Australasian Physical & Engineering Sciences in Medicine

https://doi.org/10.1007/s13246-018-0670-7

REVIEW PAPER

Development of robust, fast and efficient QRS complex detector:

a methodological review
Sandeep Raj1   · Kailash Chandra Ray1 · Om Shankar2

Received: 9 February 2018 / Accepted: 2 August 2018

© Australasian College of Physical Scientists and Engineers in Medicine 2018

Abstract
The basis and reliability for timely diagnosis of cardiovascular diseases depend on the robust and accurate detection of
QRS complexes along with the fiducial points in the electrocardiogram (ECG) signal. Despite, the several QRS detection
algorithms reported in the literature, the development of an efficient QRS detector remains a challenge in the clinical envi-
ronment. Therefore, this article summarizes the performance analysis of various QRS detection techniques depending upon
three assessment factors which include robustness to noise, computational load, and sensitivity validated on the benchmark
MIT-BIH arrhythmia database. Moreover, the limitations of these algorithms are discussed and compared with the standard
signal processing algorithms, followed by the future suggestions to develop a reliable and efficient QRS methodology. Fur-
ther, the suggested method can be implemented on suitable hardware platforms to develop smart health monitoring systems
for continuous and long-term ECG assessment for real-time applications.

Keywords  Electrocardiography · QRS complex · Computational load · Sensitivity · Real-time applications

Introduction on the surface of the body which is later analyzed by a car-

diologist. The term ECG was introduced by Augustus D.
The report of World Health Organization (WHO) places Waller (a British physiologist) in 1887 when he recorded the
the cardiovascular diseases (CVDs) as the leading cause first human ECG using a capillary electrometer [5]. In 1893,
of mortalities across the globe and will remain in the near Einthoven used an improved electrometer and a correction
future [1]. In 2008, the CVDs caused 17.3 million deaths, formula to distinguish five deflections later named as P, Q,
representing 30% of the mortalities worldwide [1]. In 2018, R, S and T [5] waves. These waves are generated by heart
the deaths in United States alone has increased to 836546 that undergoes three processes, namely atrial depolariza-
(with an average of one death in every 38 s) [1, 2]. By 2030, tion, ventricular depolarization and ventricular repolariza-
the expected number of deaths can increase up to 23.3 mil- tion resulting in the generation of P wave, QRS complex,
lion globally [3]. As a consequence of the increased rate in and T wave respectively. These different waves comprising
mortalities due to CVD’s, cardiac health research has gained the standard ECG cycle are depicted in Fig. 1 while their
significant importance by the researchers. The most common clinical significance is summarized in Table 1. In Fig. 1,
clinical technique utilized for cardiac disorder analysis is the U wave is shown, however it is seen occasionally. It is a
an electrocardiogram (ECG). An ECG is a simple, reliable, positive wave occurring after T-wave having an amplitude
low-cost and non-invasive tool commonly used to diagnose of one-fourth of the T wave. The U wave is found in subjects
cardiac disorders [4]. An ECG records the electrical signals having more prominent T waves and slow heart rates (most
originating from the myocardium by placing the electrodes frequently seen in leads V2–V4). Therefore, the genesis of
U wave is elusive. All of these waves exhibit specific char-
* Sandeep Raj acteristics (such as in time, amplitude or morphology) and
srp@iitp.ac.in carry sufficient amount of information for diagnosing CVDs.
Further, various other features such as frequency, entropy
1
Department of Electrical Engineering, Indian Institute distribution and energy, event intervals (like the RR-inter-
of Technology Patna, Patna 800013, India
val) are also extracted for reliable diagnosis. Any change in
2
Department of Cardiology, Institute of Medical Sciences, either of the features of these P, QRS, T waves may indicate
Banaras Hindu University, Varanasi 221005, India

13
Vol.:(0123456789)
 Australasian Physical & Engineering Sciences in Medicine

detailed analysis of the cardiac events is possible only if the

QRS event is detected efficiently. Therefore, the development
of fast, robust, efficient and reliable QRS detector becomes
clinically important for timely diagnosis of CVDs.
The objective of this review article lies in the thorough
analysis of QRS detection algorithms available in the litera-
ture. In order to prepare this manuscript, the literature are
searched and reported using the databases such as Google
Scholar, Scopus, web of science and the keywords searched
is QRS and detection. The analysis is divided in two stages,
i.e., the pre-processing stage and the QRS detection stage. In
Fig. 1  Cardiac cycle representing the waves
the preprocessing stage, the QRS complex feature is made
more prominent or emphasized with respect to the rest of the
cardiac abnormalities or arrhythmias leading to stroke or waves. The output of this stage is followed by the QRS detec-
sudden cardiac death. Therefore, an efficient diagnosis of tion stage where the onset and offset points are demarcated
these waves is clinically essential for reliable analysis of and the corresponding R-peak is located in the ECG signal.
heart health, such as arrhythmia classification [6–15], diag- The contribution of this study is to evaluate the performance
nosing breathing disorders [16, 17], study of cardiac func- of algorithms or techniques depending upon three criteria,
tioning during sleep and hypertension [18, 19], epilepsy [20] i.e., (a) sensitivity towards the noise, (b) computational load
and for examining various other heart disorders [21]. and (c) accuracy for both the stages. Further, the current
Among these waves, the QRS wave exhibits the most strik- study also provide suggestions to develop a fast, efficient and
ing feature in terms of morphology, amplitude and time of robust QRS detector methodology for real-time applications.
occurrence and therefore, plays a significant role in an effi- The rest of the article is summarized as follows. “MIT-
cient analysis of ECG recordings of subjects. In the past few BIH arrhythmia database” section briefly describes the most
decades, the detection of QRS complexes has been thoroughly commonly used ECG database studied in the literature. “Pre-
studied by various researchers. In fact, the study of P and T processing stage” section presents the evaluation of the pre-
wave detection is not explored much in comparison to QRS processing algorithms while “QRS detection techniques” sec-
complex detection due to the factors including low signal-to- tion presents the evaluation of the QRS detection algorithms.
noise ratio (SNR), variation in amplitude and morphology, “Performance evaluation and discussion” section presents a
and overlapping nature of P and T wave. In spite of various brief discusses the challenges of the evaluated pre-processing
studies and research works in the domain of QRS complexes and QRS detection stages on together followed by future sug-
detection, the development of a reliable universal solution is gestions to develop an efficient QRS methodology. Final sec-
still a challenge. These challenges mainly arise due to low tion presents the conclusion of the study.
SNR, variability (i.e. inter and intra) in the morphology of
QRS complex and the rest of the waves as well as the artifacts
inherent in the ECG signal. The most commonly used data- MIT‑BIH arrhythmia database
base for validating the research works has been the benchmark
Massachusetts Institute of Technology—Boston’s Beth Israel Most of the research works are evaluated and validated on
Hospital (MIT-BIH) arrhythmia database [22] (described in the benchmark Massachusetts Institute of Technology—Bos-
detail in “MIT-BIH arrhythmia database” section). However, ton’s Beth Israel Hospital (MIT-BIH) arrhythmia database

Table 1  Significance of different waves in ECG signal [5]

Segment/ wave Duration (ms) Amplitude (mV) Clinical significance

P wave 80–100 0.25 Atrial enlargement, fibrillation, flutter

PR segment 120–200 Baseline Pericarditis, heart blocks, atrial tachycardia
QRS complex 80–120 1.60 (R-wave) Ventricular hypertrophy, myocardial Infraction
QT interval 200–400 25% of R wave (Q-wave) Electrolyte abnormalities, coronary heart disease
ST segment 80–120 Baseline Coronary ischemia, myocardial infraction
T-wave 120–160 0.1–0.5 Coronary ischemia, wellens syndrome, left
ventricular hypertrophy, hyperkalemia
U-wave 20–40 0.3–0.7 Hypokalemia, hypercalcemia, thyrotoxicosis

13
Australasian Physical & Engineering Sciences in Medicine

[22] developed during 1975 and 1979 by the BIH arrhyth-

mia laboratory. The database contains two channel ambula-
Ath = 𝛼 × max{A[m]} (1)
tory ECG recordings of 47 different subjects comprising 48 where 𝛼 is the amount of ECG signal eliminated in percent-
records. A modified limb lead II (MLII) is the lead A signal age whose value vary from 0 < 𝛼 < 1 . Moreover, the value
in 45 recordings; a modified lead V1 (often V2 or V5, and of 𝛼 is optimized once before the pre-processing while the
V4 in one excerpt) is the lead B. Whereas V5 is the lead A thresholds are kept constant throughout the analysis. Vari-
signal in the other three excerpts and V2 is the lead B (two ous amplitude thresholds are employed for subsequent QRS
excerpts) or MLII (one excerpt). The heartbeats signal in detection. Morizet et al. [25] introduced a QRS scheme using
lead A have more prominent peaks than lead B signal. The Ath = 0.3 max{A[m]}, where below 30% of the peak ampli-
database includes 110109 beat labels while the data is band- tude of the signal is truncated for A[m], whereas Fraden [26]
pass filtered at 0.1 H–100 Hz. The excerpts are digitized employed Ath = 0.4 max{A[m]}.
with a sampling frequency of 360 samples per second and The main advantage of this technique is that it involves
acquired with 11-bit resolution over 10 mV range. The data- less computational load among all the existing pre-pro-
base is open-access available on-line that can be accessed cessing techniques which is due to smaller length of ECG
for performing the experiments. signal used for processing. The disadvantage being the
length of ECG segments processed are fixed and deter-
mined empirically [25–27]. If the length of ECG is longer,
the performance may degrade until it divided into shorter
lengths but the ECG segment may lose the starting and
Pre‑processing stage end of ECG beats.

In this stage, the acquired raw ECG signal is pre-processed

to remove various kinds of noise and artifacts [23] associ- First order derivative
ated with them. These various kinds of noise include the
baseline wander, artifacts due to muscle contraction, elec- Generally, a differentiator of first-order is utilized as a
trode movement and power-line interference. The pre-pro- high-pass filter which removes unwanted low-frequency
cessing stage improves the SNR of the ECG signals. Hence, noise and the baseline wander of A[m]. Moreover, it cre-
the pre-processing of the ECG signal is highly instrumental ates zero crossings at the R-peak location and modifies the
for an efficient QRS detection. Otherwise, it results in the phase in the ECG signals. Several algorithms implemented
generation of false alarms and degraded performance of the the first derivative by the following equation [28]:
QRS detector. This section presents the performance evalu- B[m] = −2A[m − 2] − A[m − 1] + A[m + 1] + 2A[m + 2]
ation of the various pre-processing techniques based on two (2)
factors, i.e. computational load and robustness to noise. A Moreover, Holsinger et al. [29] used a central finite-differ-
summary of the evaluation done is presented in Table 2 at ence approach as:
the end of this section.
For reader’s point of view, this section uses two variables, B[m] = A[m + 1] − A[m − 1] (3)
i.e. A[m] which refers to the raw input ECG signal and B[m] while Okada et al. used a backward difference scheme [30]:
refers to the output filtered signal.
B[m] = A[m] − A[m − 1] (4)
An optimal threshold is chosen and applied to B[m] along
with the first-order derivative technique for subsequent QRS
Amplitude technique complex detection within the ECG signals. The length of
ECG signal processed and the thresholds applied during the
The amplitude technique is one of the commonly used ECG analysis are fixed. The main advantage being this tech-
algorithms used for the R-peak detection within the ECG nique involves less computational load. The disadvantage
signals. Initially, a differentiation step is applied to sup- being the technique is not able to remove high-frequency
press the P and T event influence in order to highlight the noise completely.
QRS complex, which is followed by the amplitude thresh-
old. Later, this algorithm is used by Sufi et al. [24] to First and second derivative
detect the heart rate on mobiles. Moreover, the amplitude
threshold followed by the first derivative to make the slope The first and second order derivatives are calculated separately
of QRS complex more prominent. The amplitude threshold for A[m] (i.e. input ECG signal) in the QRS enhancement algo-
for a fragment of the ECG signal is determined as: rithms. These derivatives magnitudes are linearly combined

13
 Australasian Physical & Engineering Sciences in Medicine

to highlight the QRS wave region with respect to other ECG 1

B0 [m] = (A[m − 1] + 2A[m] + A[m + 1]) (14)
features. The first and second order derivatives computed by 4
Balda et al. [31] are defined as:
Further, a LPF is applied to B0 [m]
B0 [m] = |A[m + 1] − A[m − 1]| (5) m+n
1 ∑
B1 [m] = |A[m + 2] − 2A[m] + A[m − 2]| (6) B1 [m] = B [p] (15)
2n + 1 k=m−n 0
Here, B0 [m] is the first while B1 [m] is the second order deriv-
atives of A[m]. Further, both of these derivatives are linearly The input and output of this LPF are subtracted and squared,
combined as: to remove waves of low amplitude with respect to R peak as:
B2 [m] = 1.3B0 [m] + 1.1B1 [m] (7) B2 [m] = (B0 [m] − B1 [m])2 (16)
In [32], Ahlstrom et al. computed the first derivative as: Thenafter, filtering is applied to this square of difference
which makes the QRS area enlarged relative to another ECG
B0 [m] = |A[m + 1] − A[m − 1]| (8)
features:
Thenafter, the rectified first derivative is smoothed as:
{ m+n }
1 ∑
(17)
2
B1 [m] = (B [m − 1] + 2B0 [m] + B0 [m + 1]) (9) B3 [m] = B2 [m] B2 [p]
4 0 p=m−n

A rectified second derivative is then calculated:

Suppappola [55] proposed another digital filter based on
B2 [m] = |A[m + 2] − 2A[m] + A[m − 2]| (10) the multiplication of backward difference (MOBD) [6, 55,
At last, this smoothed rectified first and second derivative 56] which consists of AND-combination (i.e multiplication
are combined together as: operation) of the adjacent derivative values. A MOBD of
Mth order is defined as:
B3 [m] = B1 [m] + B2 [m] (11)
M−1
These linear combinations of derivatives are followed by a ∏
C[m] = (X[m − p] − X[m − p − 1]) (18)
proper threshold criterion for the subsequent QRS detection. p=0
The advantage of this technique is that it involves less com-
putational load. However, the computational load is more than where C[m] represents the QRS features extracted that can
first derivative algorithms. The disadvantage being the noise is be detected by the use of an proper threshold.
not reduced significantly. The length of ECG signal processed Dokur et al. [36] used two different band-pass filters and
and the parameters utilized are fixed. However, the usage of multiplied the outputs X[m] and Y[m] as:
several differentiators (i.e the advantage of each step) for pre-
processing is not justified in literature.
C[m] = X[m] × Y[m] (19)
where C[m] carries the extracted features of the QRS com-
Digital filters plex. This procedure assumes that for each filter the occur-
rence of frequency components within the pass-bands
In the literature, the digital filter techniques are efficiently characterizes each of the QRS wave. Here, the AND-com-
utilized for pre-processing the ECG signals [33–42]. Several bination executes the multiplication operation. Particularly,
algorithms are implemented to realize complex digital filters if the outputs of both the band-pass filters are ’high’ and the
[30, 33, 43–54]. Among them, the most cited are highlighted feature output (i.e AND combination) is ‘true’, then only a
here. QRS event is detected and the maximum amplitude is the
Engelse et al. [33] applied a differentiator initially to pro- R wave location.
cess the input ECG A[m] as In fact, Pan et al. [34] applied a band-pass digital filter
followed by derivative to filter and measure the slope of
B0 [m] = A[m] + A[m − 4] (12) ECG signals. A high-pass filter ( B2 [m] ) is used after a low-
Further, a digital low-pass filter (LPF) is applied to B0 [m] as: pass filter ( Y1 [n] ) to constitute a band-pass filter given as:
B1 [m] = (B0 [m] + 4B0 [m − 1] + 6B0 [m − 2] + 4B0 [m − 3] B1 [m] = 2B1 [m − 1] − B1 [m − 2] + A[m] − 2A[m − 6] + A[m − 12]
+ B0 [m − 4]) (13) (20)
Another technique based on the digital filters has been pro- The band-pass filter output is followed by the first derivative
posed by Okada [30] in which a three-point moving average which is given by:
filter is used to smoothen A[m] as: B2 [m] = 32B1 [m − 16] − (B2 [m − 1] + B1 [m] + B1 [m − 32])
(21)

13
Australasian Physical & Engineering Sciences in Medicine

Differentiation is followed by the band-pass filtered signal

( Y2 [n] ) to emphasize QRS slope, suppress the baseline wan-
der and smoothing ECG signals.
1
B3 [m] = (−B2 [m − 3] − 2B2 [m − 1] + 2B2 [m + 1] + B2 [m + 2])
8
(22)
However, the MOBD algorithm is more suitable for real-
time implementation due to its better trade-off between
computational load and accuracy. The squaring operation
amplifies the smaller differences less than the larger differ-
ences in an exponential fashion. Fig. 2  Block diagram of DWT implementation
The advantage of the digital filter based technique is that it
is able to reduce noise properly. However, its computational
load more than derivative based algorithms. The length of segment and wavelet scale varies in the literature [68–71].
ECG processed and parameters utilized are fixed. However, the selection of mother wavelet depends upon the
similarity to the QRS complex. The ECG signals are divided
Wavelet transform (WT) into 2.4 s and 11 s segments by Ahmed et al. [68] and Xiuyu
et al. [69] respectively. The scales vary from 23 to 24 and 22
The wavelet transform (WT) [57] is a mathematical tool for to 24 , which is used by Szilagyi [70] and Xu et al. [71] to
analyzing non-stationary signal localized in both time and detect the QRS complexes. Moreover, the input ECG signal
scale representation. is re-sampled at 250 Hz by Martinez et al. [72].
The advantage of WT is that it improves the signal quality
+∞
� � by choosing the coefficients of high amplitude. The disad-
a∫
1 t−b
CWT(a,b) =√ x(t) dt, a>0 (23) vantage of this technique is that it involves high computa-
a
−∞ tional load.

The continuous wavelet transform (CWT) provides a vari-

able resolution in both the time and frequency domains for Matched filters
various frequency bands by using a set of analyzing func-
tions which bears an advantage on Fourier transform (FT) The matched filter provides an optimal SNR and more essen-
and short-time Fourier transform (STFT). tially, a symmetrical output pulse waveform. Digital filter-
However, the CWT is more redundant than the discrete ing is used prior to the use of matched filters [73, 74]. The
wavelet transform (DWT) that can be deduced by discretiz- matched filter output for the filter impulse response of length
ing the scale and translation parameters. It is usually imple- M = 128 is computed as
mented using the high-pass and low-pass filters as shown in
Fig. 2. This choice of scale ( a = 2k ) and translation parameters 127
∑
( b = m(2k ) ) leads to the dyadic WT (DyWT) as: q(t) = pj × r(t − j) (27)
j=0
∞

∫
k
Wf (2 , b) = f (t)Ψ ∗2k ,b (t) dt (24) where q(t) is an output sample and r(t − j) are input sam-
−∞
ples of the matched filter. For every patient, the filter coef-
ficients pj are selected to optimize the matched filter impulse
( ) response. The filter output coefficients are chosen in such
1 t−b
Ψ2k ,b (t) = Ψ (25) a way that resembles to the bandpass-filtered QRS com-
2k∕2 2k plex. Further, the dc component of sampled QRS complex
( ) is removed and windowed and normalized to have a gain
1 t
Ψ2k ,b (t) = Ψ − m (26) of one for the matched filter used as an impulse response.
2k∕2 2k Basically, the matched filter impulse response is the time-
where j and n are integers. The DyWT is implemented reversed version of a template QRS complex. In matched
using a dyadic filter bank, in which the filter coefficients filters, the length of the template processed is fixed; while
are obtained from the mother wavelet function employed the type of filter utilized and length of the template is deter-
for analysis of non-stationary signals [58–66] like an ECG. mined empirically. However, its efficient implementation
The choice of the mother wavelet (like Haar, Daubeche- is available in [75]. The disadvantage of this technique is
sis, Mexican hat and many more), length of processed ECG

13
 Australasian Physical & Engineering Sciences in Medicine

that the analysis of ECG requires more complexity than the Afonso et al. [78] introduced finite impulse response (FIR)
derivative based algorithms. filters having fixed length i.e. 32. The filters are employed
to decompose the noisy input ECG into eight constant sub-
band frequencies. A sub-band signal within the range of
Filter banks (FB) 0–12.5 Hz is not changed, while in the range of (12.5–25 Hz)
the sub-band signal is removed outside the region of QRS
A FB typically contains a set of analysis filters. It decom- wave. The high-frequency components outside the QRS
poses the signal bandwidth into sub-band signals having uni- region are considered as noise. The sub-band signal within
form frequency bands of constant length. These sub-bands the rest of six bands of range (25–100 Hz) is considered as
provide information for processing the input signal in both zero. The main challenge is the selection of combination of
the time and frequency domain from different frequency optimal filter banks to highlight the QRS wave.
ranges [76]. The analysis filters Hj (z) bandpass the input
ECG signal A(z) [76] to produce the subband signals vj (z) as:
Hilbert transform (HT)
vj (z) = Hj (z)A(z) (28)
where j = 0, 1, … , N − 1 . The effective bandwidth of sub- Zhou et al. [79] and Nygards et al. [80] used Hilbert trans-
band vj (z) can be downsampled to decrease the total rate form (HT) for QRS detection. In the time domain, the HT
which is 𝜋∕N  . One sample is kept out from the N samples of the input signal A is:
by utilizing this downsampling process N ↓ . Hence, down-
+∞
sampled signal dj (z) is given by:
𝜋 ∫ t−𝜏
1 A𝜏
AH (t) = H{A} = d𝜏 (32)
N−1
1 ∑ −∞
dj (z) = v (z1∕N X p ) (29)
N p=0 j
In the frequency domain, the input signal can be transformed
with a filter of response:
where X = e−k(2𝜋∕N) . The sub-band vj (z) has a higher sam-
pling rate than dj (z) . The process of filtering is done using AH (j𝜔) = A(j𝜔) ⊗ H(j𝜔) (33)
downsampling at 1 / N of the input rate. This technique where ⊗ denotes the convolution operator and the transfer
reduces the computational load of filter bank algorithms function of the Hilbert transform H(j𝜔) is given by:
−k, 0 ≤ 𝜔 < 𝜋
[76] and referred as polyphase implementation. The sub- {

+k, − 𝜋 ≤ 𝜔 < 0
bands of interest are combined to form a variety features that H(j𝜔) = (34)
represent the QRS complexes [76]. For example, a sum-of-
absolute values feature can be computed using sub-bands,
The use of fast Fourier transform (FFT) reduces the compu-
j = 1, … , 4 in the range of [5.6, 22.5] Hz. Six features ( a1 ,
tational load of Hilbert transform. The HT i.e. AH [m] of the
a2 , a3 , a4 , a5 , and a6 ) are derived from these sub-bands as:
ECG signal A[m] is used to compute the signal envelope [80]
3
∑ 4
∑ j=4
∑
for band-limited signals which is given by:
| | | | | |
a1 [m] = |dj (z)|, a2 [m] = |dj (z)|, a3 [m] = |dj (z)| √
| | | | | |
j=1 j=1 j=2 Be [m] ≈ A2 [n] + A2H [m] (35)
(30)
3 4 j=4
Further, the envelope [80] is approximated which involves
∑ ∑ ∑
a4 [m] = 2
(dj (z)) , a5 [m] = 2
(dj (z)) , a6 [m] = 2
(dj (z)) less computational load as:
j=1 j=1 j=2
(31) Be [m] ≈ |A[m]| + ||AH [m]|| (36)
These features contain a range of values being proportional Then after, the envelope is low-pass filtered [80] to eliminate
to QRS wave energy. Ultimately, heuristic beat-detection the ripples and to avoid uncertainty in peak detection. More-
logic [76] is utilized to incorporate some above features rep- over, a waveform adaptive scheme is proposed to remove
resenting the QRS wave. ECG components of low frequencies. Zhou et al. [79] pro-
This technique significantly increases the SNR, which posed a method to approximate the envelope of input signal
can be considered as an advantage. The computational load based on HT given as:
of filter banks depends on four parameters, i.e. the filter
Be [m] ≈ ||B1 [m]|| + ||B2 [m]|| (37)
length, transition-band width, number of sub-bands and the
stop-band attenuation having fixed values and are deter- where B1 [m] and B2 [m] are orthogonal filter outputs given as:
mined experimentally [77]. It involves high computational B1 [m] =A[m]| − A[m − 6], (38)
load which is more than the derivative based techniques.

13
Australasian Physical & Engineering Sciences in Medicine

B2 [m] =A[m]| − A[m − 2] − A[m − 6] − A[m − 8] (39) from step 1 on the residual gp [m] , p = p + 1 . The two
Further, the noise is removed from the envelope signal conditions must be satisfied for an IMF: (a) The mean
Be [m] by using a four-tap moving average filter. A few works value of the envelopes defined by maxima and minima
[81–83] have reported the use of a first derivative before should be zero at every point. (b) The difference between
applying the HT. The ECG is differentiated which modifies number of zero crossings and number of extrema should
the phase and creates zero crossings the R-peak location. be zero or one.
Hence, it requires a transformation which rectifies the phase
to mark the true R-peak location in a signal. In [84], the The length of ECG signals processed is fixed which generates
output of Hilbert transform is followed by adaptive Fourier the IMF’s i.e. the number of IMF’s is proportional to the length
decomposition for enhancing the QRS complex in the ECG of ECG. The selection of number IMF’s is selected empiri-
signal. cally. An ensemble empirical mode decomposition (EEMD),
This technique involves high computational load and does an advanced EMD is also used to pre-process the ECG signal.
not able to reduce noise by itself. The use of FFT for the cal- This technique involves high computational load and reduces
culation of HT makes the envelope independent of the frame noise significantly.
width. During experiments moving average and digital filters
are utilized whose selection is done empirically while the Mathematical morphology
length of ECG signal processed are constant.
Chu et al. [89] proposed an enhancement technique, namely
mathematical morphology for removing the noise associ-
Empirical mode decomposition (EMD) ated with the ECG signal and latter used by Trahanias et al.
[90] for QRS detection. It depends on the idea of dilation
EMD technique is widely used for nonlinear and non-sta- and erosion. Assume that u : U → K and p ∶ P → K repre-
tionary signal analysis [85]. It decomposes a signal into a sent discrete functions, where U and P sets are denoted by
sum of intrinsic mode functions (IMFs). The EMD process U = 0, 1, … , M − 1 and P = 0, 1, … , N − 1 . K represents a
can also be utilized for adaptive filtering. As such, the com- set of integers here. Erosion of a function u [89] can be defined
bination of number of the IMFs obtained after decomposing in terms of function p as:
the ECG signal generates more prominent QRS wave. The
(u ⊖ p)[n] = min (u[n + m] − p[m]) (41)
EMD can be explained by sifting process. J modes wp[m] m=0,…, N−1
and a residual term g[m] [86–88] are obtained and given by:
where p refers to a structuring element also, and defined as
J
∑ n = 0, … , M − N  . The values of u are always smaller than
A[n] = wp + g[m] (40) function ( u ⊖ p ). Dilation of a function u [89] is defined in
p=1 terms of function p as:

The various steps involved in EMD algorithm are as follows: (u ⊕ p)[n] = minm=0,…, N−1 (u[m] − p[n − m]) (42)
where in this case n = N − 1, N, … , M − 1 . Values of u
1. Given a signal wp=1 [m] = r[m] ; with the sifting are always less than function ( u ⊕ p ). Additional steps are
rk [m] = wp [m] , k = 0. performed by combining the dilation and erosion opera-
2. Detect all extrema of input rk [m]. tions. Closing, (indicated as ∙ ) is defined as dilation after
3. Calculate the lower and upper envelopes from the max- erosion operation while Opening (indicated as ◦ ) is defined
ima and minima by using cubic spline interpolation. as erosion after dilation operations. These operators exploits
4. Compute the mean of upper and lower envelopes, the input signals, comparatively in such a way that for a
n[m] = 12 (EnvMax[m] + EnvMin[m]). sequence u, opening eliminates the peaks while closing
5. Extract the detail rk+1 [m] = rk [m] − n[m]. eliminates the negative peaks with the structuring element
6. If rk+1 [m] is an IMF, go to step 7; otherwise, iterate steps p. Chu and Delp [89] used these opening and closing opera-
2–5 on the signal rk+1 [m] , k = k + 1 . (An IMF satisfies tions [90] to suppress noise given as:
two conditions i.e (a) the number of the extrema equals
the number of zeros and (b) the upper and lower enve- [(r◦p) ∙ p] + [(r ∙ p)◦p]
r̃ = (43)
lopes should have the same absolute value.) 2
7. Extract the mode wp [m] = rk+1 [m]. where p is the structuring element. The features are gener-
8. Calculate the residual gp [m] = r[m] − wp [m]. ated for the QRS wave as
9. The extraction is finished g[m] = gp [m] if gp [m] has less ( )
than two extrema, otherwise, the algorithm is iterated [(̃r◦p) ∙ p] + [(̃r ∙ p)◦p]
d = r̃ − (44)
2

13
 Australasian Physical & Engineering Sciences in Medicine

Zhang et al. [91] utilized the first derivative (Okada’s equa- as the frequency spectrum overlaps the QRS wave. As such,
tion [30]) after multi-scale mathematical morphology filter- pre-processing algorithms based on filtering based algo-
ing to remove base-line drifts and artifacts associated with rithms are sensitive to high-frequency noise but insensitive
the A[m]. to baseline changes. Further, the amplitude and derivative
During experiments, the length of ECG segments pro- based techniques involve less computational load in their
cessed are fixed and equal [25, 26, 31–33, 92]. The fixed implementation, despite being noise sensitive. However,
length of the structuring element is used for the analysis of the various stages involved in the amplitude and derivative
A[m] i.e 3. This structuring element length is determined based algorithms are not justified for pre-processing the
empirically and shorter than the multiplication of sampling ECG signal for their validation on the MIT-BIH arrhyth-
frequency and the length of A[m] [93]. The advantage of the mia database. As such, the parameters of these techniques
multiplication operations used in literature [25, 26, 31–33, employed are purely data dependent and may yield varying
92] is not discussed. The use of low-pass filter along with results, if analyzed on a different database (or on data of dif-
this approach increases the SNR significantly. ferent patients). A brief comparison of these pre-processing
techniques is summarized in Table 2 in terms of computa-
tional load and noise sensitivity. From Table 2, it is con-
Sparsity filtering cluded that the amplitude and derivative based techniques
should be developed properly for pre-processing the ECG
The sparse representation (SR) model for a time-domain signals. Once A[m] is filtered, it is passed through the detec-
input signal a ∈ ℜn can be approximated as a ≈ D𝛼 . Here, tion stage for reliable QRS detection.
D ∈ Ren×m is a dictionary matrix 𝛼 ∈ ℜm that provides coef-
ficients representing the input signal. In SR, the input signal
is approximated as a weighted sum of each columns of the QRS detection techniques
dictionary matrix D known as atoms and their weights (as
given by the coefficients in 𝛼 ). Generally, the dictionary D The filtered ECG signal is passed through the QRS detec-
is redundant, where the number of atoms in the dictionary is tion stage. This section presents a brief description of the
greater than the length of the input signals. The coefficients QRS detection techniques used for the localization of the
𝛼 are sparse i.e. there are only few non-zero weights (coef- R-peak in the input ECG signal. Among several detection
ficients) in 𝛼 . Hence, the SR of the input signal a is approxi- algorithms include the thresholding [50, 117–120], syntactic
mated using only few atoms (with corresponding weights methods [121–123], neural networks [105, 124–126], zero-
that are not equal to zero) from the dictionary matrix D. crossing [127], hidden Markov model [128], matched fil-
In [94], second and third order derivatives of the input ters [129, 130], and singularity techniques [131–133]. These
signal were computed to smoothen the ECG signal. To detection algorithms are also evaluated on the basis of two
reduce the artifacts by solving a convex 𝓁1 optimization parameters, i.e. computational load and robustness to noise
problem where the clean input signal is modeled as a sum of which is summarized in Table 3.
two signals whose second and third-order derivatives (differ-
ences) are sparse respectively. In [95], 𝓁1-sparsity filter with Thresholding
overcomplete hybrid dictionaries is used to emphasize the
QRS complex and suppress the baseline drifts, power-line The thresholds are fixed values that are used to determine
interference and large P/T waves. In [96], the input signal is a boundary above which a R-peak is detected in A[m].
modeled as superposition of atoms which is learned from a The thresholds may be fixed or adaptive depending on the
training set plus additive random noise to remove noise and approach employed. Numerous works have been reported
other artifacts such as baseline wandering. in which the threshold based approach is utilized that is
Among all the pre-processing algorithms discussed in determined experimentally in [28, 31, 33, 34] to detect the
this section, none of them are completely efficient when all R-peak. A peak is defined as a local maximum when the sig-
kinds of noise are considered [27] for analyzing the ECG nal changes its direction within a pre-defined time interval,
signals. The amplitude and slope based techniques have a i.e. to be signal peak, the peak should exceed the threshold.
significant advantage over electromyogram (EMG) noise (i.e The approach is considered as simple while the choice of
muscle noise) and are sensitive to changes in the baseline optimal threshold is quite difficult. If the input ECG signal
of A[m]. However, the performance of these algorithms is contains maximum SNR, then it is possible to utilize the
degraded if they are applied to composite noise. Rather, a lower thresholds. In Pan et al. [34] improved the SNR by
higher performance is reported by the high-pass and cubic using bandpass filter and used the adaptive thresholds. The
spline approaches used to correct the baseline wander. The thresholds are allowed to float over the noise. The two types
filtering of the signal to remove EMG noise is more difficult of thresholds are applied to the R-peak i.e higher and lower

13
 Table 2  A brief comparison of various pre-processing techniques
Technique Algorithm Robustness to noise Computational load

Amplitude AT + FD [25, 26] + second threshold It is not able to remove noise perfectly. Therefore, This algorithm usually contains FD and a thresh-
it does not account for first order derivative old for pre-processing. The segmentation of
techniques for pre-processing the ECG (if done) and the threshold employed
mainly account for the computational load
First derivative (FD) only F D + threshold [28–30] It reduces baseline drifts and motion artifacts but This algorithm contains one equation for pre-
AT + F D [25, 26, 97] + threshold it is unable to eliminate high-frequency noise processing. Most studies used Okadas equation
FD + SD [31, 32] + threshold [91] [30]. The length of ECG signal processed and
F D + digital filtering [33] + threshold its segmentation (if done) for each record mainly
Digital filter + FD [34] + threshold account for the computational load
Math. morph. fil. + FD [98] + threshold
FD + HT [81–83] + threshold
FD + WT [91] + threshold
First and second order derivative FD + SD [31, 32] + threshold It is not able to remove noise perfectly. Therefore, This algorithm usually involves a maximum of
SD + HT [81, 83] + threshold it does not account for second order derivative four equations to pre-process the signal. The
Australasian Physical & Engineering Sciences in Medicine

techniques for pre-processing segmentation of the ECG (if done) and count of
equations mainly account for the computational
load
Digital filter FD + digital filters + threshold [33] The type and order of the filter used mainly This algorithm usually involves a maximum of
BPF + FD + threshold [34] account for the increase in SNR four equations to pre-process the signal. The
BPF + HT + threshold [99] length of ECG signal processed and its segmen-
BPF + FD + WT + threshold [91] tation (if done) for each record mainly account
BPF + matching filter + threshold [100] for the computational load
Filter banks (FB) FB + threshold [76, 101] The signal-to-noise ratio for muscular (more) and The filter banks involve huge number of multipli-
WT + FB + correlation [102] Gaussian noise is improved in relaive to mean ers which results in high computational load rela-
and median averaging techniques tive to other pre-processing techniques [101]
Matched filters Matched filters [103] The algorithm improves the SNR [106] The algorithm is computationally expensive due
Matched filters + thresholds [73, 104] to performed by each moving segment (i.e seg-
NNs (i.e. filter) + matched filter [105] ment- by-segment) with the template across the
ECG signals. However, its efficient implementa-
tions are available [75]
Wavelet transform (WT) WT + threshold [107, 108] It is unable improve the SNR. By choosing the The length of ECG segmented accounts for the
FD + WT + ZC + threshold [109] coefficients of high amplitude, the amount of trade-off between efficiency and time taken for
WT + HT + threshold [69] signal content can be increased [110] the technique [107]. WT is similar to filter banks,
WT + filter banks + correlation [110] and involves large computational load [111]
WT + neural networks [109]
WT + singularity + ZC + threshold [69]
Hilbert transform (HT) FD + HT + threshold [81–83] The SNR is not increased by itself. Rather, a filter This algorithm involves a minimum of nine steps
BPF + HT+ threshold [99] is employed which is followed by HT. A BPF to preprocess the ECG signal i.e. involves a huge
WT + HT + threshold [112] [82] of 8–20 Hz is used to eliminate the muscu- complexity. The use of FFT increases the compu-
lar noise and maximize the QRS wave tational load relative to time domain techniques.
Also, the length of ECG processed account for
increase in complexity. The use of FFT increases
the computational load relative to time domain

13
techniques. Also, the length of ECG processed
accounts for increase in computational load [113]
 Australasian Physical & Engineering Sciences in Medicine

thresholds. The higher thresholds among them are allowed

AT amplitude threshold, FD first derivative, SD second derivative, WT wavelet transform, HT Hilbert transform, BPF bandpass filter, LPF lowpass filter, HPF high pass filter, ZC zero crossing,
tational load is directly proportional to the length

This algorithm involves a high computational load

computational load that is greater than filter and

tions to preprocess the ECG signal. The compu-

This algorithm involves a minimum of nine steps
to first analyze the signal. While the lower threshold is used

The signal noise is partly considered by this algo- This algorithm involves a minimum of 15 equa-
ECG signal processed mainly account for the
to pre-process the ECG signal. The length of

of ECG processed that is greater than digital

when the no QRS complex is detected within a certain time

filter and derivative oriented techniques

which is followed by a search back technique to find the
QRS complexes back in time. The main advantage of this
approach is that it involves less computational load in com-
parison to all the detection techniques utilized. However,
derivative based techniques

this method requires specification and adjustment of numer-

ous parameters for adequate detection accuracy remains a
Computational load

challenge.

Neural networks (NN’s)

Neural networks have the ability to learn patterns in response

to newly input patterns. Those learning and self-organizing
A few starting IMFs mainly account for the QRS

abilities are appropriate for QRS-wave recognition [134],

information and filters the noise relative to dif-
ferent features and thereby increasing the SNR

rithm. Rather, a LPF is employed to increase

because the QRS-wave will change its shape according to

The signal noise is partly considered by this

the patient’s physical condition. Suzuki et al. [135] used

an ART2 (adaptive resonance theory) network employed
in this self-organizing neural-network system to detect the
QRS complex. In this approach, the category of neural net-
works should be selected and modified during analysis. The
architecture of an ART2 newtwork is shown in Fig. 3, where
Robustness to noise

LTM is the long-term memory, F1 and F2 are layers con-

necting the neurons and wi , xi , ui , qi , pi qi are the nodes that
characterize the F1 layer. A neural network with N number
algorithm
the SNR

of inputs is developed, where the sample taken from the

[115]

window is fixed for each input [136]. Garcia-Berdone’s et al.

[136] utilized 20 samples as input, thereby emphasizing that
Sparse + sparse filtering + Gaussian filter [95]

the input for NN’s should be chosen within a range of sam-

ples. In the NN hidden layer, the choice of optimal number
EMD + singularity + threshold [115, 116]

Sparse filtering + learned dictionary [96]

Math. morph. filt. + FD + threshold [98]

of neurons is difficult and determined empirically. A typical

neural network architecture is depicted in Fig. 3.
Math. morph. filt. + threshold [93]

The disadvantage of the technique is that involves high

HPF + EMD + threshold [116]

computational load and is highly noise sensitive. The aver-

age accuracy of this technique is also lesser than the thresh-
Empirical mode decomposition (EMD) EMD + threshold [114]

olding based techniques.

Hidden Markov model (HMM’s)

EMD empirical mode decomposition, FB filter banks
Algorithm

A hidden Markov model (HMM) characterizes an observed

data sequence by a probability density function which var-
ies according to the state of an underlying Markov chain. In
this approach the output function, a number of states and
Mathematical morphology filtering

transition probabilities are determined empirically. The

HMM parameters are fixed and cannot be approximated
from training data by employing maximum likelihood
methods due to the fact that data produced from the state
Table 2  (continued)

sequence remains unknown [137]. The parameters of a hid-

Sparse filtering

den Markov model are not directly estimated when the data
is unknown. A hidden Markov model is depicted in Fig.4,
Technique

where q1 , q2 , q3 , … , q6 are the number of sets of states. The

13
Australasian Physical & Engineering Sciences in Medicine

model consists of two Markov sub-sources, i.e. one for non- fragments length processed are uniform throughout the
QRS segments and one for QRS segments. analysis. Belforte et al. [121] used segment of 30-s. The
The advantage of this technique is that it provides auto- syntactic method [122] utilizes four attributes, i.e. the chord
matic estimation of all the parameters in the decision rule length, arc symmetry, arc length and degree of curvature that
stage from each ECG file undergoing analysis. However, the are computed empirically.
search for parameters involves huge computational load. The The disadvantage of the technique is that involves high
accuracy results show that a simple HMM detector achieves computational load and is highly noise sensitive. However,
accuracy which is very close to adaptive threshold based it yields a comparable accuracy with the rest of the detect-
detection techniques. ing techniques.

Syntactic techniques Singularity methods

The syntactic approach is applied after the digital deriva- Most of the ECG signal information is carried by its irregu-
tive operator [121]. The method utilizes a very simple look- lar morphology and singular points (fiducial points). In
up table for coding. The sequences of energy peaks of the mathematics, a singularity is often considered as the oppo-
derivatives of ECG waveforms corresponding to different site of smoothness and can be measured by Lipschitz expo-
leads are coded into the string of messages. For each lead nent. Using the nth derivative of a so-called smoothing func-
waveform, the strings which correspond to QRS complexes tion, the singular points can be detected by modulus maxima
are considered as a sample of positive information and are of the wavelet. In this approach thresholding is employed
processed by a grammatical inference algorithm. Analo- on individual modulus maxima of WT to reduce the white
gously the strings which correspond to non-QRS complexes noise from ECG signals. The wavelet scales are chosen
are saved and considered as a negative information, sample experimentally to search for singular points [138, 139]. The
to be eventually used in a further generalization step. Con- use of thresholds per ECG fragment is constant [138] and
sequently, two grammars are built [121]; the first one gener- computed empirically.
ates only positive sequences (corresponding to QRS events) The disadvantage of the technique is that involves high
while the second one generates sequences corresponding to computational load due to searching the singular points and
hypotheses that may or may not correspond to a QRS com- is highly noise sensitive. However, it yields an accuracy
plex. This learning algorithm infers linear grammars based (i.e. 99.22% [139]) which is approaching to the threshold-
upon formal derivatives. ing based techniques.
The syntactic method enables the detection of the QRS
wave of an ECG signal by itself [121–123]. The ECG

(a) (b)

Fig. 3  Neural network models for QRS detection

13
 Australasian Physical & Engineering Sciences in Medicine

157–159], increases the correct detection of QRS com-

plexes; however, the adjustment of multiple thresholds cho-
sen empirically is a drawback. Most of the QRS detection
presented in Table 3 perform well on the clean or filtered
signals. Rather, their performance degrades in the noisy
environment or signals containing arrhythmias. Therefore,
these QRS detection techniques lacks in providing a general-
ized solution.
Fig. 4  HMM model for QRS detection [137]

Performance evaluation and discussion

Zero‑crossing (ZC)
The R-peak detection is followed by the performance evalu-
Zero crossing methods are robust against noise and are par- ation of the subsequent algorithms. The performance evalua-
ticularly useful for finite precision arithmetic. This detection tion of various algorithms discussed in the earlier sections is
method inherits the robustness and provides a high degree of estimated on the basis of two statistical parameters i.e. sensi-
detection performance even in very noisy ECG signals. In tivity and positive predictivity. The sensitivity is defined as
this technique, the beginning of an event is identified when the rate of correctly detected events among the total number
the features of the signal (i.e. number of zero crossings per of events detected by the algorithm, while positive predic-
segment) fall below a signal adaptive threshold while the tivity refers to the rate of correctly classified events in all
end is identified when the signal rises above the threshold detected events which can be represented as:
[127, 140]. This beginning and end of the event determine
TP
the boundaries of the search interval for the temporal locali- Sensitivity (Se ) = × 100 (45a)
TP + FN
zation of the R-wave. If adjacent events are temporally very
close (multiple events), they will be combined into one
TP
single event. The beginning of the combined event is the Positive predictivity (Pp ) = × 100 (45b)
TP + FP
beginning of the first event and the end of the combined
event is the end of the last event. The threshold per segment where TP (true positives) is termed as the number of cor-
employed for determining the number of zero crossings is rectly classified events into a particular class, FN (false
fixed [127] and calculated empirically. In literature [127, negatives) refers to events of a particular class which have
141], the search for zero-crossings depend on the choice of not been detected, and FP (false positives) refers to the
wavelet scale. number of events of another class detected in a particular
The disadvantage of the technique is that involves high class. The overall performance of the existing QRS detec-
computational load and is highly noise sensitive. How- tion algorithms reported in the current study have not been
ever, it yields an accuracy (i.e. 99.70%) which is approxi- analyzed relative to computational load and noise sensitiv-
mately same with those achieved by the thresholding based ity. Further, a standard database is not used for testing these
techniques. QRS detection algorithms which makes the analysis difficult
The different algorithms involved in the detection of QRS to evaluate and compare i.e. some of works utilized different
complexes are summarized in Table 3. Among all the algo- databases or signals from patients demanding the develop-
rithms presented, the thresholding approach involve low ment of an efficient QRS detector algorithm. An algorithm
computational load. Since, the study aims to highlight the or technique can be termed as efficient, if it satisfies the
development of efficient algorithms for a robust and reli- following factors such as low computational load, evaluated
able QRS detector, the use of threshold based technique is on common standards of data and high accuracy. As such,
suggested because of its simplicity and efficiency. These the QRS detection algorithms reporting high classification
thresholds are used in time [25, 26, 153] and time-frequency performance in terms of accuracy, computational load along
domain both [154–156]. There are two types of thresholds with the other factors responsible are summarized in Table 4
used to detect the QRS complexes which include the fixed and discussed subsequently to develop a fast and robust QRS
and adaptive thresholds. The use of fixed threshold is simple detector.
and efficient for stationary input signals only having simi- A high overall performance is reported by Li et al. [160]
lar morphologies. In fact, movement of patients, baseline (records 214 and 215 are excluded) achieving a sensitivity
drift and variation in morphology of ECG signals results in of 99.89% and specificity of 99.94% respectively, evalu-
highly inaccurate detection of QRS complexes using fixed ated on the benchmark MIT-BIH arrhythmia database
thresholds. Rather, the usage of adaptive thresholding [71, [22]. The features of the different waves are extracted using

13
 Table 3  A brief comparison of QRS detection techniques described in literature
Technique Algorithm Sensitivity to noise Computational burden

Threshold It is employed in the last step in most of the R-peak The performance of this technique is degraded by This technique is the simplest among all R-peak
detection algorithms [25, 26, 28–32, 81–83, 91, noise as in the case of low SNR signals [27, 34] detection techniques. It is usually implemented by
98]. Level crossing + threshold [142]. MA filter IF and ELSE conditions and involves less compu-
+ threshold [48, 143] pulse triggered + threshold tational load [28, 31, 33, 34]
[144]. Wavelet + threshold [145, 146]. Filtering
+ threshold [47, 48]
Neural networks (NNs) WT + NNs [109] NNs are extremely noise sensitive [148]. Its perfor- The training phase adjusts the weights in NNs and
WT+ HMM [147] mance gets degraded without proper training data hence it is an iterative method [149] which makes
NNs + matched filter [105] and architecture designed it computationally expensive. The complexity also
depends on the hidden neurons, transfer func-
tion which increases the computational load of
training. Its hardware implementation requires
Australasian Physical & Engineering Sciences in Medicine

sufficient memory for training

Hidden Markov models (HMM) BPF + HMM [128, 137] The technique is sensitive in relative to varying For training a manual segmentation before the
Wavelet + HMM [147] heart rate, noise and baseline wander [150] evaluation of a record, its patient dependence and
even when the computationally efficient Viterbi
algorithm, [151] is applied makes computationally
inefficient. The HMM parameters are fixed, gener-
ally from 15 to 50 which needs to be analyzed
[128, 137]
Syntactic method The syntactic method only [121–123] This technique is noise sensitive [123] It involves high computational load in relative to
other techniques. Several parameters are meas-
ured. Moreover, efficient grammars capable of
describing syntax and semantics are required for
designing the pattern grammar model
Singularity EMD filtering + threshold [115] This technique is noise sensitive [138] The approach is more complex and computationally
WT + singularity + ZC + threshold [69] less efficient due to the time consumed in optimiz-
ing and searching stages for detecting the QRS
complex in ECG segments [69, 138]
Zero crossing (ZC) BPF + ZC [127] This technique is noise sensitive [127] The approach is computationally less efficient
WT + ZC + threshold [127] because of time consumed in searching for the
WT + singularity + ZC + threshold [69] minimum/maximum temporal localization of QRS
complex [127]
Genetic algorithms (GAs) Polynomial filter + GA [152] This technique is noise sensitive [152] The approach is computationally less efficient
because of time consumed in searching adequate
nymber of samples to be processed. It requires
1000–2000 times the speed needed for real-time
processing [152]

BPF bandpass filter, HMM hidden Markov model, MA moving average, WT wavelet transform, ZC zero crossing, GA genetic algorithm

13
 Australasian Physical & Engineering Sciences in Medicine

wavelet-based approach and singularity technique for clas- However, the processing speed of the algorithms depends on
sification of these features. However, this technique involves the operating frequency of these hardware platforms. It is to
high computational load and hence cannot be considered note that the higher is the processing speed of the hardware
superior in terms of performance. Moreover, the experi- platforms, faster is the processing and vice-versa. Some of
ments are performed by excluding some of the records from the works reported usage of mobile phones [24] to evalu-
the MIT-BIH arrhythmia database [22] to reduce the noise ate the performance of the three QRS detection techniques.
in the processed ECG signals and reported an improved per- Here, the QRS pre-processing stage consists of amplitude,
formance in the detection of QRS complexes. While several first and second order derivative algorithms, whereas the
investigators also performed their experiments by excluding detection stage consists of a thresholds only. The simplicity
paced beats [139] and ventricular flutter beats [72] from the of the combination of these methodologies can be confirmed
patient’s data. Rather, such kind of evaluation of algorithms from Table 4 in terms of computational load. It is concluded
based on the variability in the utilization of data cannot be from Table 4, the combination of first derivative with thresh-
justified. Thus, a reliable algorithm is needed for the analy- old can be considered as efficient in terms of computational
sis of ECG signals yielding better overall performance on load for detecting the QRS complexes.
the overall dataset (i.e. without excluding any fragments of While processing the ECG signals, the consumption
ECG). of power [166] will be a limitation in battery operated
In Table 4, each of the QRS detection algorithm is cat- devices. The case of classical Pan–Tompkins technique
egorized as low, medium or high in terms of its compu- [34] is an example which shows a significant power util-
tational load. The computational load of the algorithm is ity [167], though it uses first order derivative. The total
determined by computing the total number of operations computational load of Pan–Tompkins algorithm is O(mkn)
involved (in terms of addition, multiplication and differ- where ‘n’ is the number of stages through which the ECG
entiation) and the number of iterations. The algorithms signal is passed, ‘k’ is the order of the individual filters
with more number of operations (i.e., higher computa- (in this case it is 1) and m is the length of the ECG sig-
tional load) is categorized as high while the algorithms nal. When n and k are very small compared to m, the total
with lesser number of operations is termed as low. The complexity would be O(m). Due to more stages involved
algorithms having low computational load are faster and in the Pan–Tompkins algorithm, more power is required
vice-versa. Therefore, faster algorithm is more suitable in the detection of QRS complexes. In this study, the
for hardware implementation and can be used in real-time standard Pan–Tompkins algorithm is suggested as a ready
monitoring of ECG signals. Table 4 shows that the Christov made solution that can be implemented on suitable hard-
[157], Chiarugi et al. [162] and Elgendi [166] algorithms ware platforms to develop an efficient QRS detector. The
involve low computational load. In the preprocessing stage, experiments are validated on the benchmark of the MIT-
the application of first order derivative is promising, par- BIH arrhythmia database and performed using the MAT-
ticularly if it is followed by a suitable detection stage [167] LAB software package with hardware configuration of
such as dynamic and/or moving average threshold. The Intel CoreTM i5-processor CPU 3.30 GHz and 4.00 GB of
computational load of a first order derivative is O(m) i.e. RAM. The different stages involved in the Pan–Tompkins
for m length of ECG data, O(m) number of operations are algorithm is depicted in Fig. 5.
required, i.e. of linear order. Similarly, the computational The complete analysis of the QRS detection algorithms
load for second order derivative require additional O(m) depending on the factors such as noise sensitivity, com-
operations. In fact, the sole application of first order deriva- putational load and accuracy is presented in this study
tive in the preprocessing stage is noise sensitive, and hence, prior to their implementation. The algorithms employed
it must be followed by an efficient detection scheme [27]. in real-time analysis should be simple (in terms of compu-
However, the implementation of the first and second order tational load) without resulting in degraded performance
derivative schemes used for preprocessing the signal is i.e., accuracy. If the algorithm is simple, the processing
slower than the amplitude based schemes. Rather, a faster of larger databases is faster and requires less hardware
(or simple) technique cannot be considered as efficient for leading to low-power consumption and reduced cost. It
QRS detection. is also suggested to process the input data at higher oper-
Prior to the development of a fast and robust QRS detec- ating frequencies as it can be helpful to process larger
tor, these efficient algorithms are evaluated on the perfor- databases within the less amount of time. From Table 4, it
mance parameters such as noise sensitivity, computational can be concluded that the combination of first derivative
load and accuracy as mentioned in each of the earlier sec- and threshold are efficient if developed properly. Moreo-
tions. In addition, these efficient algorithms are required to ver, the Pan–Tompkins can be considered as a complete
be implemented on the suitable hardware platforms such as ready-made solution in the efficient detection of QRS com-
microcontrollers or field programmable gate arrays (FPGA). plexes which satisfies all the factors like noise sensitivity,

13
Australasian Physical & Engineering Sciences in Medicine

Table 4  Overall performance analysis of QRS detection algorithms

Previous works Year Detection technique Preprocessing technique Computational load +P Se

Pan and Tompkins [34] 1985 Thresholds (> 2) BPF + FD + squaring + MA Medium 99.56 99.76
Li et al. [160] 1995 Singularity + thresholds (> 2) WT + digital filter High 99.84 98.89
Afonso et al. [161] 1996 Thresholds (> 2) Filter banks High 99.56 99.59
Moraes et al. [139] 2002 Threshold ( = 1) LPF + FD + spatial velocity Medium 99.88 99.69
Martinez et al. [72] 2004 Thresholds (> 2) + ZC WT Medium 99.86 99.8
Chiarugi et al. [162] 2007 Thresholds (> 2) BPF + first derivative Low 99.81 99.76
Ghaffari et al. [154] 2008 Threshold ( = 1) Hybrid complex WT High 99.89 99.79
Zheng and Wu [155] 2008 Threshold ( = 1) DWT + cubic spline + interpolation High 99.59 98.68
+ MA
Ghaffari et al. [154] 2008 Threshold ( = 1) Complex frequency B-spline WT High 99.89 99.29
Arzeno et al. [81] 2008 Thresholds ( = 2) First derivative + Hilbert transform Medium 99.24 99.29
Arzeno et al. [81] 2008 Thresholds (> 2) F.D + squaring + BPF Medium 99.58 99.57
Arzeno et al. [81] 2008 Thresholds variation SD + squaring + BPF Medium 99.58 99.57
Chouhan and Mehta [163] 2008 Threshold ( = 1) Digital filters Medium 99.49 99.55
Benitez et al. [82] 2000 Threshold ( = 1) FD + HT Medium 99.31 99.13
Ghaffari et al. [154] 2008 Threshold ( = 1) WT Medium 99.72 99.91
Elgendi et al. [158] 2008 Thresholds (> 2) Digital filters Medium 99.9 97.5
Ghaffari et al. [154] 2008 Threshold WT (complex morlet) Medium 99.29 99.49
Christov [157] 2004 Thresholds (> 2) Multiple MA + FD Low 99.81 99.76
Choukari et al. [164] 2011 Thresholds ( = 2) WT + histogram + MA High 97.24 98.68
Zidelmal et al. [165] 2012 Thresholds ( = 2) WT + multiplying coefficients Medium 99.82 99.64
Elgendi [166] 2013 Thresholds ( = 2 M.A) BPF + FD + squaring Low 99.87 99.78

BPF bandpass filter, FD first derivative, LPF low pass filter, MA moving average, HT Hilbert transform, WT wavelet transform, DWT discrete
wavelet transform, SD second derivative

computational load and accuracy which is evaluated on Conclusion

all the records (without excluding of any segment) of the
MIT-BIH arrhythmia database. An efficient QRS detector This article presents a brief study of QRS complex detec-
can be integrated with the feature extraction and the clas- tion algorithms based on the literature, to figure out the
sification algorithms for arrhythmia classification [12, 168, best-suited algorithm for cardiac analysis based on the fac-
169]. Moreover, a fast and robust detector can easily be tors like robustness to noise, computational load and sensi-
employed for breathing disorders and various other cardiac tivity. For pre-processing the filtered ECG, the first-order
disorders to enhance the lifestyle of patients for CVDs. derivative is suggested because it involves less computa-
Despite of the several algorithms reported in literature tional load with high accuracy. However, this approach
their clinical utility is not discussed. It is however difficult is noise sensitive and therefore, the approach should be
for an algorithm that mentions its significance and util- followed by a suitable detection algorithm such as adap-
ity from a clinical point of view. It is best of the author’s tive thresholding. Both these techniques can be developed
knowledge that none of the discussed algorithms are firmly for detecting the QRS waves because the combina-
implemented and verified in a clinical environment or hos- tion involves less computational load and achieves higher
pitals. Hence, it is also suggested that the new algorithms accuracy suitable real-time applications. However, the
developed for robust and reliable QRS detector based on classical Pan–Tompkins approach is also a good ready-
the factors (i.e. as mentioned in the aim of the study) must made alternative which is employed in most of the works
be implemented and verified in the clinical environment. in arrhythmia classification and implemented in this study.
The developed QRS detectors based on the suggested algo-
rithms can be helpful for detecting several cardiac disor-
ders to lead a healthy and secure lifestyle.

13
 Australasian Physical & Engineering Sciences in Medicine

Acknowledgements  The authors acknowledge the Department of Sci-

ence and Technology, Government of India for sponsoring this research
work (IF 120841) under DST-INSPIRE Fellowship Scheme.

Compliance with ethical standards 

Conflicts of interest  The authors have no conflicts of interest to de-

clare.

Research involving human and animal participants  This article does

not contain any studies with human participants or animals performed
by any of the authors.

References
1. Alwan A (2014) Global status report on noncommunicable dis-
eases 2014. World Health Organization, Geneva
2. Benjamin EJ (2018) Heart disease and stroke statistics update:
a report from the American Heart Association. AHA statistical
update. Circulation 137:e67–e492
3. Goss J (2008) Projection of Australian health care expenditure by
disease, 2003 to 2033. Australian Institute of Health and Welfare,
Canberra
4. Dilaveris P, Gialafos E, Sideris S, Theopistou A, Andrikopoulos
GK (1998) Simple electrocardiographic markers for the pre-
diction of paroxysmal idiopathic atrial fibrillation. Am Heart J
135(5):733–738
5. Jenkins D (2009) A (not so) brief history of electrocardiography.
http://www.ecgli​brary​.com/ecghi​st.html
6. Benali R, Reguig FB, Slimane ZH (2012) Automatic classifi-
cation of heartbeats using wavelet neural network. J Med Syst
36(2):883–892
7. Tsipouras M, Fotiadis D, Sideris D (2002) Arrhythmia classi-
fication using the RR-interval duration signal. In: Computer in
cardiology, pp 485–488
8. Thong T, McNames J, Aboy M, Goldstein B (2004) Prediction
of paroxysmal atrial fibrillation by analysis of atrial premature
complexes. IEEE Trans Biomed Eng 51(4):561–569
9. Bashour C, Visinescu M, Gopakumaran B, Wazni O, Carangio F
(2004) Characterization of premature atrial contraction activity
prior to the onset of postoperative atrial fibrillation in cardiac
surgery patients. Chest 126(4):831S
10. de Chazal P, O’Dwyer M, Reilly R (2004) Automatic classifica-
tion of heartbeats using ecg morphology and heartbeat interval
features. IEEE Trans Biomed Eng 51(7):1196–1206
11. Krasteva V, Jekova I, Christov I (2004) Automatic detection of
premature atrial contractions in the electrocardiogram. Electro-
tech Electron E E 3(2):132–137
12. Raj S, Chand GP, Ray KC (2015) Arm-based arrhythmia beat
monitoring system. Microprocess Microsyst 39(7):504–511
13. Raj S, Ray KC (2017) ECG signal analysis using DCT-based
DOST and PSO optimized SVM. IEEE Trans Instrum Meas
66(3):470–478
14. Raj S, Ray KC, Shankar O (2016) Cardiac arrhythmia beat clas-
sification using DOST and PSO tuned SVM. Comput Methods
Programs Biomed 136:163–177
15. Raj S, Ray KC (2018) Sparse representation of ECG signals for
Fig. 5  Plot of R-peak detection within B[m] using Pan–Tompkins automated recognition of cardiac arrhythmias. Expert Syst Appl
algorithm [34] 105:49–64
16. Zapanta L, Poon CS, White D, Marcus C, Katz E (2004) Heart
rate chaos in obstructive sleep apnea in children. In: 26th Annual

13
Australasian Physical & Engineering Sciences in Medicine

international conference of the IEEE Engineering in Medicine 38. Fancott T, Wong DH (1980) A minicomputer system for direct
and Biology Society (IEMBS), vol 2, pp 3889–3892 high speed analysis of cardiac arrhythmia in 24 h ambulatory
17. Shouldice R, Brien LO, Brien CO, de Chazal P, Gozal D (2004) ECG tape recordings. IEEE Trans Biomed Eng 27(12):685–693
Detection of obstructive sleep apnea in pediatric subjects using 39. Keselbrener L, Keselbrener M, Akselrod S (1997) Nonlinear high
surface lead electrocardiogram features. Sleep 27(4):784–792 pass filter for R-wave detection in ECG signal. Med Eng Phys
18. Scholz U, Bianchi A, Cerutti S, Kubicki S (1997) Vegetative 19(5):481–484
background of sleep: spectral analysis of the heart rate variabil- 40. Leski J, Tkacz E (1992) A new parallel concept for QRS com-
ity. Physiol Behav 62(5):1037–1043 plex detector. In: 14th Annual international conference of the
19. Trinder J, Kleiman J, Carrington M, Smith S, Breen S (2001) IEEE Engineering in Medicine and Biology Society, vol 2, pp
Autonomic activity during human sleep as a function of time and 555–556
sleep stage. J Sleep Res 10(4):253–264 41. Nygards M, Hulting J (1979) An automated system for ECG
20. Jeppsen J, Beniczky S, Fuglsang FA, Sidenius P, Johansen P monitoring. Comput Biomed Res 12(2):181–202
(2017) Modified automatic R-peak detection algorithm for 42. Srnmo L, Pahlm O, Nygards M (1982) Adaptive QRS detection
patients with epilepsy using a portable electrocardiogram in ambulatory ECG monitoring: a study of performance. In: Pro-
recorder. In: 39th Annual international conference of the IEEE ceedings of IEEE computers in cardiology
Engineering in Medicine and Biology Society (EMBC), pp 43. Christov I, Dotsinsky I, Daskalov I (1992) High-pass filtering
4082-4085 of ECG signals using QRS elimination. Med Biol Eng Comput
21. Song J, Yan H, Xu Z, Yu X, Zhu R (2011) Myocardial ischemia 30(2):253–256
analysis based on electrocardiogram QRS complex. Australas 44. Laguna P, Thakor N, Caminal P, Jane R (1990) Low-pass differ-
Phys Eng Sci Med 34(4):515–521 entiators for biological signals with known spectra: application to
22. Moody G, Mark R (2001) The impact of the MIT-BIH arrhyth- ECG signal processing. IEEE Trans Biomed Eng 37(4):420–425
mia database. IEEE Eng Med Biol Mag 20(3):45–50 45. Thakor N, Webster J, Tompkins WJ (1984) Estimation of QRS
23. Luo S, Johnston P (2010) A review of electrocardiogram filtering. complex power spectra for design of a QRS filter. IEEE Trans
J Electrocardiol 43(6):486–496 Biomed Eng 31(11):702–706
24. Sufi F, Fang Q, Cosic I (2007) ECG R-R peak detection on 46. Thakor N, Zhu Y-S (1991) Applications of adaptive filtering to
mobile phones. In: 29th Annual international conference of the ECG analysis: noise cancellation and arrhythmia detection. IEEE
IEEE Engineering in Medicine and Biology Society (EMBS), pp Trans Biomed Eng 38(8):785–794
3697–3700 47. Bayasi N, Saleh H, Mohammad B, Ismail M (2014) 65-nm ASIC
25. Morizet-Mahoudeaux P, Moreau C, Moreau D, Quarante JJ implementation of QRS detector based on Pan and Tompkins
(1981) Simple microprocessor based system for on-line ECG algorithm. In: 10th International conference on innovations in
arrhythmia analysis. Med Biol Eng Comput 19(4):l497–500 information technology (INNOVATIONS), pp 84–87
26. Fraden J, Neuman M (1980) QRS wave detection. Med Biol Eng 48. Deepu C, Lian Y (2015) A joint QRS detection and data com-
Comput 18(2):125–132 pression scheme for wearable sensors. IEEE Trans Biomed Eng
27. Friesen G, Jannett T, Jadallah M, Yates S, Quint S, Nagle H 62(1):165–175
(1990) A comparison of the noise sensitivity of nine QRS detec- 49. Chen H, Chen S (2003) A moving average based filtering system
tion algorithms. IEEE Trans Biomed Eng 37(1):85–98 with its application to real-time QRS detection. In: Computers in
28. Menrad A (1981) Dual microprocessor system for cardiovascular cardiology, pp 585–588
data acquisition, processing and recording. In: Proceedings of 50. Phukpattaranont P (2015) QRS detection algorithm based on the
the 1981 IEEE international conference on industrial electronics, quadratic filter. Expert Syst Appl 42(11):4867–4877
control and instrumentation, pp 64–69 51. Kumar S, Mohan N, Prabaharan P, Soman KP (2016) Total vari-
29. Holsinger WP, Kempner KM, Miller MH (1971) A QRS preproc- ation denoising based approach for R-peak detection in ECG
essor based on digital differentiation. IEEE Trans Biomed Eng signals. In: 6th International conference on advances in comput-
18(3):212–217 ing & communications (ICACC), Cochin, India, pp 697–705
30. Okada M (1979) A digital filter for the QRS complex detection. 52. Gliner V, Behar J, Yaniv Y (2018) Novel method to efficiently
IEEE Trans Biomed Eng 26(12):700–703 create an mHealth App: implementation of a real-time electro-
31. Balda R, Diller G, Deardorff E, Doue J, Hsieh P (1977) The HP cardiogram R peak detector. JMIR mHealth uHealth 6(5):e118
ECG analysis program. In: van Bemmel JH, Willems JL (eds) 53. Jain S, Ahirwal MK, Kumar A, Bajaj V, Singh GK (2017) QRS
Trends in computer-processed electrocardiograms. North-Hol- detection using adaptive filters: a comparative study. ISA Trans
land Publishing, Amsterdam, pp 197–205 66:362–375
32. Ahlstrom ML, Tompkins WJ (1983) Automated high-speed anal- 54. Sharma T, Sharma KK (2017) QRS complex detection in ECG
ysis of Holter tapes with microcomputers. IEEE Trans Biomed signals using locally adaptive weighted total variation denoising.
Eng 30(10):651–657 Comput Biol Med 87:187–199
33. Engelse W, Zeelenberg C (1979) A single scan algorithm for 55. Suppappola S, Ying S (1994) Nonlinear transforms of ECG
QRS-detection and feature extraction. In: IEEE proceedings of signals for digital QRS detection: a quantitative analysis. IEEE
computers in cardiology Trans Biomed Eng 41(4):397–400
34. Pan J, Tompkins WJ (1985) A real-time QRS detection algo- 56. Sun Y, Suppappola S, Wrublewski T (1992) Microcontroller-
rithm. IEEE Trans Biomed Eng 32(3):230–236 based real-time QRS detection. Biomed Instrum Technol
35. Ligtenberg A, Kunt M (1983) A robust-digital QRS-detection 26(6):477–84
algorithm for arrhythmia monitoring. Comput Biomed Res 57. Doyle T, Dugan E, Humphries B, Newton R (2004) Discriminat-
16(3):273–286 ing between elderly and young using a fractal dimension analysis
36. Dokur Z, Olmez T, Yazgan E, Ersoy O (1997) Detection of ECG of centre of pressure. Int J Med Sci 1(1):11–20
waveforms by neural networks. Med Eng Phys 19(8):738–741 58. Burrus C, Gopinath R, Guo H (1998) Introduction to wavelets
37. Borjesson P, Pahlm O, Sornmo L, Nygards ME (1982) Adaptive and wavelet transforms. Prentice Hall, New Jersey
QRS detection based on maximum a posteriori estimation. IEEE 59. Strang G, Nguyen T (1996) Wavelets and filter banks. Wellesley-
Trans Biomed Eng 29(5):341–351 Cambridge Press, Wellesley

13
 Australasian Physical & Engineering Sciences in Medicine

60. Elgendi M, Jonkman M, De Boer F (2009) R wave detection 79. Song-Kai Z, Jian-Tao W, Jun-Rong X (1988) The real-time detec-
using Coiflets wavelets. In: 35th Annual Northeast IEEE bioen- tion of QRS-complex using the envelope of ECG. In: Proceed-
gineering conference, pp 1–2 ings of the annual international conference on IEEE Engineering
61. Ghaffari A, Mollakazemi M (2015) Robust fetal QRS detection in Medicine and Biology Society
from noninvasive abdominal electrocardiogram based on channel 80. Nygards M, Sornmo L (1983) Delineation of the QRS com-
selection and simultaneous multichannel processing. Australas plex using the envelope of the ECG. Med Biol Eng Comput
Phys Eng Sci Med 38(4):581–592 21(5):538–547
62. Dinh H, Kumar D, Pah N, Burton P (2001) Wavelets for QRS 81. Arzeno N, Deng Z-D, Poon C-S (2008) Analysis of first-deriva-
detection. Australas Phys Eng Sci Med 24(4):207–211 tive based QRS detection algorithms. IEEE Trans Biomed Eng
63. Song J, Yan H, Li Y, Mu K (2010) Research on electrocardio- 55(2):478–484
gram baseline wandering correction based on wavelet trans- 82. Benitez D, Gaydecki P, Zaidi A, Fitzpatrick A (2000) A new QRS
form, QRS barycenter fitting, and regional method. Australas detection algorithm based on the Hilbert transform. In: Comput-
Phys Eng Sci Med 33(3):279–283 ers in cardiology, pp 379–382
64. Yochum M, Renaud C, Jacquir S (2016) Automatic detection of 83. Arzeno N, Poon C-S, Deng Z-D (2006) Quantitative analysis
P, QRS and T patterns in 12 leads ECG signal based on CWT. of QRS detection algorithms based on the first derivative of
Biomed Signal Process Control 25:46–52 the ECG. In: 28th Annual international conference of the IEEE
65. Qin Q, Li J, Yue Y, Liu C (2017) An adaptive and time-efficient Engineering in Medicine and Biology Society (EMBS), pp
ECG R-peak detection algorithm. J Healthc Eng. https​://doi. 1788–1791
org/10.1155/2017/59805​41 84. Wang Z, Wong CM, Wan F (2017) Adaptive Fourier decompo-
66. Park J, Lee S, Park U (2017) R peak detection method using sition based R-peak detection for noisy ECG signals. In: 39th
wavelet transform and modified Shannon energy envelope. J Annual international conference of the IEEE Engineering in
Healthc Eng. https​://doi.org/10.1155/2017/49010​17 Medicine and Biology Society (EMBS), pp 987–990
67. Singh O, Sunkaria R (2017) ECG signal denoising via 85. Huang N, Shen Z, Long S, Wu M, Shih H (1998) The empiri-
empirical wavelet transform. Australas Phys Eng Sci Med cal mode decomposition and Hilbert spectrum for nonlinear and
40(1):219–229 nonstationary time series analysis. In: Proceedings of the Royal
68. Ahmed S, Al-Shrouf A, Abo-Zahhad M (2000) ECG data com- Society of London
pression using optimal nonorthogonal wavelet transform. Med 86. Oukhellou L, Aknin P, Delechelle E (2006) Railway infrastruc-
Eng Phys 22(1):39–46 ture system diagnosis using empirical mode decomposition and
69. Zheng X, Li Z, Shen L, Ji Z (2008) Detection of QRS com- Hilbert transform. In: Proceedings of IEEE international confer-
plexes based on biorthogonal spline wavelet. In: International ence on acoustics, speech and signal processing (ICASSP), vol 3
symposium on information science and engineering (ISISE), 87. Damerval C, Meignen S, Perrier V (2005) A fast algorithm for
vol 2, pp 502–506 bidimensional EMD. IEEE Signal Process Lett 12(10):701–704
70. Szilagyi S, Szilagyi L (2000) Wavelet transform and neural- 88. Safari A, Hesar H, Mohebbi M, Faradji F (2016) A novel method
network-based adaptive filtering for QRS detection. In: Pro- for R-peak detection in noisy ECG signals using EEMD and
ceedings of the 22nd annual international conference of the ICA. In: 23rd Iranian conference on biomedical engineering, pp
IEEE Engineering in Medicine and Biology Society, vol 2, pp 155–158
1267–1270 89. Chu C-H, Delp E (1989) Impulsive noise suppression and back-
71. Xiaomin X, Ying L (2005) Adaptive threshold for QRS complex ground normalization of electrocardiogram signals using mor-
detection based on wavelet transform. In: 27th Annual interna- phological operators. IEEE Trans Biomed Eng 36(2):262–273
tional conference of the IEEE Engineering in Medicine and Biol- 90. Trahanias P (1993) An approach to QRS complex detection
ogy Society (EMBS), pp 7281–7284 using mathematical morphology. IEEE Trans Biomed Eng
72. Martinez J, Almeida R, Olmos S, Rocha A, Laguna P (2004) A 40(2):201–205
wavelet-based ECG delineator: evaluation on standard databases. 91. Zhang F, Lian Y (2007) Novel QRS detection by CWT for ECG
IEEE Trans Biomed Eng 51(4):570–581 sensor. In: IEEE biomedical circuits and systems conference
73. Ruha A, Sallinen S, Nissila S (1997) A real-time microproces- (BIOCAS), pp 211–214
sor QRS detector system with a 1-ms timing accuracy for the 92. Gustafson D (1977) Automated VCG interpretation studies using
measurement of ambulatory HRV. IEEE Trans Biomed Eng signal analysis techniques, technical report R-1044. Charles Stark
44(3):159–167 Draper Laboratory, Cambridge
74. Li Y, Yan H, Hong F (2012) A new approach of QRS complex 93. Chen Y, Duan H (2005) A QRS complex detection algorithm
detection based on matched filtering and triangle character analy- based on mathematical morphology and envelope. In: 27th
sis. Australas Phys Eng Sci Med 35(3):341–356 Annual international conference of the IEEE Engineering in
75. Eskofier B, Kornhuber J, Hornegger J (2008) Embedded QRS Medicine and Biology Society (EMBS), pp 4654–4657
detection for noisy ECG sensor data using a matched filter and 94. NingIvan X, Selesnick W (2013) ECG enhancement and QRS
directed graph search. In: Proceedings of the 4th Russian-Bavar- detection based on sparse derivatives. Biomed Signal Process
ian conference on Biomedical Engineering, Zelenograd, Mos- Control 8(6):713–723
cow, Russia 95. Manikandan MS, Ramkumar B (2014) Straightforward and
76. Afonso V, Tompkins WJ, Nguyen T, Luo S (1999) ECG robust QRS detection algorithm for wearable cardiac monitor.
beat detection using filter banks. IEEE Trans Biomed Eng Healthc Technol Lett 1(1):40–44
46(2):192–202 96. Zhou Y, Hu X, Tang Z, Ahn AC (2016) Sparse representation-
77. Mengda L, Vinod A, Samson S (2011) A new flexible filter bank based ECG signal enhancement and QRS detection. Physiol
for low complexity spectrum sensing in cognitive radios. J Signal Meas 37(12):2093–2110
Process Syst 62(2):205–215 97. Arteaga-Falconi J, Al Osman H, El Saddik A (2015) R-peak
78. Afonso V, Tompkins WJ, Nguyen T, Trautmann S, Luo S (1995) detection algorithm based on differentiation. In: IEEE 9th Inter-
Filter bank-based processing of the stress ECG. In: 17th IEEE national symposium on intelligent signal processing (WISP), pp
annual engineering conference in Medicine and Biology Society, 1–4
vol 2, pp 887–888

13
Australasian Physical & Engineering Sciences in Medicine

98. Zhang F, Lian Y (2007) Electrocardiogram QRS detection conference on bioinformatics and biomedical engineering
using multiscale filtering based on mathematical morphol- (ICBBE), pp 693–696
ogy. In: 29th Annual international conference of the IEEE 116. Arafat A, Hasan T (2009) Automatic detection of ECG wave
Engineering in Medicine and Biology Society (EMBS), pp boundaries using empirical mode decomposition. In: IEEE inter-
3196–3199 national conference on acoustics, speech, and signal processing
99. Ulusar U, Govindan R, Wilson J, Lowery C, Preissl H, Eswaran (ICASSP), pp 461–464
H (2009) Adaptive rule based fetal QRS complex detection using 117. Kew H-P, Jeong D-U (2011) Variable threshold method for ECG
Hilbert transform. In: Annual international conference of the R-peak detection. J Med Syst 35(5):1085–1094
IEEE Engineering in Medicine and Biology Society (EMBC), 118. Zidelmal Z, Amirou A, Abdeslam DO, Moukadem A, Dieterlen
pp 4666–4669 A (2014) QRS detection using S-transform and Shannon energy.
100. Lin C, Hu W, Chen C, Weng C (2008) Heart rate detection in Comput Methods Programs Biomed 116(1):1–9
highly noisy handgrip electrocardiogram. In: Computers in car- 119. Dohare AK, Kumar V, Kumar R (2014) An efficient new method
diology, pp 477–480 for the detection of QRS in electrocardiogram. Comput Electr
101. Zhang F, Wei Y, Lian Y (2009) Frequency-response mask- Eng 40(5):1717–1730
ing based filter bank for QRS dection in wearable biomedical 120. Lin Z, Wang B, Chen H, Zhang Y, Wang X (2017) Design and
devices. In: IEEE international symposium on circuits and sys- implementation of a high quality R-peak detection algorithm.
tems (ISCAS), pp 1473–1476 In: China semiconductor technology international conference
102. Vai M-I, Zhou L-G (2004) Beat-to-beat ECG ventricular (CSTIC), pp 1–3
late potentials variance detection by filter bank and wavelet 121. Belforte G, De Mori R, Ferraris E (1979) A contribution to the
transform as beat-sequence filter. IEEE Trans Biomed Eng automatic processing of electrocardiograms using syntactic
51(8):1407–1413 methods. IEEE Trans Biomed Eng 26(3):125–136
103. Kaplan D (1990) Simultaneous QRS detection and feature extrac- 122. Ciaccio E, Dunn S, Akay M (1993) Biosignal pattern recognition
tion using simple matched filter basis functions. In: Proceedings and interpretation systems. 2. Methods for feature extraction and
of computers in cardiology, pp 503–506 selection. IEEE Eng Med Biol Mag 12(4):106–113
104. Hamilton P, Tompkins W (1988) Adaptive matched filtering for 123. Trahanias P, Skordalakis E (1990) Syntactic pattern recognition
QRS detection. In: Proceedings of annual international confer- of the ECG. IEEE Trans Pattern Anal Mach Intell 12(7):648–657
ence of the IEEE Engineering in Medicine and Biology Society, 124. Vijaya G, Kumar V, Verma H (1998) ANN-based QRS-complex
pp 147–148 analysis of ECG. J Med Eng Technol 22(4):160–167
105. Xue Q, Hu Y, Tompkins WJ (1992) Neural-network-based adap- 125. Hu YH, Tompkins WJ, Urrusti JL, Afonso VX (1993) Applica-
tive matched filtering for QRS detection. IEEE Trans Biomed tions of artificial neural networks for ECG signal detection and
Eng 39(4):317–329 classification. J Electrocardiol 26:66–73
106. Lu Y, Xian Y, Chen J, Zheng Z (2008) A comparative study to 126. Strintzis M, Stalidis G, Magnisalis X, Maglaveras N (1992)
extract the diaphragmatic electromyogram signal. In: Interna- Use of neural networks for electrocardiogram (ECG) feature
tional conference on BioMedical Engineering and Informatics extraction recognition and classification. Neural Network World
(BMEI), vol 2, pp 315–319 3:313–328
107. Dinh H, Kumar D, Pah N, Burton P (2001) Wavelets for QRS 127. Kohler B, Hennig C, Orglmeister R (2003) QRS detection using
detection. In: Proceedings of 23rd annual international confer- zero crossing counts. Appl Genomics Proteomics 2:138–145
ence of the IEEE Engineering in Medicine and Biology Society, 128. Coast D, Stern R, Cano G, Briller S (1990) An approach to car-
vol 2, pp 1883–1887 diac arrhythmia analysis using hidden Markov models. IEEE
108. Szilagyi L (1999) Wavelet-transform-based QRS complex detec- Trans Biomed Eng 37(9):826–836
tion in on-line Holter systems. In: 21st Annual conference and 129. Dobbs S, Schmitt N, Ozemek H (1984) QRS detection by tem-
the 1999 annual fall meeting of the Biomedical Engineering plate matching using real-time correlation on a microcomputer.
Society Engineering in Medicine and Biology, BMES/EMBS, J Clin Eng 9:197–212
conference, proceedings of the first joint, vol 1 130. Ebenezer D, Krishnamurthy V (1993) Wave digital matched
109. Shyu L-Y, Wu Y-H, Hu W (2004) Using wavelet transform and filter for electrocardiogram preprocessing. J Biomed Eng
fuzzy neural network for VPC detection from the Holter ECG. 15(2):132–134
IEEE Trans Biomed Eng 51(7):1269–1273 131. Di Virgilio V, Francaiancia C, Lino S, Cerutti S (1995) ECG fidu-
110. Alesanco A, Olmos S, Istepanian R, Garcia J (2003) A novel cial points detection through wavelet transform. In: 17th Annual
real-time multilead ECG compression and de-noising method conference of the IEEE Engineering in Medicine and Biology
based on the wavelet transform. In: Computers in cardiology, pp Society, vol 2, pp 1051–1052
593–596 132. Rao K (1997) DWT based detection of R-peaks and data com-
111. Bothe H (1997) Neuro-fuzzy-methoden. Springer, Berlin pression of ECG signals. IETE J Res 43(5):345–349
112. Zhang F, Lian Y (2009) Wavelet and Hilbert transforms based 133. Kadambe S, Murray R, Boudreaux-Bartels G (1999) Wavelet
QRS complexes detection algorithm for wearable ECG devices transform-based QRS complex detector. IEEE Trans Biomed Eng
in wireless body sensor networks. In: IEEE biomedical circuits 46(7):838–848
and systems conference (BioCAS), pp 225–228 134. Abibullaev B, Seo H (2011) A new QRS detection method
113. Zhou H-Y, Hou K-M (2008) Embedded real-time QRS detection using wavelets and artificial neural networks. J Med Syst
algorithm for pervasive cardiac care system. In: 9th International 35(4):683–691
conference on signal processing (ICSP), pp 2150–2153 135. Suzuki Y (1995) Self-organizing QRS-wave recognition in
114. Tang J, Yang X, Xu J, Tang Y, Zou Q, Zhang X (2008) The ECG using neural networks. IEEE Trans Neural Network
algorithm of R peak detection in ECG based on empirical mode 6(6):1469–1477
decomposition. In: 4th International conference on natural com- 136. Garca-Berdons C, Narvez J, Fernndez U, Sandovalm F (1997) A
putation (ICNC), vol 5, pp 624–627 new QRS detector based on neural network. IWANN: biologi-
115. Hongyan X, Minsong H (2008) A new QRS detection algorithm cal and artificial computation: from neuroscience to technology.
based on empirical mode decomposition. In: 2nd International Springer, Heidelberg, pp 1260–1269

13
 Australasian Physical & Engineering Sciences in Medicine

137. Cost A, Cano G (1989) QRS detection based on hidden Markov 153. Chen H, Chan H (2006) A real-time QRS detection method
modeling. In: Proceedings of annual international conference on based on moving-averaging incorporating with wavelet denois-
IEEE Engineering in Medicine and Biology Society, images of ing. Comput Methods Programs Biomed 82(3):187–195
the twenty-first century, pp 34–35 154. Ghaffari A, Golbayani H, Ghasemi M (2008) A new mathemati-
138. Ayat M, Shamsollahi M, Mozaffari B, Kharabian S (2009) ECG cal based QRS detector using continuous wavelet transform.
denoising using modulus maxima of wavelet transform. In: Comput Electr Eng 34(2):81–91
Annual international conference of IEEE Engineering in Medi- 155. Zheng H, Wu J (2008) Real-time QRS detection method. In:
cine and Biology Society (EMBC), pp 416–419 10th International conference on e-health networking, applica-
139. Moraes J, Freitas M, Vilani F, Costa E (2002) A QRS complex tions and services (HealthCom), pp 169–170
detection algorithm using electrocardiogram leads. In: Comput- 156. Fard P, Moradi M, Tajvidi M (2008) A novel approach in R peak
ers in cardiology, pp 205–208 detection using hybrid complex wavelet (HCW). Int J Cardiol
140. Manikandan MS, Soman KP (2012) A novel method for detect- 124(2):250–253
ing R-peaks in electrocardiogram (ECG) signal. Biomed Signal 157. Christov II (2004) Real time electrocardiogram QRS detection
Process Control 7(2):118–128 using combined adaptive threshold. Biomed Eng Online 3(1):28
141. Mallat S, Hwang W (1992) Singularity detection and processing 158. Elgendi M, Mahalingam S, Jonkman M, de Boer F (2008) A
with wavelets. IEEE Trans Inf Theory 38(2):617–643 robust QRS complex detection algorithm using dynamic thresh-
142. Ravanshad N, Rezaee-Dehsorkh H, Lotfi R, Lian Y (2014) A olds. In: International symposium on computer science and its
level-crossing based QRS-detection algorithm for wearable ECG applications (CSA), pp 153–158
sensors. IEEE J Biomed Health Inform 18(1):183–192 159. Elgendi M, Jonkman M, De Boer F (2009) Improved QRS detec-
143. Varon C, Caicedo A, Testelmans D, Buyse B, Huffel S (2015) A tion algorithm using dynamic thresholds. Int J hybrid Inf Technol
novel algorithm for the automatic detection of sleep apnea from 2:65–80
single-lead ECG. IEEE Trans Biomed Eng 62(9):2269–78 160. Li C, Zheng C, Tai C (1995) Detection of ECG characteristic points
144. Zhang X, Lian Y (2014) A 300-mv 220-nw event-driven ADC using wavelet transforms. IEEE Trans Biomed Eng 42(1):21–28
with real-time QRS detection for wearable ECG sensors. IEEE 161. Afonso V, Tompkins WJ, Nguyen T, Luo S (1996) Filter bank-
Trans Biomed Circuits Syst 8(6):834–843 based ECG beat detection. In: Proceedings of the 18th annual
145. Alavi S, Saadatmand-Tarzjan M (2013) A new combinatorial international conference of the IEEE Engineering in Medicine
algorithm for QRS detection. In: 3rd International conference and Biology Society, bridging disciplines for biomedicine, vol
on computer and knowledge engineering (ICCKE), pp 396–399 3, pp 1037–1038
146. Bouaziz F, Boutana D, Benidir M (2014) Multiresolution wave- 162. Chiarugi F, Sakkalis V, Emmanouilidou D, Krontiris T, Varanini
let-based QRS complex detection algorithm suited to several M, Tollis I (2007) Adaptive threshold QRS detector with best
abnormal morphologies. IET Signal Process 8(7):774–782 channel selection based on a noise rating system. In: Computers
147. Krimi S, Ouni K, Ellouze N (2008) An approach combining in cardiology, pp 157–160
wavelet transform and hidden Markov models for ECG segmen- 163. Chouhan V, Mehta S (2008) Detection of QRS complexes in
tation. In: 3rd International conference on information and com- 12-lead ECG using adaptive quantized threshold. Int J Comput
munication technologies: from theory to applications (ICTTA), Sci Network Secur 8:155–163
pp 1–6 164. Chouakri S, Bereksi-Reguig F, Taleb-Ahmed A (2011) QRS
148. Clifford G, Azuaje F, McSharry P (2006) Advanced methods and complex detection based on multi wavelet packet decomposi-
tools for ECG data analysis. Artech House, Inc, Norwood tion. Appl Math Comput 217(23):9508–9525
149. Dokur Z, Imez T (2001) ECG beat classification by a novel 165. Zidelmal Z, Amirou A, Adnane M, Belouchrani A (2012) QRS
hybrid neural network. Comput Methods Programs Biomed detection based on wavelet coefficients. Comput Methods Pro-
66(2–3):167–81 grams Biomed 107(3):490–496
150. Cheng W, Chan K (1998) Classification of electrocardiogram 166. Elgendi M (2013) Fast QRS detection with an optimized knowl-
using hidden Markov models. In: Proceedings of the 20th annual edge-based method: evaluation on 11 standard ECR databases.
international conference of the IEEE Engineering in Medicine PLoS ONE 8(9):e73557
and Biology Society, vol 1, pp 143–146 167. Elgendi M, Eskofier B, Dokos S, Abbott D (2014) Revisiting
151. Coast D (1993) Segmentation of high-resolution ECGs using QRS detection methodologies for portable, wearable, battery-
hidden Markov models. In: IEEE international conference on operated, and wireless ECG systems. PLoS ONE 9(1):1–18
acoustics, speech, and signal processing (ICASSP), vol 1, pp 168. Raj S, Luthra S, Ray KC (2015) Development of handheld car-
67–70 diac event monitoring system. IFAC-PapersOnLine 48(4):71–76
152. Poli R, Cagnoni S, Valli G (1995) Genetic design of optimum 169. Raj S, Ray KC (2018) A personalized arrhythmia monitoring
linear and nonlinear QRS detectors. IEEE Trans Biomed Eng platform. Sci Rep 8(11395):1–11. https​://doi.org/10.1038/s4159​
42(11):1137–1141 8-018-29690​-2

13