00083591

7/29/2019 00083591

1/10

IEEE 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. Vo. 18. NO 8. AUGUST I Y Y I 785Applications of Adaptive Filtering to ECG Analysis:Noise Cancellation and Arrhythmia Detection

Nitish V . Thakor, Seiiior Member, IEEE , and Y i-Sheng Zhu, Se~iiorMomher, IEEE

Abstract-Several adaptive fil ter structures are proposed fornoise cancellation and arrhythmia detection. T he adaptive fil-ter essentially minimizes the mean-squared error between aprimary input, which is the noisy E CG, and a reference input,which is either noise that is correlated in some way with thenoise in the primary input or a signal that is correlated onlywith ECG in the primary input. Different filter structures arepresented to eliminate the diverse forms of noise: baseline wan-der, 60 Hz power line interference, muscle noise, and motionartifact. An adaptive recurrent filter structure is proposed foracquiring the impulse response of the normal QRS complex. Theprimary input of the filter is the ECG signal to be analyzed,while the reference input is an impulse train coincident withthe QR S complexes. T his method is applied to several ar rhyth-mia detection problems: detection of P-waves, premature ven-tricular complexes, and recognition of conduction block, atrialfibr illation, and paced r hythm.

I . INTRODUCTIONMBU LA TORY electrocardiogram (ECG) recordingA s now routinely used to detect infrequent, asympto-matic arrhythmias or to monitor effects of cardiac drugsor surgical procedures. Recently, microprocessor-basedevent recorders have been developed that carry out on-line signal processing, data reduction. and arrhythmia de-tection [ ] , [2]. Computational power of the micropro-cessor allows us to implement digital filters for noise can-cellation and arrhythmia detection. For example. L ynn[3], [4] and Thakor and Didier [ 5]have developed integercoefficient and quantized coefficient digital filters, respec-tively, for real-time execution by microprocessors. Ahl-

strom and Tompkins [6]describe similar filters for real-time ECG signal processing.Adaptive fil tering technique has been shown to be use-ful in many biomedical applications. T he basic idea be-hind adaptive filtering has been summarized by Widrowet al . [7] and used in a variety of ECG processing appli-cations [8], [9]. One simple but important application isin 60Hz powerline interference cancellation. A referenceM anuscript received April 24. 1989: revised April 2. 1990. This workwas supported by Research Career Development A ward HL.01509 from theNational Institutes of Health and a Presidential Y oung I n\ estigator A ward(ENG 8451491 froin the National Science Foundation.N . V . Thakor is with the Department of Bi omedical Engineering. T heJ ohns Hopkins School of Medicine. Baltimore. MD 2 1205.Y .-S. Zhu was with the Department of Biomedical Enginecring. TheJ ohns Hopkins School of M edicine. B altimore. M D 21705. H e isnnw with

the Department of Electrical Engineering. University of Science and Tech-nology of China, (USTC). Bei,jing. China.IEEE Log Number9101193.

signal representing powerline interference from some partof the body (other than the ECG recording area) may beused to cancel powerline interference from the ECG. An-other application is in fetal ECG recording. Y elderman eral . [8] used the idea that the mothers own ECG recordedfrom one of the conventional leads can be used as a cor-related noise source for adaptive cancellation. T o improvethe signal-to-noise ratio, multiple channels are employedfor adaptive fi ltering. Dufault and Wilcox [101employedmultiple surface leads to discriminate P-waves. Anotherinteresting application is cancellation of cardiogenic in-terference from an impedance plethysmographic signal.Sahakian and Kuo [ l ] employed ECG signal as the ref-erence input to the adaptive filter to cancel the cardiogenicartifact from the thoracic impedance signal. A similar ideais used by Zhu and Thakor [ 121 to detect P-waves. How-ever. a more comprehensive scheme is needed for noisecancellation and arrhythmia detection in ambulatory ECG.The first aim of this paper is to demonstrate adaptivefilter application in noise cancellation. We develop spe-cialized filter structures for cancellation of noise arisingfrom diverse sources. The second aim is to show how anadaptive recurrent filter structure detects cardiac arrhyth-mias. Our idea is to build an impulse response of the QRScomplex and to detect as arrhythmias the signals whoseimpulse response deviates from normal.

11. A D A PT I V EILTER TRUCTURESA. Basic Adaptive Filtering Structure

Fig. I (a) shows a fi lter with a primary input that is anECG signal s I with additive noise n l , while the referenceinput is noise tz2, possibly recorded from another genera-tor of noise n2 that is correlated in some way with nl . Ifthe filter output is .v and the filter error is E = (sI +nl )- , thenE 2 = ( SI + f11) - 24(s1f !?I) +Y 2=(HI - )? + s + 2s1n1- 2ys,. (1)

Since signal and noise are uncorrelated, the mean-squarederror (M SE) is(2)[E2]=E[(n1- ) ? ] +E [ s : ] .

Minimizing the M SE results in a filter error output that isthe best least-squares estimate of the signal s I . he adap-0018-92949110800-078S~0l0 1991 IEEE

_ _ _ _ _ _ _ -~

7/29/2019 00083591

2/10

786

sl+nl----a-?IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL . 38, NO . 8. AUGUST 1991

(b)Fig. 1. Two adaptive fi lter structures. T ype I(a): the reference input isnoise nz correlated with noisen , ; the desired signal appears at E . Type I @ ) :The reference input is signal s2 correlated with signal s i : the desired siganlappears at y .

tive filter extracts the signal, or eliminates noise, by it-eratively minimizing the M SE between the primary andthe reference inputs.Fig. l(b) il lustrates another situation where the ECG isrecorded from several electrode leads. The primary inputs I +nl is a signal from one of the leads. A referencesignal s2 s obtained from a second lead that is noise free.The signal s I can be extracted by minimizing the M SEbetween the primary and the reference inputs. Using aprocedure similar to (1) we can show that

(3)Minimizing the MSE results in a fil ter output y that is thebest least-squares estimate of signal sI .B. The Least-Mean Squares (LMS) Algori thm

The LMS algorithm [7] is an iterative technique forminimizing the M SE between the primary and the refer-ence inputs. Usually a transversal filter structure is em-ployed and the fil ter coefficients or weights are obtainedusing the LMS algorithm. The L MS algorithm is writtenas

E[E*]= E[(s ,- y12] +~ [ n : ] .

(4)where W, = [w lk w2, wj, . . wnklT s a set of filterweights at time k, X , = [xl, x2, . . . xj , * * xn,]' is theinput vector at time k of the samples from the referencesignal, dk is the desired primary input from the ECG tobe filtered, y k is the filter output that is the best least-squares estimate of d,

E , = dk - k.Parameter p is empirically selected to produce conver-gence at a desired rate; the larger its value, the faster theconvergence. The time constant for convergence is =

1 /(4pcy) where CY is the largest eigenvalue of the autocor-relation matrix of the reference signal [6]. T his parametermay not be so large that it causes excessive misadjustmentor instability. To ensure stabil ity, CY >p >0.

Fig. 2. Schematic of adaptive recurrent filter. Primary input si(;) is thesample at time k for the P-QRS-Tcomplex i. Reference input is an impulsesequence (indicated as0. 0, I . . .0,0)coincident with recurrences of theQRS complexes. Filter output yi is the desired impulse response. Error t,is used to adapt filter weights W,.

C. The Adaptive Recurrent Fil ter (ARF )The objective of the ARF technique is to adapt fil tercoefficients, or weights, so that the impulse response ofthe desired signal i s acquired. Let the P-QRS-T signalcomplex span k =0 . ( J - 1) samples, and thereforethe transversal fi lter will require L weights. The referencesignal is an impulse coincident in time with the first sam-

ple of the signal complex. Each recurrence i = 1, 2, - .of the signal complex results in a new reference impulseand a new update of all the filter weights (F ig. 2). Thedesired impulse response is obtained by minimizing theM SE between the primary and the reference inputs.For the adaptive recurrent f il ter, the reference vector isxh = [O, 0, 1 * * * 01. Therefore,

wh+ = WI; f 2pEk. (2)At each time step only one filter weight is adapted. Allthe fil ter weights are adapted once each recurring cycle.D. Reference Impulse Detection

To implement the ARF, we must first begin by identi-fying a reference impulse train coincident with the QRScomplexes. T he reference impulse is located in such amanner that the fil ter weights span the entire QRS-T com-plex. This may be accomplished by placing the impulseat the very beginning of the QRS complex. This is veryconveniently done when a pacemaker is being used; thereference impulse sequence is obtained by detecting thepacemaker spike. For nonpaced rhythms, we must detectthe QRS complex. QRS detection is a very common firststep in all arrhythmia detection algorithms, and can becarried out in hardware [131or in software[141, [151. Thereference impulse is now coincident with each occurrenceof the QRS complex. T he actual fi lter weights are onceagain selected so as to span the entire QRS complex.

111 NOISE AN CELLAT I O NN AMBU LATO R YCGNoise in ambulatory recordings is contributed both bybiologic and environmental sources. Examples of envi-ronmental noise are 60 (or 50) Hz and its harmonics gen-erated by power l ines, radio-frequency and electrosurgi-

7/29/2019 00083591

3/10

THAKOR AND ZHU: ADAPTIVE FILTERING TO EC G AN AL Y S I S

T A B L E IF I LTER ESIGNUMMARY

787

Fi lter Primary Reference Input ResultNoi se Ty pe Type Input at ExampleEC G +noise Constant t Fig. 3Common-mode E Fig. 4Baseline wander 1(a)60Hz 1 a)

Motion artifact 1 a) ECG +noise Impulse v Fig. 6signalEM G l(b) aVf aVr-aVI Y Fig. 5sequence

cal noise, and instrumentation noise [16]. Examples ofbiologic interference are: baseline drift and wander, elec-tromyogram (E M G), and motion artifact [17]. T he firstaim of this paper is to present a step-by-step approach tocancellation of thpse diverse noise contributions to am-bulatory ECG. Table I summarizes all the designs andapplications considered in this paper.A . Baseline Wander Reduction

Van Alste and Schilder [181 describe an efficient finiteimpulse response (F IR) notch fil ter that is rather effectiveat removing baseline wander and power line interference.The adaptive fi lter to remove baseline wander is a specialcase of notch filtering, with the notch at zero frequency(or dc). Only one weight is needed, and the reference in-put is a constant with a value of 1 (Table I). T his fi lterhas a "zero" at dc and consequently creates a notch witha bandwidth of ( p / n )* s where f s is the sampling rate.Frequencies in the range of 0-0.5 Hz should be re-moved to reduce baseline drift. If the sampling rate is 500sample/s, the convergence parameter p should be smallerthan 0.003.Fig. 3 shows the result obtained by the adap-tive baseline canceller. Parameter p may be dynamicallyadjusted to obtain the desired low-frequency response.Note that this filter will produce some distortion of theST-segment since low-frequency components of ECG areattenuated. Si nce the selected value of p is small, thisfilter converges slowly and therefore cannot track abrupttransients produce by motion artifacts.

'

B. Adaptive 60U .CancellerFumo and Tompkins [191 and Sahakian and Fumo [20]describe filter designs that subtract a 60Hz sinusoid fromECG. Widrow et al. [7] describe a filter employing twoweights so that in-phase and out-of-phase components ofthe60Hz can be cancelled. In general, however, the pow-erline noise is not a pure 60 Hz (or 50 Hz) sinusoid, butis distorted. T herefore, we suggest the use of the true in-tefering signal as a reference. The common-mode signal,

usually recorded at the right leg reference electrode, istruly correlated with the noise in the ECG recording. Theprimary input to the filter is the ECG signal to be fi ltered,and the reference input is the common-mode signal (T ableI). Fig. 4shows an example of 60 Hz cancellation.C. Multilead Canceller for EMG Noise

EMG noise has a broad bandwidth which sometimesoverlaps that of the ECG [22]. Simple low-pass filtering,therefore, is not adequate. O ur idea is to employ morethan one ECG lead. Since electrodes are usually placed atdifferent locations, the EM G noise from various leads iqaybe uncorrelated. We ensure uncorrelated inputs to the fil -ter by selecting two orthonormal ECG leads. T he standardEC G lead system employs three limb leads-I, 11, and111-as well as three augmented leads-aVr, aV1, and aVf.From the analysis of the cardiac vector, we note that theaVr-aV1 vector is orthonormal to aVf. Noise in ortho-normal leads is expected to be uncorrelated. T he prima@

~ ~

7/29/2019 00083591

4/10

788 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL 38. NO R. AUGUST 1991

Fig. 4. 60 Hz fil ter: (a) original ECG with 60Hz noise; (b) filtered ECG:(c) separated 60 H z noise.

Fig. 5. EM G filter: (a) primary input, the ECG in lead aVf ; (b) referenceinput, derived from aVr-aVI: (c) filter error t , (d) f ilter output y.

input to the filter is the ECG signal f rom lead aVf, whilethe reference is the signal aVr-aV1 (T able I ). Fig. 5showsthe results of adaptive cancellation of E M G.beginning of each P-QRS-T complex (Table I ). The ad-aptation takes place only for the samples spanning the sig-nal complex, and subtraction of this complex from theECG leaves the motion artifact as residue (F ig. 6).Notethat since the fi lter does not adapt between QRS com-dexes. the baseline between complexes is simply inter-. Motion Artifact Cancellation . .

Motion artifact is usually the most difficult ofThispolated. This clearly results in some signal distortion. Theresulting ECG is not suitable for diagnostic quality dis-noise tobe eliminated from ambulatory ECGis because its spectrum completely overlaps that of theECG, and its morphology often resembles that of the P ,Play but the noise-free QRS comPlexes at the Outputsuch as heart rate measure-e in

QRS, and T waves [17], [21]. M ost linear fi ltering ap- ments and arrhythmia detection.proaches fail to solve this problem. The adaptive recur-rent filter [12] is useful in cancelling noise from signalsthat have a repetitive morphology. The primary input tothis filter is the ECG signal with motion artifact, and thereference input is an impulse that is coincident with theE. Two-Stage Mul tichannel Filter

In ambulatory E CG all forms of noise may occur simul-taneously and unpredictably. Fig. 7(a) illustrates a com-

7/29/2019 00083591

5/10

THAKOR AND ZHU: ADAPT IVE FtLTER lNG TO ECG ANALYS IS 789

Fig. 6. Application of recurrent fil ter to motion artifact cancell ation: (a)EC G with motion artifact; (b) impulse sequence coincident withQR S com-plexes; (c) filter error e, the motion artifact, (d) filter output.

h I x3

(b)Fig. 7. (a) Two-stage multiinput adaptive filter: W1: filter for baseline re-moval; W2 filter for 60 Hz cancellation; W3: fil ter for EM G and motionartifact cancellation. (b) Cancellation of baseline wander, 60Hz, EMG,and motion artifact from ambulatory ECG: d: ECG signal to be filtered;yl : filter estimate of baseline wander; y2: filter estimate of the common-mode 60 Hz noise; c l : EC G without 60Hz and baseline wander; 13: im-pulse coincident with QR S complex; 2: EM G and motion artifact; y3:filtered ECG.

7/29/2019 00083591

6/10

790 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 38, NO . 8, AUGUST 1991

bination f il ter structure. The primary input is the ECGsignal recorded f rom a subject who is jogging. The noiseincludes 60Hz, baseli ne wander, E M G noise, and motionaqifact. The first stage of the filter separates baseli newander y l and 60 Hz noise y2 [Fig. 7(b)]. T he secondstage is a recurrent fi lter (Section 11-C) to remove EM Gnoise and motion arti fact. A two-stage fi lter is necessary:if we did not remove the baseline and 60 Hz noise first,there might be false QRS detections, and these wouldcause the impulse sequence to the recurrent fi lter to beincorrect. The outputof the second stagey3 is the desiredsignal free of noise.

IV . A R R H Y T H M I AETECT IONAdaptive fi ltering can be applied to arrhythmia analy-sis. This application is facili tated by two facts: 1)the ECGsignal is usually characterized by a well-defined P-QRS-Tcomplex, and 2) the signal compexes are recurrent wi theach heart beat. Under normal situations, the morphologyremains stable from beat to beat (although there may besome minor variations). Any significant departure in mor-phology indicates presence of an arrhythmia. U nder nor-mal circumstances the P-QRS-T complex remains wellsynchronized. If this normal sequence is disrupted, as inthe case of many arrhythmias, the adaptive filter picks outuncorrelated components in the sequence.

A . Adaptive Cancellation of the QRS-T ComplexL et us first examine the abil ity of the ARF to acquirethe impulse response of the QRS-T complex. Fig. 8 illus-trates the waveform of a normal ECG, coincident impul-ses obtained after QRS detection, and a gradual adaptationof the transversal fi lter. T he result is an excell ent cancel-lation of the QRS-T complex, leaving behind the P-wavesequence as the filter error. If a large enough convergenceparameter p is selected, the A RF should adjust to smallbeat-to-beat variations in the QRS morphology. P-wavescan subsequently be used by algorithms to detect atrialarrhythmias.

B. Detection of Ectopic BeatsEctopic beats are usually characterized by a morphol-ogy that is distinct from that of normally conducted QRScomplexes. Fig. 9shows an ECG signal with infrequentectopic beats that are abnormally conducted and, there-fore, have different morphologies. The A RF first adaptsto the impulse response of the normally conducted QRScomplexes, resulting in a minimal f il ter error. A subse-quent ectopic beat results in a significant misadjustment.The filter error, or residue, clearly delineates the ectopicbeat. T he ARF quickly reacquires the impulse response

of the normal complexes, and as a result adaptation errorfor subsequent beats is minimal.

R

.....A--.-----. /Lv.-.m-:a-.---P PFig. 8. (a) A normal ECG signal with well defined P-Q RS-T sequence. (b)Each reference impulse I coincides with the begining of each Q RS-T com-plex. (c) The filter error output. The A RF learns the impulse response ofthe QRS-T complex, and cancels the learned complex from the ECG at eachsuccessive beat. After several beats the adaptation is compl ete, sothat theQRS-T complex is completely cancelled, leaving P-waves as residue (orthe filter error output).

C. P-Wave Detection i n Conduction BlocksP-waves, in view of their small amplitudes, are verydifficult to detect. In conduction disorders, the P-wavesare dissociated from the QRS complexes, making the de-tection problem even more difficult [ o], [121. In fact,P-waves may occasionally and randomly overlap the QRScomplexes in the case of second or third degree blocks.Fig. 10 shows that the A RF first achieves complete can-cellation of the QRS-T complex as described earli er. T headaptation error now prominently displays the P-wave se-quence. Since the P-wave sequence is uncorrelated withthe QRS-T complex, the adaptive filter considers it asnoise and does not adapt to i t (even when there is a perfectoverlap of a P-wave and a QRS complex). T he presentresults are by no means perfect. Further signal processingmay be required, especiaIly when EM G or motion artifactis present.

Ywye..LL.w... I YVL.a.....6 uwl. lMN, -.-._I . l d d l , 1 I D M ~ ~ ~ ~ ~.J V DOJ J - ~ J J I . U J L I ~ U L w 1771U L I I G iiisLiL uicUI c i c u r i a i ariuElectronics kneineers. Inc. All rights reser ved. Second-class mt ae e oaid at New L ork. NY and at ad&tional mailing o?fifices. Postmster: Send address

7/29/2019 00083591

7/10

THAKOR AND ZHU: ADAPT IVE F ILTER ING TO EC G AN AL Y S I S 79

(b)i \i \ r r

.... n I

Fig. 9. Detection of ectopic beats by the ARF . (a) EC G signal. (b) Ini-pulses coincident with QRS complexes. (c) Filter error output. The filterfirst acquires the impulse response of the normal QRS complex, so that theQRS complex is graduall y cancelled (top set of traces). When the filterinput contains ectopic beats (middle and lower set or traces), the filter pro-duces a very large adaptation error. A5 a result, the filter error output clearlydelineates the ectopic beats.

D. Detection of Atrial FibrillationAtrial fibrillation exhibits no apparent P-waves, and in-stead a fluctuating wave pattern is seen throughout thebaseline. The QRS complexes may not recur at regularintervals. Using the A RF, we acquire the impulse re-sponse of the QRS complex and adaptively cancel the QRScomplexes from the ECG. T he residue comprises the atrialsignal. SIocum et al. [23] suggest calculating the auto-correlation function to extract the rhythm of the atrialwaveform. Fig. 11 shows the autocorrelation of the ECGsignal, the ventricular signal, and the adaptation errorcomprises principally of the atrial signal. Atrial f ibril la-tion is thus identified from the differing autocorrelationfunctions of the atrial and the ventricular rhythms.

E. Paced RhythmsWhen a patient has a pacemaker implanted, the surfaceECG carries a pacemaker spike artifact. The pacemakerspike is only a few milliseconds wide and hence is readily

Fig. 10. Detection of conduction block. (a) ECG signal with P-wavesmarked. As a result of the conduction block, the P-waves and the QRScomplexes are dissociated. (b)A series of impul ses that are coincident withthe QRS complexes. T his serves as the reference input to the filter. ( c) Thefil ter error output. T he A RF adapts only the impulse response of theQRScomplexes, and cancels these from the ECG signals. This results in P-wavesas filter error. Even when P-waves and QRS complexes overlap (partiallyor fully) they are delineated.

detected. The ARF is triggered by the pacemaker spike,and the impulse response of the paced rhythm is acquiredby adapting the fi lter weights (Fig. 12). Occasionally thepacemaker fails to initiate paced rhythm (owing to leadfailure or the altered pacing threshold of the heart). In thatcase the paced and the nonpaced signal complexes exhibitdifferent morphologies. T he ARF , triggered by a non-paced beat, registers a large adaptation error. This tech-nique can be used to monitor pacemaker performance andfailure.V . DISCUSSION

There are several advantages to the adaptive fil teringapproaches just described. T he most signif icant feature ofthese fi lters is that they allow estimation of the underlyingsignal in the absence of apriori knowledge of the statis-

7/29/2019 00083591

8/10

792

\/

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 38. NO . 8. AUGUST 1991

Fig. 11 . Detection of atrial fibrillation. (a) Composite atrial fibrillationsignal. (b) Fi lter error (mainly atrial rhythm). (c) Fi lter output (mainlyventricular rhythm). A daptive cancell ation of the QRS complex leaves be-hind the baseline signal consisting mainly of atrial fibrillation waves. Dif-ferent rhythms of ventricular and atrial signals can be distinguished fromthe autocorrelation functions. (d) A utocorrel ation function of the ventric-ular rhythm. (e) A utocorrelation function of the atrial rhythm.

tical or spectral properties of the signal and noise. Thesefilters are easy to implement on modem microprocessorswith numeric capabiliti es. An ideal application of thesefilters is in ambulatory arrhythmia monitoring [ I ] , (21.

R I

h##(c)

Fig. 13. Distortion arising from i naccurate placement of the reference im-pulse. (a) Correct cancell ation of the QRS complexes with P wave as theresidue. (b) Leftward movement (marked #) of the impulse results in in-accurate cancellation of the impulse response of theQRS complex. (c) Dis-tortion resulting from a rightward movement (marked ##) of the impulse.

The ARF acquires the impulse response of the normalQRS complex. This fi lter structure is superior to the con-ventional adaptive fil ter that continuously adapts since fil -ter weights degrade between signal complexes. Previ-ously, Ferrera and Widrow [9]proposed a time-sequencedfilter in which a large number of separate transversal fi l-ters are constructed for each sample within the signalcomplex. T he principal limitation of the time-sequencedfilter is that it requires a great deal of memory.Two limitations of the ARF technique are apparent. Thefirst problem is that the reference impulse must be exactlycoincident with the signal complex. When this is doneindependently by detecting QRS complexes, there is thepossibility of uncertainty and error. QRS detection, es-pecially in the presence of noise and artifact, can be in-accurate. Fig. 13shows the fi lter error and resulting dis-tortion in the impulse response of the QRS complex, dueto inaccurate placement of the reference impulse. Thesecond problem arises from unusual beat-by-beat varia-

7/29/2019 00083591

9/10

THAKOR A N D ZHU: ADAPT IVE F ILTER ING TO ECG ANALYS IS 793

tion that may prevent complete adaptation from takingplace. On the one hand, abrupt variation (as in Fig. 9) canbe detected from a large filter error, and even can be usedto detect ectopic beats. On the other hand, gradual vari-ations in QRS morphology (due, for example, to sinus ar-rhythmia) may lead to incomplete adaptation and insuffi-cient filter error for detection purposes. The convergenceparameter p must be carefully selected to track rapid orslow changes as desired.Noise cancellation requires different strategies for dif-ferent sources. A n adaptive fi lter with constant or unityreference input is used to cancel baseline wander. Imple-mentation of the baseline wander fi lter is considerablysimpler than in previous designs [181. Still, the lower cor-ner frequency may be high for some applications. Thisfi lter is inappropriate when the ST -segment shape mustnot be distorted at all. Powerline noise (60 Hz) can becancelled by a variety of methods, including simple ana-log and digital notch fil ters. T he adaptive filter does notoffer a significant advantage, except that the notch band-width can be easily control led by a single convergenceparameter. Our new idea here is to use a common-modesignal as the reference, so that the reference is highly cor-related with noise in the ECG. EM G noise is usually re-duced by conventional low-pass fi ltering. T he adaptivefilter helps remove EMG components whose spectrummay overlap that of the ECG. Our idea is to take advan-tage of the lack of correlation in the noise from ortho-normal ECG leads. The multichannel fi lter. of course,does not work effectively when only a single lead is avail-able or when EM G noise originates at all the electrodes.Motion artifact is the most difficult problem, and the re-current fi lter only partially solves i t. Because of the rela-tively slow convergence time of the filter, large nonsta-tionary motion artifacts that overlap with theQRS complexcannot be cancelled. The recurrent fil ter also distorts thesignal somewhat since adaptation takes place only overthe QRS complex samples spanning the filter. Conse-quently, this filter is more suited to applications such asrhythm analysis in ambulatory monitoring and less suitedto diagnostic ECG analysis.The adaptive recurrent fi lter is employed in variousforms for ectopic beat and arrhythmia detection: P-wavedetection, ectopic beat detection, atrial arrhythmia anal-ysis, and pacemaker evaluation. Arrhythmia detection isa complex problem, often involving diverse morphologiesand rhythms that vary among different subjects as well asover time for the same subject. A data adaptive algorithm,therefore, is desirable. T he proposed adaptive filters onlyform one step in a complex strategy to detect arrhythmias.In view of their minimal computational burden, these fil-ters should find application in microcomputer-based ar-rhythmia monitors [2]. Additional pattern recognition al-gorithms [23] may be required to detect a vast range ofarrhythmias normally encountered in diagnostic ECGanalysis [11.

REFERENCES[ I ] N. V . T hakor. From Hol ter monitors to automatic implantable de-fibrillators: Developments in ambulatory arrhthymia monitoring.IEEE Trans. Biotned. Eng., vol. BM E-31, pp. 770-778. 1984.121 N. V . Thakor, J . G. W ebster, and W. J . Tompkins, Design, imple-mentation and evaluation of microcomputer-based ambulatory ar-rhythmia monitor, Med. Bid. Eng. Cornput., vol. 22, pp. 151-159,1984.[3] P. A . Lynn. R ecursive digital fi lters for biological signals, Med.Bid. Eng.. vol. 9, pp. 37-43. 1979.[4] -. Transversal resonator digital filters: Fast and flexible on lineprocessors for biological signals. Mrd. B ol . Eng. Cotput., vol.21. pp. 718-730. 1983.1-51 N . V. T hakor and D. M oreau, Design and analysi s of quantizedcoeffi cient digital fil ters: A pplication tobiomedical signal processingwith microprocessors, Med. Bid. Eng. Cornput., vol. 25. pp. 18-25. 1987.[6] M. A . A hlstrom and W . J . Tompkins, Digital filter for real-timeECG signal processing using microprocessors. IEEE Truns. Bi omed.Eng., vol. B M E-32, pp. 708-713. 1985.[7] B. Widrow, J . R . Glover, J . M. McCool, et al., A daptive noisecancelling: principles and applications. Proc. IEEE. vol. 63, pp.1692-1716, 1975.181 M . Yelderman, B . Widrow, J . M . Ciotfi, E . Hesler. andJ . A . L eddy.E CG enhancement by adaptive cancellation of electrosurgi cal inter-ference, lEEE Trans. Biomcd. Eng.. vol. B M E-30, pp. 392-398,1983.[9] E. R. Ferrera and B. W idrow, Fetal electrocardiogram enhancementby time-sequenced adaptive filtering, IEEE Traris. Biomecl. Eng.,

vol. BM E-29, pp. 458-460. 1982.[l o] R . A . D uFault and A . C . Wil cox. D ual channel P-w ave detectionin the surface EC G via the L M S algori thm, in Proc. IEEE/dth Annu.Con5 Eng. Med. B ol . Soc., 1986, pp. 325-328.[ I ] A . Sahakian and K . H. Kuo. Cancelling cardiogenic artifact inimpedance pneumography, in Proc. 9/hAnn. Cong. Erig. Med. Bid.Soc., 1985, pp. 855-859.(121 Y . Zhu and N . V . Thakor, P-wave detection by adaptive cancell a-tion of QRS-T complex, in /roc / EEE /8 t h Anri. Co)if: Erig. Mrcl.Biol. Soc.. 1986. pp. 329-331.1131 N. V . Thakor, J . G . Webster. and W . J . Tompkins, Optimal QRSdetector, Mcd. B ol. Eizg. Cor7ipu/.,vol. 21, pp. 343-350, 1983.[141 J . Pan and W . J . Tompkins. A real-time QRS detection algorithmlEE E Trans. Bi omed. Eng.. vol. B M E-32, pp. 230-236, 1985.[IS] 0. Pahlni and L . Sornmo, Software QRS detection in ambulatorymonitoring-a review, Med. Eiol. Eng. Cotnpur., vol. 22, pp. 289-297. 1984.[I61 J . C . Huhta and J. G. W ebster. I nterference in biopotential record-ing, in Eiornedicul Electrode T echnology: Theory aiid Practice. H.A . M iller and D. C . Harrison, E ds. New Y ork: Academic. 1974.

pp. 129-142.[ 171 -. 60-Hz interference in electrocardiography. lEEE Trans.Biorned. Oig. , vol. BM E-20. pp. 91-101. 1973.181 J . A . Van Alste and T. S. Schilder. Removal of baseline wanderand power-line interference from the ECG by an efficient FIR filterwith a reduced number of taps, IEEE Tram. Bi omed. Eng.. vol.BME-32. pp. 1052-1062, 1985.191 G. S. Furno and W . J . Tompkins, A learning filter for removingnoise interference. IEE E Trans. Bi omed. Gi g . . vol. BME-30, pp.234-235. 1983.101 A . V . Sahakian and G . F . Furno. A n adaptive filter for distortedline-frequency noise, Biorricd. Sc i . Iristrum., vol. 19, pp. 47-52,1983.1211 N . V . Thakor, J . G. Webster, and W . J . Tompkins, Estimation ofQRS complex power spectra for design of QRS filters, IEEE Trcrns.Biotned. Eng. , vol. BM E-31, pp. 702-706, 1984.1221 J . Slocum, A . Sahakian, and S. Swiryn, Computer detection ofatrioventricular dissociation f rom surface electrocardiograms duringwide QRS complex tachycardias, Circ.ulation, vol. 72 , pp. 1028-1036, 1985.(231 Q. Cheng, H. S. L ee, and N . V . Thakor. EC G waveform analysisby significant point extraction. 11. Pattern matching, Cotnput.Bi omed. Res., vol. 20, pp. 428-442. 1987.

7/29/2019 00083591

10/10

194 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL . 38. NO. 8. AUGUST 1991

Nitish V. Thakor (S'78-M '81 -SM '89) receivedthe B.Tech. degree in electrical engineering fromthe Indian Institute of Technology. Bombay, in1974 and the Ph.D. degree in electrical and com-puter engineering from the University of Wi scon-sin. Madison, in 1981For two years he worked as an Electronics En-gineer for Phil ips India Between 1981 and 1983he served on the faculty of Northwestern Univer-sity. and is currently on the faculty of the Biomed-ical Engineering Department of The J ohns Hop-kins Medical S chool, B altimore, M D. Hi s current research interests arecardiovascular and neurosensory instrumentation, signal processing. andmicrocomputer applications in medicine He also servesas a consultant tomedical instrumentation industri es in these areasDr Thakor is a recipient of two research awards a Research CareerDevelopment A ward from the National Insti tutes of Health anda Presiden-tial Y oung Investigator A ward from the National Science Foundation

Y i-Sheng Zhu (M '88-SM '90) was born in Chinain 1945 He graduated from the Department ofElectrical Engineering, University of Science andTechnology of China (USTC). Beijing. in 1968Since 1978 he has been on the faculty of Elec-trical Engineering at the USTC where his teachingand research i nterests have been in digital signalprocessing, information theory, microcomputer-based medical instrumentation, pattern recogni-tion, and VLSI applications Between 1986 and1988 he was a visiting scholar in the Department

Dr. Zhu is a recent recipient of an award from the National Scienceof Biomedical Engineering, The J ohns Hopkins Uni versity, Baltimore, M DFoundation of China