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/JIOT.2017.2670022, IEEE Internet of
Things Journal
Real-Time Signal Quality-Aware ECG Telemetry
System for IoT-Based Health Care Monitoring
Udit Satija, Student Member, IEEE, Barathram.Ramkumar, and M. Sabarimalai Manikandan, Member, IEEE
Abstract—In this paper, we propose a novel signal quality-
aware IoT-enabled ECG telemetry system for continuous car-
diac health monitoring applications. The proposed quality-aware
ECG monitoring system consists of three modules: ECG signal
sensing module; automated signal quality assessment module;
and signal-quality aware ECG analysis and transmission module.
The main objectives of this paper are: design and development
of a light-weight ECG signal quality assessment method for
automatically classifying the acquired ECG signal into acceptable
or unacceptable class and real-time implementation of proposed
IoT-enabled ECG monitoring framework using ECG sensors,
Arduino, Android phone, Bluetooth and cloud server. The pro-
posed framework is tested and validated using the ECG signals
taken from the MIT-BIH arrhythmia and Physionet Challenge
databases and the real-time recorded ECG signals under dif-
ferent physical activities. Experimental results show that the
proposed SQA method achieves promising results in identifying
the unacceptable quality of ECG signals and outperforms existing
methods based on the morphological and RR interval features
and machine learning approaches. This study further shows
that the transmission of acceptable quality of ECG signals can
significantly improve the battery lifetime of IoT-enabled devices.
The proposed quality-aware IoT paradigm has great potential for
assessing clinical acceptability of ECG signals in improvement of
accuracy and reliability of unsupervised diagnosis system.
Keywords—Internet of things (IoT), wearable medical devices,
wireless healthcare monitoring, m-Health, ECG signal quality
assessment, medical body area networks
I. I NTRODUCTION
Internet-of Things (IoT)-driven health and wellness moni-
toring systems enable remote and continuous monitoring of
individuals, with applications in chronic conditions such as
obesity, hypertension, diabetes, hyperlipidemia, heart failure,
asthma, depression, elderly care support, preventive care and
wellness [1]- [11]. The IoT paradigms can play a significant
role in improving the health and wellness of subjects by
increasing the availability and quality of care, and dramat-
ically lowering the treatment costs and frequent travel. The
IoT-driven healthcare system employs networked biosensors
to simultaneously collect multiple physiological signals and
wireless connectivity to share/transmit gathered signals directly
to the cloud diagnostic server and the caregivers for further
analysis and clinical review. Further, the IoT-enabled remote
monitoring applications can significantly reduce travel, cost
and time in long-term monitoring applications.
Udit Satija, Barathram.Ramkumar, and M. Sabarimalai Manikandan
are with School of Electrical Sciences, Indian Institute of
Technology Bhubaneswar, Odisha, India. e-mail: us11@iitbbs.ac.in,
barathram@iitbbs.ac.in, msm@iitbbs.ac.in
Copyright (c) 2017 IEEE. Personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained from
the IEEE by sending a request to pubs-permissions@ieee.org.
2327-4662 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
In the health and wellness monitoring environments, the IoT
has emerged as one of the most powerful information gathering
and sharing paradigms for personalized healthcare systems [1],
[2], [6], ambient assisted living [7], user’s posture detection
[8], activity recognition [9], [13]–[15]. D. He, and S. Zeadally
[1] presented the RFID authentication schemes for internet of
things in healthcare environment using elliptic curve cryptogra-
phy. L. Catarinucci et al. [2] proposed a novel IoT-aware smart
hospital system (SHS) for automatic monitoring and tracking
of patients, personnel, and biomedical devices within hospitals
and nursing institutes. R. C. Alves [3] proposed wireless sen-
sor networks (WSNs) and IoT-based physical therapy system
for real-time assessing joint angle and vital signs to assist
physical therapists. In wireless healthcare monitoring systems,
the communication network longevity is a major challenge
for energy constrained power supplies of tiny sensor nodes
in WBAN [4], [5]. Some studies showed that about 80% of
total energy is consumed only for communication purpose [5].
T. Ishtaique ul Huque et al. [4], illustrated the importance of
effective body node coordinator (BNC) placement within a
wireless body area network (WBAN) to maximize the network
longevity. Three different BNC placement algorithms were
presented considering different features of available energy
efficient routing protocols in a WBAN. O. Omeni et al. [5]
proposed energy efficient medium access protocol for wireless
medical body area sensor networks. B. Mortazavi et al. [6]
presented an approach to integrate contextual information
within the data processing flow in order to improve the
quality of measurements. O. Yurur et al. [8] proposed a light-
weight online classification method to detect smarthpone user’s
postural actions, such as sitting, standing, walking, and running
by analyzing the data acquired from the smartphone equipped
accelerometer sensors. Ahmadi et al. [9] proposed a novel
ambulatory motion analysis method using wearable inertial
sensors to accurately assess all of activities of athlete for injury
and to enhance athlete performance. Xin Qi et al. [13] proposed
RadioSense, a system that exploits wireless communication
patterns for body sensor network (BSN) activity recognition.
II. M OTIVATION AND R ELATED W ORK
In real-world applications, most IoT-enabled devices operate
on limited power batteries for prolonged time periods. The
energy consumption becomes a crucial design consideration to
enhance the entire network lifetime [11]. Therefore, controlling
data exchange can minimize the energy expenditure and further
improve network utilization. P. Kolioset et al. [11] studied the
potential of data-driven event triggering in a real-world public
transport tracking scenario. The study showed that the data-
driven event triggering has great potential to improve energy
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/JIOT.2017.2670022, IEEE Internet of
Things Journal

1
efficiency and network utilization [11]. A long-term continuous

Amp.
0.5 (a)

cardiac health monitoring system highly demands more battery 0

0 1 2 3 4 5 6 7 8 9 10
power for real-time transmission of electrocardiogram (ECG) 10

signals and increases bandwidth and treatment costs and diag-

Amp.
0 (b)

nostic server traffic load. The event-triggering is a promising −10

0 1 2 3 4 5 6 7 8 9 10
paradigm that has the potential to adequately address the 5

Amp.
aforementioned key challenges. In our previous works, an 0 (c)

−5
automated low-complexity robust cardiac event change detec- 0 1 2 3 4 5 6 7 8 9

tion method was proposed for long-term healthcare monitoring 2

Amp.
1 (d)
applications [16]. Literature studies showed that the battery 0

lifetime of IoT-enabled devices is highly influenced by the 0 1 2 3 4 5 6 7 8 9 10

1
power consumption of the communication network utilization

Amp.
(e)
0
for continuously sharing the sensor data to cloud server and −1
0 1 2 3 4 5 6 7 8 9 10
the on-device embedded processor for event detection and 0.5
modeling in real-time. Therefore, intelligent solutions are

Amp.
0 (f)

demanded to improve battery lifetime of IoT devices and to −0.5

0 1 2 3 4 5 6 7 8 9 10
reduce diagnostic traffic load and bandwidth utilization costs. Time (s)

In practice, the ECG signals are corrupted by various kinds Fig. 1: Illustrates ECG signals corrupted with different kinds
of artifacts and noise including, flat line (FL) due to the of artifacts and noise. The ECG signals are taken from the
electrode disconnection, baseline wander (BW) due to the Physionet ECG databases. (a)-(c) shows the records with pause
respiration, abrupt change (AC) due to the physical activities, and abrupt baseline wander. (d) ECG signal corrupted with
muscle artifacts due to the muscle contraction, power line baseline wander. (e) and (f) ECG signals corrupted with severe
interference (PLI) and recording instrument noise [17]. For muscle artifacts.
visual inspection, the noisy ECG signals taken from the
standard ECG databases are shown in Fig. 1. Under physical
activity conditions, the quality of ECG signals is shown in
Fig. 2. The ECG signals are often corrupted by extraneous
disturbances due to muscular activity (i.e., electromyographic
(EMG) or muscle noise) that are ubiquitous in wearable cardiac
health monitoring devices under ambulatory and high-intensive
exercise ECG recordings [36]. Consequently, the muscle arti-
facts degrade the signal quality, spectral resolution and results
in large amplitudes. For example, the small amplitude local
waves (P, T, U and small QRS complex) are obscured by
the larger amplitude muscle artifacts. Thus, it is difficult (a) Subject conditions while ECG signal
recording.
for the physicians to visualize and locate the presence or
absence of these low-amplitude local waves that can provide 2000

significant clinical information required to diagnose certain

Amp.

0 (i)

heart abnormalities. Moreover, it is also difficult to perform −2000

0 1 2 3 4 5 6 7 8 9 10
more accurate and reliable measurement of morphological Time (s)

parameters such as amplitude, duration, pause interval, timings, 5000

polarity, and shapes of the local waves that are most important
Amp.

0 (ii)

for accurate heart rate monitoring, arrhythmia detection and −5000

0 1 2 3 4 5 6 7 8 9 10
heartbeat pattern recognition [35]. Further, transmission of Time (s)

bad quality ECG signals to the cloud server can reduce
lifetime of on-device battery in body area networks as well as
increases false alarm rates in unsupervised health monitoring
applications.
A. Existing Signal Quality Assessment Methods
In remote health monitoring, signal quality assessment
(SQA) is an intermediate step between ECG signal acquisi-
tion and transmission for clinical diagnosis. The SQA plays
a prominent role in applications which include automated
arrhythmia recognition, heart rate variability (HRV) analysis,
emotional recognition, biometric authentication, and unsuper-
vised health monitoring. In general, the SQA method grades
(b) Recorded ECG signals during (i) sitting position (ii)
movement or muscle activity.
Fig. 2: Illutrates the quality of ECG signals recorded under
different conditions of subject.
quality of an acquired ECG signal into five groups: excel-
lent, good, adequate, poor, and unacceptable [18]. However,
most SQA methods grade the ECG signal quality into two
groups: acceptable vs. unacceptable (or good vs. bad). Various
SQA methods were proposed based on energy-concavity index
(ECI), correlation [26], power in different sub-bands of ECG

2327-4662 (c) 2017 IEEE. Personal use is 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/JIOT.2017.2670022, IEEE Internet of
Things Journal
components, higher-order moments and spectral energy [18],
correlation and diversity approach [27], signal representation in
spectral-domain [28], cross-covariance among different ECG
signals [29], linear prediction [30], correlation-based kors
matrix [31] and regularity matrix [32], and variation of am-
plitudes in spectral-domain [19]. The SQA method includes
two major steps: feature extraction and signal quality ratings
using heuristic rules [19] and machine learning approaches
[20]. Most existing methods extract the PQRST morphological
features, RR interval features [21], PQRST shape regularity
[22], proportion of R-peak amplitude and noise-amplitude
[23], [24], consistency of QRS wave [25] for assessing the
acceptability of ECG signals.
Existing SQA methods highly demand robust methods for
accurate and reliable detection and measurement of morpho-
logical and RR-interval features from noise-free and noisy
ECG signals. Although the ECG morphology feature based
methods have shown promising results in noise-free ECG
recordings, accuracy and robustness of QRS complex detection
and waveform delineation methods are significantly degraded
in the presence of severe muscle artifacts and other external
noise. Machine learning SQA methods demand a large col-
lection of different patterns of ECG signals and various types
of ECG sources for accurate modeling of signal and noise
patterns. However, existing SQA methods may demand more
computational power.
B. Objectives and Contributions
Signal quality-aware paradigm is a promising technological
solution for the IoT that has great potential to significantly
improve the resource utilization efficiency (including, bat-
tery power consumption, network utilization, bandwidth and
treatment costs, cloud server traffic load) and to improve
the accuracy and reliability of unsupervised cardiac health
monitoring and diagnostic systems in wearable medical body
area network environments. With signal quality-aware IoT
paradigm, signal processing/event detection and communica-
tion modes can be put to sleep when unexpected level of
background noise and artifacts are detected from ECG signals
that can significantly alter the morphological features of the
ECG signal. Processing and analyzing severity of ECG noise
sources can play a significant role in improvement of resource
utilization efficiency by reducing high number of computation
and communication actions on severely corrupted ECG signals
which exhibit significant variations in morphological features
of the local waves.
In this paper, we propose a novel signal quality-aware ECG
telemetry system for IoT-enabled cardiac health monitoring
in wearable medical body area networks. The main objective
of this paper is to present a light-weight real-time signal
quality assessment technique for improving battery lifetime of
IoT-enabled wearable devices and reducing the cloud server
traffic load, bandwidth and treatment costs. Overall, this paper
includes the following key contributions:
• Design and development of an automatic and fast signal
quality assessment technique for grading the acquired
ECG signals into acceptable and unacceptable.
2327-4662 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
• Demonstrates the feasibility of real-time implementation
of the proposed signal quality-aware IoT framework
using ECG sensors, Arduino, Android-smartphone, and
communication networks (including, Bluetooth, WiFi,
cellular networks)
• Demonstrates the applicability of the proposed quality-
aware IoT paradigm in improvement of battery power
consumption as compared to the traditional continuous,
periodic and event-change triggering solutions that are
currently being used in state-of-the-art medical body
area networks.
• Demonstrates the significant distortions of the local
waves such as P-wave, QRS-complex and T-wave of the
ECG recordings under resting, ambulatory and physical
activity environments.
III. P ROPOSED S IGNAL Q UALITY-AWARE I OT-E NABLED
ECG T ELEMETRY F RAMEWORK
The main modules of our signal quality-aware (SQA)-IoT
framework are illustrated in Fig. 3. It consists of three modules:
(i) ECG signal sensing module, (ii) automated signal quality
assessment module, and (iii) signal-quality aware ECG analy-
sis and transmission module. In this paper, we mainly focus on
design and real-time implementation of automated ECG signal
quality assessment method and validation of the effectiveness
of the proposed SQA-IoT framework under resting, ambulatory
and physical activity conditions.
The proposed automated ECG signal quality assessment
(ECG-SQA) method consists of three steps such as flat-line
(or ECG signal absence) detection, abrupt baseline wander
extraction, and high-frequency noise detection and extraction
to compute the signal quality index (SQI) for assessing the
clinical acceptability of ECG signals. In this work, the ECG-
SQA is implemented based on the discrete Fourier transform
(DFT)-based filtering, turning points and decision rules.
A. Baseline Wander Removal and Abrupt Change Detection
In practice, the ECG signals are corrupted by baseline wan-
ders that are mainly caused by respiratory activity, body move-
ments, skin-electrode interface, varying impedance between
electrodes and skin due to poor electrode contact and per-
spiration [34]. Under physical activities, the baseline wanders
can degrade the ECG signal quality and may severely affect
the PQRST complexes. The presence of baseline wanders with
abrupt changes makes determination of characteristic points of
local waves of ECG beats more difficult. The frequency of the
baseline wander lies below 0.8 Hz (upto 1 Hz during stress)
[34].
In this work, we perform baseline wander detection using
the discrete Fourier transform (DFT) based filtering approach.
Let x[n], n = 0, 1, 2, . . . M − 1 be a discrete-time ECG signal.
The DFT of x[n] is computed as
N −1
X −jnπk
X[k] = x[n]e N (1)
n=0
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/JIOT.2017.2670022, IEEE Internet of
Things Journal

Medical Data
Data Acquisition Signal Quality Assessment Analysis
Data
Compression And Visualization

Feature
Extraction

Turning Point
Event/Pattern
Add
Noise Sequence Based Threshold Recognition
ECG leads
ECG Absence (or FL) Detection Quality-Aware Signal Analysis
& Diagnostic Model Doctor
Yes
ECG Data DFT-based LF Noise Maximum Absolute SQI Acceptable
Extraction Amplitude Threshold Computation SQI = 0 Cloud
Acquisition
Diagnostic
Abrupt Baseline Wander Detection Center
No Bluetooth

Segment into Non- Turning Point Unacceptable Wi-Fi

overlapping Blocks Measurements

HF Noise Detection Internet

Hospital
Signal Quality Assessment Framework Signal Quality-Aware
Communication Network

Fig. 3: Proposed signal quality-aware IoT framework for energy and resource-efficient ECG telemetry system.

X[k] denotes the kth DFT coefficient. The baseline component (Q) of 50%. Then kth block of b(n) is obtained as
is extracted from the DFT coefficients having the frequency  
Pk
values below 1 Hz. The DFT coefficient index k for the bk (n) = b +n n = 1, 2, · · · L (5)
component of F Hz is computed as k = b FFNs c. The baseline 2
wander signal is obtained as where bk (n) is the kth block of b(n) and L is the number
b[n] = x[n] − x̃[n] (2) of blocks. Let ak be the local dynamic range of bk (n). Then
relative dynamic amplitude difference is computed as
where b(n) is the baseline wander signal and x̃(n) is the
baseline wander removed signal which is computed as di = ai+1 − ai , i = 1, 2, . . . L − 1 (6)
N −1 Then the estimated di is compared with a predefined threshold
1 X jnπk
x̃(n) = X̃(k)e N (3) for detection of abrupt baseline drift event. The detection
N n=0 decision rule is given by

X̃ denotes the thresholded DFT coefficient vector which is 1 max{|d|} > γ
SQIABW = (7)
obtained as X̃(k) = [0, . . . , 0, X[k + 1], . . . X[N − k − 0 otherwise
1], 0, . . . , 0]. Fig. 4 shows the baseline wander signal extracted
from the ECG signal. From the results, it can be noted that where SQIABW denotes the signal quality index (SQI) for the
the baseline wander exhibits high amplitude variation for short presence of abrupt baseline drift (ABW) event. Here, the γ
duration. Thus, the extracted signal is further processed to is chosen as 0.2 mV for detecting abrupt baseline wander
discriminate the abrupt baseline drift from slowly varying which can distort the ST segment and other low-frequency
baseline wanders. For detecting the presence of baseline, we components of the ECG signal. The ECG signal with abrupt
use dynamic amplitude range of b(n) computed and compared baseline wander event is classified as an unacceptable. The
with predefined threshold of 0.1 mV which is chosen based ECG signal with slowly varying baseline wander is classified
on normal minimum amplitude of P-wave [17]. The dynamic as an acceptable since it can be removed from the signal
amplitude range of b(n) is computed as without significantly distorting the PQRST complexes.

Ag = max{|b[n]|} (4) B. ECG Signal Absence Detection

where Ag is the dynamic amplitude value. If Ag exceeds
a predefined threshold, the presence of either only baseline
wander or abrupt baseline drift is detected. In some scenarios,
the baseline component may be removed without distorting
the ECG signal information but a removal of abrupt baseline
drift will be challenging task without distorting the components
of the ECG signal [34]. Thus, the baseline signal is further
processed to detect the presence of abrupt baseline drift.
The baseline wander signal b(n) is first segmented into
blocks with a block size (P ) of 500 ms and block overlap
Due to the disconnection of electrodes with skin and the
electronic component saturation, sensing device exhibits the
absence of ECG signal information in the acquired signal. In
practice, we observe that the recording shows the presence of
zero amplitude flat line (ZFL), only baseline wander (OBW),
and the long pause with physiological and external noises.
Existing approaches were developed for detection of ZFL
event. In this work, we present a novel approach for detecting
aforementioned noise events. Our approach is based on turning
points (TP) which can be computed as mentioned in Algorithm

2327-4662 (c) 2017 IEEE. Personal use is 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/JIOT.2017.2670022, IEEE Internet of
Things Journal

10
as follows:
5
• Add uniformly distributed random noise of very small
Amp.
(a)
0 amplitude into the x̃[n], i.e.,
−5
0 2 4 6 8 10 s[n] = x̃[n] + aw[n] (8)
2

1
The random noise is added into the signal to increase
a number of turning points. The amplitude scaling a is
Amp.

0 (b)

−1 set to 0.01 mV. In the absence of ECG signal, the noisy

−2
0 2 4 6 8 10
signal s(n) results in a large number of turning points.
Time (s) • Compute the number of turning points (tp ) for the noisy
signal s(n) with a threshold.
Fig. 4: Illustrates the detection of abrupt baseline wander. (a)
• Apply decision rule for detecting the absence of ECG
The ECG signal corrupted with abrupt baseline wander taken
signal (or flat line segment),
from MIT-BIH arrhythmia database record number 116. (b)
The baseline wander signal b(n) obtained from equation (2).

1 tp > 0.65M
SQIFL = (9)
0 otherwise
If SQIFL = 1 then the segment is classified as a flat line
Algorithm 1 Turning point computation algorithm
or ECG signal absence segment. Otherwise the segment
Input: x(n) → ECG signal, γH → Threshold
is further processed for detecting the presence of high
Step 1: For turning point calculation, we first compute both positive and negative slopes
frequency noises.
which can
 be computed as,
1 :
 if (x(n) − x(n − 1) > 0) In practice, a presence of sudden long pause event may be
SL = 0 : if (x(n) − x(n − 1) = 0) seen in the real-time ECG recordings due to the abnormal
health condition of subjects. Therefore, the signal quality index

−1 : if (x(n) − x(n − 1) < 0)
Step 2: Then, for computing the local maxima and minima count all the positive and
information is stored/transmitted to the diagnostic server for
negative slopes respectively. That is,
each processed ECG segment. Furthermore, subject is imme-
NSL = 0; I = [ ];
diately altered when an absence of ECG signal is detected,
for(i = 1 : length(SL))
like in the conventional ECG monitoring systems.
if((SL(i) − SL(i − 1) < 0)||(SL(i) − SL(i − 1) > 0))
NSL = NSL + 1; I = [I, i]
end C. HF Noise Detection
end In practice, the high frequency (HF) noises such as muscle
Here, NSL is total number of turning points and I contains indices of turning points. artifacts, powerline interference, motion artifacts, pause and
Step 3: Now, compute the turning points (which are greater than γH ) and their locations instrument noise are introduced in the acquired ECG signal.
as, These HF noises can obscure the local waves (P, T, U and
z = []; z[1] = 0; small amplitude QRS complex) of the ECG signal. Thus, it is
for(j = 2 : length(I)) difficult to perform more accurate and reliable measurement of
z[j] = x(I(j)) − x(I(j − 1)); morphological parameters such as amplitude, duration, pause
end interval, timings, polarity, and shapes of the local waves.
tp = 0; tploc = [ ]; //tp → turning points; tploc = location of turning Removal of HF noises especially, muscle noise is quite chal-
points lenging without distorting the local waves of the signal. Thus,
for(i = 1 : length(I) − 1) this paper presents a simple HF noise detection based on the
if(|z(i)| > |Thr|) turning points and thresholding rule, which is described as:
tploc = [tploc , I(i)]; • Normalize the ECG signal with maximum amplitude
tp = tp + 1; value.
end • Segment the normalized signal x̃(n) into non-
end overlapping blocks of block size (P ) of 2 s.
x̃k (n) = x̃(Pk + n) n = 1, 2, · · · P (10)
1. The turning point with zero threshold can provide estimates where k = 0, 1, · · · , M1 − 1, and M1 = b M
P c
of local maximum and local minimum points in the ECG • Compute the number of turning points and the location
signal. The acquired signal may be considered as random if of the turning points for each block x̃k (n) using the
number of turning points exceeds the (2/3)th of length of a Algorithm 1. For measurements of number of turning
signal [37]. This is the basis for ECG signal absence detection points, an amplitude threshold γH is set to 0.05 mV
(or flat line detection). To compute the turning points, the based on acceptable level of HF noise which cannot
value of γH is set to zero. In absence of ECG signal, the distort the small-amplitude local waves of ECG signal.
recorded segment is classified as an “unacceptable”. The flat • Find the minimum distance between two turning points
line detection (or ECG signal absence detection) algorithm is i.e., rk = min{tplock (i + 1) − tplock (i)}, f or i =

2327-4662 (c) 2017 IEEE. Personal use is 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/JIOT.2017.2670022, IEEE Internet of
Things Journal

1, 2, · · · L1 , where, L1 is the length of tplock in kth

block.
• Apply decision rules for detecting the ECG segment with
HF noises,

1 R1 || R2
SQIHF = (11)
0 otherwise
where R1 is true if a total number of turning points in (a) ECG signal corrupted with low-frequency and high-
any block (tpk ) exceeds 5% of the number of samples frequency noises.
(P ∗ F s, where, F s is the sampling frequency of the
signal) in that block and R2 is true if the distance
between two turning points rk is less than two.
If SQIHF = 1 then the segment is classified as a HF
noise segment. Otherwise the segment is classified as a high-
frequency noise-free ECG segment.

D. ECG Signal Quality Grading

In this work, we perform the grading of acquired ECG (b) ECG signal corrupted with abrupt change and pause.
signal based on the decision scores obtained for the detection
of abrupt baseline wander, ECG signal absence, and high Fig. 5: Screenshot of the output of the proposed method for
frequency noises. The ECG signal is graded with three classes bad quality ECG signal obtained under extensive activity.
such as Good, Intermediate and Bad based on the HF noise
score. The description of ECG signal quality rating is summa-
rized in Table I. The presence of flat line and abrupt baseline
wander may result in noisy clinical features. Based upon
assessment results, it is noted that the some morphological
features and RR intervals can be measured from the ECG
signal with some level of HF noises. Thus, we grade the
noisy ECG signal into intermediate and bad. The value of
γH is chosen as 0.1% of the maxima or minima in the HF
noise detection stage for detecting severe noise or bad quality
signal. While the noise level threshold γH is set to 0.05%
of the maxima or minima for grading the noisy signal into Fig. 6: Real-time recording and assessing ECG signal quality.
intermediate class. Finally, signal quality grading (SQG) is
implemented as

Good SQI = 0 is evaluated under different kinds of single and combined ECG
SQG = Intermediate SQI = 0.5 (12) noises. The proposed algorithms are implemented on Android
phone for real-time evaluation.

Bad SQI > 0.5
where overall signal quality index (SQI) is computed as
A. Test Databases and Performance Metrics
SQI = SQIABW + SQIFL + SQIHF (13) To test the performance of the proposed automated ECG
In this study, good quality indicates the ECG signal with signal quality assessment method, a wide variety of ECG
acceptable level of noise which does not distort the morpholog- signals with different PQRST morphological patterns are taken
ical features while intermediate and bad indicate the medium from the standard ECG databases including MIT-BIH arrhyth-
and high level of noise or poor quality, respectively. We believe mia (MITBIHA) [38] and Physionet Challenge database [39].
that the intermediate class of ECG signals may be considered The MITBIHA database contains 48 two-channel records of
for application-specific ECG signal analysis systems. noise-free and noisy ECG signals with varying ECG waveform
morphologies. The ECG signals are digitized with a sampling
rate of 360 samples/second and 11-bit resolution [38]. The
IV. R ESULTS AND D ISCUSSION Physionet challenge database includes acceptable and unac-
In this Section, the proposed signal quality-aware IoT frame- ceptable quality of ECG signals which are sampled at a rate
work is evaluated and validated using a wide variety of ECG of 500 samples/second [39]. The noise portions of the ECG
recordings under resting, ambulatory and physical activity en- signals are labeled for performance evaluation. Both synthet-
vironments and standard benchmark metrics. The accuracy of ically created noisy ECG signals and real-time noisy ECG
the proposed automated ECG signal quality assessment method signals are used for evaluating the proposed SQA method.

2327-4662 (c) 2017 IEEE. Personal use is 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/JIOT.2017.2670022, IEEE Internet of
Things Journal

TABLE I: Description of signal quality rating based on the noise information.

Noise Class Rating Grading Signal Parameters/Threshold Description
ABW 0 Good max{|d|} < γ; relative dynamic Ensures the preservation of shape of ST segment such as ST elevation and depression.
amplitude difference
1 Bad max{|d|} > γ; relative dynamic Abrupt baseline wander distorts the shape of ST segment. Removal of abrupt baseline is
amplitude difference challenging task with preservation of morphological shapes of the signal.
0 Good tp < 0.65N ; tp → turning points, Represents noise-free ECG signal with preservation features of local waves.
M → length
ECG Absence/ 1 Bad tp > 0.65N ; tp → turning points, Indicates the absence of ECG signal in the record.
FL M → length
0 Good tpk < 0.05 ∗ P ∗ F s&&rk > 2; Represents the noise-free ECG signal with preservation of all local waves.
Count tp for 0 < γH < 0.05 ∗ Ak
HF Noises 0.5 Intermediate tpk > 0.05 ∗ P ∗ F s||rk < 2; RR interval and some morphological features can be measured with acceptable level of
Count tp for γH > 0.05 ∗ Ak HF noise which may be removed using signal processing techniques by ensuring the
preservation of morphological features of the signal.
1 Bad tpk > 0.05 ∗ P ∗ F s||rk < 2; Under severe HF noisy environments, it is very difficult to ensure the preservation of
Count tp for γH > 0.1 ∗ Ak morphological features such as amplitude, duration, polarity and shapes of P-wave, QRS-
complex and T-wave.

The noise sources include abrupt baseline wander, muscle
artifacts, powerline interference, slowly varying baseline, flat
line with noise and without noises, instrument noise, additive
white Gaussian noise, and motion artifacts. The performance
of the proposed SQA method is evaluated using sensitivity (Se)
which is computed as
TP
Se(%) = × 100 (14)
FN + TP
where TP denotes the true positive that represents correctly
detected segments by the method and FN denotes the false
negative that represents missed segments by the method.
B. Real-time Validation of SQA-IoT ECG Telemetry System
For real-time performance analysis, experiments were con-
ducted with twenty subjects and Smartphone. The experimental
set-up for real-time recording and assessing ECG signal quality
is shown in Fig. 6. The 15-min duration ECG signals are
recorded for each of the subjects under different physical
activity conditions. The subjects are in sitting position for 10-
min duration and then are asked to perform activities including
walking, jumping, running and fast hand movements for 5-min
duration. The purpose of this experiment is to collect motion
artifacts under real-time Holter recording. In this experiment,
the ECG signals are continuously transmitted to the Smart-
phone via Bluetooth. The proposed signal quality assessment
method is implemented on Smartphone with Lollipop 5.0
android version, 1.2 GHz quad-core Qualcomm Snapdragon
410 MSM8916 processor and 2 GB RAM. The android Smart-
phone application framework transmits not only the acceptable
quality of ECG signals but also sends the physical activity
information of the subject. The physical activity recognition is
performed using the features of the acceleration signals from
the triaxial accelerometer sensor embedded on the Smartphone
[33]. Evaluation results show that the Smartphone application
framework has capability to process real-time ECG signals and
send the noise-free ECG signals to the cloud server or doctor.
C. Influence of Physical Activity on ECG Quality
To investigate the influence of the physical activity on
the ECG signal quality, the ECG signals are recorded for
longer-duration. The ECG signals are recorded for four activity
scenarios including sitting, walking, jogging, and body move-
ments. The acquired ECG signals are shown in Fig. 7. Fig.
5 shows the bad quality ECG signal that was collected from
the subject during intensive activity condition. The recorded
ECG signal shows that it is difficult to extract clinical features
more accurately and reliably. The acquired ECG signal in
Fig. 5 (b) shows the presence of abrupt baseline and pause
under physical activity condition. Fig. 8 shows the quality of
the ECG segments recorded during aforementioned physical
activities. The proposed system grades the ECG signal quality
into three grades: good, intermediate and bad. The experiment
study shows that the quality of the acquired ECG signals is
good during sitting or normal and walking scenarios. The ex-
perimental results further show that the recorded ECG signals
are severely corrupted with high-frequency noise and artifacts
during jogging and other high extensive activities. Thus, the
recognition of physical activities can play a significant role
in understanding the context of ECG recordings. Further, it
can significantly reduce the processing and transmission power
when the ECG signals are acquired under high intensive
activity scenarios.
D. Performance of SQA Methods
The overall evaluation results for the ECG segments taken
from MITABIHA and Physionet Challenge databases and real-
time ECG signals acquired for twenty subjects are summarized
in Table II. From results, it is noted that the proposed SQA
method can achieve an average sensitivity (Se) of more than
98% and 99%, respectively for noise-free and noisy ECG sig-
nals taken from the well-known MITBIHA database. The SQA
method achieves an average Se of more than 99% and 94%,
respectively for the noise-free and noisy ECG signals taken
from the Physionet Challenge database. The SQA method
yields an average Se of more than 93% and 97%, respectively
for noise-free and noisy ECGs under real-time ECG recordings
during physical activities. For performance comparisons, three
SQA methods are implemented based on the QRS detection
features and template matching [21], QRS detection and RR-
interval features and heuristics rules [23], and statistical fea-
tures and QRS complex features with support vector machine

2327-4662 (c) 2017 IEEE. Personal use is 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/JIOT.2017.2670022, IEEE Internet of
Things Journal

0.5 TABLE II: Performance of the proposed SQA method.

Amp.

(a)
0
Database Signal Type Total True False Se
−0.5 Segments Positive Negative (%)
0 1 2 3 4 5 6 7 8 9 10
1 MITBIHA Noise-free ECGs 413 407 6 98.55
Noisy ECGs+FL 995 889 6 99.40
0.5
Amp.

(b) Physionet Noise-free ECGs 391 390 1 99.74

0 Challenge Noisy ECGs 311 293 18 94.21
−0.5 Only FL 75 75 0 100.00
0 1 2 3 4 5 6 7 8 9 10
Our Records Noise-free ECGs 1200 1118 82 93.17
0.4
Noisy ECGs 600 583 17 97.17
0.2
Amp.

(c)
0

−0.2
0 1 2 3 4 5 6 7 8 9 10
TABLE III: Performance comparison of SQA methods
0 SQA Method Signal TS TP FN Se (%) Avg. Time*
Hamilton-Tompkins based Clean 2004 1582 422 78.94 230 ms
Amp.

(d)
−0.5
QRS detector and Noisy 1906 1322 584 69.36
−1 Template matching [21]
0 1 2 3 4 5 6 7 8 9 10 11 12
Time (s)
R-peak and noise amplitudes Clean 2004 1598 406 79.74 270 ms
temporal features [23] Noisy 1906 1409 497 73.92
Fig. 7: Recorded ECG signals under different activities. (a) Sub-band power in ECG and Clean 2004 1827 177 91.17 350 ms
higher-order moments [18] Noisy 1906 1544 362 81.01
sitting, (b) walking, (c) jogging, and (d) intensive muscle Clean 2004 1915 89 95.56 7 ms
activity. Proposed Method Noisy 1906 1865 41 97.85
Note: * Computational time is calculated in MATLAB software

200
Good second ECG signal is processed using the proposed SQA
180 Intermediate
Bad
method for determining noise-free ECG signal (‘accepatble’)
160
and noisy ECG signals (‘unaccepatble’). Then, the commu-
Number of ECG Segments

140 nication network is activated for transmitting the acceptable

120
100
80
60
40
20
0
Sitting
Fig. 8: Demonstrates quality assessment of the ECG signals
collected during various physical activities.
classifier [18]. Table III summarizes the performance of four
SQA methods. From the results, it is noted that the proposed
SQA method outperforms other existing methods based on the
morphological and RR intervals features and machine learning
approaches. As compared with signal processing techniques of
exiting methods, the proposed automated SQA is simple and
thus suitable for assessing the quality of recorded ECG signals
in real-time environments.
E. Power Consumption and Computational Delay
To demonstrate the battery power saving, the proposed
algorithms are implemented on Android phone for real-time
acquiring, processing, and transmitting the acceptable quality
ECG signal. In this experiment, it is noted that the 15 min
normal ECG signal consumes 10% battery power of android
phone. For this experimental study, an one hour ECG signal
is divided into blocks with duration of 10 second. Each 10
quality of ECG signals to the cloud server or the another
Android phone for further clinical evaluation and review. The
experimental results demonstrate that the power consumption
is reduced upto 33% in transmitting only 241 good quality
segments out of 360 ECG segments and discarding 119 bad
quality segments including intermediate and bad quality using
the proposed SQA method. To study the computation load
Walking Jogging Extensive Muscle Activity
Subject Positions of the proposed method and the existing methods, the three
SQA methods are implemented in MATALB software. The
computational times for the feature extraction and quality
testing are summarized in Table III. The proposed SQA method
has an average computational time of 7 ms for processing of 10
second ECG segment. In the future directions, we further study
the effectiveness of the proposed signal quality-aware IoT
framework in improvement of resource utilization efficiency
(including, battery power consumption, network utilization,
bandwidth and treatment costs, cloud server traffic load) under
resting, ambulatory and physical activity conditions.
V. C ONCLUSION
In this paper, we present a novel signal quality-aware IoT-
enabled ECG telemetry system for cardiac health monitoring
applications. This paper proposes a light-weight ECG signal
quality assessment (ECG-SQA) method for automatically as-
sessing the quality of acquired ECG signals under resting,
ambulatory and physical activity environments. Experimen-
tal results demonstrate that the proposed ECG-SQA outper-
forms other existing methods based on the morphological
and RR interval features and machine learning approaches.
The experimental study further demonstrates that the ECG
signals are severely distorted under more intensive physical

2327-4662 (c) 2017 IEEE. Personal use is 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/JIOT.2017.2670022, IEEE Internet of
Things Journal
activities. Real-time evaluation results further show that the
proposed quality-aware ECG telemetry system significantly
reduces the battery power consumption by transmitting the
acceptable quality of ECG signals and putting the IoT devices
to sleep mode for the unacceptable ECG signals. From this
study, we believe that the proposed quality-aware IoT-enabled
cardiac health monitoring framework has significant potential
in improving the resource utilization efficiency of IoT-enabled
devices and the reliability of unsupervised ECG signal analysis
and diagnosis system by reducing the false alarm rates under
severe noisy ECG recordings.
R EFERENCES
[1] D. He, and S. Zeadally, “An analysis of RFID authentication schemes
for internet of things in healthcare environment using elliptic curve
cryptography,” IEEE Internet Things J., vol. 2, no. 1, pp. 72-83, 2015.
[2] L. Catarinucci, et al., “An IoT-aware architecture for smart healthcare
systems,” IEEE Internet Things J., vol. 2, no. 6, pp. 515-526, 2015.
[3] R. C. Alves, et al., “Assisting physical (hydro) therapy with wireless
sensors networks,” IEEE Internet Things J., vol. 2, no. 2, pp. 113-120,
2015.
[4] M. T. I. ulHuque, et al., “Body node coordinator placement algorithms
for wireless body area networks,” IEEE Internet Things J., vol. 2, no. 1,
pp. 94-102, 2015.
[5] O. Omeni, et al., “Energy efficient medium access protocol for wireless
medical body area sensor networks,” IEEE Trans. Biomed. Circuits Syst.,
vol. 2, no. 4, pp. 251-259, Dec. 2008.
[6] B. Mortazavi, et al., “Context-Aware data processing to enhance quality
of measurements in wireless health systems: an application to MET
calculation of exergaming actions,” IEEE Internet Things J., vol. 2, no. 1,
pp. 84-93, 2015.
[7] C. Tsirmpas, et al., “A new method for profile generation in an internet
of things environment: an application in ambient-assisted living,” IEEE
Internet Things J., vol. 2, no, 6, pp. 471-478, 2015.
[8] O. Yurur, C. H. Liu, and W. Moreno, “Light-Weight online unsupervised
posture detection by smartphone accelerometer,” IEEE Internet Things J.,
vol. 2, no. 4, pp. 329-339, 2015.
[9] A. Ahmadi, et al., “Toward automatic activity classification and move-
ment assessment during a sports training session,” IEEE Internet Things
J., vol. 2, no. 1, pp. 23-32, 2015.
[10] S. Amendola, R. Lodato, S. Manzari, C. Occhiuzzi, and G. Marrocco,
“RFID technology for IoT-based personal healthcare in smart spaces,”
IEEE Internet Things J., vol. 1, no. 2, pp. 144-152, 2014.
[11] P. Kolios, C. Panayiotou, G. Ellinas, and M.Polycarpou, “Data-Driven
event triggering for IoT applications,” IEEE Internet Things J., 2016.
[12] F. Bsching, and L. Wolf, “The rebirth of one-time pads secure data
transmission from BAN to sink,” IEEE Internet Things J., vol. 2, no. 1,
pp. 63-71, 2015.
[13] X. Qi, G. Zhou, Y. Li, G. Peng, “RadioSense: Exploiting wireless
communication patterns for body sensor network activity recognition,”
33rd IEEE Real-Time Systems Symposium (RTSS), pp. 95-104, 2012.
[14] D. Tao, Y. Wen, and R. Hong, “Multi-column Bi-directional long short-
term memory for mobile devices-based human activity recognition,” IEEE
Internet Things J., 2016.
[15] P. Casale, M. Altini, and O.Amft, “Transfer learning in body sensor
networks using ensembles of randomized trees,” IEEE Internet Things J.,
vol. 2, no. 1, pp. 33-40, 2015.
[16] U. Satija, B. Ramkumar, and M. S. Manikandan, “Robust cardiac event
change detection method for long-term healthcare monitoring applica-
tions,” IET Healthcare Technol. Lett., 2016.
[17] U. Satija, B. Ramkumar, and M. S. Manikandan, “A simple method for
detection and classification of ECG noises for wearable ECG monitoring
devices,” IEEE Int. Conf. Signal Process. Integrated Net. (SPIN), 2015.
2327-4662 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
[18] G. D. Clifford, J. Behar, Q. Li, and I. Rezek, “Signal quality indices and
data fusion for determining clinical acceptability of electrocardiograms,”
Physiol. Meas., vol. 33, no. 9, pp. 1419-1437, 2012.
[19] B. E. Moody, “Rule-based methods for ECG quality control,” Comput.
Cardiol., vol. 38, pp. 361-363, 2011.
[20] J. Behar, J. Oster, Q. Li, and G. D. Clifford, “ECG signal quality during
arrhythmia and its application to false alarm reduction,” IEEE Trans.
Biomed. Eng., vol. 60, no. 6, pp. 1660-1666, 2013.
[21] C. Orphanidou et al., “Signal quality indices for the electrocardiogram
and photoplethysmogram: derivation and applications to wireless moni-
toring,” IEEE J. Biomed. Health Informatics, vol. 19, no. 3, pp. 832-840,
2015.
[22] P. X. Quesnel, A. D. C. Chan, and H. Yang, “Signal quality and false
myocardial ischemia alarms in ambulatory electrocardiograms,” IEEE Int.
Symp. Med. Meas. Appl. (MeMeA), pp. 1-5, June 2014.
[23] D. Hayn, B. Jammerbund, and G. Schreier, “QRS detection based ECG
quality assessment,” Physiol. Meas., vol. 33, no. 9, pp. 1449-1462, 2012.
[24] D. Hayn, B. Jammerbund, and G. Schreier, “ECG quality assessment
for patient empowerment in mHealth applications,” Comput. Cardiol., pp.
353-356, Sept. 2011.
[25] L. Johannesen, and L. Galeotti, “Automatic ECG quality scoring
methodology: mimicking human annotators,” Physiol. Meas., vol. 33, no.
9, pp. 1479-1490, 2012.
[26] H. Naseri, and M. R. Homaeinezhad, “Electrocardiogram signal quality
assessment using an artificially reconstructed target lead,” Computer
Methods Biomech. Biomed. Eng., vol. 18, no. 10, pp. 1126-1141, 2015.
[27] F. J. Martinez-Tabares, J. Espinosa-Oviedo, and G. Castellanos-
Dominguez, “Improvement of ECG signal quality measurement using
correlation and diversity-based approaches,” 34th IEEE Int. Conf. Eng.
Med. Bio. Soc. (EMBC), 2012.
[28] D. Tobon Vallejo, T. Falk, and M. Maier, “MS-QI: A modulation
spectrum-based ECG quality index for telehealth applications,” IEEE
Trans. Biomed. Eng., 2014.
[29] E. Morgado, et al., “Quality estimation of the electrocardiogram using
cross-correlation among leads,” Biomed. Eng. Online, vol. 14, no. 1, p. 1,
2015.
[30] K. Noponen, M. Karsikas, S. Tiinanen, J. Kortelainen, H. Huikuri, and
T. Seppnen, “Electrocardiogram quality classification based on robust best
subsets linear prediction error,” Comput. Cardiol., pp. 365-368, Sep. 2011.
[31] A. C. Maan, et al., “Assessment of signal quality and electrode
placement in ECGs using a reconstruction matrix,” Comput. Cardiol., pp.
289-292, Sep. 2011.
[32] H. Xia, et al., “Computer algorithms for evaluating the quality of ECGs
in real time,” Comput. Cardiol., vol. 38, pp. 369-372, 2011.
[33] S. S. Kambhampati, V. Singh, M. S. Manikandan, and B. Ramkumar,
“Unified framework for triaxial accelerometer-based fall event detection
and classification using cumulants and hierarchical decision tree classifier,”
IET Healthcare Technol. Lett., vol. 2, no. 4, pp. 101-107, August 2015.
[34] U. Satija, B. Ramkumar, and M. S. Manikandan, “A robust sparse signal
decomposition framework for baseline wander removal from ECG signal,”
IEEE TENCON, Singapore, 2016.
[35] M. S. Manikandan, and K. P. Soman, “A novel method for detecting
R-peaks in electrocardiogram (ECG) signal,” Biomed. Signal Process.
Control, vol. 7, no. 2, pp. 118-128, 2012.
[36] U. Satija, B. Ramkumar, and M. S. Manikandan, “A unified sparse
signal decomposition and reconstruction framework for elimination of
muscle artifacts from ECG signal,” IEEE ICASSP, pp. 779-783, 2016.
[37] R.M. Rangayyan, “Biomedical signal analysis: A case study approach,”
Wiley, 2009
[38] G. B. Moody, R. G. Mark, “The impact of the MIT-BIH Arrhythmia
Database,” IEEE Eng. Med. Biol., vol. 20, no. 3, pp. 45-50, 2001. http:
//physionet.org/physiobank/database/mitdb
[39] Physionet. PhysioNet/Computing in Cardiology Challenge 2011.
http://www.physionet.org/challenge/2011/