See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/327298268

A Novel R Peak Detection Method for Mobile Environments

Article  in  IEEE Access · August 2018

DOI: 10.1109/ACCESS.2018.2867329

CITATIONS READS

0 75

4 authors, including:

Miran Lee Suh-Yeon Dong

Ritsumeikan University Sookmyung Women's University
7 PUBLICATIONS   2 CITATIONS    11 PUBLICATIONS   112 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Miran Lee on 05 November 2018.

The user has requested enhancement of the downloaded file.

 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2018.2867329, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

A Novel R Peak Detection Method for Mobile

Environments
MIRAN LEE1, DAJEONG PARK1, SUH-YEON DONG2, and INCHAN YOUN1, 3
1
Center for Bionics, Korea Institute of Science and Technology, Seoul, 02792 South Korea
2
Department of Information Technology Engineering, Sookmyung Women’s University, Seoul, 04310 South Korea
3
KHU-KIST Department of Converging Science and Technology, KyungHee University, Seoul, 02792 South Korea

Corresponding author: Inchan Youn (iyoun@kist.re.kr) and Suh-Yeon Dong (suhyeon.dong@gmail.com).

This research was supported by the Bio & Medical Technology Development Program of the National Research Foundation (NRF) funded by the Korean
government, MSIP (2014M3A9D7070128) and by the Next-generation Medical Device Development Program for Newly-Created Market of the NRF funded
by the Korean government, MSIP (2015M3D5A1066100).

ABSTRACT With the development of the portable electrocardiogram (ECG) sensor, R peaks can be
monitored during various physical activities in mobile environment. However, such ECG signals contain real-
world noise, complicating the accurate detection of R peak. In this paper, we propose a novel approach for R
peak detection in ECG signals with real-world noise by using signal envelop filtering (SEF) and Shannon
energy envelope with the Savitzky–Golay filter (SEE-SG). SEF was designed to diminish abrupt peaks due
to movement by adjusting its amplitude based on an automatically determined threshold. SEE-SG was used
to enhance the QRS-complex. We first designed the experimental protocol for physical activity (sit, walk,
and ascend), namely Normal ECG in a Physical Activity Database (NPADB), and validated our method by
using the ECG data of 23 subjects. Moreover, to avoid overtuning on our database, we validated the proposed
method using two public databases: the MIT-BIH QT database (QTDB) and Noise Stress Test Database
(NSTDB). The experimental results show promising performance on QTDB (SE and P+: 99.96% and ER:
0.08%), NSTDB (SE: 94.07%, P+: 93.45, and ER: 12.57%), and NPADB (SE: 99.95%, P+: 99.78%, and ER:
0.27%). We also assessed our method by comparing the absolute differences of R–R intervals and heart rates
between actual and detected R peaks. In conclusion, as the proposed method achieved noise robustness and
reliable performance on R peak detection during physical activities, it can be used to monitor a user’s ECG
and cardiac health information accurately in daily life.

INDEX TERMS R peak detection, electrocardiogram, signal envelop filtering, Shannon energy envelope.

I. INTRODUCTION
The electrocardiogram (ECG) is a tool used to assess the
electrical activity of the heart [1], [2]. Various components
of the ECG, such as the P, Q, R, S, and T waves, provide
significant information for the diagnosis of heart diseases
and people’s health [3], [4]. As the R peak is the most
prominent among these waves, it is used as a reference point
for estimating the location of other waves. Moreover,
statistical information regarding successive R peaks is
frequently used for the assessment of heart rhythm’s
regularity and cardio-respiratory rate [5], [6]. Thus, detecting
accurate locations of R peaks is the first and most important
step in ECG signal processing.
The exact detection of R peaks is difficult, owing to the
interference of various noises, such as motion artifacts,
electrode contact noise, respiration, muscle contraction, and
power line interference that may be present in ECG signals [7],
[8]. As wearable ECG sensors that allow noninvasive and
inexpensive procedures are being widely used [9], users can
move freely outside the lab environment. However, data
obtained from such uncontrolled environments encounter
severe motion artifacts caused by the user’s physical
movement. These motion-related artifacts complicate the
detection of R peaks. The effect of these artifacts can weaken
the R peak spectrum, causing confusion in the P, Q, R, S, and
T waveforms [10], or resulting in a large amplitude of noise
signals in ECG signals [7], [10].
For the past few years, several studies have increased efforts
to solve this problem by using the MIT-BIH Noise Stress Test
database (NSTDB), which contains these motion-related
artifacts, such as motion artifacts, baseline wandering, and
muscle noise [11]. Li et al. [8] proposed a context-aware QRS

VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2018.2867329, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

detector according to the noise level. Merah et al. [12] TABLE 1. Description of databases used in this study.
proposed a stationary wavelet transform filter to reduce
# Total Duration/subject
the large amplitude of noises in ECG signals. Sangketkit et Database Category
beats
# Sub
(min)
al. [13] proposed a dual-threshold method with a band-pass Normal ECG Sit 8,385 23 5
filter. Benitez et al [14] proposed a band-pass FIR filter with in Physical Walk 10,775 23 5
Hilbert transform to remove muscular noise and maximize the Activity
(NPADB) Ascend 1,4048 23 5
QRS complex. Dohare et al. [15] used dual-median filter to MIT-BIH
correct the baseline drift and variable window width for R QT Database N/A 86,892 82 15
peak detection, and Plesnik et al. [16] proposed a phase-space (QTDB)
SNR
reconstruction by using a band-pass filter for R peak detection 4,265 2 30
24 dB
of ECGs with noise. All these methods worked successfully MIT-BIH 18 dB 4,265 2 30
Noise Stress
and were validated using the NSTDB; however, this database 12 dB 4,265 2 30
Test Database
is synthesized, thus artificially adding noises on clean ECG 6 dB 4,265 2 30
(NSTDB)
0 dB 4,265 2 30
signals. A method that functions well in a real-world situation
-6 dB 4,265 2 30
has not yet been devised. * Sub: subject; SNR: signal-noise-ratio;
Recently, some studies acquired their own database of
ECGs including physical activities and proposed a peak
detection method. Chen et al. [17] tested their algorithm for A. NPADB
detecting physical activity (walking and running) but did not First, we collected normal ECG data during physical activities
report on the accuracy of R peak detection. Lang et al. [18] and to detect R peak in ECGs contaminated by real-world noises.
Tantiger et al. [19] also attempted to detect R peaks during Twenty-three subjects (16 males and 7 females, mean ±
physical activities by considering a limited number of standard deviation age of 27.2 ± 3.2 years) participated in three
participants and a limited data size. Nevertheless, they did not physical activities: sitting, walking, and ascending. Subjects
evaluate their methods by using public databases to verify were asked to perform each task on a treadmill for 5 min, as
generalization. The aforementioned studies show that R peak shown in Fig. 1. The treadmill speed for the walking and
detection during physical activities has just recently begun. ascending tasks was set to 4 and 2.7 km/h, and the angle for
In this study, we propose a novel method of R peak the ascending task was set to 20° from the ground.
detection during physical movements. To develop a robust and For acquiring ECG data, we used the T-REX TR100A
reliable peak detection method, we utilized our own database (Taewoong Medical, Gimpo, South Korea), which is a
as well as various public databases. We first acquired an ECG wearable ECG sensor with a disposable patch-type electrode
database by collecting data during various physical activities, attached roughly 3–5 cm below the left chest; the sensor was
such as sitting, walking, and ascending, which frequently placed over the patch. This device was designed to observe
occur in daily life. Then, we developed a method based on long-term ECG signals with a battery capable of measuring for
signal envelop filtering (SEF) to weaken the amplitude of up to 48 h. The adhesive of the patch is harmless to human
abrupt peaks caused by physical movement. In addition, the skin, and the electrodes of lead 1 (left), lead 2 (right), and
QRS complex was enhanced using Shannon energy envelop ground are implemented as one easily attached patch. As the
(SEE) with a Savitzky–Golay (SG) filter. By using the filtered frequency range of ECG signals is normally 0.05–100 Hz [21],
signal, R peaks were detected with respect to the adaptive the sampling rate of ECG data is set to 256 Hz. The entire
threshold. The proposed algorithm was verified using our procedure of the experiment was approved by the institutional
assembled database and further tested using two public review board of the Korea Institute of Science and Technology
databases to assess its generalization ability. (2017-030). Owing to the policy of protecting human subjects’
The remainder of this paper is organized as follows. Section privacy, the recorded ECG data are not publicly available.
2 introduces our database, and Section 3 presents the proposed
R peak detection method using ECG signals with real-world
noises of physical activity. Detection performance of the
proposed method is demonstrated with various ECG databases
in Section 4. Section 5 discusses the outcome of the
experiments, and Section 6 draws the study conclusion.

II. DATABASE
To validate the proposed method for ECG R peak detection, (a) (b) (c)
we used three databases: a database we assembled of normal
ECGs in physical activities (NPADB), and two public FIGURE 1. Physical activities of (a) sitting (b) walking (speed: 4.0
databases, MIT-BIH QT Database (QTDB) [20] and NSTDB km/h) and (c) ascending (speed: 2.7 km/h; angle from the ground:
20°) conducted during the experiment.
[11]. Table 1 lists the detailed explanations of each database.

VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2018.2867329, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

B. MIT-BIH DATABASE adaptive threshold, as depicted in Fig. 2. During physical

We also used the QTDB to verify the proposed peak detection
method in a public domain. QTDB provides 105 records for
15-min experts from the MIT-BIH Arrhythmia Database [22],
the European Society of Cardiology ST-T Database [23], and
several other ECG databases collected at Boston’s Beth Israel
Deaconess Medical Center, and were selected to include a
wide variety of QRS and ST-T morphologies for QT detection
algorithms [20]. Of the total 105 recordings, only 82
recordings with beat annotations were selected.
In addition, for validating the noise sensitivity of the
proposed method, we used the NSTDB, comprising 12 half-
hour ECG recordings of two subjects in the MIT-BIH
Arrhythmia Database (Record nos.: 118 and 119), and three
types of noise were independently added: baseline wandering,
muscle artifacts, and electrode motion artifacts [11]. After the
first 5 min of clean ECG, noises were alternately and
separately added with a window size of 2 min. For each
subject recording, six signal-to-noise ratios (SNRs) of 24, 18,
12, 6, 0, and −6 dB were provided.
III. METHOD
An ECG signal is usually contaminated by several types of
noise, such as muscle noise, moving artifacts, and baseline
wandering. This, the effect of noise on signals recorded during
physical activity will be very severe. For this reason,
preprocessing and peak detection methods that were
developed for ECG signals recorded without movement, in a
laboratory environment, may not be enough for applying to
signals during physical activities. Therefore, we propose a
novel preprocessing and peak detection method that works
well in mobile environments.
The proposed method consists of four steps: baseline
correction, SEF, SEE calculation, and R peak detection using
Step 3
Start
Shannon energy envelope calculation
Calculating normalized squared
ECG signals
differentiation
Step 1 Generating Shannon energy (SE)
Baseline correction
Enveloping Shannon energy
Applying method filter with Savitzky-Golay filter (SEE)
activity, an ECG typically undergoes baseline wandering,
resulting in less accurate peak detection. Thus, we first applied
a median filter to remove the baseline drift. Next, we created
an SEE for simultaneously detecting R peak and diminishing
abrupt peaks. Then, we defined the QRS location using this
envelope with a SG filter and finally detected the R peak by
using an adaptive threshold. The detailed methodology for
each step will be described as follows. Also, additional
graphical materials are provided as a supplementary figure
(Figure S1). For each physical activity, the output of each
stage in our algorithm can be found in Figure S1.
A. BASELINE CORRECTION
Baseline wandering can be frequently observed in an ECG
signal because of respiration and movement [24], which
makes R peak detection in ECGs during physical activity even
more difficult. In baseline correction, the median filter is a
nonlinear filter, which is frequently used due to lack of
distortion in the absence of baseline wandering and its
computational efficiency [15], [25]. This filter simply
subtracts the median of the ECG signal using a predefined
window. We used two window widths (200 and 600 ms) as
selected in [26]. First, the filter with the window width of 200
ms targets the QRS complexes and P waves, while the second
filter with a window width of 600 ms targets the T wave. The
baseline estimates from two median filters are then subtracted
from the original ECG to finally obtain the baseline-corrected
signal.
B. SEF
Envelope filtering has been frequently used in ECG signal
processing to enhance QRS complex in [27]-[30]. However, it
can also emphasize abrupt peaks simultaneously. These abrupt
peaks can cause false negatives, causing the actual R peak to
be neglected, and can be seen more frequently in physical
activity. Therefore, we additionally implemented the SEF to
diminish unwanted abrupt peaks from ECG signals in physical
activity. To do so, we modified the analytic SEF algorithm
(ASEF) used to remove spikes in EEG signals [31]. We first
extracted envelope 𝑚(𝑡) of real signal 𝑠(𝑡) ; this is the
instantaneous amplitude of 𝑠(𝑡) written as

Step 4
𝑠(𝑡) = 𝑚(𝑡) ∙ 𝑐𝑜𝑠 (𝑡). (1)
Step 2

Signal envelope filtering

Peak detection Then, we filtered 𝑚(𝑡) by using a band-pass filter with a
Applying adaptive threshold cut-off frequency of 10–40 Hz to filter the QRS complexes,
using low-pass filter
Finding the optimal threshold and the filtered envelope was denoted as 𝑚𝑓𝑖𝑙𝑡 (𝑡). We defined
Generating bi-level signal threshold 𝑡ℎ(𝑡) by using the filtered envelope 𝑚𝑓𝑖𝑙𝑡 (𝑡), its
Modifying the amplitude
of abrupt peak average ̅̅̅̅̅̅,
𝑚𝑓𝑖𝑙𝑡 and arbitrary constant 𝛿, as follows:
Output R peak location
𝑡ℎ(𝑡) = 𝑚𝑓𝑖𝑙𝑡 (𝑡) + 𝛿 ∙ ̅̅̅̅̅̅.
𝑚𝑓𝑖𝑙𝑡 (2)
End
At each time point t, we calculated the final filtered
FIGURE 2. Flow chart of the proposed method in four steps: baseline envelope 𝑀𝑓𝑖𝑙𝑡 (𝑡) according to the magnitude of 𝑚(𝑡). As
correction, signal envelope filtering to diminish the abrupt peaks,
Shannon energy envelope calculation for QRS enhancement, and shown in (3), the value of 𝑚(𝑡) was selectively replaced with
peak detection. 𝑚𝑓𝑖𝑙𝑡 (𝑡) . If 𝑚(𝑡) > 𝑡ℎ(𝑡) , m(t) is replaced by 𝑡ℎ(𝑡) .

VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2018.2867329, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

4.0 4.0 4.0

δ=0 δ = 0.2 δ = 0.8 SE P+ ER

3.5 3.5 3.5

98 18
ECG amplitude (mV)

3.0 3.0 3.0

Performance of SE and P+ (%)

2.5 2.5 2.5

2.0 2.0 2.0 16

96
1.5 1.5 1.5

1.0 1.0 1.0 14

ER (%)
0.5 Original 0.5 Original 0.5 Original 94
Filtered (δ=0) Filtered (δ=0.2) Filtered (δ=0.8)
0 0 0
0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5 12
Time (sec) Time (sec) Time (sec)

(a) (b) (c) 92

10
1
Value of Cxz and σ*

0.8 Csz
σ* 90 8
0.6 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0

0.4
Parameter δ

0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0
FIGURE 3. Example of the signal envelope filter with 𝛅. (a) Graph of
original ECG with filtered (𝛅 = 0) signals; (b) filtered (𝛅 = 0.2) signals;
(c) filtered (𝛅= 0.8) signals; (d) 𝑪𝒔𝒛 (black-line with squared marker)
and 𝛔∗ (red-line with circle marker) with respect to 𝛅.
Otherwise, 𝑚(𝑡) is retained. Then, the final filtered signal
𝑧(𝑡) is obtained by multiplying 𝑀𝑓𝑖𝑙𝑡 (𝑡) with 𝑐𝑜𝑠 (𝑡):
𝑡ℎ(𝑡),
𝑀𝑓𝑖𝑙𝑡 (𝑡) = {
𝑚(𝑡),
𝑧(𝑡) = 𝑀𝑓𝑖𝑙𝑡 (𝑡) ∙ 𝑐𝑜𝑠(𝑡).
To find the optimal threshold that diminishes abrupt peaks
while maintaining fiducial points, we compared the envelope
shapes by varying parameter 𝛿 in (2). If parameter 𝛿 is
extremely small, the envelope at almost every time point is
filtered, that is, fiducial points can also diminish (Fig. 3 (a)).
While  is extremely large, the threshold increases
considerably and the envelope is not affected by the filter (Fig.
3(c)).
To determine the optimal 𝛿, we used index of coherence
𝐶𝑠𝑧 and an index based on the standard deviation of the filtered
signal 𝑧(𝑡) for each 𝛿 in the range 0 ≤  ≤ 2 (𝛿 has been
varied by 0.1 units). First, coherence 𝐶𝑠𝑧 (𝑓) is the ratio of the
squared magnitude of cross spectral density ( 𝑃𝑠𝑧 ) and the
product of each auto spectral density (𝑃𝑠𝑠 and 𝑃𝑧𝑧 ) at each
frequency 𝑓. We computed the average 𝐶𝑠𝑧 with respect to 𝑓
for each value of 𝛿. This value indicates how well filtered
signal 𝑧(𝑡) corresponds to raw signal 𝑠(𝑡) . Thus, as 𝛿
filter that passes high-frequency components (QRS complex;
10–40 Hz) and attenuates low-frequency components (P and
T waves; 5–10 Hz). Then, we squared the first-order
differentiation of 𝑧(𝑡) to obtain a unipolar signal for peak
detection. The squared differentiation of the filtered signal is
denoted as 𝑆𝐷𝑛 and can be computed as
𝑆𝐷𝑛 = (𝑧𝑛+1 − 𝑧𝑛
where 𝑛 is a sample index. We normalized the value of 𝑆𝐷𝑛
to scale its range from 0 to 1 by using a min–max
FIGURE 4. Testing performances with respect to parameter 𝛅 (SE:
sensitivity, P+: positive predictivity, and ER: detection error rate).
Parameter δ
(d) increases, 𝐶𝑠𝑧 approaches 1. Another index, denoted as σ∗ ,
was defined using the standard deviation of the filtered signal
𝑧(𝑡) . We computed the standard deviation of the filtered
signal for each value of 𝛿 and normalized the values of the
standard deviation to be in the range 0–1. Then, we obtained
the final index σ∗ by subtracting the normalized standard
deviation from 1. As 𝛿 increased, the filtered signal
approached closer to the original signal (where 𝐶𝑠𝑧
𝑖𝑓 𝑚(𝑡) ≥ 𝑡ℎ(𝑡) approaches 1). In this case, σ∗ approaches 0 because the
. (3)
𝑖𝑓 𝑚(𝑡) < 𝑡ℎ(𝑡) standard deviation of 𝑧(𝑡) increased, owing to unfiltered
abrupt peaks. In contrast, as 𝛿 approached close to 0, the
(4) filtered signal’s standard deviation decreased relatively such
that σ∗ approached 1 (where 𝐶𝑠𝑧 approaches 0). Therefore, we
could observe the point at which these two indices crossed
each other within the range of 𝛿 defined earlier. We assumed
the value of 𝛿 near this crossing point to represent the
appropriate adjustment of abrupt peaks, and tested it with
example data (using data excerpted from NPADB, excluding
the validation of performance in the results), as shown in Fig.
4.
C. SEE CALCULATION
To detect R peaks, it is necessary to identify the location of the
QRS complex. We generated the SEE using the SG filter to
emphasize peak information. More specifically, we aimed to
emphasize the QRS complex while attenuating the P and T
waves. To do this, we first differentiated the filtered signal
𝑧(𝑡) in the first order before calculating the SEE. As noted in
[27], the first-order differentiation has an effect equivalent to
that of a high-pass
normalization method. We denoted this unipolar signal as 𝑆𝐷 ̂𝑛 .
Then, 𝑆𝐸𝑛 can be obtained following the Shannon energy
function [32] as follows.
𝑆𝐸𝑛 = −(𝑆𝐷 ̂ 𝑛 ).
̂ 𝑛 ) log(𝑆𝐷 (6)
Within one cycle of an ECG signal, the obtained 𝑆𝐸𝑛 has
three bumps due to Q, R, and S peaks, as shown in Fig. 5 (b).
)2 . (5)
Next, 𝑆𝐸𝑛 was smoothed using an SG filter to allow this
envelope to smoothly cover all of these bumps, as shown in
Fig. 5 (c). Unlike the standard smoothing filter, the SG filter

VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2018.2867329, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

1000
(a) Original ECG signal z(t) if 𝑆𝐸𝐸𝑛 > 𝐴𝑇𝑛 , 𝐵𝑛 = 1, else 𝐵𝑛 = 0. Therefore, bi-level signal
𝐵𝑛 = 1 during the QRS complex of the ECG signal, and 0
0
otherwise. Finally, the R peak was detected as the maximal
-1000
0 0.5 1 1.5 2 2.5 3 3.5 4
amplitude point in each QRS region where 𝐵𝑛 = 1, as shown
0.4
(b) Shannon energy SE in Fig. 5 (d).
According to the aforementioned processing, we adjusted
0.2
the abrupt peak by using the SEF and enhanced the QRS
0
0 0.5 1 1.5 2 2.5 3 3.5 4 complex based on SEE with the SG filter. Fig. 6 (b) shows the
1
(c) Shannon energy envelope actual R peaks that are not revealed by the abrupt peaks
without SEF, while Fig. 6 (c) shows that they are noticeably
0.5
enhanced after SEF application.
0
0 0.5 1 1.5 2 2.5 3 3.5 4

1000
(d) Original ECG and bi-level signal
0 0.5
-1000
0 0.5 1 1.5 2 2.5
FIGURE 5. Example of Shannon energy envelope and bi-level signal
((a) Original ECG signals; (b) Shannon energy SE; (c) Shannon
energy envelope with Savitzky-Golay filter (blue-solid line) and an
adaptive threshold (magenta-dot line); (d) Original ECG (blue-solid
line) and bi-level signal (magenta-solid line) with annotation R peaks
(red-circle).
has been reported to be advantageous for peak preservation
using a simple least square fitting [33], [34]. We denoted the
filtered envelope as 𝑆𝐸𝐸𝑛 . The window width of the SG filter
was selected as 35, which is roughly twice the full width at
half of maximum of the QRS complex. Furthermore, the
degree of the SG filter was selected as 5, and all of these
parameters were selected based on the validation using our
preliminary data. Filtered envelope 𝑆𝐸𝐸𝑛 was then used to
define the threshold for the location of the QRS complex.
D. PEAK DETECTION USING AN ADAPTIVE
THRESHOLD
We used the filtered envelope as the adaptive threshold for
QRS localization. Adaptive threshold 𝐴𝑇𝑛 was calculated
based on low-pass filtering at 3 Hz of envelope 𝑆𝐸𝐸𝑛 from the
previous step. This step was especially necessary for the ECG
signals obtained during physical activities, to reliably remove
the unexpected bumps in P or T waves due to physical
movement. Then, 𝑆𝐸𝐸𝑛 was converted to bi-level signal 𝐵𝑛 ;
(a)
w/o SEF
R R R R R R
NSD
1
IV. RESULTS
A. PERFORMANCE MEASURES
0
3 3.5 4 As a measure of R peak detection performance, sensitivity
(SE), positive predictivity (P+), and error-detection rate (ER)
were considered in this study [17], [35], [36]. To compute
these measures, the number of true beats (TB), true positives
(TP), false positives (FP), and false negatives (FN) were
counted for each record. TB shows the actual R peak, and its
annotation was determined by four medical experts. More
specifically, the R peak is determined to be TB when more
than three quarters of the experts labeled it as TB.
SE is the ratio of the R peaks (TP) correctly detected using the
proposed method to the true beats (TB) by the given
annotation. P+ is the ratio of the true positive to all the detected
R peaks (i.e., TP and FP), and ER is the ratio of false detections
(FP and FN) over the TBs by the given annotation [17]. These
measures are calculated using (7)–(9).
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 (𝑆𝐸) = 𝑇𝑃/(𝑇𝑃 + 𝐹𝑁) × 100 (7)
𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑣𝑖𝑡𝑦 (𝑃+)
(8)
= 𝑇𝑃/(𝑇𝑃 + 𝐹𝑃) × 100
𝐷𝑒𝑡𝑒𝑐𝑡𝑖𝑜𝑛 𝑒𝑟𝑟𝑜𝑟 𝑟𝑎𝑡𝑒 (𝐸𝑅)
(9)
= (𝐹𝑃 + 𝐹𝑁)/𝑇𝐵 × 100
For determining the detected R peaks as TPs, the time range
of 150 ms around the annotated R peak is considered to be a
TP region [17], [37]. In other words, any detected peak within
Abrupt peak
this time range is considered as the correct peak (TP). This
determination rule was applied to all three databases for
objectivity.
In addition, the differences of the means of R–R intervals
w/ SEF (DRRI) and heart rate (DHR) have been used as performance
R R R R R R
measures for evaluating the proposed R peak detection method.
Both DRRI and DHR are calculated using (10)–(11),

NSD

respectively.
SEE

SEE

𝐷𝑅𝑅𝐼 (𝑚𝑠) = |mRRI(actual) − mRRI(detected)| (10)

(b) (c)

FIGURE 6. Comparison of processes using the signal envelop filter

(SEF): (a) original ECG signal; (b) Normalized squared differential 𝐷𝐻𝑅 (𝑏𝑝𝑚) = |(60/mRRI(actual) )
(NSD) and Shannon energy envelope (SEE) without SEF; (c) NSD (11)
− (60/mRRI(detected))|
and SEE with SEF.

VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2018.2867329, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

TABLE 2. Results of R peak detection using the proposed method in QTDB and NSTDB.

Database SNR (dB) # TB SE (%) P+ (%) ER (%)

QTDB N/A 86,892 99.96 (±0.08) 99.96 (±0.11) 0.08 (±0.14)
NSTDB 24 4,265 99.95 (±0.07) 99.88 (±0.02) 0.17 (±0.05)
18 4,265 99.86 (±0.08) 99.88 (±0.04) 0.26 (±0.12)
12 4,265 99.58 (±0.47) 99.83 (±0.12) 0.59 (±0.59)
6 4,265 98.11 (±1.24) 97.07 (±1.62) 4.85 (±2.89)
0 4,265 90.20 (±0.45) 87.73 (±2.06) 22.43 (±2.79)
-6 4,265 76.75 (±0.14) 76.32 (±2.90) 47.14 (±3.72)
* Numbers in parentheses indicate standard deviations.

where mRRI (actual) and mRRI (detected) represent the
means of R–R intervals from actual R peaks and detected R
peaks using the proposed method, respectively.
B. PERFORMANCE OF THE MIT-BIH DATABASE
The performance of the R peak detection method in QTDB
and NSTDB is summarized Table 2. In the case of the QTDB,
the total number of TBs is 86,892. The averages of SE, P+,
and ER are 99.96%, 99.96%, and 0.08%, respectively. Among
these, the worst SE, P+, and ER were 99.57% (sele0114),
99.50% (sel847), and 0.68% (sel232), respectively. The
NSTDB contains noise-contaminated ECG recordings with
varying SNRs, with the average SE, P+, and ER as 94.08%,
TABLE 3. Performance of ECG R peak detection in “sit” class of NPADB using proposed method.
93.45%, and 12.57%, respectively. In the NSTDB, the best
performances of SE, P +, and ER were 99.95% (SNR: 24 dB),
99.88% (SNR: 24 and 18 dB), and 0.17% (SNR: 24 dB),
respectively, while their respective worst performances were
76.75 (SNR: −6 dB), 76.32 (SNR: −6 dB), and 47.14 (SNR:
−6 dB).
C. PERFORMANCE OF NPADB
Finally, we verified the proposed method in our database,
which comprises ECG recordings, obtained during physical
activities, and thus may be contaminated with real-world noise.
Tables 3–5 report the resulting performances of using the
proposed R peak detection method during the tasks of sitting,

Rec. TB TP FP FN SE P+ ER Rec. TB TP FP FN SE P+ ER
S1 386 386 0 0 100.00 100.00 0.00 S13 405 405 0 0 100.00 100.00 0.00
S2 428 428 0 0 100.00 100.00 0.00 S14 295 295 0 0 100.00 100.00 0.00
S3 379 379 0 0 100.00 100.00 0.00 S15 387 387 0 0 100.00 100.00 0.00
S4 398 398 0 0 100.00 100.00 0.00 S16 339 339 1 0 100.00 99.71 0.29
S5 368 368 0 0 100.00 100.00 0.00 S17 410 410 0 0 100.00 100.00 0.00
S6 322 322 0 0 100.00 100.00 0.00 S18 354 354 0 0 100.00 100.00 0.00
S7 296 296 0 0 100.00 100.00 0.00 S19 368 368 0 0 100.00 100.00 0.00
S8 310 310 0 0 100.00 100.00 0.00 S20 373 373 0 0 100.00 100.00 0.00
S9 418 418 0 0 100.00 100.00 0.00 S21 325 325 0 0 100.00 100.00 0.00
S10 286 286 0 0 100.00 100.00 0.00 S22 351 350 0 1 99.72 100.00 0.28
S11 443 443 0 0 100.00 100.00 0.00 S23 371 371 0 0 100.00 100.00 0.00
S12 373 373 0 0 100.00 100.00 0.00 Total 8,385 8,384 1 1 99.99 99.99 0.03

TABLE 4. Performance of ECG R peak detection in “walk” class of NPADB using proposed method.

Rec. TB TP FP FN SE P+ ER Rec. TB TP FP FN SE P+ ER
S1 470 469 2 1 99.79 99.58 0.64 S13 514 514 0 0 100.00 100.00 0.00
S2 547 547 0 0 100.00 100.00 0.00 S14 416 415 6 1 99.76 98.57 1.68
S3 460 460 0 0 100.00 100.00 0.00 S15 488 487 2 1 99.80 99.59 0.61
S4 511 511 1 0 100.00 99.80 0.20 S16 440 440 13 0 100.00 97.13 2.95
S5 491 491 0 0 100.00 100.00 0.00 S17 502 502 2 0 100.00 99.60 0.40
S6 410 410 1 0 100.00 99.76 0.24 S18 430 430 0 0 100.00 100.00 0.00
S7 428 428 1 0 100.00 99.77 0.23 S19 467 467 4 0 100.00 99.15 0.86
S8 477 477 1 0 100.00 99.79 0.21 S20 495 495 0 0 100.00 100.00 0.00
S9 484 483 0 1 99.79 100.00 0.21 S21 384 384 13 0 100.00 96.73 3.39
S10 364 364 0 0 100.00 100.00 0.00 S22 404 403 1 1 99.75 99.75 0.50
S11 563 563 1 0 100.00 99.82 0.18 S23 489 488 0 1 99.80 100.00 0.20
S12 541 540 0 1 99.82 100.00 0.18 Total 10,775 10,768 48 7 99.93 99.52 0.55

VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2018.2867329, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

TABLE 5. Performance of ECG R peak detection in “ascend” class of NPADB using proposed method.

Rec. TB TP FP FN SE P+ ER Rec. TB TP FP FN SE P+ ER
S1 593 593 1 0 100.00 99.83 0.17 S13 631 630 1 1 99.84 99.84 0.32
S2 685 685 0 0 100.00 100.00 0.00 S14 545 544 4 1 99.82 99.27 0.92
S3 606 606 0 0 100.00 100.00 0.00 S15 601 600 0 1 99.83 100.00 0.17
S4 668 668 0 0 100.00 100.00 0.00 S16 561 560 4 1 99.82 99.29 0.89
S5 651 651 0 0 100.00 100.00 0.00 S17 600 600 1 0 100.00 99.83 0.17
S6 531 530 0 1 99.81 100.00 0.19 S18 593 592 0 1 99.83 100.00 0.17
S7 592 592 0 0 100.00 100.00 0.00 S19 583 583 0 0 100.00 100.00 0.00
S8 688 687 2 1 99.85 99.71 0.44 S20 604 604 0 0 100.00 100.00 0.00
S9 642 642 1 0 100.00 99.84 0.16 S21 488 488 8 0 100.00 98.39 1.64
S10 497 496 0 1 99.80 100.00 0.20 S22 551 551 0 0 100.00 100.00 0.00
S11 750 750 1 0 100.00 99.87 0.13 S23 638 638 0 0 100.00 100.00 0.00
S12 750 750 0 0 100.00 100.00 0.00 Total 14,048 14,040 23 8 99.94 99.82 0.24

TABLE 6. Absolute differences in mRRI (DRRI) and mHR (DHR) computed by R peaks using proposed method and ground truth.

Absolute difference (between actual and detected R peaks)

Database Category DRRI (ms) DHR (bpm)
Mean SD Mean SD
Sit 0.59 1.57 0.04 0.10
NPADB Walk 4.20 4.57 0.58 0.58
Ascend 3.10 7.32 0.75 1.87
QTDB N/A 0.23 0.11 0.91 0.12
SNR: 24 dB 0.73 0.45 1.2 1.21
SNR: 18 dB 0.62 0.35 1.01 1.42
SNR: 12 dB 2.70 3.28 1.13 0.64
NSTDB
SNR: 6 dB 8.86 4.19 17.34 0.32
SNR: 0 dB 22.23 19.7 38.46 0.91
SNR: −6 dB 22.66 7.34 36.09 8.67
* N/A: not applicable; mRRI: mean of R–R intervals; mHR: mean of heart rate

walking, and ascending, respectively. The performance
measures for sitting, which contain relatively clean ECG data,
were 99.99% of the average SE and P+, and 0.03% of the
average ER for 8,385 beats. In contrast, the average SE and P+
slightly dropped to 99.93% and 99.52% for the walking task,
and 99.94% and 99.82% for the ascending task, respectively.
Moreover, the average ERs slightly increased to 0.55% and
0.24% for the walking and ascending tasks, respectively.
D. COMPARISON OF THE RESULT OF THE R–R
INTERVAL AND HEART RATE
We additionally assessed the means of R–R intervals (mRRI)
and heart rate (mHR) by using the proposed R peak detection
method. Table 6 confirms DRRI and DHR between actual R
peaks and the R peaks detected using the proposed method. In
the case of NPADB, the means of DRRI were 0.59 (sitting),
4.20 (walking), and 3.10 (ascending) ms. In addition, the
means of DHR were 0.04 (sitting), 0.58 (walking), and 0.75
(ascending) bpm, and the range of difference is less than 1 bpm.
In the case of QTDB, the mean of DRRI and DHR were 0.23
(±0.11) ms, and 0.91 (±0.12) bpm, respectively. In NSTDB,
the means of DRRI and DHR increased as SNR decreased
from 24 to −6 dB.
V. DISCUSSION
In this paper, we proposed a novel R peak detection method
using SEF and SEE-SG on an ECG signal with real-world
noise. From its reliable performances on three real and
synthesized databases, our method showed its usefulness in a
mobile environment. However, some issues, such as
parameter choice in SEF and performance comparison
according to each database’s characteristics, still need to be
discussed.
A. PARAMETER CHOICE
We proposed the combined method of using SEF to adjust
abrupt peaks in the ECGs with real-world noise and using
SEE-SEG to enhance QRS-complex regions. In the process of
developing our method, we used a sample dataset excerpted
from NPADB, and excluded this dataset for the performance
validation in the results.
In the SEF, we mentioned earlier that our method was
motivated from the ASEF algorithm used to remove spikes in

VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2018.2867329, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

an EEG signal, as proposed by Melia et al. [31]. The most
remarkable difference about their method compared with our
method is not the removal but the weakening of the amplitude
of the abrupt peak. As their method was aimed at eliminating
the noise of the EEG signals, the amplitude of the abrupt peak
diminished to be m𝑓𝑖𝑙𝑡 (t) when it was greater than the
threshold. In contrast, our method weakens the amplitude of
the abrupt peak to threshold 𝑡ℎ(𝑡) to maintain the shape of the
R peak amplitude.
Owing to this critical difference with the method by Melia
et al. [31], parameter δ could not be determined in a different
manner. We determined optimal parameter δ by using both
𝐶𝑥𝑧 and 𝜎 ∗ to create threshold 𝑡ℎ(𝑡). Fig. 3 (d) shows that
optimal parameter δ = 0.2 when the difference between 𝐶𝑥𝑧
(0.43) and 𝜎 ∗ (0.39) is the lowest in the case of the sample
dataset. As shown in Fig. 3(b), by using this optimal parameter
δ (0.2), which preserves the phase and shape of the original
ECG signal, only the amplitude of the abrupt peak is weakened.
Thus, we conclude that the value of δ when the difference
between 𝐶𝑥𝑧 and 𝜎 ∗ is the lowest is optimal in SEF. In other
words, the value of δ is not fixed to a specific value because
ECG signal characteristics may vary depending on the subject
and recording environment. Instead, our algorithm
automatically determines the optimal δ for each record based
on the aforementioned criterion.
B. PEAK DETECTION IN NORMAL ECG
QTDB includes ECGs with annotations by cardiologists and
is close to a normal ECG database as the percentage of beats
marked “normal” is 92%.
We summarized other studies, validating their algorithms
using QTDB in Table 7 [12], [37]-[42]. The table shows that
our method outperformed other methods on QTDB in terms of
all three types of performance measures, that is, SE, P+, and
ER.
In addition, we tested the proposed method for normal R
peak detection using MIT-BIH Arrhythmia Database
(MITDB). Ten records (100, 101, 103, 115, 122, 123, 205, 220,
230, and 234) from healthy persons containing normal ECG
signals in MITDB have been used [34] for validating our
method. As a result, these records included over 98.9% normal
ECG beats of the total beats and showed good average
performances (SE: 99.99±0.02; P+: 99.93±0.08; ER:
0.08±0.08 %).
Furthermore, we also measured the output signal-to-noise
ratio (𝑆𝑁𝑅𝑜 ) of our algorithm using record 103 in MITDB.
Record 103 was selected because it contained clean and
normal ECG signals in the MITDB, and it was generated with
the same protocol [11] as synthesized records 118 and record
119 of NSTDB. Six types of noise signals (SNR: -6, 0, 6, 12,
18, and 24 dB) were synthesized in record 103 and used data

TABLE 7. Comparisons of the methods using QTDB and NSTDB.

Publication Preprocessing QRS enhancement and R peak detection # TB SE (%) P+ (%) ER (%)
QTDB
Median filter + Signal envelope
Present study Shannon energy + Savizky–Golay filter 86,892 99.96 99.96 0.08
filtering
Band-pass filter (0.4-40 Hz) +
Ghaffari et al. [37] Multiple higher order moments metric (MHOM) 86,892 99.96 99.92 0.12
Wavelet transform
Rezgui et al. [38] Empirical mode decomposition Selection of IMFs for R-peak 86,892 99.95 99.91 0.14
Local maxima, minima and zero crossings of the
Merah et al. [12] Stationary wavelet transform 86,892 99.94 99.89 0.18
fourth SWT coefficient detail
Band-pass filter (5-20 Hz) + Modified Hilbert transform-based algorithms
Ghaffari et al. [39] 86,892 99.92 99.89 0.19
Hilbert transform (ECGMHT)
Multiscale approach proposed
Martinez et al. [40] Wavelet transform 86,892 99.92 99.88 0.20
by Li et al. [43].
Wavelet-based multiresolution Signal mirroring + local maximum location and
Qin et al. [41] 86,995 99.83 99.90 N/R
analysis adaptive threshold selection
Thungtong [42] Wavelet transform Selection of the optimal wavelet detail coefficients 83,292 99.63 99.78 0.59

NSTDB
Nonlinearly scaled curve length transformation
Li et al. [8] Low-pass filter N/R 98.64 77.61 N/R
(S.S.A)
Local maxima, minima and zero crossings of the
Merah et al. [12] Stationary wavelet transform 25,590 95.30 93.98 10.81
fourth SWT coefficient detail
Sign removal and smoothening using moving
Sangketkit et al. [13] Band-pass filter (5-22 Hz) N/R 94.83 89.83 N/R
average filter
Median filter + Signal envelope
This study Shannon energy + Savizky–Golay filter 25,590 94.08 93.45 12.57
filtering
Benitez et al. [14] Band-pass filter (8-20 Hz) Hilbert transform N/R 93.48 90.60 N/R

Dohare et al. [15] Median filter Sixth power of signal + variable window width 25,590 88.09 88.93 N/R

Plesnik et al. [16] Band-pass filter (0.5-100 Hz) Phase-space reconstruction 25,590 72.01 86.73 64.86

* N/R: not reported; #TB: number of true beat; SE: sensitivity; P+: positive predictivity; ER: detection error rate

VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2018.2867329, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

containing noise signal (2 minutes) to measure SNR. 𝑆𝑁𝑅𝑜
was calculated using formula (12) [34].
1 𝑁−1
∑ |𝑠̂ (𝑛)|2
𝑆𝑁𝑅o = 10 log10 ( 𝑁 𝑛=0
1 𝑁−1
∑ |𝑠(𝑛) − 𝑠̂ (𝑛)|2
𝑁 𝑛=0
where 𝑁 is the number of samples. 𝑠(𝑛) and 𝑠̂ (𝑛) denote
the clean signal and denoised signal, respectively. As a result,
𝑆𝑁𝑅𝑜 is as 1.96, 6.74, 12.38, 17.75, 22.25, and 27.24dB when
synthesized with -6, 0, 6, 12, 18, and 24 dB of noise signals,
respectively, and then it was finally proved that the SNR was
improved using our algorithm by removing noise signals.
C. PEAK DETECTION IN NOISY ECG
NSTDB includes ECGs with three types of noise; thus, we
selected this database to evaluate the noise sensitivity of the
peak detection method. Moreover, one of the recordings
comprises 100% arrhythmia and the other contained 22.35%
abnormal beats according to the beat annotation provided in
the database. Therefore, the performance on this database can
assess abnormal-beat detection ability as well as noise
sensitivity.
Several studies used this database and are summarized in
Table 7. Among them, the study by Li et al. [8] achieved the
best SE of 98.64% but the worst P+ of 77.61%. It was
presumed that there were many FPs in their method.
Sangketkit et al. [13] reported 0.8% higher SE than our
method, but they only considered the ECGs of 24, 12, and 0
dB SNRs. Moreover, their value of P+ was lower than that
obtained using our proposed method. Benitez et al. [14],
Dohare et al. [15], and Plesnik et al. [16] also reported their
detection results using this database. The method proposed by
Merah et al. [12] showed reasonably good performance in
terms of all three performance measures and slightly
outperformed our method. Through these comparisons,
) (12)
although our algorithm was not aimed to detect arrhythmic
ECG beats, we could achieve promising results on noisy as
well as abnormal ECGs.
D. PEAK DETECTION IN MOVING ECG
We compared our method with previous studies for R peak
detection during physical activities (Table 8). Similar to our
study in which we developed our own database (NPADB),
some of the previous studies also used their own databases of
ECGs in moving environments.
Chen et al. [17] proposed their algorithm to reduce noise
during running and compared RRIs with commercial ECG
sensors. However, they did not report the quantitative result of
R peak detection during running and evaluated their algorithm
with relatively little data (10 RRIs of 10 persons). Similar to
our study, Lang et al. [18] and Tantinger et al. [19] represented
good performances of peak detection in physical activity by
modifying the filter used in Kohler’s algorithm [44]. However,
they did not evaluate their algorithms using any public
database, and thus the reliability of their algorithms is not
validated yet.
Compared to these aforementioned methods, our method
achieved better performance in a real moving environment in
terms of all three types of performance measures in R peak
detection. In addition to the results obtained using public
databases, our method proved that it is possible to detect R
peaks sufficiently well for ECG with real-world noise due to
various physical activities.

TABLE 8. Comparison of the methods using own DB.

Publication Preprocessing QRS enhancement and R Activity # Sub Performance

Band-pass filter (0.5-17
Chen et al. [17]
Hz)
Band-pass filter (0.67-
Lang et al. [18] 40 Hz) + IIR forward-
backward filtering
Tantinger et al. Linear phase FIR filter
[19] (18-35 Hz)
Median filter + Signal
This study
envelope filtering
* N/R and Sub denotes not reported and subject, respectively
peak detection
- Result of RRI and HR
Enhancement mask to the
mRRI: 0.74 ms
filtered signal + Searching - Run 10
(the difference of mRRI between their algorithm
QRS fiducials points
and commercial sensor)
- R peak detection
Zero-crossing QRS detection - Run (10 km/h)
SE: 99.2
by Kohler et al. [44] + - Ergometer (130 Watt) 10
P+: 97.2
threshold * Duration: each 5 min
- Sit - R peak detection
- Walk (2 km/h) SE: 99.94%
Zero-crossing QRS detection
- Run (8 km/h) P+: 98.77%
by Kohler et al. [44] + 10
- Cycle (150 Watt)
threshold
- Rowing
* Duration: each 2 min
- R peak detection
SE: 99.95%
P+: 99.78%
ER: 0.27%
Shannon energy + Savizky - Sit
Golay filter + adaptive - Ascend (2.7 km/h, 20°)
23 - Result of RRI and HR
threshold using low-pass - Walk (4.5 km/h)
DRRI: 2.63 ms
filter * Duration: each 5 min
DHR: 0.46 ms
(the difference of mRRI (DRRI) and mHR (DHR)
from between actual R peaks and detected R
peaks)

VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2018.2867329, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

VI. CONCLUSION [8] H. Li and J. Tan, “Body Sensor Network Based Context
This study introduced a novel approach for automatic R peak Aware QRS Detection,” in Proc. IEEE-EMBC, New
York, NY, USA, 2006, pp. 1-8.
detection in ECG during physical activities. The proposed [9] M. Hooshmand et al., “SURF: Subject-adaptive
method is based on the SEF and SEE with Savitzky–Golay, unsupervised ECG signal compression for wearable
and can adjust the amplitude of abrupt peaks caused by fitness monitors,” IEEE Access, vol. 5, pp. 19517-19535,
Sep. 2017, DOI: 10.1109/ACCESS.2017.2749758.
physical movement and enhance the QRS-complex region. [10] D. A. Tong, K. A. Bartels, and K. S. Honeyager,
We developed and validated our method using the ECG “Adaptive reduction of motion artifact in the
data of 23 subjects recorded during various physical activities electrocardiogram,” in Proc. EMBC-BMES, Houston,
Texas, USA, 2002, pp. 1403-1404.
such as sitting, walking, and running. To avoid overtuning on [11] G. B. Moody, W. Muldrow, and R. G. Mark, “A noise
a specific database, we also validated our method on MIT-BIH stress test for arrhythmia detectors,” Comput Cardiol, vol.
QTDB and MIT-BIH NSTDB. Consequently, the proposed 11, no. 3, pp. 381-384, Jan. 1984.
method reported reliable performance on all three databases: [12] M. Merah, T. A. Abdelmalik, and B. H. Larbi, “R-peaks
detection based on stationary wavelet transform,”
QTDB (SE and P+: 99.96% and ER: 0.08%), NSTDB (SE: Comput Meth prog bio, vol. 121, no. 3, pp. 149-160, Oct.
94.07%, P+: 93.45, and ER: 12.57%), and NPADB (SE: 2015, DOI: doi.org/10.1016/j.cmpb.2015.06.003.
99.95%, P+: 99.78%, and ER: 0.27%). In addition, the results [13] P. Sangketkit and S. Wongsa, “Robustness Evaluation of
a Dual-Threshold QRS Detection Method for Wearable
of the absolute difference of the mRR and mHR between ECG Recorders,” in Proc. ECTI-CON, Phuket, Thailand,
actual R peaks and R peaks detected using the proposed 2017, pp. 230-233.
method were obtained for NPADB (2.63 ms; 0.46 bpm; when [14] D. S. Benitez et al., “A new QRS detection algorithm
based on the Hilbert transform,” in Proc. Computers in
sitting, walking, and ascending), QTDB (0.23 ms; 0.91 bpm), Cardiology, Cambridge, MA, USA, 2000, pp. 379-382.
and NSTDB (9.63 ms; 15.87 bpm; total average of six noise [15] A. K. Dohare, V. Kumar, and R. Kumar, “An efficient
levels). new method for the detection of QRS in
electrocardiogram,” Comput Electr Eng, vol. 40, no. 5, pp.
Along with the development of mobile ECG sensors, our 1714-1730, Jul. 2014, DOI:
algorithm can be used to detect R peaks even during physical doi.org/10.1016/j.compeleceng.2013.11.004.
activities to extend its applicability to mobile healthcare in [16] E. Plesnik et al., “Detection of the electrocardiogram
daily lives. Future work will be undertaken to improve our fiducial points in the phase space using the Euclidian
distance measure,” Med Eng Phys, vol. 34, no. 4, pp. 524-
method for applicability to various ECG waveforms (e.g., 529, May. 2012. DOI:
atrial premature beat, ventricular beat, and paced beat) in a 10.1016/j.medengphy.2012.01.005.
variety of activity states. [17] C. L. Chen and C. T. Chuang, “A QRS Detection and R
Point Recognition Method for Wearable Single-Lead
ECG Devices,” MDPI Sensors, vol. 17, no. 9, pp. 1969-
REFERENCES 1987, Aug. 2017, DOI: 10.3390/s17091969.
[18] N. R. Lang et al., “Filter and Processing Method to
[1] M. Rakshit and S. Das, “An efficient wavelet-based Improve R-Peak Detection for ECG Data with Motion
automated R-peaks detection method using Hilbert Artefacts from Wearable Systems,” in Proc. CinC, Nice,
transform,” Biocybern Biomed Eng, vol. 37, no. 3, France, 2015, pp. 917-920.
pp.566-577, Feb. 2017, DOI: [19] D. Tantinger et al., “Evaluation of QRS detection
doi.org/10.1016/j.bbe.2017.02.002. algorithm implemented for mobile appli-cations based on
[2] C. Venkatesan et al., “ECG signal preprocessing and ECG data acquired from sensorized garments,” Biomed
SVM classifier-based abnormality detection in remote Tech (Berl), vol. 57, no. SI-1 Track-F, pp. 635-638, Aug.
healthcare application,” IEEE Access, no. 6, pp. 9764- 2012, DOI: 10.1515/bmt-2012-4342.
9773, Jan. 2018, DOI: 10.1109/ACCESS.2018.2794346. [20] P. Laguna et al., “A Database for Evaluation of
[3] S. Poungponsri and X. H. Yu, “An adaptive filtering Algorithms for Measurement of QT and Other Waveform
approach for electrocardiogram (ECG) signal noise Intervals in the ECG,” in Proc. Computers in cardiology,
reduction using neural networks,” Neurocomputing, vol. Lund, Sweden, 1997, pp. 673-676.
117, no. 6, pp. 206-213, Oct. 2013, DOI: [21] C. Saritha, V. Sukanya, and Y. N. Murthy, “ECG signal
10.1016/j.neucom.2013.02.010. analysis using wavelet transform,” Bulg. J. Phys, vol. 35,
[4] H. Rabbani et al., “R peak detection in electrocardiogram no. 1, pp. 68-77, Feb., 2008.
signal based on an optimal combination of wavelet [22] G. B. Moody and R. G. Mark, “The impact of the MIT-
transform, Hilbert transform, and adaptive thresholding,” BIH Arrhythmia Database,” IEEE Eng Med Biol Mag,
J Med Signals Sens, vol. 1, no. 2, pp. 91-98, May. 2011. vol. 20, no. 3, pp. 45-50, May, 2001, DOI:
[5] S. Evans et al., “Heart rate variability as a biomarker for 10.1109/51.932724.
autonomic nervous system response differences between [23] A. Taddei et al., “The European ST-T Database: standard
children with chronic pain and healthy control children,” for evaluating systems for the analysis of ST-T changes
J Pain Res, vol. 6, pp.449-457, Jun. 2013, DOI: in ambulatory electrocardiography,” Eur Heart J, vol. 13,
10.2147/JPR.S43849. no. 9, pp. 1164-1172, Sep. 1992, DOI:
[6] M. Bsoul, H. Minn, and L. Tamil, “Apnea MedAssist: doi.org/10.1093/oxfordjournals.eurheartj.a060332.
Real-time Sleep Apnea Monitor Using Single-Lead ECG,” [24] E. Spanò et al., “Low-Power Wearable ECG Monitoring
IEEE Trans Inf Technol Biomed, vol. 15, no. 3, pp. 416- System for Multiple-Patient Remote Monitoring,” IEEE
427, May, 2011, DOI: 10.1109/TITB.2010.2087386. Sens J, vol. 16, no. 13, pp. 5452-5462, May, 2016, DOI:
[7] G. M. Friesen et al., “A comparison of the noise 10.1109/JSEN.2016.2564995.
sensitivity of nine QRS detection algorithms,” IEEE T [25] A. Jayant, T. Singh, and M. Kaur, “Different Techniques
bio-Med Eng, vol. 37, no. 1, pp. 85-98, Jan. 1990, DOI: to Remove Baseline Wander from ECG signal,” IJERMT,
10.1109/10.43620. vol. 2, no. 6, pp. 16-19, Jun. 2013.

VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2018.2867329, IEEE Access
Author Name: Preparation of Papers for IEEE Access (February 2017)

[26] P. De Chazal, M. O’Dwyer, and R. B. Reily, “Automatic [43] C. Li, C. Zheng, and C. Tai, “Detection of ECG
Classification of Heartbeats Using ECG Morphology and characteristic points using wavelet transforms,” IEEE T
Heartbeat Interval Features,” IEEE T bio-Med Eng, vol. bio-Med Eng, vol. 42, no. 1, pp. 21-28, Jan. 1995, DOI:
51, no. 7, pp. 1196-1206, Jun. 2004, DOI: 10.1109/10.362922.
10.1109/TBME.2004.827359. [44] B. U. Kohler, C. Henning, and R. Orglmeister, “QRS
[27] J. S. Park, S. W. Lee, and U. Park, “R Peak Detection Detection Using Zero Crossing Counts,” Int J Genomics
Method Using Wavelet Transform and Modified Proteomics, vol. 2, no. 2, pp. 138-145, Sep. 2003.
Shannon Energy Envelope,” J Healthc Eng, vol. 2017, no.
4901017, pp. 1-14, Jul. 2017, DOI:
10.1155/2017/4901017.
[28] D. Benitez et al., “The use of the Hilbert transform in
ECG signal analysis,” Comput Biol Med, vol. 31, no. 5,
pp. 399-406, Sep. 2001, DOI: 10.1016/S0010- MIRAN LEE received her B.S. degree in
4825(01)00009-9. computer science from Incheon National
[29] M. E. Nygårds and L. Sörnmo, “Delineation of the QRS University, and M.S. degree in electronics
complex using the envelope of the ECG,” Med Biol Eng engineering from Inha University. She works in
Comput, vol. 21, no. 5, pp. 538-547, Feb. 1983, DOI: Center for Bionics at Korea Institute of Science and
10.1007/BF024423. Technology, Seoul, Rep. of Korea.
[30] F. Zhang and Y. Lian, “QRS detection based on Her research interests include biomedical
morphological filter and energy envelope for applications interface, pattern recognition, bio-signal
in body sensor networks,” J Signal Process Syst, vol. 64, processing (EMG, EEG, EMG, and ECG), and bio-
no. 2, pp. 187-194, Nov. 2009, DOI: 10.1007/s11265- inspired systems.
009-0430-8.
[31] U. Melia et al., “Filtering and thresholding the analytic
signal envelope in order to improve peak and spike noise
reduction in EEG signals,” Med Eng Phys, vol. 36, no. 4,
pp. 547-553, Apr. 2014, DOI:
DAJEONG PARK received her B.S. degree
in Biomedical engineering from Inje University,
10.1016/j.medengphy.2013.11.014.
and M.S. degree in Bio-convergence Engineering
[32] M. S. Manikandan and K. P. Soman, “A novel method for
from Korea University. She works in Center for
detecting R-peaks in electrocardiogram (ECG) signal,”
Bionics at Korea Institute of Science and
Biomed Signal Process Control, vol. 7, no. 2, pp. 118-128,
Technology, Seoul, Rep. of Korea.
Mar. 2012, DOI: 10.1016/j.bspc.2011.03.004.
Her current research interests include bio-
[33] N. Rastogi and R. Mehra, “Analysis of Savitzky-Golay
signal processing, biomedical system, and pattern
filter for baseline wander cancellation in ECG using
recognition.
wavelets,” IJESET, vol. 6, no. 1, pp. 15-23, Aug. 2013.
[34] S. Agrawal and A. Gupta, “Fractal and EMD based
removal of baseline wander and powerline interference
from ECG signals,” Comput Biol Med, vol. 43, no. 11,
pp. 1889-1899, Jul. 2013, DOI:
10.1016/j.compbiomed.2013.07.030. SUH-YEON DONG received the B.S., M.S.,
[35] J. Arteaga-Falconim, H. AI Osman, and A. EI Saddik, “R- and Ph.D. degrees from the School of Electrical
peak detection algorithm based on differentiation,” in Engineering, Korea Advanced Institute of Science
Proc. WISP, Siena, Italy, 2015, pp. 1-4. and Technology, Daejeon, Korea, in 2010, 2011,
[36] B. U. Kohler, C. Henning, and R. Orglmeister, “The and 2016, respectively. She is currently an assistant
principles of Software QRS detection,” IEEE Eng Med professor at the department of Information
Biol Mag, vol. 21, no. 1, pp. 42-57, Aug. 2002, DOI: Technology Engineering, Sookmyung Women’s
10.1109/51.993193. University, Seoul, Rep. of Korea.
[37] A. Ghaffari et al., “Segmentation of Holter ECG waves Her current research interests include machine
via analysis of a discrete wavelet-derived multiple learning based bio-signal processing and cognitive
skewness-kurtosis based metric,” Ann Biomed Eng, vol. neuroscience.
38, no. 4, pp. 1497-1510, Jan. 2010, DOI:
10.1007/s10439-010-9919-3.
[38] D. Rezgui and Z. Lachiri, “EMD Method for automatic
ECG fiducial points detection,” in Proc. IPAS, INCHAN YOUN received his M.S. degree in
Hammament, Tunisia, 2016, pp. 1-5. mechanical engineering from Pittsburgh
[39] A. Ghaffari et al., “Parallel processing of ECG and blood University, and Ph.D. degree in Biomedical
pressure waveforms for detection of acute hypotensive Engineering from Tulane University. He is a
episodes: a simulation study using a risk scoring mode,” principal research scientist in Center for Bionics at
Comput Methods Biomech Biomed Engin, vol. 13, no. 2, Korea Institute of Science and Technology, Seoul,
pp. 197-213, Mar. 2009, DOI: Rep. of Korea.
10.1080/10255840903099711. His research interests include biomechanics,
[40] J. P. Martinez et al., “A wavelet-based ECG delineator, rehabilitation system development, man-machine
evaluation on standard databases,” IEEE T bio-Med Eng, interface (nerve signal interface and imaging
vol. 51, no. 4, pp. 570-581, Mar. 2004, DOI: interface), theragnosis (diagnosis of arthritis).
10.1109/TBME.2003.821031.
[41] Q. Qin et al., “An adaptive and time-efficient ECG R-
peak detection algorithm,” J Healthc Eng, vol. 2017, no.
5980541, pp. 1-14, Sep. 2017, DOI:
10.1155/2017/5980541.
[42] A. Thungtong, “A robust algorithm for R peak detection
based on optimal discrete wavelet transform,” in Proc.
JCSSE, Nakhon Si Thammarat, Thailand, 2017, pp. 1-6.

VOLUME XX, 2017

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
View publication stats http://www.ieee.org/publications_standards/publications/rights/index.html for more information.