ( i ) , X ( i + l ) , e - . , X ( j + 1)I mplitude threshold.

2 ) AF2 : This algorithm is an adaptation of the analogQRS detection schem e developed by Fraden and Neuman[ 5 ] . A threshold is calculated as a fraction of the peakvalue of the ECG .Amplitude threshold

= 0 . 4 ma x [ X ( n ) ]The raw data is then rectified:

0 < n < 8191.Y O ( n ) = x ( n ) f X ( n ) 2 0Y O ( n ) = - X ( n ) i f X ( n ) < 0

0 < n < 81910 < n < 8191

The rectified ECG is passed through a low level clipper:Y1 ( n ) = Y O ( n ) if Y O ( n ) 2 amplitude thresholdY1 ( n ) = amplitude threshold if Y O ( n )

< amplitude threshold.The first derivative is calculated at each point of theclipped, rectified array:

1 < n < 8190A QRS candidate occurs when a point in Y 2 ( n ) exceedsthe fixed constant threshold,

Y 2 ( i ) > 0.7.3) AF3: The concept for this algorithm was taken fromGustafson [ 6 ] . The first derivative is calculated at eachpoint of the ECG:

Y 2 ( n ) = Y l ( n + 1) - Y l ( n - 1 )

Y ( n ) = X ( n + 1 ) - X ( n - 1 ) 1 < n < 8190.The first derivative array is then searched for points whichexceed a constant threshold:

form only the first 1. 2 s of data is considered, but in thisexperiment the entire 32 s of ECG is used.Slope threshold = 0.70 max [ Y ( n ) ] 2 < n < 8189.

The first derivative array was searched for points whichexceed the slope threshold. The first point that exceedsthe slope threshold is taken as the onset of a QRS candi-date:Y ( i ) > slope threshold.2) FD2: This algorithm is a modification of an earlydigital QRS detection schem e developed by Holsinger [8].In the original form it had an unacceptably high incidenceof false positives, therefore it was substantially modified.The first derivative is calculated for the ECG.

Y ( n ) = X ( n + 1 ) - X ( n - 1 ) 1 < n < 8190.This array is searched until a point is found that exceedsthe slope threshold:

Y ( i ) > 0 .4 5A QRS candidate occurs if another point in the next threesample points also exceeds the threshold:

Y ( i + 1 ) > 0.45, orY ( i + 2 ) > 0 . 4 5 , orY ( i + 3 ) > 0.45.

This technique of requiring multiple, possibly nonadja-cent sample points to define a QRS candidate generallyallows the use of higher thresholds since noise often re-duces the value of the first derivative for one or more ofthe sample points of the R-wave. The w idth of the regionand the number of points w hich must exceed the thresholdare determined by the tuning procedure as specified ear-lier.Y ( i ) I . 1 5 .

Then the next three derivative values Y ( i + I ) , y ( i +2 ) , an d Y ( i + 3 ) must also exceed 0 . 1 5 .If the above conditions are me t, point i can be classifiedas a QRS candidate if the next two sample points havepositive slope amplitude products:Y ( i .+ 1 ) X ( i + 1) and Y ( i + 2 ) X ( i + 2 ) > 0 .

E. Algorithms Based on First and Second Derivative1) F s I : This algorithm is a simplification of the QRSdetection scheme presented by Balda [ 9 ] . The absolutevalues of the first and second derivative are calculatedfrom the ECG:

Y O ( n ) = A B S [ X ( n + 1 ) - X ( n - l ) ]2 < n < 8189

D . Algorithms Based on First Derivative Only Y l ( n ) = A B S [ X ( n + 2 ) - 2 X ( n ) + X ( n - 2 ) ]I ) F D I : This algorithm was adapted from one devel- 2 < n < 8189.oped by M enard [ 7 ] . The first derivative is calculated foreach point of the ECG, using the formula specified by These tw o arrays are scaled an d then sum med:

Menard: Y 2 ( n ) = 1 . 3 Y O ( n ) + 1 . 1 Y l ( n ) 2 < n < 8189.Y ( n ) = - 2 X ( n - 2 ) - X ( n - 1 ) + x ( n + 1 ) This array is scanned until a threshold is met or exceeded:+ 2 X ( n + 2 ) 2 < n < 8189. Y 2 ( i ) 2 1.0.The slope threshold is calculated as a fraction of the max-imum slope for the first derivative array. In the original Once this occurs, the next eight points are compared tothe threshold. If six or more of these eight points meet or


9/14

FRIESEN rr a l . : NOISE SENSITIVITY OF QR S ALGORITHMS 93

exceed the threshold, the criteria for identification of a Y l ( n ) = YO(n) + 4YO(n - 1 ) + 6YO(n - 2 )QR S candidate is met.2 ) FS2: This algorithm was adapted from the QR S de - + 4YO(n - 3 ) + YO(n - 4 )tection scheme developed by Ahlstrom and Tompkins[ lo] . The rectified first derivative is calculated from theECG:YO(^) = A B S [ X ( ~+ 1 ) - X (n - I ) ]

Two thresholds are use d, equal in magnitude but oppositein polarity. The output of the low-pass filter is scanneduntil a point with amplitude greater than the positivethreshold is reached. This point is the onset of a 160 mssearch region. The number of alternate threshold cross-ings is used to classify the initial crossing as either a base-< n < 8188. line shift, a QR S candidate, or a s noise:The rectified first derivative is then smoothed: If Y1 ( i ) > 2 1 . 0 , then search region onset = i .If no other threshold crossings occur within the 160 mssearch region, the occurrence is classified as a baselineshift. Otherwise, the following three conditions are tested:

Y l ( n ) = [YO(n - I ) + 2YO(n) + YO(n + 1)]/43 < n < 8188.

The rectified second derivative is calculated:Y2(n) = ABS[X(n + 2 ) - 2 X ( n ) + X (n - 2 ) ] Condition 1: If Y l ( i + j ) < -21 .0 0 < j < 40

3 < n < 8188. Condition 2: If Y l ( i + j ) < - 2 1 . 0 0 < j < 40,The rectified, smoothed first derivative is added to therectified second derivative: andY l ( 1 + k ) > 2 1 . 0 j < k < 40

0 < j < 40 ,Y3(n) = Y l ( n ) + Y2(n) Condition 3: If Y l ( i + j ) < - 2 1 . 03 < n < 8188. andThe maximum value of this array is determined and scaledto serve as primary and secondary thresholds: Y l ( i + k ) > 21 .0 j < k < 40Primary threshold = 0 .8 ma x [Y3 (n)]

3 < n < 8188and

Y l ( i + 1 ) < -21 .0 k < 1 < 40.Secondary threshold = 0.1 ma x [Y3 ( n ) ] If any of the above conditions apply, the occurrence isclassified as a QR S candidate. If additional thresholdcrossings occur, the occurrence is classified a s noise.2) DF2: This algorithm is an adaptation of OkadasQR S detection algorithm [121. The first stage smooths theECG using a three-point moving average filter:

YO(n) = [X(n - 1 ) + 2 X ( n ) + X (n + 1 ) ] / 4

3 < n < 8188.The array of the summed first and second derivatives isscanned until a point exceeds the primary threshold. Inorder to be classified as a QR S candidate, the next sixconsecutive points must all meet o r exceed the secondarythreshold:Y3 ( i ) > = primary threshold, and

Y3(1 + l ) , Y3(i + 2 ) ,1 < n < 8190.

* , Y3(i + 6 ) The output of the moving point averaging filter is passedthrough a low-pass filter.> secondary threshold.n f mIn the original version of this algorithm, the second de-

in the same manner as the first derivative. The perfor-mance of the original algorithm was erratic in the pres-ence of noise and was substandard when compared to thecurrent form as specified above.

Y I ( ~ ) [ 1 / ( 2 m + l ) ] C YO(^)rivative of the rectified ECG was smoothed, presumably k = n - mm < n < 8191 - m.

The difference between the input and output of the low-pass filter is squared:F, Algorithms Based on Digital F ilters Y2(n) = (YO(n) - Y l ( n ) ) * m < n < 8191 - m.1) DF1: This algorithm is an adaptation of the one de-veloped by Engelese and Zeelenberg [ l]. The ECG ispassed through a differentiator with a 6 2 . 5 Hz notch filter. The squared difference is filtered:

YO(n) = X ( n ) - X (n - 4 ) 4 < n < 8191.The differentiated, filtered data is then passed through adigital low-pass filter. m < n < 8191 - m.


10/14

94 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL 37. N O . I. JANUARY 1990

A fourth array is formed using the following formula:~ 4 ( n ) ~ 3 ( n ) if Y YO(^) - YO(^ - m ) ]

YO(^) - YO(^ + m ) ] > oY 4 ( n ) = 0 otherwise.

The maximum value of this array is determined and scaledto form the threshold:Threshold = 0.125 max [ Y 4 ( n ) ]

m < n < 8191 - m .A QRS candidate occurs when a point in Y 4 ( n ) exceedsthe threshold:

If Y4 ( n ) > threshold then, QRS candidate.Okada suggests setting m equal to three for best results;however, i t was determined using the tuning procedurethat larger values of m result in improved performance forseveral types of noise. As m increases the performanceincreases along with computational demands. The im-provement in performance begins to fall off at values ofm greater than four while the computational demands con-tinue to increase. T he value of m was set to six in orderto give good performance while keeping the computa-tional requirements at a reasonable level.

V I. RESULTSThe results from this experiment are l isted in Tables I-V. Each table gives the results for all of the algorithmsfor all levels of a given type of noise.

A . Electromyographic NoiseThe first table gives the results for the effects of elec-tromyographic noise. This noise type has broad-band fre-quency characteristics which overlap the frequency spec-trum of the QR S complex. The amplitude of the QR S

complex is, however, considerably greater than the noise.The ability of algorithm AF2 to correctly identify all ofth e QR S complexes without incurring any false positivescan be attributed to this fact. T he noise is effectively elim-inated in AF 2 on the basis of its amplitud e characteristics.The two other algorithms based on amplitude and slopedo not perform nearly a s well. At the maxim um noise cor-ruption level, neither AF1 or AF 3 is able to locate morethan 65 percent of the QR S complexes and each have morefalse positives than QR S complexes detected.The failure of algorithms AF1 and AF3 to achieve theresults obtained by AF2 is a consequence of the methodfor selecting the value of the scaling constant. The v alueof the scaling constant is obtained via the tuning proce-dure while the algorithm op erates on the com posite noise-corrupted data. The composite noise has both low- andhigh-frequency components. Changes in baseline result inthe selection of a low value for the scaling constant, oth-erwise QRS complexes on elevated baseline would beclipped. The use of the composite data in the tuning pro-cedure requires a comprom ise in the selection of co nstants

for all of the algorithms. Th e results fo r a given noise typ ecan often b e improved if the algorithm parameters are se-lected specifically for that noise type instead of the com-posite noise, as is the case in this study.A few gen eral comments seem appropriate at this po int.Altering the algorithm parameters may reduce the numberof false positives generated by a given algorithm for agiven noise type. In many cas es, the number of detectedQRS complexes will also drop. If an algorithm is unableto detect all of the QR S complexes without generatingfalse positives f or any set of parameters, th en it has a poorability to discriminate between that n oise type and the QR Scomplex. Each algorithm has a maximum potential to dis-criminate between noise and the Q R S , which depends onthe optimum selection of parameters. The best choice ofparameters for the composite noise may not yield the b estresults for any other type of noise.Algorithm AF3 does not use amplitude in the samemanner as AF1 o r AF2. Algori thms AF1 and AF2 makeuse of the large amplitude of the R-wave whereas AF3only requires that the product of slope and amplitude bepositive. It is not poss ible to reconfigure th is algorithm tofilter the noise on the basis of amplitude as it is with theprevious two algorithms, AF 1 and AF2.Algorithms based on derivatives only have the worstperformance of all. Electromyographic noise has first andsecond derivative characteristics that are similar to thoseof the QR S complex. A lgorithm FD1 is able to detect all37 QRS complexes, but it has more than five times thatmany false positives. O ther algorithms in this class dem-onstrate an even poorer ability to discriminate betweenQRS and noise. FD2 is the worst, locating 32 of 37 QRScomplexes w ith 41 7 false positives.The digital filters have lower levels of false positivesthan the algorithms based on first or first and second de-rivatives, but they are unable to locate all of the QRScomplexes. It is possible that the Q RS complexes are sup-pressed along with the noise by the filtering algorithms.B. Powerline Interference

The results for the 60 Hz powerline interference, l istedin Table 11, are considerably better than th ose for the EM Ginterference. The frequency spectrum for the powerlineinterference consists only of the 60 Hz fundamental andfirst harmon ic. Fiv e of the algorithm s are able to correctlylocate all of the 37 QR S complexes. Four of these, AF2 ,FD 1 , DF1, and DF2 a re ab le t o do so without generatingany false positives. Only on e algorithm gives false posi-tives at the highest noise level, although a second did atintermediate noise levels.Powerline in terference at the 100 percen t level is ableto effectively mask all QR S complexes from algorithmsAF3 and FS1. These algorithms give no false positivesand do not detect any QRS complexes at this noise level.In some systems, such as patient monitoring, this char-acteristic may be advantageous since the prolonged ab-sence of output from a detector is a clear indication of a


11/14

FRIESEN et al . : NOISE SENSITIVITY OF QR S ALGORITHMS 95TABLE IE L E C T R O M Y O G R A P H I CEM G)

% QRS Detected N o. of False Posit ives0 25 50 75 100 0 25 50 75 100

AF 1AF2AF 3FD 1FD 2FS 1FS2DF 1DF 2

10 010 010 010 010 010010 0100100

10010 010 010 010 0924310097

8610 09510 095763010 078

73 65 0 0100 100 0 065 51 0 897 97 0 092 86 0 081 62 0 4727 27 0 095 84 0 062 62 0 0

30200693221200

190308625733 54011

48031

20 741 727 8643613

TABLE I1P O W E R L I N EN T E R F E R E N C E% QR S Detected No. of False Posit ives

0 25 50 75 100 0 25 50 75 100AF 1AF 2AF 3FD 1FD 2FS 1FS2DF 1DF 2

10 0 100 59 59 19 0 0 0 0100 100 100 100 10 0 0 0 0 010 0 10 0 19 0 0 0 0 0 010 0 100 100 100 100 0 0 0 0100 loo 59 59 59 0 0 0 010 0 78 100 22 0 0 29 7 0 010 0 81 100 10 0 10 0 0 0 0 810 0 10 0 10 0 100 100 0 0 0 010 0 100 100 10 0 100 0 0 0 0

0000001500

problem requiring human intervention. This may be pref-erable to a system which gives low numbers of false pos-itives s ince this condition m ay be more difficult to detect.However, an algorithm which can discriminate between60 Hz noise and QR S complexes would always be pref-erable if it has the other required characteristics.C. Respiration

Baseline drift due to respiration presents a lesser chal-lenge to all of the algorithms except those based on am-plitude and first derivative. Th e results f or this noise typeare l isted in Table 111. All of the other algorithms giveideal performance. Low-frequency noise of this type isapparently ignored by algorithms based exclusively onderivatives or by algorithm s using digital filters.Algorithms which utilize amplitude are compromisedby the change in amplitude of the QR S complexes due tolarge changes in baseline. Since absolute amplitude is usedas a criterion for identification, QR S complexes at or neara minimum (trough) of the respiratory cycle may not havesufficient amplitude to meet the criterion. Since these al-gorithms operate on amplitude rather than frequency, theQR S complexes are filtered along with the noise.D. Baseline Shift

Abrupt baseline shift is a slightly greater challenge thanthe baseline drift due to respiration. The results for this

noise type are listed in Table IV . Although four of thealgorithms are able to locate all QR S complexes withoutgenerating any false positives, a majority of the algo-rithms exhibit some decrease in performance. Only twoalgorithms, FD1 and DF2, have any false positives. Bothdetect all QR S complexes correctly and have approxi-mately 20-30 percent as many false positives a s correctdetections.DFl suffers only a very slight decrease in performanceat the highest noise level by missing one QR S complex.Only FS2 has significant difficulty locating QR S com-plexes, missing all of them at higher levels of baselineshift. AF2 also experiences a lesser degree of difficulty.This is a consequence of the change in amplitude of sev-eral of the QR S complexes as in the case of baseline driftdue to respiration. The other algorithms based on ampli-tude are not affected since the maximum change in am-plitude is less for the abrupt baseline shift as com pared tothe baseline drift due to respiration.E. Composite Noise

Composite noise was as great a challenge as electro-myographic noise since only one algorithm, DF1, wasable to detect all QR S complexes without any false posi-tives. Refer to Table V for the results. Algorithms FD1and DF2 gave reasonable results at the 75 percent noiselevel. Algorithms AF 2, F D2, and FS1 w ere very sensitive


12/14

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 37. NO. I . JANUARY 1990

TABLE I11R E S P I R A T I O NR E S P )% QR S Detected No. of False Positives

0 25 50 75 100 0 25 50 75 100AF 1AF 2AF 3FD 1FD 2FS 1FS 2DF 1DF 2

100 100 100 76 65 0 0 0 0100 100 65 57 54 0 0 0 0100 100 100 86 76 0 0 0 4100 100 100 100 100 0 0 0 0100 100 100 100 100 0 0 0 0100 100 100 100 100 0 0 0 0100 100 100 100 100 0 0 0 0100 100 100 100 100 0 0 0 0100 100 100 10 0 100 0 0 0 0

TABLE IVB A S E L I N EH I F T% QR S Detected N o. of False Positives

0 25 50 75 100 0 25 50 75 100A F 1AF 2AF 3FD 1FD 2FS 1FS 2D F 1DF 2

10010010010010010010 0100100

~

IO 0100100100100100100100100

100 100 100 0 0 0100 59 59 0 0 0100 100 100 0 0 0100 100 100 0 0 6100 100 100 0 0 0100 100 10 0 0 0 0

3 0 0 0 0 0100 97 97 0 0 010 0 100 100 0 0 0

000

1200009

0006000013

TABLE VC O M P O S I T EOISE% QR S Detected No . of False Positives

0 25 50 75 100 0 25 50 75 100A F 1AF 2A F3F D 1FD 2FS 1FS 2DF 1DF 2

100100100100100100100100100

100100100100100704 3100100

958492100

1005924100100

81735410092762210095

7 0 0 0 0 370 0 0 0 1332 0 0 2 497 0 0 0 186 0 0 0 1378 0 20 283 32 119 0 0 0 610 0 0 0 0 084 0 0 0 0

775646227 22000

to the maximum noise level, resulting in a very high rateof fa lse positives.Although the individu al noise types are scaled down be-fore they are combined to form the composite noise, theresults are generally worse than for the individual noisetypes. The combination of the various noise types has asynergistic effect decreasing the performance. Compositenoise is a more realistic model of the noise problem w hichwould be expected in a clinical setting.F. QRS Detection Delay

Now let us briefly discuss the delay time in QR S detec-tion. In most cases the number of sample points delay inQR S detection remained relatively constan t at two to three

samples despite increases in noise level. Two exceptionsshould be no ted. Algorithms FS 1 an d FS 2 had the lowestdetection delay for the clean ECG, 0.0 an d 1.0 samplepoints, respectively. At low-noise levels for either EMG,powerline or composite noise, the delay increases to arange of 4 o 8 sample points. This suggests that thesetwo algorithms have difficulty distinguishing between highfrequency noise and the QR S onset. Since the delay forthe other algorithms is relatively con sistent, the actual on-set can be approximated by s ubtracting a fixed delay fromthe candidate onset. It may b e preferable to search for thepeak which is more precisely defined (using techniquessuch as template matching) in systems which require ac-curate fiducial point determination.


13/14

FRIESEN er al . : NOISE SENSITIVITY OF QR S A LG O R ITH MS 97

VII. CONCLUSIONNo single algorithm evaluated in this study is clearlysuperior for all of the types of noise considered. Fo r manyapplications DF 1 would be the obvious choice since it canbest handle comb inations of noise if the contribution fromeach of the individual noise types can be limited. The ef-fectiveness of algorithm DF1 can be in part attributed to

the powerline notch filter which could also be used in apreprocessing stage for any of the algorithms.Algorithms based on amplitude and slope are most im-mune to EM G noise. Sin ce this type of noise presents thegreatest challenge, these algorithms have a significant ad-vantage. These algorithms are sensitive to changes inbaseline which accounts for the decrease in performancewhen subjected to composite noise. If chang es in baselinecan be corrected by high pass filtering and/or a cubicspline technique [131, these algorithms would offer thehighest performance. Filtering EMG noise is more diffi-cult due to the frequency spectrum overlap with the QRScomplex, therefore algorithms which are insensitive tobaseline changes but sensitive to high frequency noisehave less potential than the amplitude-slope algorithm orreliable performance in a clinical setting.The ability of the algorithms to recognize differentforms of normal or abnormal QRS complexes or to ignorelarge peaked T-waves was not tested. Most instrumenta-tion systems utilize more than one channel (ECG lead)and each channel presents a different waveform. Th e sig-nal used in this study was obtained from a limb lead (11).No attempt was made to run the algorithms on a signalwhich would represent the signal obtained from other EC Gleads.The limitations described above do not seriously com-promise the usefulness of these results for many applica-tions. The performance of these algorithms was in partattributable to the tuning p rocedure. T his procedure wouldnot be practical in a clinical setting, but ad aptive mecha-nisms could be developed to configure the algorithms dur-ing an initial learning phase to a patients unique QRScomplex. It is probable that multichannel systems couldhave even higher performance, perhaps by the incorpo-ration of a voting scheme or signal processing methods.This study provides useful information about the noiseproperties of the nine algorithms. These results will provevaluable to those who use QRS algorithms similar to theones presented in this study.

ACKNOWLEDGMENTTh e authors wish to acknowledge the support of the En-gineering Experiment Station at Auburn University and

the North Carolina Biotechnology Center. T hey also wishto thank D r. J. Buchanan of the University of North Car-olina at Chapel Hill for his thoughful comments and re-view of the manuscript.REFERENCES

[ I ] John G. Webs ter , Ed . , Medical Instrumentation-Application andDesign. Boston: Houghton Mifflin, 1978.

[2 ] G . S . Furno and W. J . Tompkins, A learning filter for reducing noiseinterference, IEEE Trans. Biomed. E n g . , vol . BME-30, pp. 234-235, 1983.[3 ] W. F . G anong , Review ofMedica 1 Physiology, 1 l t h ed . Los Al tos,CA: Lange Medical , 1983.[4] P. M. Mahoudeaux et a l . , Simple microprocessor-based system foron-l ine ECG analysis , Med. Biol . Eng. Cornput ., vol . 19, pp. 497-500 , 1981 .[5] J . Fraden and M. R. Neuman, QRS wave detect ion, Med. Biol .Eng. Comput . , vol. 18, pp. 125-132, 1980.[6] D. Gustafson e t a l . , Automated VCG interpretat ion studies usingsignal analysis techniques, R-1044 Charles Stark Draper Lab.,Cambr idge , MA, 1977 .[7] A. Menrad e t a l . , Dual microprocessor system for cardiovasculardata acquisi t ion, processing and recording, in Proc. 1981 IEEE Inr.Con5 Industrial Elect. Contr. Instrument., 1981, pp. 64-69.[8] W. P. Holsinger e t a l . , A QRS prep rocess or based on digital differ-ent iat ion, IEEE Trans. Biomed. E n g . , vol . BME-18, pp. 212-217,1971.[9] R. A. Balda e t a l . , The HP ECG analys i s p rogram, Trends inComputer-Processed Electrocardiograms, J . H. vanBemnel and J . L.Willems, Eds. North Holland, 1977, pp. 197-205.[ l o ] M. . Ahlstrom and W. J . Tompkins, Automated high-speed anal-ysis of hol ter tapes with microcomputers , IEEE Trans. Biomed.E n g . , vol . BME-30 , pp. 651-657, O ct . 1983.[ l I ] W. A. H. Engelse and C. Zeelenberg, A single scan algori thm forQRS-detection and feature extract ion, IEEE Comput. C a r d . , LongBeach: IEEE Com puter Society, 1979, pp. 37-42.[I21 M . Okada, A digital fi lter for the QRS complex detect ion, IEEETrans. Biomed. Eng ., vol . BME-26, pp. 700-703, D ec. 1979.[13] C. R. Meyer and H . N. Keiser, Electrocardiogram basel ine noiseest imation and removal using cubic spl ines and state-space com pu-tation techniques, Compur. Biomed. Res., vol . 10, pp. 459-470,1977.[ I41 C. S . Weaver e t a l . , Digital filtering with applications to electro-cardiogram processing, IEEE Trans. Audio Electroacoust., vol . AU-16, pp. 350-389, Sept . 1968.[15] H. K. Wolf, J . D. Sherwood, and J . D. Kanon, The effect of noiseon the performance of several ECG programs, Proc. Comput. Car-d io l . , pp. 303-304, 19 76.[16] G . J . H . Ui j en , J . P . C. de Weerd , and A. J. H . Vendrik, Accuracyof QRS detection in relation to the analysis of high-frequency com-ponents in the electrocardiogram, Med. Biol . Eng. Compur . , vol.17, pp. 492-502, July, 1979.[ I71 C. N . Mead e t a l . , Development and evaluat ion of a new QRS de-tector/del ineator, Proc. Comput. Card io l . , pp. 251-254, Sept. 1979.[18] L. Sornmo et a l . , A mathematical app roach to QRS detect ion,Proc. Comput. Card io l . , pp. 205-208, 1980.[I91 A. L. Goldberg and V. Bhargava, Peak-to-peak ampli tude of the

high-frequency QRS: A simple, quantitative index of high-frequencypotent ials , Comput . Biomed. Res., vol . 14, pp. 399-406, 1981.[20] L. Sornmo et a l . , A method of evaluat ing QRS shapes features usinga mathematical m odel for the ECG, IEEE Trans. Biomed. Eng ., vol .[21] A. L. Goldberg and V. Bhargava, Computerized measurement ofthe first derivativ e of the QRS comp lex: Theo retical and practical con-siderat ions, Comput. Biomed. Res., vol . 14, pp. 464-471, 1981.[22] S . Blumlein e t a l . , Detection of signal associated with noise, Proc.

Comput. Card io l . , pp. 339-342, 1981.[23] T. Fancott e t a l . , Design considerat ions for noise immunity in theconcordia high speed ambulatory ECG tape analysis system, P roc.Comput. Card io l . , pp. 343-346, 1981.[24] M. N ygirds and L. Sornm o, A QRS del ineat ion algori thm with lowsensitivity to noise and morphology changes, Proc. Comput. Car-d io l . , pp. 347-350, 1981.[25] C. N . Mead er al . , A frequency-domain-based QRS classificationalgori thm, Proc. Comput. Card io l . , pp. 351-354, 1981.[26] F . E . M. rekelmans and C . D . R. de Vaal , A QRSdetect ion scheme

for mult ichannel ECG devices, Proc. Comput. Card io l . , pp. 437-440 , 1981 .[27] R. A. A. F. van Dam, F. E. M. Brekelmans, and J . S . Duisterhout ,A high performance m icroprocessor-based arrhythmia monitor,Proc. Comput. Card io l . , pp. 449-452, 1981.[28] J . L. Talmon and A. Hasman, A new approach to QRS-detect ionand typification, Proc. Cornput. Card io l . , pp. 479-482, 1981.[29] P. 0 . Borjesson et a l . , Ada ptive QR S detection based on maximuma posteriori est imation, 1EEE Trans. Biomed. En g., vol . BME-29,pp. 341-351, 1982.

BME -28, pp. 713-717, O ct . 1981.


14/14

9 8 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. V OL. 37. NO . I , J A N U A R Y I Y YO

1351

[361

A . V . Sahak ian , W. J . Tompkins , B. M. Tom pkins , and J . K. Kreul ,A microprocessor-based arrhythmia m onitodrecorder for the oper-at ing and recovery room s, M e d . Insrrumenr., vol . 17, pp. 131-134,Mar . -Apr . 1983 .N . V . Thakor , J . G. Webs ter , and W. J . Tompkins , Opt imal QRSdetect ion, M e d . Biol. Eng. Cornput., vol . 21, pp. 343-350, May1983 .-, Est imation of QRS complex power spectra for design of a QRSfi l ter, IEEE Trans. Biomed. E n g . , vol . BME-31, pp. 702-706, NOV.1984.J. Pan and W. J . Tompkins, A real-t ime QRSdetect ion algori thm,IEEE Trans. Biomed. E n g . , vol . BM E-32, pp. 230-236, M ar. 1985.L. Sornmo, 0 . Pah lm, and M. Nygi rds , Adap t ive QRS detec t ion :A study of performance, IEEE Trans. Biomed. En g., vo l . BME-32 ,pp. 392-401, June 1985.M. L . Ahl s t rom and W. J . Tompkins, Digital filters for real-timeECG signal processing using microprocessors , IEEE Trans. Biomed.E n g . , vol . BME-32, pp. 708-713, Sept . 1985.J . A. van AlstC and T. S . Schilder, Removal of base-l ine wanderand power-line interference from the ECG by an efficient FIR filterwith a reduced number of taps, IEEE Trans. Biomed. E n g . , vol .BME-32, pp. 1052-1060, Dec. 1985.P. S . Hamilton and W. J . Tompkins, Quanti tat ive invest igat ion ofQRS detect ion rules using the M IT/BIH arrh ythmia database, IEEETrans. Biomed. E n g . , vol . BME -33, pp. 1157-1165, De c. 1986.

abama, Mobi l e .

Gary M. Friesen was born in Long Beach, CA,on May 7, 1957. He received the B.E.E. degreefrom Auburn Universi ty, Auburn, AL. in 1980.From 1981 until 1985 he pursued graduate studiesin physiology and electrical engineering at Au-burn University in the areas of the pathophysiol-ogy of hypovolemic shock and medical instru-mentation.H e is currently completing the requirements forthe M.S.E. E. degree from Auburn Universi ty andthe M. D. d egree from the University of South Al-

Manal Afify Jadallah was born in Alexandria,Egypt , on Se ptember 7, 1964. She received theB.S.E.E. degree from North Carol ina State Uni-versi ty, Raleigh. in 1987 .She is currently working for the Research andDevelopment Department of Alcatel NetworkSystems and pursuing a Masters degree in electri -cal engineering. H er work on QRS algori thms wasdone while pursuing the baccalaurete degree. Sheis now involved in microcontroller board level de-sign for telecommunication systems under devel-opment at Alcatel .

Stanford L . Yates was born in Cincinnat i , OH ,on July 28, 1962. He received the B.S.E.E. de-gree from North Carol ina State Universi ty, Ra-leigh, in 1987.Since then, he has worked for Bell NorthernResearch in Research Triangle Park, NC . His cur-rent interests include real-time computing, medi-cal electronics, and digi tal s ignal processing.Mr. Yates is a member of Phi Kappa Phi andTau Beta Pi .

Stephen R. Quint (Savailable at the time of 69-M69-M82), photograph and biography not publication.

H. Troy Nagle (S66-M7O-SM74-F83) wasborn in Boonevil le, MS, on August 31 , 1942 . Hereceived the B.S.E.E. and M.S.E.E. degrees fromthe Universi ty of Alabama, Universi ty, in 1964and 19 66, the Ph.D. degree from Auburn Univer-si ty, Auburn, AL, in electrical engineering in1968, and the M.D. degree from the Universi ty ofMiami School of Medicine, Coral Gables, FL, in1981 .He is a Professor of Electrical and ComputerEngineering at North Carol ina State Universi ty,Raleigh, and Research Professor in the Biomedical Engineering Cumcu-lum at the Universi ty of North Carol ina, Chapel Hil l . He is widely pub-Thomas C . Jannett (M80) received the B S de- l ished in data acquisi t ion and control systems. He is coauthor of textbooksgree in engineer ing and the M.S . degree in in digital local desig n and sampled-da ta-control system s His current area sbiomedical engineering i n 1979 and 19 81, respec- of research are faul t-tolerant systems deslgn, design for testabi li ty, andtively, from The Universi ty of Alabam a, Birming- medical instrumentationham, and the Ph.D . degree in electrical engineer- Dr Nagle is a member of Phi Kappa Phi , Tau Beta Pi , Eta Kappa Nu,ing from Auburn Universi ty, Auburn, in 1986 Sigma Xi, and Omicron Delta Kappa. He is l is ted in W h o s Who in Amer-He is Assistant Professor of Electrical Engi- i ca , and is the recipient of the IEEE Centennial Medal He was Presidentneering and Biomedical Engineering at The Uni- of the Industnal Electronics Society in 1984 and 1985, Vice-president forversi ty of Alabama at Birmingham. His research Area Activit ies of the Computer Society in 198 6, and a member of theinterests include automated drug del ivery sys- IEEE Board of Directors represent ing the Computer Society dunng 1987terns, medical instrumentat ion, and system iden- and 1988 He is current ly the Vice President for Technical Activi t ies of thetification. IEEE

9 qrs detectionalgorithms

Documents