A

Ma

b

a

ARRAA

KBQES

1

dcEibd[choetasstotwa

1d

Biomedical Signal Processing and Control 7 (2012) 118– 128

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

j o ur nal homep a ge: www.elsev ier .com/ locate /bspc

novel method for detecting R-peaks in electrocardiogram (ECG) signal

.Sabarimalai Manikandana,∗, K.P. Somanb

Department of Electronics and Communication Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, Tamilnadu, IndiaCenter for Excellence in Computational Engineering and Networking, Amrita Vishwa Vidyapeetham, Coimbatore, Tamilnadu, India

r t i c l e i n f o

rticle history:eceived 22 November 2010eceived in revised form 3 March 2011ccepted 4 March 2011vailable online 2 April 2011

eywords:

a b s t r a c t

The R-peak detection is crucial in all kinds of electrocardiogram (ECG) applications. However, almostall existing R-peak detectors suffer from the non-stationarity of both QRS morphology and noise. Tocombat this difficulty, we propose a new R-peak detector, which is based on the new preprocessingtechnique and an automated peak-finding logic. In this paper, we first demonstrate that the proposedpreprocessor with a Shannon energy envelope (SEE) estimator is better able to detect R-peaks in case ofwider and small QRS complexes, negative QRS polarities, and sudden changes in QRS amplitudes over

iosignal processingRS detectionCG analysisignal detection

that using the absolute value, energy value, and Shannon entropy features. Then we justify the simplicityand robustness of the proposed peak-finding logic using the Hilbert-transform (HT) and moving average(MA) filter. The proposed R-peak detector is validated using the first-channel of the 48 ECG records ofthe MIT-BITH arrhythmia database, and achieves average detection accuracy of 99.80%, sensitivity of99.93% and positive predictivity of 99.86%. Various experimental results show that the proposed R-peakdetection method significantly outperforms other well-known methods in case of noisy or pathological

signals.

. Introduction

Automatic detection of the R-peaks in a long-term electrocar-iogram (ECG) signal is the most important step for diagnosis ofardiac disorders, heart-rate variability analysis, biometric, andCG coding systems [1–3]. The performance of these systems heav-ly rely on the accuracy of the R-peak detector. Many methodsased on the derivatives [6], digital filters [7–10], linear pre-iction, wavelet transform [1,12–14], mathematical morphology15,16], and empirical mode decomposition (EMD) [17], geometri-al matching [18], neural networks [19] and hybrid approach [11]ave been developed for the detection of R-peaks. Methods basedn the filtering techniques and decision rules are computationallyfficient and hence ideal for any automatic ECG analysis [9]. Most ofhe methods include a preprocessing or feature extraction stage and

decision stage [4]. Generally, preprocessing stage applies variousignal processing techniques to accentuate the QRS complex anduppress noises but most of them have some drawbacks. In [7], theradeoff between misses and false detections relies on the selection

f bandwidth of the filter and size of the moving-window integra-or. The WT-based QRS detector has the choice problem of motheravelet and scales to obtain QRS events. Although the EMD-based

pproach in [17] can overcome the choice problem of basis function,

∗ Corresponding author. Tel.: +91 0422 2656422.E-mail address: msm.sabari@gmail.com (M.Sabarimalai Manikandan).

746-8094/$ – see front matter © 2011 Elsevier Ltd. All rights reserved.oi:10.1016/j.bspc.2011.03.004

© 2011 Elsevier Ltd. All rights reserved.

selection of a set of intrinsic mode functions (IMFs) is very difficultunder noisy environments. The performance can be improved bydesigning more effective filtering and better threshold adjustmentprocedures [17]. However, it is hard to design a single compre-hensive preprocessing technique for achieving simultaneous QRSenhancement and noise reduction effectively in practice. Therefore,most of the works paid attention on constructing suitable decisionrules based on the preprocessing results.

The decision stage of the filtering approaches commonly con-sists of a set of heuristic rules and a set of tactics for finding theapproximate location of candidate R-peak. These thresholds areadapted periodically based upon amplitudes, durations and RR-intervals of past detected R-peaks. In such a case, performancerelies highly on the accurate estimation of initial parameters in thelearning phase. Most of methods had higher false-negative (FN) andfalse-positive (FP) detections in case of small and wider QRS com-plexes. Therefore, additional decision rules have been developedfrom a positive view but they have negative effects. For example,in [9], failure instances of the five QRS detection methods withand without use of secondary threshold were investigated by usingthe MIT-BITH arrhythmia database records 108 (negative R-peaksand high noises) and 208 (wide-QRS complexes). Experimental

results showed that the detection accuracy of the Method I, mod-ified Method I, and Method III is very poor for negative R-peaks inrecord 108 [9]. Furthermore, Method I and Method III fail to detectsmall and wider QRS complexes in record 208. In this case, modifiedMethod I, Method II and modified Method II had better detections

dical Signal Processing and Control 7 (2012) 118– 128 119

wtfwolstoctarcRcoto

tlficldntmp2ktS

2

iSRawreSttEiztzWltlle

2

vc

g gg g

Bandpass Filtering

(BPF)

Amplitude Normalization

(AN)

Shannon Energy (SE) Computation

Hilbert Movin Aver a e

ECGFirst-Order Forward Differencing (FOFD)

Zero-Phase Filtering

Stage 1: Linear Digital Filtering

Stage 2: Smooth SE Envelope Extraction

Stage 3: Peak-Finding Logic

x[n] f[n] d[n]

[n]

s[n]

Transformation

Positive Zero Crossing Point (ZCP) Detection

Stage4: Identification of the real R-Peak in the ECG within 25 samples

Filtering

z[n]

M.Sabarimalai Manikandan, K.P. Soman / Biome

ith use of secondary threshold but it results in higher FP detec-ions for noisy ECG signal. Moreover, many methods had higher FNor ECG with small and wider QRS complexes. Therefore, two rulesith adaptive amplitude-dependent and time-dependent thresh-

lds were widely adopted to reject or include identified R-peaksocated at tm and tn: (i) if tn − tm < 0.2 s (refractory period) and (ii)earch back if tn − tm > 1.5RRavg. These rules may improve detec-ions for regular rhythms but some rules may be in conflict withthers. Furthermore, searchback mechanism cannot be halted inase of irregular rhythms with varying QRS complexes. Moreover,here are often lots of thresholds defined in heuristic rules. Thebove detection issues clearly indicate that it is hard to find a properule set in case of (i) wide QRS complexes, (ii) low-amplitude QRSomplexes, (iii) negative QRS polarities, (iv) sudden changes inR intervals, (v) sudden changes in QRS amplitudes, (vi) suddenhanges in QRS morphologies, and (vii) sharp P- and T-waves. Allf the aforementioned constraints on the preprocessor and heuris-ic rules of the existing QRS detectors demonstrate that detectionf R-peaks is still very challenging task.

In this paper, we propose a new preprocessor which is based onhe Shannon energy envelope estimator and a simple peak-findingogic using the Hilbert-transform (HT) and moving average (MA)lter to combat the detection problem of unusually shaped QRSomplexes and noises. The rest of this paper is organized as fol-ows. In Section 2, the four-stage R-peak detection methodology isescribed in detail. We here introduce our new preprocessor and aovel automatic peak-finding scheme without using any amplitudehreshold and prior knowledge of detected R-peaks. The experi-

ental results to demonstrate the effectiveness of the proposedreprocessor and peak-finding scheme are presented in Sections.2 and 2.3. We experimentally evaluate our method using the well-nown MIT-BIH arrhythmia database, and a comparison betweenhe proposed methodology and other approaches is presented inection 3. Finally, in Section 4, we discuss and conclude our study.

. The proposed R-peak detection methodology

The block diagram of the proposed R-peak detection algorithms shown in Fig. 1. It consists of four stages, namely, digital filtering,hannon energy envelope extraction, peak-finding logic, and true-peak locator. The first stage of the proposed algorithm includes

bandpass filter, an amplitude normalization and first-order for-ard difference operation to emphasize the QRS complex and to

emove the noise in ECG signal. In the second stage, Shannon energystimation and zero-phase filtering are applied to obtain a smoothhannon energy (SE) envelope that plays the most critical role inhe proposed algorithm. We can observe that major local maxima inhe SE envelope represent approximate locations of the R-peaks inCG signal. In the third stage, the proposed peak-finding techniques developed based on the Hilbert transformation, drift removal andero-crossing point (ZCP) detection. The proposed technique iden-ifies accurate locations of the local maxima by detecting positiveero-crossing points in Hilbert-transformation of the SE envelope.

e here show that the positive ZCP in Hilbert sequence of the enve-ope, s[n], indicates local peaks of the s[n] and the MA filter reduceshe effect of ZCP drift in positive and negative directions. Finally,ocations of the local maxima are used as guides to find accurateocations of the R-peaks in ECG signals. The detailed discussions onach stage in Fig. 1 are presented in the following subsections.

.1. Noise suppression and enhancement of QRS complex

In the realistic environments, ECG signals may be corrupted byarious kinds of noise, including power line interference, electrodeontact noise, motion artifacts, muscle contraction and baseline

Time In stants of R-peaks

Fig. 1. Block diagram of the proposed four-stage R-peak detection methodology.

drift and also have large P and T waves. Therefore, the digital filter-ing stage is constructed using bandpass filter (BPF) and first-orderdifferentiation to accentuate the QRS complex, and to reduce noiseand the influence of the P and T waves. The 4th order Chebyshevtype I bandpass filter is designed for the bandwidth of [6 18] Hz.Here, the filter is applied in both the forward and reverse direc-tions to avoid the phase distortion. Then, the filtered signal, f[n],is differentiated to provide information about the slope of the QRScomplexes. The differentiation of the filtered ECG is implementedas

d[n] = f [n + 1] − f [n], (1)

which acts as a high-pass filter. The differentiation stage furtherreduces the interference of tall P- and T-waves. The output of thedifferentiator is a bipolar signal and thus a rectification is requiredto simplify the detection of negative R-peaks. After band-pass filter-ing and differentiation, we normalize the differentiated ECG (dECG)signal, d(n), by

d̃[n] = d[n]

maxNn=1(|d[n]|) , (2)

where N denotes the number of samples in ECG segment.

2.2. Shannon energy and smooth envelope extraction

After differentiation, the dECG signal is passed through a non-linear transformation to obtain positive peaks regardless of polarityof QRS complexes. The main objective of transformation is to usesingle-sided threshold mechanism and to enhance the QRS com-plexes [5]. In literature, the squaring transformation is widely used

but it considerably diminishes the magnitude of the candidate R-peaks of low-amplitude QRS complexes and wide QRS complexes[9]. Here, we study the performance of different nonlinear transfor-mation techniques using the low-amplitude QRS complexes, widerQRS complexes, and noisy ECG signals. The absolute value, energy

120 M.Sabarimalai Manikandan, K.P. Soman / Biomedical Signal Processing and Control 7 (2012) 118– 128

−1

0

1

2O

rig

inal

EC

GS

ign

al

−1

−0.5

0

0.5

No

rmal

ized

dE

CG

0.5

1

Sh

ann

on

En

erg

yE

nve

lop

e

1

0.5

En

erg

yE

nve

lop

e

1

0.5

0

Ab

solu

te E

nve

lop

e

0 1 2 3 4 5 6 7 8 9 10

0.5

1

time (sec)

En

tro

py

En

velo

pe

−1

0

1

0 1 2 3 4 5 6 7 8 9 100

0.5

1

time (sec)

1

0.5

0

0.5

1−1

−0.5

0

0.5

0

0.5

1

(a) (b)

F ed ECc candie s.

vnE

a

e

s

s

Ttbhbr

pdiT3FedaieRTaiwsls

ig. 2. Illustrates the performance of different envelopes extracted for a differentiatomponents and the energy (squarer) approach decreases the magnitude of R-wavexhibits the small difference between the successive peaks than the other envelope

alue, Shannon entropy value, and Shannon energy value of theormalized dECG signal are computed using the following usingqs. (3)–(6)[20]

(n) = |d̃[n]|, (3)

[n] = d̃2[n], (4)

e(n) = −|d̃[n]| log(|d̃[n]|), (5)

[n] = −d̃2[n] log(d̃2[n]). (6)

hen, the filtering is applied to obtain peaks corresponding tohe QRS-complex portions and smooth out the spikes and noiseursts that is implemented using a rectangular impulse response,(k) of length M. Here, the filtering operation is performed inoth the forward and reverse directions, i.e., conv(s, h) → s′ →everse the data → s′

r → conv(s′r , h) → s′′ → reverse the data → y.

This zero-phase filtering provides sharp peaks around QRS com-lex regions and smooth out spurious spikes. The smoothnessepends on the filter length L. Generally, the L should be approx-

mately the same as the duration of possible wider QRS complex.he length of the filter is found empirically. For sampling rate of60 samples/s, the filter length of 55 samples is found in this study.ig. 2 illustrates the performance of Shannon energy, Shannonntropy, energy and absolute value envelopes of the normalizedECG signals. Experimental results show that the Shannon entropypproach stresses the effect of low noise components that resultsn the envelope with artifacts and more unwanted peaks. Thenergy (squarer) approach decreases the amplitude of the small-wave candidates much more than that of high-amplitude ones.his approach results in the envelope with large high/low peakmplitude ratio. The Shannon entropy and the energy approaches

ncrease the difficulty of the R-wave peak detection task. Therefore,

e here introduce Shannon energy approach to obtain a positiveignal. The Shannon energy envelogram (SEE) approach has the fol-owing advantages over conventional approaches: (i) it results inmall deviations between the successive peaks; (ii) it reduces the

G (dECG) signal. The Shannon entropy approach emphasizes the effect of low noisedates. The Shannon energy approach magnifies medium-amplitude values and thus

effects of low-value noise components; and (iii) it produces smoothsharp local maxima which clearly indicate the time instants of R-peaks. Therefore, the Shannon energy envelope of a normalizeddifferentiated ECG (dECG) signal, d̃[n], is computed by using an Eq.(6). The major advantages of SE approach show that it can lead to themuch better R-peak detection performance in the presence of smalland wider QRS complexes, and non-stationary noise. Experimentalresults, shown in Fig. 2, demonstrate that major local maxima in asmooth Shannon energy envelope, s[n], indicate approximate loca-tions of the R-peaks in a ECG signal. Hence, the smooth Shannonenergy envelope are processed further to detect R-peaks.

2.3. Hilbert transform-based peak-finding logic

The detection of a R-wave peak is achieved by comparingthe envelope of a ECG signal against a fixed/adaptive amplitude-dependent and RR interval-dependent thresholds. Most R-peakdetection methods use a similar approach to determine the thresh-old. The detection thresholds are adapted periodically based uponamplitudes, durations and RR-intervals of past detected R-peaks. Insuch a case, performance relies highly on the accurate estimationof initial parameters in the learning phase. Almost all methods useadditional decision rules for the reduction of false-positive detec-tions and introduce secondary thresholds to detect missed R-peaks.These heuristic rules may improve detections for regular rhythmsbut some rules may be in conflict with others. Furthermore, search-back mechanism with secondary threshold cannot be halted in caseof irregular rhythms with varying QRS complexes and noise. More-over, there are often lots of thresholds defined in heuristic rules. Inthis paper, we introduce a novel automatic peak-finding logic by

exploiting the property of the Hilbert transform (HT) and using themoving average (MA) filter.

The main objective of this paper is to demonstrate the use of theHilbert-transform for reducing the complexity of the local maximafinding task. The Hilbert transform is widely used for analyzing the

M.Sabarimalai Manikandan, K.P. Soman / Biomedical Signal Processing and Control 7 (2012) 118– 128 121

−0.3 −0.15 0 0.15 0. 3−0.5

−0.25

0

0.25

0.5

0.75

1.0am

plitu

de

r(t)=1/(1+t2)

H[r(t)]=t/(1+t2)

zero−crossing point

candidate R−wave model

Hilbert transform ofthe candidate R−wave

candidate R−peak

Fl

ir

x

Tbb

X

T

x

wFirHtilw

asQipttelIstitfvfiamdt

0 0.5 1 1.5 2 2.5 3 3.5 4

0

0.5

1

0 0.5 1 1.5 2 2.5 3 3.5 4−0.5

0

0.5

1

rn

HT[rn] R−peak

0 0.5 1 1.5 2 2.5 3 3.5 4−0.5

0

0.5

1

time (sec)

TP TP

FN TP FN

odd−symmtery function

TP TP

TPa

b

c

TP TP TP TP TP TP TP

TP TP TP TP

TP TP TP TP

TP

TP: True Positive FN: False Negativ e

after the removal of lo w−frequency drift ofthe odd−symmtery function as shown in plot (b )

time (sec)

ig. 3. Illustration of the maximum finding task. The maximum value of the enve-ope function r(t) corresponds to the zero-crossing point of the r̂(t).

nstantaneous amplitude and frequency of the signal. The HT of aeal signal x(t) is defined as

ˆ(t) = H[x(t)] = 1�t

× x(t) = 1�

∫ ∞

−∞

x(�)t − �

d�. (7)

he Hilbert transform defined in the time domain is a convolutionetween 1/�t and x(t). The Fourier transform (FT) of the x̂(t) is giveny

ˆ (f ) = F[

1�t

× x(t)]

= F[

1�t

]F[x(t)] = −jsgn(f )X(f ). (8)

hen, the Hilbert transform of the signal x(t) can be computed as

ˆ(t) = IFT[X̂(f )] where X̂(f ) ={

jX(f ) f < 0−jX(f ) f > 0

(9)

here IFT denotes the inverse Fourier transform, and X(f) is theourier transform of the signal x(t). To demonstrate the max-ma finding task using the Hilbert transform, the even function,(t) = 1/(1 + t2), is considered as a R-wave envelope model. Theilbert transform of the even function, r(t), is given by r̂(t) =

/(1 + t2). The graphical representation of the r(t) and r̂(t) is shownn Fig. 3. It can be observed that the maximum value of the enve-ope function r(t) corresponds to the zero crossing point of the r̂(t)

hich is referred as odd-symmetry function (OSF).In practice, the strength of peaks in the filtered ECG is high

round the R-wave portions. The amplitudes of the peaks in themoothed envelogram are smaller for the low-amplitude and wideRS complexes. Therefore, effectiveness of the peak-finding logic

s studied using different envelopes. Fig. 4(a) clearly shows that theositive zero-crossing (or negative-to-positive transition) points ofhe odd-symmetry function correspond to locations of the peaks inhe envelogram, and the performance of the peak-finding logic isxcellent. Fig. 4(b) indicates the failure instance of the detectionogic due to the low-frequency drift in the odd-symmetry function.t can be observed that the zero-crossing points of the third andixth peaks are shifted in positive and negative directions, respec-ively. The degree of shift in a zero-crossing point and its directions directly related to the relative peak amplitude and spacing of thewo maxima and the sign of the peak difference. In practice, low-requency (LF) drift may arise because of the large peak-amplitudeariation and the baseline shift of the peaks. In such cases, the peak-

nding logic may fail to detect small amplitude candidate R-peaks,nd removal of low-frequency drift is most important. Therefore,oving average (MA) filter is used in this work. The low-frequency

rift is effectively removed by subtracting the output of the MA fil-er from the original input. The result is shown in Fig. 4(c) for the

Fig. 4. Effectiveness of the HT-based peak-finding logic for different envelopes: (a)normal ECG with small relative peak-amplitude (RPA) variation and RR = 0.55 s. (b)Abnormal ECG with large RPA variation and RR = 0.55 s. (c) Performance of the logicafter the removal of LF drift shown in (b).

drifted odd-symmetry function in Fig. 4(b). It clearly shows thatthe method had perfect detections. The choice of the filter lengthis important for computing the low-frequency drift that mainlydepends on the heart rate. For our sample rate of 360 samples/s,the MA filter length (L) is 900 samples (2.5 s). We noted that theperformance of the peak-finding logic is better when the differ-ence between two successive peak amplitudes is very small. Butamplitudes of candidate R-peaks depend on temporal and spectralparameters of the QRS complex. However, the proposed Shannonenergy-based preprocessor provides small deviations between thesuccessive candidate R-peaks under time-varying QRS morpholo-gies.

The various signal transformations at different stages of the pro-posed method using the ECG segment taken from the record 205 ofthe MIT-BIH arrhythmia database are shown in Fig. 5. The outputwaveforms of the bandpass filter and first-order forward derivativeoperation are shown in Fig. 5(b) and (c), respectively. The Shannonenergy envelogram of the normalized dECG and the output of theHilbert transformation of the envelogram are presented in Fig. 5(d)and (e), respectively. We can observe that the zero-crossing pointsare drifted in negative direction due to the large amplitude differ-ence between the two successive peaks. Here, the low-frequencywaveform drift is eliminated by subtracting the output of the mov-ing average (MA) filter from the input waveform shown in Fig. 5(e).The output of LF drift removal stage is shown in Fig. 5(f). Finally,the zero-crossings accompanied by negative to positive transitionare detected and used as guides to locate candidate R-peaks in theShannon energy envelogram. The output of positive zero-crossingpoint (ZCP) detector is shown in Fig. 5(g), which shows that the loca-tions of the impulses correspond to the peaks of candidate R-wave.The low-amplitude R-peaks of the record 205 can be successfullydetected only when the output of LF drift eliminator [Fig. 5(f)]is processed. Experiments clearly demonstrate that the Shannonenergy approach and Hilbert-transform based peak-finding logicplay a key role in the detection of R-peak. The proposed deci-sion stage does not require any information about the previouslydetected R peaks.

2.4. True R-peak detection

Experiments show that the locations of candidate R-peaks dif-fer slightly from the time instants of true R-peaks in an originalECG signal. Therefore, a simple true R-peak locator is incorporated

122 M.Sabarimalai Manikandan, K.P. Soman / Biomedical Signal Processing and Control 7 (2012) 118– 128

−0.5

0

0.5

x[n

]−0.1

0

0.1

f[n

]

2

4

6

SE

E, s

[n]

−4−2

02

HT

of

s[n

]

−2

0

2

z[n

]

0

1

ZC

Ps

1 2 3 4 5 6 7 8 9 10−0.5

0

0.5

ime (s

Fin

al,

R−p

eaks

−0.50

0.51

no

rm. d

[n]

(g)

(e)

(f)

(d)

(c)

(b)

(a)

(h)

F ) OutpO e ZCP

tniFt

3

McostnctwdibpCttfiTo2

rbmR

t

ig. 5. Proposed R-peak detector: (a) ECG signal, x[n], taken from a record 205. (butput of Hilbert transformer. (f) Output of LF drift eliminator. (g) Output of positiv

hat finds the correct time instants of the R-peaks in a ECG sig-al by searching the largest amplitude within ±25 samples of the

dentified location of the candidate R-peak in the previous step.ig. 5(h) shows the original signal and the detected R-peaks usinghe proposed four-stage methodology.

. Results and discussion

The proposed R-peak detection method is evaluated using theIT-BIH arrhythmia database. It contains 48 half-hour of two-

hannel ECG recordings sampled at 360 Hz with 11-bit resolutionver a 10 mV range. The ECG records from this database includeignals with acceptable quality, sharp and tall P and T waves, nega-ive QRS complex, small QRS complex, wider QRS complex, muscleoise, baseline drift, sudden changes in QRS amplitudes, suddenhanges in QRS morphology, multiform premature ventricular con-ractions, long pauses and irregular heart rhythms. In this study,e consider the entire ECG recordings since the proposed methodoes not require any learning phase. Episodes of ventricular flutter

n ECG record 207 are excluded from the performance analysis foretter comparison with the other detection methods [1,9]. The pro-osed algorithm was implemented on a 2.4-GHz Intel core 2 QuadPU using MATLAB version 7.1, and was tested on the ECG signalsaken from the first channel of the MIT-BIH arrhythmia database. Inhis study, the smoothing filter length (M) and the moving averagelter length (L) of the detector are set as (M, L) = (55, 900) samples.he average processing time required for performing our methodn each 30 min ECG data in the MIT-BIH database is approximately.24 s.

From the detection results, we calculated three quantitativeesults: true-positive (TP) when a R-peak is correctly detectedy the proposed algorithm, false-negative (FN) when a R-peak isissed, and false-positive (FP) when a noise spike is detected as

-peak. To evaluate the performance of the proposed detection

ec)

ut of bandpass filter. (c) Output of differentiator. (d) Output of SEE estimator. (e) finder. (h) Output of true R-peak locator.

algorithm, the sensitivity (Se), the positive predictivity (+P), and thedetection error rate (DER) can be computed by using the followingequations, respectively

Se = TPTP + FN

× 100%, (10)

+P = TPTP + FP

× 100%, (11)

DER = FP + FNTP

× 100%. (12)

The overall performance of the method is measured in terms of thedetection accuracy which is defined as

Accuracy(Acc) = TP/(TP + FP + FN) × 100%. (13)

The R-peak detection rates for the first-channel (each) of 48 ECGrecordings of the MIT-BIH arrhythmia database are summarizedin Table 1. The proposed method produces 79 false-negative (FN)beats and 140 false-positive (FP) beats for a total detection failureof 219 beats. But the individual detection accuracies of the ECGrecords vary from 98.99% to 100% depending on the characteristicsof normal and pathological ECG signals, and different noises.

In the MIT-BIH arrhythmia database, ECG records 104, 105,108, 200, 203, 210, and 228 contain high-grade noise and artifact.Records 108, 111, 112, 116, 201, 203, 205, 208, 210, 217, 219, 222,and 228 include severe baseline drifts and abrupt changes. Records201, 202, 203, 219 and 222 exhibits various irregular rhythmic pat-terns. Records 201, 219 and 232 include long pauses up to 6 s induration. Records 108 and 222 contain tall sharp P waves. Record113 has tall sharp T waves. For these ECG recordings, the number

of false positive detections is more in all the algorithms. Records200, 203 and 233 contain multiform ventricular arrhythmia, nega-tive QRS polarity and sudden changes in QRS morphology. Record208 has wider premature ventricular contractions (PVCs). Record223 exhibits sudden changes in QRS amplitudes. Records 116 and

M.Sabarimalai Manikandan, K.P. Soman / Biomedical Signal Processing and Control 7 (2012) 118– 128 123

Table 1Performance evaluation of the proposed R-wave detection method using first channel of the MIT-BIH arrythmia database with the detection parameters such as smoothingfilter length (M) and MA filter length (L)), (M, L) = (55, 900) samples.

ECG record Total (beats) FN (beats) FP (beats) DER (%) Se (%) +P (%) Accuracy (%)

100 2273 0 0 0 100 100 100101 1865 2 4 0.322 99.89 99.79 99.68102 2187 0 0 0 100 100 100103 2084 0 0 0 100 100 100104 2229 0 14 0.628 100 99.38 99.38105 2572 8 18 1.011 99.69 99.30 98.99106 2027 0 2 0.099 100 99.90 99.90107 2137 0 0 0 100 100 100108 1763 4 12 0.908 99.77 99.32 99.10109 2532 0 0 0 100 100 100111 2124 0 0 0 100 100 100112 2539 0 0 0 100 100 100113 1795 0 3 0.167 100 99.83 99.83114 1879 0 0 0 100 100 100115 1953 1 3 0.205 99.95 99.85 99.80116 2412 16 8 0.995 99.34 99.66 99.01117 1535 0 0 0 100 100 100118 2278 0 3 0.132 100 99.87 99.87119 1987 0 0 0 100 100 100121 1863 0 2 0.107 100 99.89 99.89122 2476 0 0 0 100 100 100123 1518 0 0 0 100 100 100124 1619 0 0 0 100 100 100200 2601 0 6 0.231 100 99.77 99.77201 1963 0 1 0.051 100 99.94 99.94202 2136 0 7 0.328 100 99.67 99.67203 2980 11 5 0.537 99.63 99.83 99.46205 2656 0 0 0 100 100 100207 1862 0 0 0 100 100 100208 2955 8 5 0.44 99.73 99.83 99.56209 3005 0 0 0 100 100 100210 2650 6 3 0.339 99.77 99.89 99.66212 2748 9 0 0.328 99.67 100 99.67213 3251 0 0 0 100 100 100214 2262 2 3 0.221 99.91 99.87 99.78215 3363 0 0 0 100 100 100217 2208 1 2 0.136 99.95 99.91 99.86219 2154 0 2 0.093 100 99.91 99.91220 2048 0 0 0 100 100 100221 2427 0 0 0 100 100 100222 2483 0 0 0 100 100 100223 2605 1 0 0.038 99.96 100 99.96228 2053 6 7 0.633 99.71 99.66 99.37230 2256 0 2 0.088 100 99.91 99.91231 1571 2 9 0.700 99.87 99.43 99.30232 1780 0 18 1.011 100 99.00 99.00233 3079 2 0 0.065 99.94 100 99.94234 2753 0 1 0.036 100 99.96 99.96

Overall 79 140 0.205 99.93 99.86 99.79

Table 2Comparison of the numbers of false-positives (FPs) and false-negatives (FNs) for specific records of the MIT-BIH arrythmia database.

Rec. no. Number of false-positive (FP) detections Number of false-negative (FN) detections

Ref. [8] Ref. [17] Ref. [14] Ref. [16] Our method Ref. [8] Ref. [17] Ref. [14] Ref. [16] Our method

104 3 20 0 7 14 7 1 0 1 0105 53 35 15 7 18 22 14 21 19 8106 1 5 0 21 2 2 0 2 20 0108 50 68 2 10 12 47 9 62 2 4113 2 6 1 10 3 1 0 682 11 0116 4 – 20 4 8 25 – 0 27 16121 1 15 1 13 2 0 0 0 0 0200 3 47 1 4 6 2 3 2 9 0203 14 23 79 3 5 61 95 19 7 11208 9 2 13 3 5 19 19 3 10 8209 2 0 0 2 0 2 0 1 9 0210 2 8 5 16 3 41 23 2 5 6221 1 4 0 4 0 1 2 4 8 0222 40 5 27 1 0 37 0 12 0 0223 0 28 0 4 0 2 1 0 22 1228 19 38 1 10 7 6 22 14 2 6232 3 26 0 14 18 0 0 17 2 0233 0 1 1 7 0 3 7 0 8 2Total 207 331 166 140 103 278 196 841 162 62

124 M.Sabarimalai Manikandan, K.P. Soman / Biomedical Signal Processing and Control 7 (2012) 118– 128

0 5 10 15 20 25

−0.5

0

0.5

1

EC

G, x

[n]

0 5 10 15 20 25

0.2

0.4

0.6

0.8

smoo

th S

EE

, s[n

]

0 5 10 15 20 25

−0.5

0

0.5

1

outp

ut o

f H

T, z

[n]

0 5 10 15 20 25

−0.5

0

0.5

1

Det

ecte

d P

eaks

time (sec)

(a)

(b)

(c)

)

F line da

2Edicef

F

(d

ig. 6. Performance of the proposed R-peak detector for the ECG signal with baseccuracy of 100%.

08 contain smaller QRS complexes than the others. For theseCG recordings, most of the algorithms had higher false-negative

etections. However, the proposed algorithm achieves a significant

mprovement in the detection of R-peak under time-varying QRSomplex morphology and different kinds of noise and artifacts. Theffectiveness of the proposed method in terms of the number ofalse negatives and false positives is shown in Table 2. The wave-

0 5 10

−0.5

0

0.5

1

EC

G, x

[n]

0 5 10

0.2

0.4

0.6

0.8

smoo

th S

EE

, s[n

]

0 5 10

−0.5

0

0.5

1

Det

ecte

d P

eaks

time

0 5 10−1

−0.5

0

0.5

outp

ut o

f H

T, z

[n]

(d)

(a)

(b)

(c)

ig. 7. Detection performance for the ECG with continuously varying QRS complex morp

rift and low-amplitude QRS complexes (Record 103). The method has a detection

forms of the different stages of the proposed method using the ECGsegments taken from first-channel of the different recordings of the

MIT/BIH database are shown Figs. 6–12. In each of these figures,waveform depicted in (a) is the original ECG signal, x[n]. The wave-form depicted in (b) is the smooth Shannon energy envelogram(SEE) of the signal d̃[n]. The waveform (c) is the final output of theHilbert transformation and low-frequency drift removal process.

15 20 25

15 20 25

15 20 25(sec)

15 20 25

hology, sudden changes in beat-to-beat RR-interval, and tall T waves (Record 106).

M.Sabarimalai Manikandan, K.P. Soman / Biomedical Signal Processing and Control 7 (2012) 118– 128 125

0 5 10 15 20 25

0

0.5

1

EC

G, x

[n]

0 5 10 15 20 25

0.2

0.4

0.6

0.8

smoo

th S

EE

, s[n

]

0 5 10 15 20 25

0

0.5

1

Det

ecte

d P

eaks

time (sec)

0 5 10 15 20 25−1

−0.5

0

0.5

outp

ut o

f H

T, z

[n] (b)

)

(a)

(c)

F r contd

TRftdr

(d

ig. 8. Detection performance for the ECG signal with wide premature ventriculaetects low-amplitude QRS complexes and wide PVCs.

he waveform depicted in (d) shows the time instants of detected-peaks using the proposed method. The detection performance

or the ECG signal with the high sharp P-waves, negative polari-ies, and more noises is summarized in Table 2. Most of the R-peaketection methods have poor detection performance for the noisyecords 104, 105, and 108. The proposed method has a total detec-

0 5 10

−0.5

0

0.5

1

EC

G, x

[n]

0 5 10

0.2

0.4

0.6

0.8

smoo

th S

EE

, s[n

]

0 5 10 −1

−0.5

0

0.5

outp

ut o

f H

T, z

[n]

0 5 10

−0.5

0

0.5

1

Det

ecte

d P

eaks

time (d)

(c)

(b)

(a)

Fig. 9. Detection performance for the ECG with different QRS mor

ractions (PVCs), and small QRS complexes (Record 208). The method successfully

tion failure of 56 beats (44 FP beats and 12 FN beats). The digitalfilter-based method [8] and the EMD-based method [17] produced

a total detection failure of 182 beats (106 FP beats and 76 FN beats),and of 147 beats (123 FP and 24 FN beats), respectively. For the noisyrecords, the detection performance of the proposed method is bet-ter than the existing methods, except the 3M filtering method [16]

15 20 25

15 20 25

15 20 25

15 20 25 (sec)

phologies, small QRS complexes and artifacts (Record 223).

126 M.Sabarimalai Manikandan, K.P. Soman / Biomedical Signal Processing and Control 7 (2012) 118– 128

0 5 10 15 20 25

−0.5

0

0.5

1

EC

G, x

[n]

0 5 10 15 20 25

0.2

0.4

0.6

0.8

smoo

th S

EE

, s[n

]

0 5 10 15 20 25−1

−0.5

0

0.5

outp

ut o

f H

T, z

[n]

0 5 10 15 20 25

−0.5

0

0.5

1

Det

ecte

d P

eaks

time

(a)

(b)

(c)

(d)

with p

wbfmt

Fs

Fig. 10. Detection performance for the ECG signal

ith a total detection failure of 46 beats (24 FP beats and 22 FNeats). The numerous long pauses up to 6 s in duration are mainlyound in record 232 that yields more false positive detections in

ost of the methods. The detection performance of the method forhe RR-interval greater than 2 s is shown in Fig. 11. The method

0 5 10

−0.5

0

0.5

1

EC

G, x

[n]

0 5 10

0.2

0.4

0.6

0.8

smoo

th S

EE

, s[n

]

0 5 10−1

−0.5

0

0.5

outp

ut o

f H

T, z

[n]

0 5 10

−0.5

0

0.5

1

Det

ecte

d P

eaks

time

(a

(

(

(

FP

ig. 11. Detection performance for the ECG with noise and numerous long pauses up to 6egment with long RR-intervals greater than 2 s.

(sec)

aced beats and severe muscle noise (Record 104).

has a failed detection of 18 FP beats. Fig. 8 shows the detectionperformance for the record 208 which has wide PVCs, and smallQRS complexes. For this record, most of the filtering based methodshad highest FN. To detect the missed ECG beats, many methods usea set of decision rules with adaptive thresholds estimated based

15 20 25

15 20 25

15 20 25

15 20 25 (sec)

)

b)

c)

d)

FP FP

s in duration (Record 232). The method has the highest false positives for the ECG

M.Sabarimalai Manikandan, K.P. Soman / Biomedical Signal Processing and Control 7 (2012) 118– 128 127

0 5 10 15 20 25−1

0

1

EC

G, x

[n]

0 5 10 15 20 25

0.2

0.4

0.6

0.8sm

ooth

SE

E, s

[n]

0 5 10 15 20 25

−0.5

0

0.5

1

outp

ut o

f H

T, z

[n]

0 5 10 15 20 25−1

−0.5

0

0.5

1

Det

ecte

d P

eaks

time (sec)

(a)

(b)

drift

(c)

very small QRS complexes

FPs

the number of FNs is high

F w-amd an the

oppdtdwppp

TR

ig. 12. Detection performance for the ECG with severe baseline drift and very loetections similar to the other methods. The overall detection accuracy is better th

n the information (amplitude, RR-interval and duration) of therevious detected R-peaks. The method with secondary thresholdroduces several spurious peaks. To reduce the number of falseetections, the refractory period of 200 ms and T-wave discrimina-or (200–360 ms) are used in most of the methods. These additionalecision rules may yield the poorest results in the case of ECG signal

ith irregular heart rhythm and continuously varying QRS com-lex morphology. Furthermore, these methods require the learninghase. The performance comparison in Table 2 shows that the pro-osed method produces 103 FP beats and 62 FN beats, which are

able 3-peak detection performance comparison with other methods.

Ref. Methodology

Preprocessing stage Primary threshold, �pt Secondarthreshold

[8] Bandpass filter, derivativesquarer and movingwindow integrator

Threshold coefficient ×predicted peak value

0.3 × �pt

[9] Bandpass filter, derivativeand squarer

Based on root mean square(RMS) of segment

0.5 × �pt

[1] Quadratic spline wavelet Four thresholds based onRMS value of the WT at thecorresponding scales

Searchbathreshold

[14] Coiflet wavelet, squaringand moving averaging

Threshold based on theSNR computed betweenwavelet coefficients at 1stand 5th level

None

[16] Multiscale morphologyfiltering, differentiationand multi-frameaccumulation

Threshold based on themaximum of thetransformed ECGwaveform

None

[17] Empirical modedecomposition anddenoising

Four thresholds based onmodulus maxima andaverage of past values

0.5 × �pt

Our method Bandpass filter, derivativeShannon energy andSmoothing

Hilbert transform-basedpeak-finding logic (noamplitude thresholds)

None

(d)

plitude QRS complexes (Record 116). The method has higher false-negative (FN) methods in [8,17,14,16].

lower than that of the other methods reported in [8,17,14], and[16]. Various experiments demonstrate that the proposed methodwith simple HT-based peak-finding logic provides better detectionperformance without using any amplitude threshold in the deci-sion stage to determine location of the R-peaks in ECG signal. Inthis work, we can observe that the proper decision rule provides

much higher detection performance.

Finally, the overall performance of proposed method is com-pared with six R-peak detectors reported in the literature. Basedon the results of Table 3, an average accuracy of 99.80%, a sensitiv-

FP (beats) FN (beats) Se (%) +P (%) Acc (%)

y Blanking

Refractory(200 ms) andT-wave removal

248 340 99.69 99.77 99.46

200–360 ms forT-wave removal

447 467 99.57 99.59 99.17

cks

Refractory (200 ms) 153 220 99.80 99.86 99.66

None 214 1333 99.78 99.80 98.59

None 204 213 99.81 99.80 99.62

refractory (300 ms) 467 244 99.77 99.56 99.33

None 140 79 99.93 99.88 99.80

1 dical S

iaabBdrQcimtaop

4

hapSatadttQtpadewbramsaettT3

[

[

[

[

[

[

[

[

[

28 M.Sabarimalai Manikandan, K.P. Soman / Biome

ty of 99.93%, and a positive predictivity of 99.86% are obtainedgainst the first-channel of the ECG recordings of the MIT-BIHrrhythmia database. The proposed method achieved significantlyetter performance than the other detection methods [7–17].ased on the detection results, we concluded that the properesign of preprocessing and decision stages can improve the accu-acy of detection of R peaks in ECG recordings with negativeRS polarities, low-amplitude R peaks, and wide PVCs, suddenhanges in QRS amplitudes, sharp P and T waves, sudden changesn QRS morphologies, and irregular heart rhythm. The proposed

ethod does not require additional decision rules with sets ofhresholds based on the running estimates of the signal peaksnd noise peaks, the average RR interval and rate limits, a setf tactics for blanking and T-wave discrimination, and traininghase.

. Conclusion

In this paper, a novel, effective, and four-stage methodologyas been developed for the automated detection of R-peaks inn ECG signal. The proposed preprocessor is based on a band-ass filter, first-order forward derivative, amplitude normalization,hannon energy estimation and zero-phase filtering with rect-ngular impulse response that provides a smooth envelogram ofhe ECG signal. The decision stage is based on Hilbert transformnd moving average filter, which is a new and simple strategy toetermine the location of the R-wave peak. Experiments showedhat the SE-based preprocessor significantly increases the detec-ion accuracy for ECG recordings with low-amplitude, and wideRS complexes. Also, the Hilbert transform-based peak-finding

echnique simplifies the determination of locations of the R-eaks. This approach does not require any amplitude thresholdnd prior knowledge of the past detected R-peaks. The stan-ard MIT-BIH arrhythmia database (48 ECG records of 30 minach) is used to test the effectiveness of the proposed methodhose detection performance was measured in terms of the num-

er of false positives, true positives and false negatives for eachecord. The detection results obtained are presented, discussednd compared with the existing R-peak detection methods. Theethod achieves average detection accuracy of 99.80%, a sen-

itivity of 99.93%, and a positive predictivity of 99.86%. Despitell these various morphologies of the QRS complexes and differ-

nt kinds of noise and artifacts contained in the ECG signals ofhe database, the proposed method achieves much higher detec-ion rates than those produced by the other existing methods.he average time required for processing each sample is about.45 �s.

[

[

ignal Processing and Control 7 (2012) 118– 128

Acknowledgments

The authors would like to thank Professor Robert Allen, Editor-in-Chief, and the anonymous referees for their valuable suggestionsand comments.

References

[1] J.P. Martínez, R. Almeida, S. Olmos, A.P. Rocha, P. Laguna, A wavelet-based ECGdelineator: evaluation on standard databases, IEEE Trans. Biomed. Eng. 51 (4)(2004) 570–581.

[2] Y. Gahi, M. Lamrani, A. Zoglat, M. Guennoun, B. Kapralos, K. El-Khatib, Biomet-ric Identification System Based on Electrocardiogram Data, New Technologies,Mobility and Security, Tangier, 2008, pp. 1–5.

[3] M.S. Manikandan, S. Dandapat, Wavelet threshold based TDL and TDR algo-rithms for real-time ECG signal compression, Biomed. Signal Process. Control03 (2008) 44–66.

[4] B.U. Kohler, C. Hennig, R. Orglmeister, The principles of software QRS detection,IEEE Eng. Med. Biol. Mag. 21 (2002) 42–57.

[5] O. Pahlm, L. Sörnmo, Software QRS detection in ambulatory monitoring—areview, Med. Biol. Eng. Comput. 22 (1984) 289–297.

[6] Y.-C. Yeha, W.-J. Wanga, QRS complexes detection for ECG signal: the differenceoperation method, Comput. Methods Prog. Biomed. 91 (2008) 245–254.

[7] J. Pan, W.J. Tompkins, A real time QRS detection algorithm, IEEE Trans. Biomed.Eng. 32 (3) (1985) 230–236.

[8] P.S. Hamilton, W.J. Tompkins, Quantitative investigation of QRS detection rulesusing the MIT/BIH arrhythmia database, IEEE Trans. Biomed. Eng. 33 (1986)1157–1165.

[9] N.M. Arzeno, Z.-D. Deng, C.-S. Poon, Analysis of first-derivative based QRS detec-tion algorithms, IEEE Trans. Biomed. Eng. 55 (2) (2008) 478–484.

10] S. Benitez, P.A. Gaydecki, A. Zaidi, A.P. Fitzpatrick, The use of the Hilbert trans-form in ECG signal analysis, Comput. Biol. Med. 31 (2001) 399–406.

11] C. Meyer, J.F. Gavela, M. Harris, Combining algorithms in automatic detectionof QRS complexes in ECG signals, IEEE Trans. Inf. Technol. Biomed. 10 (3) (2006)468–475.

12] L. Romero, P.S. Addison, N. Grubb, R-wave detection using continuous waveletmodulus maxima, IEEE Proc. Comp. Cardiol. 30 (2003) 565–568.

13] F. Abdelliche, A. Charef, R-peak detection using a complex fractional wavelet,in: International Conference on Electrical and Electronics Engineering, Bursa,2009, pp. 267–270.

14] M. Elgendi, M. Jonkman, F. De Boer, R wave detection using Coiflets wavelets,in: IEEE 35th Annual Northeast Bioengineering Conference, Boston, MA, 2009,pp. 1–2.

15] Y.L. Chen, H.L. Duan, A QRS complex detection algorithm based on mathemati-cal morphology and envelope, in: Proceedings of the 27th Annual InternationalConference of the IEEE EMBS, Shanghai, China, 2005, pp. 4654–4657.

16] F. Zhang, Y. Lian, QRS detection based on multi-scale mathematical morphologyfor wearable ECG devices in body area networks, IEEE Trans. Biomed. CircuitsSyst. 3 (4) (2009) 220–228.

17] X. Hongyan, H. Minsong, A new QRS detection algorithm based on empiricalmode decomposition, in: Proceedings of the 2nd International Conference onBioinformatics and Biomedical Engineering, 2008, pp. 693–696.

18] K.V. Suárez, J.C. Silva, Y. Berthoumieu, P. Gomis, M. Najim, ECG beat detection

using a geometrical matching approach, IEEE Trans. Biomed. Eng. 54 (4) (2007)485–489.

19] B. Abibullaev, H.D. Seo, A New QRS detection method using wavelets and arti-ficial neural networks, J. Med. Syst. (2010), doi:10.1007/s10916-009-9405-3.

20] H. Liang, S. Lukkarinen, I. Hartimo, Heart sound segmentation algorithm basedon heart sound envelogram, IEEE Proc. Comp. Cardiol. 24 (1997) 105–108.