Cardiovascular Engineering and Technology, Vol. 2, No. 4, December 2011 ( 2011) pp.

408–425
DOI: 10.1007/s13239-011-0065-3

An Efficient R-peak Detection Based on New Nonlinear Transformation
and First-Order Gaussian Differentiator
P. KATHIRVEL,2 M. SABARIMALAI MANIKANDAN,1 S. R. M. PRASANNA,3 and K. P. SOMAN2
1
Department of Electronics and Communication Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham,
Coimbatore 641105, India; 2Center for Excellence in Computational Engineering and Networking, Amrita Vishwa
Vidyapeetham, Coimbatore, Tamilnadu, India; and 3Department of Electrical and Electronics Engineering, Indian Institute
of Technology Guwahati, Guwahati, Assam, India
(Received 18 May 2011; accepted 27 September 2011; published online 12 October 2011)

Associate Editor Ajit P. Yoganathan oversaw the review of this article.

Abstract—In this paper, we present a reliable and efficient tude, wide), very big change in amplitudes of adjacent
automatic R-wave detection based on new nonlinear trans- R-peaks, irregular heart rates, and noisy ECG signals.
formation and simple peak-finding strategy. The detection
algorithm consists of four stages. In the first stage, the Keywords—ECG signal analysis, QRS complex detection,
bandpass filtering and differentiation operations are used to
ECG beat recognition, Heart rate variability analysis, Shan-
enhance QRS complexes and to reduce out-of-band noise. In
the second stage, we introduce a new nonlinear transforma- non energy, Gaussian differentiator, Peak finding logic.
tion based on energy thresholding, Shannon energy compu-
tation, and smoothing processes to obtain a positive-valued
feature signal which includes large candidate peaks corre-
sponding to the QRS complex regions. The energy thres- INTRODUCTION
holding reduces the effect of spurious noise spikes from
muscle artifacts. The Shannon energy transformation ampli- Automatic R-peak detection algorithm is essential
fies medium amplitudes and results in small deviations for many types of ECG applications including heart
between successive peaks. Therefore, the proposed nonlinear
transformation is capable of reducing the number of false- rate monitor (HRM), ECG wave delineator, fetal heart
positives and false-negatives under small-QRS and wide- rate monitor (fHRM), cardiac sound separator, patient
QRS complexes and noisy ECG signals. In the third stage, we authentication and ECG coder. Various experiments
propose a simple peak-finding strategy based on the first- showed that overall performance of all these ECG
order Gaussian differentiator (FOGD) that accurately iden- applications depends highly on accurate detection of
tifies locations of candidate R-peaks in a feature signal. This
stage computes convolution of the smooth feature signal time instants of the R-wave peaks in the ECG sig-
and FOGD operator. The resultant convolution output has nal.6,10,19,23,29 The typical ECG beat consists of
negative zero-crossings (ZCs) around the candidate peaks of P-wave, ‘‘QRS complex‘‘, T- and U-wave. The R-wave
feature signal due to the anti-symmetric nature of the FOGD is the most prominent wave in each cardiac cycle of an
operator. Thus, these negative ZCS are detected and used as ECG signal. Therefore, development of robust auto-
guides to find locations of real R-peaks in an original signal
at the fourth stage. Unlike other existing algorithms, the matic R-wave peak detector is essential but it is still a
proposed algorithm does not use search back algorithm and challenging task due to irregular heart rates, various
learning phase. The proposed algorithm is validated using the amplitude levels and shapes of QRS morphologies, and
standard MIT-BIH arrhythmia database and achieves an various kinds of noise and artifacts.
average sensitivity of 99.94% and a positive predictivity of In the literature, many QRS detection methods have
99.96%. Experimental results show that the proposed algo-
rithm outperforms other existing algorithms in case of been developed based on the first and second deriva-
different QRS complex morphologies (negative, low-ampli- tives, digital filters (DFs), linear prediction (LP), two-
pole recursive filter, maximum a posteriori (MAP)
estimation, wavelet transform (WT), filter-banks, Hil-
bert transform (HT), higher-order statistics (HOS),
Address correspondence to M. Sabarimalai Manikandan,
Department of Electronics and Communication Engineering, Amrita
template matching (TM), geometrical matching (GM),
School of Engineering, Amrita Vishwa Vidyapeetham, Coimba- matched filters, mathematical morphology, multiscale
tore 641105, India. Electronic mail: msm.sabari@gmail.com mathematical morphology (3M) and empirical mode
408
1869-408X/11/1200-0408/0  2011 Biomedical Engineering Society

Efficient R-peak Detection 409 decomposition (EMD). EMD. the differentiated signal d[n]. logic and searchback algorithm to locate R-peaks. cases. FN and FP detections for ECG signals with small- wavelets.25 0.34–38 Generally. and genetic is studied using the ECG record 104 taken from the algorithms. . and (d) energy envelope ee[n]. syntactical rules. Example illustrating the output of stages of the conventional preprocessing algorithm. (c) squared (or energy) signal e[n]. and x[n].25 e[n]. 1.1 0. The decision stage generally includes peak-detection Generally.05 0 d[n] −0. hidden Markov models (HMM). which shows the original signal cause that has lower computational complexity. the squared signal requires less memory space. tant role in detection performance.12–16. Therefore. decision stage based on the preprocessing results for tuate the QRS feature and to suppress the noises and achieving better detection rates. zero-crossing (ZC). Although different many researchers have attempted to construct suitable signal processing techniques were employed to accen. mathematical morphological opera. effectiveness of the traditional squaring transformation tor machine (SVM). the works. support vec.1–9. nonlinear transformation plays an impor. many methods had a relatively high number of niques such as bandpass filters.05 0 1 2 3 4 5 6 7 8 9 Time (sec) FIGURE 1.05 −0.22–28. and the feature signal ee[n].25 The preprocessing stage is generally complexes under the high-amplitude ones. QRS and wide-QRS complexes and noise. Therefore. (a) Original ECG signal x[n].2 0. derivatives filter banks. In such constructed using one or more signal processing tech. tors and squaring transformation. (b) difference of the filtered ECG (dECG) signal d[n]. artifacts but most of them have some drawbacks. The output results are digital filtering based QRS detector is widely used be.1 0 1 2 3 4 5 6 7 8 9 −3 (c) x 10 10 8 6 e[n] 4 2 0 1 2 3 4 5 6 7 8 9 (d) 0. neural net. The (a) 1 x[n] 0 −1 0 1 2 3 4 5 6 7 8 9 (b) 0.15 ee[n] noise peaks 0. the MIT-BIH arrhythmia database.18–20.20. illustrated in Fig. It is noticed that the The QRS detectors based on filtering methods feature signal has more noise peaks and also the consists of two stages: the preprocessing stage and the squarer buries the peaks of small-QRS and wide-QRS decision stage.30–32.

Several different docode for implementation of the different stages. ‘‘Evaluation Methods’’ section describes the character- monly consists of sets of heuristic decision rules and a istics of validation ECG database. the power line interference. and the algorithm assessment procedure. These medical rules may im. choices of bandpass filters with different passband Stage 1: QRS enhancement and noise reduction Stage 2: New nonlinear transformation x[n] bandpass f [n ] d[n] e th [n] Shannon energy squaring and filtering differentiation computation and thresholding ECG [5 20] Hz smoothing Stage 4:True R-peak Locator Stage 3: New peak-finding logic s[n] r[m] finding locations of true p[m] detecting zero-crossings z[n] convolution of R-peak in ECG using accompanied by positive to envelope s[n] with a locations of candidate negative transition Locations of FOGD function wd [n] R-peak detected.14. decision stage of the filtering-based methods com. and various kinds of noise and artifacts. and Discussion’’ section reports the results of the pro- dependent thresholds for detecting R peaks in the posed algorithm for the well-known MIT-BIH arrhyth- feature signal. dramatic effect on the QRS complex. including pseu. muscle contractions. we present an effective and efficient Bandpass Filtering and Differentiation automatic detection algorithm that identifies the time- location of the R-peaks in ECG signals. and learning phase.30.2 s (refractory period) and (ii) search back automated algorithm for detection of R-peaks of a if tn  tm > 1. most of the algorithms use similar decision rules and threshold determination procedure. other sources. The frequency The rest of this paper is organized as follows. (vii) sharp P/T waves. ECG signals are often algorithm is simple that does not require additional obscured by various kinds of noise and artifacts from decision rules with sets of thresholds based on the run. (vi) sudden changes in QRS morphologies. Common ECG noise sources such as ning estimates of the signal peaks and noise peaks. In literature. we describe a simple and robust tn  tm < 0.5RRavg. mia database and provides comparisons between the dated periodically based upon amplitudes. corrupt ECG signals significantly. conclusions are drawn in last section. and RR-intervals of detected R-peaks in the previous Finally. (v) sudden changes in QRS amplitudes. searchback algorithm with PROPOSED ALGORITHM DESCRIPTION AND adaptive amplitude-dependent and time-dependent THEORY thresholds were widely adopted to reject or include identified R-peaks located at tm and tn: (i) if In this section. and RR-interval. intervals. baseline average RR interval and rate limits. Moreover. A preprocessing stage emphasizes the QRS complex components and a new peak-finding stage identifies the location of candidate R-peak.20. In this paper. durations proposed algorithm and other existing algorithms. (iii) – New Peak-Finding Technique negative QRS polarities.26.35 These thresholds are up. a set of tactics to drift due to respiration. (iv) sudden changes in RR – Finding Location of True R-Peaks. ‘‘Proposed content of a QRS complex is essentially in the fre- Algorithm Description and Theory’’ section describes a quency range 5–30 Hz. benchmark parame- set of tactics with fixed or/and adaptive amplitude. and abrupt baseline shift can reject or include identified R-peaks. (ii) low-amplitude QRS complexes. . The proposed In the realistic environments. ECG segment. long-term ECG signal. duration-dependent. ‘‘Results dependent.410 KATHIRVEL et al. The above detection of the following steps: issues clearly indicate that it is hard to find a proper set – Bandpass Filtering and Differentiation of decision rules and thresholds in case of (i) wide QRS – New Nonlinear Transformation complexes. ters. p[m] R-peaks FIGURE 2.4. Figure 2 shows a block dia- prove detections for regular rhythms but some rules gram of our R-peak detection algorithm that consists may be in conflict with others.33 Heart pathologies may have a four-stage detection algorithm in detail. Block diagram showing the architecture and individual stages of the new detection algorithm for R-peak detection in ECG signals.

After filtering. s[n] 0. and (f) the Shannon energy envelope s[n].2 signal f[n] is implemented as s reveals that most of the frequencies present in the d½n ¼ f½n þ 1  f½n: ð1Þ QRS complex lies below 20 Hz. Efficient R-peak Detection 411 frequencies for enhancement of QRS complex and We construct the 15th-order FIR bandpass DF using suppression of P/T waves and noises have been the least squares approach.002 (d) 0. Figure 3b shows that the output of the (a) 1 x[n] 0 −1 (b) 0.01 e [n] 0. 3b.05 d[n] 0 −0.01 noise components 0. we motion artifacts and P/T waves are concentrated up to apply first-order forward differentiation to emphasize 5 Hz. (b) the difference of the filtered ECG (dECG) signal d[n].05 and 0.05 −0. Spectral analysis of the various complex. power. (e) the normalized thresholded energy signal e~th ½n.1 (c) 0. The original ECG and the difference of the complexes) and reduce the effect of P/T waves.002 0 1 2 3 4 5 6 7 8 9 (e) 1 normalized e [n] 0.25 Most of energies of noises generated by of the passband are 6 and 20 Hz. respectively. sufficient P/T wave and noise suppression up large slope and high-frequency content of the QRS to 5 Hz is essential. filtered ECG (dECG) signals are shown in Figs.006 e[n] 0. where cut-off frequencies described. motion artifacts and muscle noise. In this work.006 th 0. (d) the thresholded energy signal eth[n]. (a) The original ECG signal x[n]. we chose the passband such that they maximize the energy of The derivative operation reduces the effect of large P/ different QRS complexes (narrow-QRS and wide-QRS T waves. the differentiation of the filtered ECG QRS morphologies with duration between 0. So. . 3a and line interference. Example illustrating the output of stages of the proposed nonlinear transformation technique. (c) the squared (or energy) signal e[n]. Here.5 th (f) envelope.1 0 0 1 2 3 4 5 6 7 8 9 Time (sec) FIGURE 3.

2. thresholded energy signal is first normalized as eral spurious noise peaks and thus had a large number of eth ½n false positives. thresholding process and Shannon energy transfor. otherwise. and (ii) it buries the peaks of small-QRS e~th ½n ¼ . ð6Þ and wide-QRS complexes under the high-amplitude maxN n¼1 ðjeth ½njÞ ones. ð4Þ locations of candidate R-peaks in the SE envelope. Therefore. By comparing Figs. This Squaring and Adaptive Thresholding smoothing process is designed to generate peaks cor- The dECG signal d[n] is first squared to obtain a responding to the QRS-complex portions. The feature signal that is illustrated in Fig. . m ¼ 1. . window is symmetric about bM 2 c þ 1 and its first-order . . Thus. In litera. The effectiveness of this nonlinear transforma- problems in case of negative QRS complexes. 1 ðmM2 Þ w½m ¼ e2 r2 . mental results demonstrated that it improves detection tain positive peaks regardless of polarity of QRS accuracy under ECG signal with smaller. In such cases. 3f. eth ½n ¼ ð3Þ e½n. e½n<g of true R peaks in a ECG signal. 3c and 3d.412 KATHIRVEL et al.30 The experi- ture. .. these  candidate peaks are processed further to find locations 0. a new nonlinear transformation based on squar- ing. 3. 2. in this e~th ½n is computed as work.. In this thresholding process. the traditional squaring transformation: (i) it produces sev. 3. 4b. M ð8Þ ment. the squaring transformation is widely used to ob. . We can squaring is implemented as clearly notice that the locations of candidate R-peaks in the SE envelope s[n] correspond to approximate e½n ¼ d2 ½n: ð2Þ locations of the R peaks in the ECG waveform shown The thresholding function is defined as in Fig. differentiator is a bipolar signal. which gives the slope at each sample. where vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi The First-Order Gaussian Differentiator u N u1 X 1X N re ¼ t ðe½n  lÞ2 and l¼ e½n: ð5Þ In this work. This section describes a new and simple peak-find- puted as ing strategy using the first-order Gaussian differentia- tor (FOGD) operator that automatically finds the g ¼ 0:5  re . M  1 the number of false positive detections under noisy ð9Þ ECG signal with long pauses. 4a and the corresponding FOGD function is Since squarer buries the peaks of small-QRS and shown in Fig. the energy values e[n] New Peak-Finding Logic Using the First-Order smaller than a threshold parameter g are set to zero Gaussian Differentiator and other values are retained. . 3a. in this proposed algorithm. The pro- positive-valued signal and then adaptive thresholding posed nonlinear transformation generates a smooth is performed on the energy (or squarer) values e[n]. We can observe that the Gaussian wide-QRS complexes under the high-amplitude ones. and the FOGD is computed as we can observe that the thresholding process effectively eliminates spurious noise spikes and tends to reduce wd ½m ¼ w½m þ 1  w½m. the adaptive- threshold parameter g for each ECG segment is com. The Shannon energy values are smoothed using the finite-impulse response (FIR) filter with a rectangular New Nonlinear Transformation impulse response h[k] of length L to reduce the effect of multiple peaks around QRS complex regions.and wider- complexes. Here. To compute Shannon energy. m ¼ 1. e2th ½n log e~2th ½n: s½n ¼ ~ ð7Þ mation is designed to combat such problems effectively. The outputs of the squaring and thresholding stage is shown in Fig. Therefore. tion was studied in our previous work. a rectification we thus use Shannon energy transformation which process (or nonlinear transformation) is employed to results in small deviations between the successive R obtain a positive-valued signal that eliminates detection peaks. the M-point Gaussian window w[m] is N n¼1 N n¼1 defined as 2 where N is the number of samples in a test ECG seg. 3. The 901-point Gaussian window with spread r = 36 is shown in Shannon Energy Computation and Smoothing Fig. most of the methods failed to detect and then the Shannon energy of the normalized signal smaller and wider-QRS complexes. Figure 1 illustrates that disadvantages of QRS complexes.

a simple algorithm finds the loca- tion of maximum peak value in the segment. long pauses and irregular heart negative ZCs are detected and used as guides to find rhythms. baseline drift. It contains 48 and negative ZCs because of the anti-symmetric nat. it can be observed that 0. This process is repeated for all the detected locations p[m]. 5a. The pseudocode for this function is shown in Table 1 and its output is illustrated in Fig. Then. a seg- sample number. 5f. Therefore. 5c. w d[m] 0 −0.27 Positive ZC is defined as 360 Hz with 11-bit resolution over a 10 mV range. this stage uses locations of all candidate 100 200 300 400 500 600 700 800 900 R-peaks detected at the previous stage. The a ZC associated with a positive slope that can be ECG records from this database include signals with obtained if the signal moves from a negative to a acceptable quality. 5e illustrates that the detected time instants are slightly shifted from the (b) instants of real R peaks in the ECG signal. Efficient R-peak Detection 413 (a) peak point Detecting Negative Zero-crossings 1 In this work. (a) The 901-point Gaussian window with spread didate R-peak ðp½m  K2 Þ is extracted from the original r 5 36 and (b) the FOGD wd[m]. it is a negative ZC. half-hour of two-channel ECG recordings sampled at ure of the FOGD function. First. The shows that the Gaussian window function has a peak output of this stage is shown in Fig. In this study. Episodes of ventricular flut- . wider to a negative value. multiform premature ventricular con- energy envelope s[n] shown in Fig. respectively and also its slope is of QRS complexes and noise.015 To find time instants of real R-peaks in an original ECG signal.015 zero−crossing point in the next stage. sharp and tall P and T waves. If the signal z[n] moves from a positive negative QRS complex.01 candidate R-peaks are used as guides to locate real 0. we EVALUATION METHODS perform the convolution operation on the SE envelope s[n] and the FOGD sequence wd[m] that is computed as Characteristics of Validation Database X 1 Many of the existing algorithms were not evaluated z½n ¼ wd ½k s½n  k: ð10Þ against a standard database and also not measured the k¼1 detection performance using the benchmark parame- The convolution output of the FOGD function wd[m] ters.6 tn+1. the detected time instants of 0.8 the sign of the ZC function z[n] at time instants tn and w[m] 0. quire any learning phase.005 R-peaks. sudden negative slope. This is the basis for the proposed peak-finding logic. ECG signal. performance of the proposed algorithm using the well- ferred to as ZC function z[n] that has both positive known MIT-BIH arrhythmia database. Therefore. 5d. the convolution output is re. 0. m ment centered around each location of detected can- FIGURE 4. In this work. 5d and 5b. derivative is an antisymmetric function. Therefore. positive value. The results at m ¼ bM 2 c þ 1: Figure 4b shows that the slope of show that the locations of real R-peaks are accurately FOGD is positive and negative for 1  m  bM 2 c and detected regardless of varying amplitudes and shapes bM2 c þ 1  m  M. The output shown in Fig. which has a QRS complex.01 −0. tractions (PVCs). we consider the entire ECG locations of R-peaks in the original ECG signal recordings since the proposed algorithm does not re- shown in Fig.005 Finding Location of Real R-Peaks −0. sudden changes in QRS ZCs indicate locations of the peaks in the Shannon morphology. 5b.20 This makes the experimental results difficult to with a Shannon energy envelope s[n] is shown in compare and to assess. we evaluate the Fig.4 comparing Figs. For detecting candidate R-peaks in SE envelope. negative ZCs are detected by checking 0. small QRS complex. where ºÆß denotes the floor function. Figure 5c demonstrate that negative changes in QRS amplitudes. muscle noise. Figure 4a Table 2 list the pseudocode for this function. By 0. zero at m ¼ M 2 .2 the proposed peak-finding technique accurately iden- 0 tifies time locations of candidate R-peaks in the SE 100 200 300 400 500 600 700 800 900 envelope.

(a) 1 x[n] 0 −1 (b) 0. input signal is a convolution output. (c) the convolution output of SE envelope s[n] with the FOGD function wd[m]. the length of the smoothing detection methods. Algorithm Implementation and Assessment Function: p=DetectNegativeZerocrossing The proposed algorithm is implemented in MAT- Inputs LAB on a 2. p:=Locations of negative ZCs. and (f) detected time-instants of real R-peaks in the ECG signal. z[n].1 0. Algorithm pseudocode: DetectNegativeZerocrossing. N:= Size of signal (in samples).15 s[n] 0. (e) detected time instants of candidate R-peaks with the original ECG signal. By considering .23 filter is 0. The z:= Input signal. if (sign(z[n]) > 0) & & (sign(z[n + 1]) < 0). of M may merge two consecutive R-peaks in the ECG p½k :¼ n // storing location of negative ZC.05 (c) 0. determine the length of the smoothing filter.4-GHz Intel core 2 Quad CPU. The larger value for n = 1 to N  1. The duration of normal QRS and wide QRS complexes is ter in ECG record 207 are excluded from the perfor. (d) the time instants of negative ZCs that correspond to locations of peaks in the SE s[n].5 0 (e) ECG signal P[m] with 1 0 −1 (f) 2 R−peaks Detected 1 0 −1 0 1 2 3 4 5 6 7 8 9 Time (sec) FIGURE 5.125 times the sampling rate.05 z[n] 0 −0. signal. Example illustrating the output of stages of the proposed nonlinear transformation and peak-finding logic techniques. The choice of smoothing filter length (L) and the Begin Gaussian window length (M) is very important that Initialize k = 1. (a) The original ECG signal x[n].05 (d) 1 p[m] 0. algorithm is tested on the ECG signals taken from the Outputs first channel of the MIT-BIH arrhythmia database.4. may affect the detection performance. the average of lower and upper end. (b) the Shannon energy envelope s[n].2 s. Here. usually between 0.05 and 0. TABLE 1. In this work. For the sampling rate mance analysis for better comparison with the other of 360 samples per second. increment by 1.414 KATHIRVEL et al. limits of QRS complex duration is considered to End.

zero-padding process. The proposed algorithm achieves an detected by the proposed algorithm. storage.6. While different noises. TP þP ¼  100%. 2. we report the experimental results usual duration of computer evaluated ECG records is and provide comparisons between proposed algorithm 10-s. vary from 99. The proposed algorithm pro- signal. 3. In many ECG signal analysis. 3…M. M)=(45. Based on obtained against the first-channel of the ECG the previous works in the field of computer-aided ECG recordings of the MIT-BIH arrhythmia database. positive predictivity and the detection error else ðp½M þ K2Þ<N rate (DER) are computed as zpd:=zerosð1. SegStartPosition :¼ p½m  K2 .19. FP þ FN End.90%.23 we therefore process an ECG The proposed algorithm achieved better perfor- segment with duration of 10-s. The average time mance than the other detection algorithms reported required to process 10-s ECG segment is 6. The +P indicates the per- p½m :¼ p½m þ K2 . which is computed as window length (L) is computed as 2.M: centage of correctly detected true R-peaks to total zpd:=zerosð1. The performance of the algorithm is validated by comparing the results of automatic annotations against RESULTS AND DISCUSSION cardiologist annotations. . (FP) beats for a total detection failure of 114 beats. Table 3 summarizes and hence a number of diagnostic parameters such as sensitivity and positive predictivity of the proposed amplitude. . K2Þ. and shape of the ECG algorithm for test ECG signal taken from first-channel beat. TP end.11.94%. Therefore. DER ¼  100%: ð13Þ TP The overall performance is measured in terms of a the upper limits of heart rates in practice.10. Se ¼  100%.96% are (FP) when a noise spike is detected as R-peak. SegEndPosition :¼ p½m þ K2 . zero-padding process. the smoothing filter length (L) TP Accuracy (Acc) ¼  100%: ð14Þ and the Gaussian window length (M) of the detector TP þ FP þ FN are set as (L. R-peaks by the algorithm to total number of R-peaks Begin if p½1< K2 of original ECG signal.21 to 100% depending on the charac- Then. increment by 1. sess the performance of the proposed algorithm. 2. a sensitivity of (FN) when a R-peak is missed. ð11Þ TP þ FN for m = 1 to M. signal analysis. The sen- y ½1 to ðlengthðx Þ þ K2 Þ :¼ ½zpd x. and a positive predictivity of 99. and R-peaks are verified by visual inspection. Advancement of Medical Instrumentation (AAMI): M:= Total number of detected peaks. which typically contain a number of heartbeats and other existing algorithms. interval. in Table 4. m ¼ 1. we adopt the following the benchmark Inputs parameters reported by the Association for the y[n]:= Filtered ECG signal. 10 s of an ECG signal is con. In data compression for transmission.21 The In this section.4-ms. in this work. sitivity. The Outputs Se indicates the percentage of correctly detected true r[m]:=Locations of real R-peaks in ECG signal. ð12Þ ECGSeg :¼ y ½SegStartPosition to SegEndPosition. TP þ FP ½maxvalue maxindex :¼ maxKk ðjECGSeg[k]jÞ. But sidered. performing the visual inspection on the automatic The different stages and the overall performance annotation results. and (iii) false-positive 99. Efficient R-peak Detection 415 TABLE 2. duration.17. (ii) false-negative average accuracy of 99. n = 1. we calculate three quantities: of nine R-peak detection algorithms are summarized (i) true-positive (TP) when a R-peak is correctly in Table 4. In this study. the Gaussian detection accuracy. the first stage is to (each) of 48 ECG recordings of the MIT-BIH locate the peaks of the QRS complexes in the 10-s ECG arrhythmia database. Algorithm pseudocode: FindingRealRpeaks. each ECG record is the individual detection accuracies for ECG records divided into ECG segments with duration of 10-s.5 times the sam- pling rate. r½m :¼ maxindex þ SegStartPosition. m = 1. . The basis of standard clinical electrocardiography is the 10-s 12-lead ECG. duces 76 false-negative (FN) beats and 38 false-positive and retrieval of ECGs. each ECG segment is processed and the detected teristics of normal and pathological ECG signals. K2Þ. y ½1 to ðlengthðx Þ þ K2 Þ :¼ ½x zpd. Benchmark Parameters Function: r=FindingRealRpeaks In this work. number of peaks marked by the algorithm. sensitivity (Se) and positive predictivity (+ P) to as- K:=window size. 2. …N p[m]:=Locations of detected candidate R-peaks. 3.900) samples.

96 99. 116. 203.42 shows a detection failure rate of 0.197% (4 beats). The ECG records 104. 111.80 most algorithms produced a large amounts of false 107 2137 0 0 0 100 100 100 positives.585 99.94 99.97 ure of 2.537 99. and 232 often demonstrates that the proposed algorithm produces create more detection errors with other algorithms. 228.02% (41 beats) and of 1. 203 and 233 contain multiform 108 1763 1 4 0.95 100 99. 217. To test the robustness of our algorithm. respectively 217 2208 2 0 0.89 100 99. 209. Records 108 and 222 103 2084 0 0 0 100 100 100 104 2229 0 9 0. 219 and 232 include 102 2187 0 0 0 100 100 100 long pauses up to 6 s in duration.57 contain large T waves. 203.80 100 99. 202. Recently.and MMWT-based algorithms had DERs of 215 3363 0 0 0. it has been shown that the multiscale 202 2136 1 0 0. In the case of the very big changes 221 2427 0 0 0 100 100 100 222 2483 0 0 0 100 100 100 in shape of QRS complex and amplitude of QRS 223 2605 0 0 0 100 100 100 complex (record 106 shown in Fig.68% (34 beats).97 100 99. rithms. 7).96 99. Re- cords 108.93 99.01% (26 beats) and 1. 205. 108.428 99.46 rithm36 and the mathematical morphology and wavelet 205 2656 3 0 0. The DF-based algorithm14 and the 122 2476 0 0 0 100 100 100 EMD-based algorithm15 produced a total detection 123 1518 0 0 0 100 100 100 124 1619 0 0 0 100 100 100 failure of 530 beats (325 FP beats and 205 FN beats). 210.415 99.89 233 3079 1 0 0. 210.49 100 99. 201.72 ventricular arrhythmia. and 117 105 2572 3 8 0. our algorithm 228 2053 2 10 0. Performance of the proposed algorithm with the significant improvement is achieved with our algo- MIT-BIH arrythmia database (detection parameters: smooth- ing filter length (L) and Gaussian window length (L)). most of the algorithms had higher false- 116 2412 18 1 0. 201.90 duration are mainly found in record 232 that yields more false positive detections in most of the algo- Detection parameters: smoothing filter length (L) and Gaussian window length (L). 222. the 3M filter- 213 3251 0 0 0 100 100 100 214 2262 0 1 0. 106 2027 4 0 0.91 whereas the proposed algorithm produces 8 FP beats 219 2154 0 0 0 100 100 100 (0. 106.94 99. and 228 include severe baseline drifts and abrupt changes in QRS morphology.77 99. 200. 112.416 KATHIRVEL et al.96 99.25 99.051 99. 219 and 222 exhibits various irregular 101 1865 0 0 0 100 100 100 rhythmic patterns.098 99.95 100 99.053 100 99.91 100 99.117%) for a DER of 220 2048 0 0 0 100 100 100 0. For these ECG 115 1953 0 0 0 100 100 100 recordings. 108.197 99.95 mathematical morphology (3M) filtering based algo- 203 2980 14 2 0.21 117 1535 0 0 0 100 100 100 negative detections.89 99.311%) and 3 FN beats (0. Records 116 and 208 contain smaller 114 1879 0 1 0. Records 201.112 100 99.95 99. 208. negative QRS polarity and 109 2532 0 0 0 100 100 100 111 2124 0 0 0 100 100 100 sudden changes in QRS morphology.88 99. respec- 201 1963 1 0 0. 200 2601 0 0 0 100 100 100 and of 506 beats (209 FP and 297 FN beats). 200. 203. it is necessary to compare 210 2650 11 0 0. 116.90 99. Thus.032 99. 11.496 76 38 0.788 99.60 contain large sharp P waves. For these ECG recordings.404 100 99. For 230 2256 0 0 0 100 100 100 this record. 203. and 228 contain high-grade noise and artefact.284 99.113 99. 231.89 207 1862 0 0 0 100 100 100 transform (MMWT) based algorithm37 achieve the 208 2955 15 0 0. 221. L) =(45.091 99. TABLE 3. 223.53 99. ings 104.900) samples.508 99. ECG Total FN FP DER Se +P Accuracy record (beats) (beats) (beats) (%) (%) (%) (%) 219. L) 5 (45. 222. Records 201. (M. rithm.51 99. Records 200. Records 111.49 lowest total DER and outstanding false positive and 209 3005 0 0 0 100 100 100 false negative rates.58 the performance of these algorithms. but two false positive beats and achieves outstanding false . 105.044 100 99. 105.58 100 99.96 ing. 113. 100 2273 0 0 0 100 100 100 202.60 99. The comparison result of 30-min record 232 in Table 5 208.95 QRS complexes than the others.047 99. 113. Record 208 has 112 2539 0 0 0 100 100 100 wider PVCs. The numerous long pauses up to 6 s in Overall 109. 234 2753 0 0 0 100 100 100 respectively. the non-stationary noise is added to the ECG signal. (M.428% (11 beats).900) samples).69 99. 210.95 tively. with RR-interval greater than 4-s are shown in Fig. Record 223 exhibits sudden changes in 113 1795 0 0 0 100 100 100 QRS amplitudes. 118 2278 0 0 0 100 100 100 The effectiveness of the proposed algorithm in terms 119 1987 0 0 0 100 100 100 of the number of false negatives and false positives is 121 1863 0 0 0 100 100 100 shown in Table 5. The detection results of our algorithm for the ECG signal The MIT-BIH arrhythmia database ECG record.67% (43 beats). the 3M filtering-based algorithm and the 231 1571 0 0 0 100 100 100 MMWT-based algorithm showed total detection fail- 232 1780 0 2 0.000 100 100 100 1. For the high- 212 2748 0 0 0 100 100 100 grade noise and artifacts (Record 105).

W=30-ms where feature signal s[n] Md is the highest value of the modulus maxima in ith segment.33 ðdÞ rule fi constructing the eðdÞ ¼ 0:3ð0:7Mi þ 0:3AðdÞ Þ of 300 ms ðld Þ ðld Þ ðld Þ layer: the detection layer.18 Mi 23 Quadratic spline wavelet Four thresholds for QRS detection: Searchback with Refractory period 153 220 99.57 99. Ad is the average value of the R-wave amplitudes in d layer. four scales iqrs ¼ RMSðW2i x ½nÞ. if RMSi < 0.5 9 median of 0.6RMSi.375(A i + A i+1 ).80 98.59 removal fi apply SNR computed between coiflet wavelet fi wavelet coefficients squaring fi moving at 1st & 5th level coefficients averaging fi feature at 1st & 5th level signal s[n] 417 . 3 lowered thresholds of 0. the h(ld) 2 = 0.77 99.39 Mi1.17 derivative fi squaring h(i) pt = 0.6Ai .5 9 hpt T-wave removal 447 467 99.64 99.77 99. i ¼ 1. 2.2 s. b. 15 EMD fi soft-thresholding Four thresholds: window size (W) hst = 0.66 decomposition.2 s 4qrs ¼ 0:5RMSðW24 x½nÞ Thresholds to check significance: cqrspre ¼ 0:06maxðjW22 x½njÞ.18 Mi & & Mi £ 2Mi1 0. negative average slope Pa = peak amplitude.59 99.18 Mi & & Mi > 2Mi1 signal s[n] = 1. TABLE 4. Ald & Ald are the two R-wave amplitudes in (i  1)th segment within Efficient R-peak Detection that leakage period in ld layer. a. Methodology Ref.46 derivative fi squaring fi of eight past peak levels if RR > 1. 3 Bandpass filtering Automatic thresholds P calculation: Search back with <263 ms 393 253 99. Preprocessing stage Primary threshold.56 99. W=10-ms (ld) (ld) the verification layer. h1 ¼ 0:25ðAi þ Aiþ1 Þ.5 9 hpt Refractory period 467 244 99. Sn = negative slope. and c are the for each five consecutive coefficients that are found empirically samples 9 Baseline drift & noise Threshold based on the None None 214 1333 99.80 99.2–0. if RMSi > 0.3 9 hpt Refractory period 248 340 99.41 (four filters) fi derivative fi THVp ½m ¼ að15 Pm l¼m4 Sp ½lÞ adaptive thresholds squaring fi feature THVn ½m ¼ bð15 m l¼m4 Sn ½lÞ P signal s½n ! computing THVa ½m ¼ cð15 m l¼m4 Pa ½lÞ positive average slope and Sp = positive slope. T-wave moving-window integration fi of past RR intervals removal feature signal s[n] 4 Bandpass filtering fi Finding the threshold of segment i: hst = 0.77 99. if RMSi > 0.78 99. hpt Secondary threshold.39Mi. cqrspost ¼ 0:09maxðjW22 x ½njÞ.36 s transformation fi feature = 0.86 99. Preprocessing and decision stages of R-peak detection algorithms and comparison of their overall detection performance.69 99. hst Blanking FP (beats) FN (beats) Se (%) P+ (%) Acc (%) 14 Bandpass filtering fi hpt = threshold coefficient 9 median hst = 0. W=20-ms (ld) (ld) leakage layer fi h3 = 0.

small QRS complex.14 Hongyan None None None and Minsong. These decision rules with amplitude-dependent threshold may yield the poorest results in the case of ECG signal with irregular heart rhythm. the performance of many algorithms highly rely on an accurate measurement of initial Secondary threshold. In our previous work in Sabarimalai Manikandan and Dandapat.. The Shannon energy & smoothing fi smoothing fi feature signal s[n] waveform (c) is the smooth Shannon energy envelo- derivative fi normalization gram (SEE) of the thresholded signal e~th ½n: The SEE is multiple-frame accumulation Preprocessing stage filtering fi derivative fi derivative fi squaring fi fi Shannon energy fi used as the feature signal. and wide QRS complex.80 99. Maxi < 3 Methodology output: time instants output:time instants phase. The waveform depicted in (b) is the differenced bandpass filtered signal.15 and Zhang and Lian.418 KATHIRVEL et al. RR- interval and duration thresholds estimated for the previous detected R-peaks.37 In this work.81 99. 3 £ Maxi £ 5 average RR interval and rate limits. In this proposed algorithm. It was seen that the algo- FN (beats) rithm with secondary threshold produces several spu- 213 79 76 rious peaks.96 abrupt change in amplitude of QRS complex. Continued. and noise and artifacts. the proposed algorithm does not use additional deci- sion rules with sets of thresholds based on the running HT of s½n ! LF drift removal fi Convolution of s[n] with wd ½n ! estimates of the signal peaks and noise peaks. continuously Blanking varying QRS morphology. The waveform (d) is the fi feature signal s[n] convolution of the feature signal s[n] with FOGD Bandpass filtering fi Bandpass filtering fi absolute of v½n ! function wd[n]. hpt = 0.27Maxi. The waveform depicted in (e) shows the feature signal s[n] normalization fi Three morphology thresholding fi time instants (marked in red circle) of detected R-peaks using the proposed algorithm.62 99. Experiments show that the proposed FOGD-based 36 30 .90 rithms. almost all algorithms used sets of Se (%) 99. The performance comparison in Table 5 shows that the proposed algo- rithm has a total detection failure of 106 beats (36 FP TABLE 4. moving average (MA) filter and ZC detector.9 Hamilton and Tompkins. The waveforms of the different stages of the pro- of ith segment posed algorithm using the ECG segments taken from of R-peaks of R-peaks first-channel of the different recordings of the MIT- BIH arrhythmia database are shown in Figs. Ref. d[n]. Many algorithms have large number of false-nega- P+ (%) tive (or missed detections in case of ECG signals with 99. positive values compared to recently reported algo- Acc (%) 99.8 Elgendi et al. a set of rules for Maxi =maximum value = 0.30 the deci- sion stage includes HT. Maxi > 5 blanking and T-wave discrimination.3 Darrington. 6–12.36. In each of these figures. x[n]. beats and 70 FN beats) and outperforms the other algorithms reported in Adnane et al. and training hpt = 0.93 99.88 99.80 99.15Maxi. It can been seen that the ECG FP (beats) record 203 has many shortest RR-intervals of less than 38 204 140 360 ms. hst parameters in the learning phase. Thus. and (ii) the negative ZCs detection. the refractory period of 200 ms and T-wave discriminator (200–360 ms) are widely used in most of the algorithms.1Maxi.. the negative ZCs detection fi negative ZCs detection fi real R-peaks detection fi real R-peaks detection fi Primary threshold. waveform depicted in (a) is the original ECG signal. To detect missed R-peaks. the decision stage includes two steps: (i) the convolution of Our work the FOGD function wd[m] with the Shannon energy envelope s[n].94 heuristic decision rules with adaptive amplitude. None None None Furthermore.

No. Number of false-positive (FP) detections Number of false-negative (FN) detections Rec. <0.05 mV 9 2 13 3 5 18 1 0 19 19 3 10 30 39 9 15 209 Bursts of noise 2 0 0 2 1 29 2 0 2 0 1 9 0 17 12 0 Efficient R-peak Detection 210 Small QRS complexes. noise & artifacts 2 8 5 16 8 10 11 0 41 23 2 5 29 54 1 11 221 Wide QRS complexes 1 4 0 4 0 16 0 0 1 2 4 8 4 12 3 0 222 Irregular heart rates & small QRS 40 5 27 1 3 3 1 0 37 0 12 0 0 13 6 0 223 Abrupt change in amplitude of R-peak 0 28 0 4 0 4 6 0 2 1 0 22 1 12 1 2 228 Severe noise & very big change in 19 38 1 10 76 14 20 10 6 22 14 2 6 15 8 0 amplitudes of adjacent R-peaks 231 Irregular heart rates 0 2 0 7 0 0 22 0 0 0 331 1 0 9 5 0 232 Numerous long pauses up to 6 s 3 26 0 14 20 31 29 2 0 0 17 2 0 8 5 0 Total 209 325 166 131 338 619 174 36 297 205 1239 165 212 503 161 70 419 . 50 68 2 10 55 178 14 4 47 9 62 2 28 48 2 1 & severe noise and artifacts 113 Sharp-tall T-wave & baseline drifts 2 6 1 10 0 0 5 0 1 0 682 11 1 9 5 0 116 Very small QRS (Amp. noise & artifacts 208 Wide-QRS & small-QRS <0.05 mV) 4 – 20 4 3 1 0 1 25 – 0 27 20 28 21 18 200 Multiform PVCs with noise and artifact 3 47 1 4 47 122 0 0 2 3 2 9 0 102 2 0 201 Junctional escape beats 3 3 1 2 0 0 0 0 19 10 66 4 3 21 10 1 202 Irregular heart rates and low-QRS 0 5 1 2 0 1 0 0 3 6 3 6 2 9 17 1 203 Very big change in adjacent QRS-shape. negative QRS complexes. very small QRS (<0.04 mV). heart rates. 14 23 79 3 27 88 2 2 61 95 19 7 53 45 23 14 QRS-amplitude. Comparison of numbers of false-positives and false-negatives for pathological and noisy records of the MIT/BIH database. TABLE 5. ECG signal quality DF14 EMD15 WT9 3MM36 Slope3 Maxima8 MMWT37 Our DF14 EMD15 WT9 3MM36 Slope3 Maxima8 MMWT37 Our 104 Multiform PVCs & severe muscle noise 3 20 0 7 63 25 13 9 7 1 0 1 12 25 2 0 105 High-grade noise and artifacts 53 35 15 7 28 51 38 8 22 14 21 19 6 28 5 3 106 Abrupt changes in QRS morphology 1 5 0 21 02 28 10 0 2 0 0 20 17 9 24 4 108 Sharp-tall P-wave.

04 0. Our algorithm produces 0 missed beats and 0 FP beats for a total of 2187 true beats.02 0 2 4 6 8 10 12 14 16 18 (d) 0.5 0 2 4 6 8 10 12 14 16 18 (b) 1 0.5 R−peaks Detected 1 0.06 0.05 z[n] 0 −0. and tall T waves (Record 106). Performance of the proposed algorithm for ECG record 102 with varying QRS complex morphologies.5 d[n] 0 −0.5 0 2 4 6 8 10 12 14 16 18 (c) 0.05 0 2 4 6 8 10 12 14 16 18 (e) 1.5 d[n] 0 −0.05 z[n] 0 −0. Our algorithm produces 04 missed beats and 0 FP beats for a total of 2027 true beats.420 KATHIRVEL et al.08 0.05 0 2 4 6 8 10 12 14 16 18 (e) R−peaks 2 Detected 1 0 −1 0 2 4 6 8 10 12 14 16 18 Time (sec) FIGURE 7.5 0 −0. Detection performance for the ECG signal with continuously varying QRS complex morphology.08 s[n] 0. Detection accuracy is 100%.5 −1 0 2 4 6 8 10 12 14 16 18 (c) 0.1 0.12 0. .5 0 −0.06 0.04 0.5 0 2 4 6 8 10 12 14 16 18 Time (sec) FIGURE 6.02 0 2 4 6 8 10 12 14 16 18 (d) 0. (a) 2 1 x[n] 0 −1 0 2 4 6 8 10 12 14 16 18 (b) 0. (a) 1 x[n] 0. sudden changes in beat- to-beat RR-interval.1 s[n] 0.

06 0.5 −1 0 5 10 15 20 25 (c) 0. <0.05 0 2 4 6 8 10 12 14 16 18 (d) 0. Efficient R-peak Detection 421 (a) 2 very small QRS complexes (Amp. Detection performance for the ECG record 116 with very small QRS complexes (Amp.05 mV) x[n] 0 −2 0 5 10 15 20 25 (b) 0. .1 0.08 s[n] 0. Detection performance for the ECG record 228 with very big change in amplitudes of adjacent R-peaks and severe noise. Our algorithm produces 10 FP beats and 2 FN beats for a total of 2053 true beats.04 0.5 −1 0 2 4 6 8 10 12 14 16 18 (c) 0.02 0 5 10 15 20 25 (d) 0.1 z[n] 0 −0.05 0 5 10 15 20 25 (e) 4 R−peaks Detected more number of false negatives (FNs) 2 0 −2 0 5 10 15 20 25 Time (sec) FIGURE 8.05 mV).1 0 2 4 6 8 10 12 14 16 18 (e) Detected R−peaks 3 falsely detected beat 2 1 0 −1 0 2 4 6 8 10 12 14 16 18 Time (sec) FIGURE 9.15 s[n] 0.05 z[n] 0 −0.5 0 d[n] −0. < 0. The algorithm produces 18 missed beats for a total of 2412 true beats.1 0. (a) 3 2 x[n] 1 0 −1 0 2 4 6 8 10 12 14 16 18 (b) 0.5 d[n] 0 −0.

1 0. Detection performance for the ECG record 108 with large P-waves and severe muscle noise.15 0.05 0 2 4 6 8 10 12 14 16 18 (d) 0. (a) 1 0.5 0 2 4 6 8 10 12 14 16 18 (b) 0.2 Detected R−peaks 0 −0.422 KATHIRVEL et al.1 s[n] 0.6 −0.5 0 −0.5 d[n] 0 −0.2 −0.4 −0.2 −0. (a) 0.5 d[n] 0 −0. Our algorithm produces 2 FP beats and 0 FN beats for a total of 1780 true beats.5 x[n] 0 −0.1 0 2 4 6 8 10 12 14 16 18 (e) 1 Detected R−peaks 0.5 −1 0 2 4 6 8 10 12 14 16 18 (c) 0.6 −0.8 0 2 4 6 8 10 12 14 16 18 (b) 0. Detection performance for the ECG record 232 with noise and numerous long pauses up to 6 s.2 x[n] 0 −0.1 z[n] 0 −0. .8 0 2 4 6 8 10 12 14 16 18 Time (sec) FIGURE 10.15 s[n] 0.5 0 2 4 6 8 10 12 14 16 18 Time (sec) FIGURE 11.1 z[n] 0 −0.05 0 2 4 6 8 10 12 14 16 18 (d) 0.1 0 2 4 6 8 10 12 14 16 18 (e) 0. Our algorithm produces four FP beats and one FN beats for a total of 1763 true beats.5 −1 0 2 4 6 8 10 12 14 16 18 (c) 0.4 −0.

1 0. Detection performance for the ECG record 203 with very big change in adjacent QRS-shape. amy. Our algorithm produces 2 FP beats and 14 FN beats for a total of 2980 true beats.05 0 2 4 6 8 10 12 14 16 18 (d) 0. and smoothing degraded due to missed or falsely detected beats. small QRS commonly. peak-finding technique achieves higher detection cardiac sound separator. wider QRS complex. automatically detecting R-peaks in a ECG signal. Efficient R-peak Detection 423 (a) 1 x[n] 0 −1 −2 0 2 4 6 8 10 12 14 16 18 (b) 0. and noise and artifacts. and bizarrely shaped QRS complexes usually having wider duration of greater than 120 ms. The proposed shape. many short RR-interval (RR <360 ms). ECG coder.15 s[n] 0. QRS-amplitude and heart rates. 203. This is a very difficult record.1 0 2 4 6 8 10 12 14 16 18 (e) Detected R−peaks 1 0 −1 −2 0 2 4 6 8 10 12 14 16 18 Time (sec) FIGURE 12. rates in the case of ECG signals with sharp and tall P loidosis. long by heart attack. Shannon energy computation. multiform PVCs. misclassification of beats is more problematic the number of false-positives and false-negatives in the in heart rate variability. QRS-amplitude. The PVCs are caused changes in QRS morphology. The 208. sudden changes in QRS amplitudes. and provides the signifi. case of ECG signal with small-QRS and wide-QRS . The Dandapat. ECG wave delineator. sudden delay of ventricular activation.04 mV). This nonlinear transformation process reduces example. Most and T waves. and heart rates. wide. and haemochromatosis. algorithm consists of new nonlinear transformation QRS complexes and very big change in adjacent QRS. very small QRS (Amp. prob. algorithm has lower false-positive and false-negative ably due to myocardial infiltration-for example. <0.1 z[n] 0 −0. and are characterized by premature pauses and irregular heart rhythms. In the literature. patient authentication and accuracy. For process. many methods had more false negative This paper presents a new and simple algorithm for detections for the ECG records namely 200. sarcoidosis. muscle noise. The overall performance of the various QRS CONCLUSION detection methods is summarized in Table 5. and simple peak-finding techniques. and 228 that include small-QRS complexes. The accuracy nonlinear transformation is based on energy thres- of computer-aided cardiac diagnostic system may be holding. cant computational advantage as compared to that of Various experiments prove that the proposed non- Hilbert transform-based peak technique reported in linear transformation and simple peak-finding tech- our previous work in Sabarimalai Manikandan and niques improves the detection performance. negative QRS complex.5 d[n] 0 −0.5 −1 0 2 4 6 8 10 12 14 16 18 (c) 0.30 The most common ECG abnormality is experimental results demonstrate that the proposed the presence of low amplitude QRS complexes. easy to implement. baseline conduction over the right or left bundle due to block or drift. fHRM. 201. a wide QRS complex is due to abnormal complex. even for humans.

4654–4657. Med. (2) it uses a single thresholding wavelet transform. The proposed algorithm produces 10 El-Segaier. and A. and M. 2002. QRS complex rule. IEEE complex fractional wavelet. Z. D. 31:399–406. Meas. Biomed. Biomed. 2005. B. Hennig. Ghasemi. doi:10. M. S. J. Eng. Zsuzsanna. 42:21–28. Eng. In: International Conference Trans. 2004. De Boer. Eng.. The use of the Hilbert transform in ECG signal Li. Biomed.. G. Murray. phase. Towards real time QRS detection: a fast mines locations of candidate R-peaks in the QRS method using minimal pre-processing. proposed peak-finding technique automatically deter. and a positive pre- pp. C. Zheng. A. and H.. and J. Detecting QRS complexes of ACKNOWLEDGMENTS two-channel ECG signals by using combined wavelet en- The authors would like to thank Editor-in-Chief.. and R. A. China. S. Kligfield. IEEE Trans. Comput. L. Wide. Computer-based detection and total 109. I. IEEE Trans. puterized Electrocardiology. C. Naka. and A. for his continuous encouragement and the anonymous 17 Hughes.1007/s10916-009-9405-3.. Biomed. IEEE Martnez. Khler.424 KATHIRVEL et al. 8 Darrington. Kozmann. PhD thesis. 2009. IEEE Society endorsed by the International Society for Com- Trans. China. Srnmo. R wave from the MIT-BIH arrhythmia database and achieves detection using Coiflets wavelets. P. The Conference of the IEEE EMBS. IIEEE Trans. In: IEEE 35th Annual Northeast Bioengineering Conference. Lukkarinen. S.. Quantitative inves- previous segment. Bursa. The princi- 267–270. K. Methods Programs Biomed. B. Detection of ECG charac- analysis. E. Bartels. F. H. Orglmeister. University of Oxford. 93(1):20– Association Electrocardiography and Arrhythmias Com- 31. Council on Clinical Cardiology. Tsipouras. . Rocha. Biomed. Inf.. Heart Rhythm 4(3):394–412. 18 Kadambe. IEEE Eng. dictivity of 99. Charef. Development of QRS electrocardiogram: part I. Trans. 2009. 20(4):1221–1228. and W. 2008. Deng. B. A new QRS detection Mag. 49(10):1094–1101. a set of medical tactics to reduce tigation of QRS detection rules using the MIT/BIH false-positives and detect missed beats. 2005. G. Biomed. 33:1157– 1165. L. A new QRS detection for automated ECG signal analysis applications. An automated methodology for fetal heart rate 1 Abdelliche. pp. ples of software QRS detection. Shanghai. Bailey. 2010. the proposed algorithm is more suitable 15 Hongyan. and Y. 2008. tions for the standardization and interpretation of the 3 Adnane. Biomed. Med. Experimental results analysis of heart sound murmur.. Wang. Eng.. P. 2010. 1986. P. 11 Gerencsr. R-peak detection using a extraction from the abdominal electrocardiogram. M. The proposed R-peak detection algo. 55(2):478–484. and noises. and learning arrhythmia database. Rev. Biol. IEEE Trans. A new (1) it is a one-pass algorithm without any further mathematical based QRS detector using continuous searchback algorithm. Olmos. L. 38 false positive beats and 76 false negative beats for a Sepponen. Choi. show that this algorithm outperforms other existing 33(7):937–942. B. interval analysis. Tompkins. and H.. 1995. L...-E. J. M. Tai. we also developed envelope. Philipson. Wavelet transform-based QRS complex detector. D. Gaydecki. Y. pp. Signal feature signal. Eng... 1999. Nait-Ali. M. 20 on Electrical and Electronics Engineering. pp. and C. Sig- does not require sets of thresholds based upon the R nal Process. In: Proceedings of the 2nd International Conference on Bio- informatics and Biomedical Engineering. 693– 696. Lilja. L. 2009. and F. Analysis of College of Cardiology Foundation. 1–2. and the Heart Rhythm first-derivative based QRS detection algorithms. Pesonen. 16 Huang. 51(4):570– algorithm based on mathematical morphology and 581. Beijing. The main advantages of the proposed algorithm Eng.94%. of the SPSA method in ECG analysis. S. REFERENCES 19 Karvounis. S. Biol. In: International Conference on Bioinformatics and Biomedical Engineering. Seo. and E. C. Comput.. F. Technol. Jiang. and (4) it detection using empirical mode decomposition. algorithm based on empirical mode decomposition. 34:81–91. J. Recommenda- Med. Jonkman.-U. R. X. 1–4. Thus. M. 2001. Gettes. 9 rithm is validated using all 48 half-hour ECG records Elgendi. A. 50(3):808–812. Bio- 6 Biel. compared with the other existing algorithms are that 12 Ghaffari. O. J. 2007. peaks and RR intervals that are determined in the 14 Hamilton. ECG med. 2008. P. and G. Laguna. 2002. and K. Pettersson. 2006. In: Proceedings of 27th Annual International a simple peak-finding technique using the FOGD. 23 analysis: a new approach in human identification. an average sensitivity of 99. tropy. and P. Boston. and S. MA. 13 Hadj Slimane. Control 1:169–176. 46(7):838–848. 2007. complexes. Eng. J. Instrum.. and H. Z. IEEE Trans. S. Fitzpa..-D. L. K. Zaidi.-S. N. 2001. pp. The electrocardiogram and its detection algorithm designed for wearable cardiorespira. (3) it uses simple peak-finding strategy. Golbayani. P. and P. and A. Comput. Minsong. Eng. 11(6):628–638. O. Almeida. Z. N. 5 Benitez. A wavelet-based ECG delineator: evaluation on 7 Chen. In this work. M. and C.496 true ECG beats.. Fotiadis. 2009. mittee. Duan. Probabilistic models for automated ECG referees for their valuable suggestions and comments. Poon. Biomed. 22 trick. 2 Abibullaev. 21:42–57. A QRS complex detection standard databases.96%. Digit. P. L. technology: a scientific statement from the American Heart tory system. J. Ann.. R. and H. A. the American 4 Arzeno. P. The use algorithms in case of pathological and noisy signals. C. R.. 21 method using wavelets and artificial neural networks. 2006. Process. teristic points using wavelet transforms.. Syst. Electr.

Eng. Signal Process. Efficient R-peak Detection 425 24 32 Meyer. and N. R. filter. Lian. and S. Biomed. Xu.03. M. A novel phological filter and energy envelope for applications in method for detecting R-peaks in eletrocardiogram (ECG) body sensor networks. Dandapat. E. ambulatory monitoring—a review. Proceedings of the Third International Conference on Eng. 1984. 624–627.. Dandapat. IEEE Trans. Syst. 2010. 40(2):201–205. 1993. M. J. and L. 2009. Circuits Syst.1007/s11265-009-0430-8. Combining algorithms Tang. and W. O. P. In: Fourth International 25 Pahlm.. 2008. 140–145. R-wave detec. J. Eng. and X. X. Eng.-J. IEEE Methods Programs Biomed. tion using continuous wavelet modulus maxima. P. Yeha... 1985. Biomed. Wanga. 54(4):485–489. 1–4. H. G. Eng. A new QRS detector 31 Surez.. Biomed.. Yang. Tang.-C. and X. Thakor.1016/j. Addison. 30:565–568. M. Software QRS detection in Conference on Natural Computation. Med. Com. Q. Zhang. Zhang. Harris. N. 2009. ECG beat detection using a geometrical matching International Conference on Signal Processing (ICSP). and W. for ECG signal: the difference operation method. Biomed. 27 34 Prasanna. Finding pitch Trahanias. 31:702–706. doi: signal. Zhao. and Y. J.2011. pp. 2005. empirical mode decomposition. J. Biol. IEEE Trans. based on empirical mode decomposition. C. 22:289–297. Srnmo. V. and Y. Biomed. Technol. J. In: IEEE using mathematical morphology. V. QRS detection based on mor- 30 Sabarimalai Manikandan. 2006. Biomed.. Control 3(4):220–228. IEEE Trans. Y. 38 10. A real time QRS detection mation of QRS complex power spectra for design of a QRS algorithm. and M. . F. C. Tompkins. L.. An approach to QRS complex detection markers using first order Gaussian differentiator. S. 91:245–254. Cardiol.. 2008. F. Comput. in automatic detection of QRS complexes in ECG signals.. Zhang.. Wavelet scale mathematical morphology for wearable ECG devices threshold based TDL and TDR algorithms for real-time in body area networks. Zou. J.. Signal Process. and W.bspc. P. Y. Inf. 32(3):230–236. and S. J. Grubb. W. Esti- 26 Pan. 33 put. Comput. approach. Gomis. Zhu. S. Signal Process.. Gavela. Y.. ECG signal compression. Control. and A. 1984. 5. 2007. In: IEEE 10th Najim. J. F. IEEE Trans. 37 3:44–66. Chen. Biomed. and M. Vol. Tompkins. 2008. K. doi: 10. 10(3):468–475. pp. Biomed. 35 Intelligent Sensing Information Process. Silva. Lian. 36 Proc.004. 2011. Berthoumieu. Webster. The algorithm of R Peak detection in ECG based on IEEE Trans. Subramanian. IEEE Trans. 2003. J. QRS detection based on multi- 29 Sabarimalai Manikandan.. pp. QRS complexes detection 28 Romero.