Internet of Things 9 (2020) 100166

Contents lists available at ScienceDirect

Internet of Things
journal homepage: www.elsevier.com/locate/iot

Research article

Energy-Efficient IoT-Health Monitoring System using

Approximate Computing
Avrajit Ghosh a,∗, Arnab Raha b,1, Amitava Mukherjee c
a
Department of Electronics and Telecommunication Engineering, Jadavpur University, India
b
Intel Corporation, CA, USA
c
Department of Computer Science and Engineering, Adamas University, India

a r t i c l e i n f o a b s t r a c t

Article history: Wireless Body Sensor Nodes (WBSN) are frequently used for real time IoT-based health
Received 27 September 2019 monitoring of patients outside the hospital environment. These WBSNs involve bio-sensors
Revised 20 January 2020
to capture signals from a patient’s body and wireless transmitters to transmit the collected
Accepted 21 January 2020
signals to a server located in private/public cloud in real time. These WBSNs include hard-
Available online 30 January 2020
ware for processing of signals before being transmitted to the cloud. Simultaneous oc-
Keywords: currence of all these processes inside energy constrained WBSNs results in considerable
Wireless Body Sensor Nodes amount of power consumption, thus limiting their operational lifetime. Due to the inher-
IoT-based Health Monitoring ent error-resilience in signal processing algorithms, most of these data reaching the servers
Low Power Hardware Prototype are redundant in nature and hence of not much clinical importance. Transmission and stor-
ECG Signal age of these excess data result in inefficient usages of transmission bandwidth and storage
Discrete Wavelet Transform capabilities. In this paper, we develop a real time encoding scheme that performs iterative
Sparse Encoding thresholding and approximation of wavelet coefficients for sparse encoding of bio-signals
Approximate Computing
(ECG signals), thereby reducing the energy and bandwidth consumption of the WBSN. The
encoding scheme compresses bio-signals (ECG signals), while still maintaining the clini-
cally important features. We optimize various process parameters to model a low power
hardware prototype for the implementation of our algorithm on a real time microcontroller
based IoT platform that operates as an end-to-end WBSN system in real time. Experimen-
tal results show a system-level energy improvement of 96% with a negligible impact on
signal quality (2%).
© 2020 Elsevier B.V. All rights reserved.

1. Introduction

In the present world, the growth of cardiovascular diseases and interests in personal healthcare fuel up the requirement
for a system that provides accurate and non-stop medical supervision [1,2]. For chronic cardiovascular patients, the tradi-
tional healthcare systems entail two categories of monitoring, which are i) holter monitoring that requires a smaller time
range of about two days, and ii) event monitoring that spans across a week. These monitoring systems involve bulky and
obtrusive sensor devices like Electrocardiogram (ECG) sensors that limit patient autonomy [3]. Moreover, recording over

∗
Corresponding author.
E-mail addresses: avrajitghosh1001@gmail.com (A. Ghosh), arnabraha1989@gmail.com (A. Raha), amukh10@gmail.com (A. Mukherjee).
1
Arnab Raha contributed to this paper as an independent contributor. The views and opinions expressed in this article are those of the authors’ own
and do not necessarily reflect the views of Intel Corporation.

https://doi.org/10.1016/j.iot.2020.100166
2542-6605/© 2020 Elsevier B.V. All rights reserved.
2 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166

Fig. 1. Wearable body sensor node architecture.

multiple channels and in multiple sessions with sampling rate varying from 125-500 Hz, leads to a huge amount of data
generation [4]. Thus continuous medical supervision is unsustainable for traditional healthcare system due to increasing
costs resulting from energy consumption by the sensor nodes (devices), data transmission, and larger storage requirement
[5].
Wireless Body Sensor Nodes (WBSNs) have emerged as a feasible/portable solution to enable the long term acquisition
of bio-signals, outside of a hospital environment [6,7]. WBSNs are embedded sensor appliances consisting of low power
nodes aiding in the acquisition of ECG signals [8,9]. Smart WBSNs incorporate various innovative signal processing applica-
tions to reduce the ECG data to be transmitted [10]. The ambulatory ECG data collected is affected by noise like powerline
interference, electrode motion, muscle noise, baseline drift, etc [11]. WBSNs, operating on battery, are usually heavily energy-
constrained that requires the development of energy and storage-efficient embedded signal processing and data compression
algorithms. These processing algorithms should therefore be simple and of low complexity [12]. Also, the overall system la-
tency required to process and send data should be minimal as delay cannot be tolerated at the time of crisis [1,13].
A typical WBSN comprises the following three subsystems depicted in Fig. 1 ([14,15]):
• Sensing Subsystem: This comprises biosensors and analog to digital (ADC) converters. Analog biosignals acquired through
the biosensors are converted to digital form for processing in the next subsystem.
• Processing Subsystem: Noise filters and signal processing units with embedded algorithms for signal compression and en-
ergy optimization, form this block. Digital biosignals from the sensing subsystem are processed and prepared for trans-
mission by the next subsystem.
• Communication Subsystem: This comprises the Ultra Low-power WiFi module which is used to transmit the processed
signal data to a medical server.
As demonstrated in Fig. 1, the processed digital signal, along with the other relevant details of a patient are transmitted
to a medical server. Finally, this server sends the patient’s data to the doctor or healthcare provider or any other authority
under concern [16].
In order to perform the desired processing, the ensuing smart WBSNs embed low complexity computing capabilities that
have non-negligible energy footprints [17]. Note that the design considerations for WBSNs considered in this paper are a)
power consumption, which includes both transmission power and execution power, b) portability, c) real-time availability,
d) energy efficiency and e) scalability. Energy consumption of the system depends upon the sum of the energy of the indi-
vidual components in the hardware platform such as the sensing, processing, and communication subsystems as shown in
Fig. 1. However, the energy efficiency can be examined as the energy bottleneck through embedded digital signal process-
ing (DSP) algorithms[18]. Real-time availability depends upon the time complexity of the DSP algorithms and the hardware
latencies [5]. In the case of healthcare application, a real-time system is actually a soft real-time system in which some
latency is permitted [19]. Increasing the energy efficiency and performance are the two major driving factors considered
while designing a real-time IoT based health monitoring system. For energy-tolerant application such as ECG signal analysis,
guaranteeing 100% correctness is not always mandatory and we leverage this error-resilience to gain system-level energy
significantly in this work. In order to achieve this gain in the system level energy, this work first investigates the error toler-
ances of our proposed algorithms 1 and 2 (refer to Section 4) through regulation/optimization of certain design parameters
that are subsequently used to formulate an energy saving methodology using approximate computing. The experimental re-
sults demonstrate the significant system-level energy gains that can be achieved with negligible impact on the clinical value
of the transmitted signal.
A WBSN based device continuously monitors ECG signals from a patient’s body. This continuous monitoring creates ex-
cessive volumes of data that, in turn, requires larger storage space and incurs higher transmission cost and power consump-
tion. Also, this large amount of data requires extra processing time to be executed by the Processing Subsystem of the WBSN
device, resulting into massive power consumption. Thus, frequently data compression algorithms are utilized for sparse en-
coding of this data. As a result of sparse encoding, the processing requirements decrease, which result into less power
dissipation in the Processing subsystem. As the number of transmitted bits (that forms the packets) is decreased, the trans-
mission is decreased too. At the same time, since sparse encoding algorithms are lossy in nature, the signal reconstruction
at the receiving end will not be exact, and will result into some deviation between the original and reconstructed signals.
 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166 3

Fig. 2. Illustrative Figures (a)-(d): Variation of total time and power consumption for various values of introduced errors. Error increases with an increase
of sparsity from (a) to (d).

However, in clinical domain, maintaining a very low percentage of root mean square deviation (PRD) among the original and
reconstructed signals, is a matter of utmost importance. For example in this case, let x be the original ECG signal and xrecon
be the compressed version of x, then PRD is a measure of deviation of xrecon from x and is given by Eq. 1. If PRD will be high,
certain subtle clinically important features in the signal will be lost; this will be inappropriate in this case. Increasing the
sparsity of the encoded signal would increase the PRD and in in turn would lessen the power consumption, storage require-
ment, transmission time and system’s computing overhead. The sparsity achieved by the reconstructed signal is measured
by the Compression Ratio (CR) which defined as the ratio of number of bits required to encode the original signal (Bo ) to
the number of bits required to encode the compressed signal (Bc ) as described in Eq. 2. If the reconstructed signal becomes
too erroneous, it will have a high PRD and hence it will lose its clinical importance to some extent. Although, it is important
to reduce the PRD, occasionally with a very small increase in it, we can get significant levels of energy saving. Sometimes
a very small PRD may not be even required to obtain correct ECG analysis, and in those cases the PRD can be traded-off
for much higher amounts of system-level energy savings because of improved CR. By solving an optimization problem, this
paper aims at establishing a trade-off between the PRD and the compression ratio/energy consumed.
x − xrecon 2
P RD(% ) = × 100 (1)
x2
Bo
CR = (2)
Bc
To explain our idea, we use illustrations using real experimental data as shown in Fig. 2 that shows both microcontroller
execution time (mctime ) (energy consumption of the processing subsystem) and transmission time (txtime ) (that impacts the
communication subsystem energy energy) decrease together with an increase in error bound (er ) (PRD is obtained by repre-
senting (er ) in %) between the original and reconstructed signals. The value of er is increased from 0.0 to 0.6 that is depicted
in Fig. 2a to Fig. 2d. With an increase in er , the sparsity of the encoded signal increases, as a result of which mctime decreases.
4 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166

As er increases, few numbers of bits (or packets) are transmitted and hence, txtime decreases. Note that mctime and txtime are
directly related to the microcontroller execution energy (mcenergy ) and transmission energy (txenergy ), respectively. However,
as we see, both mcenergy and txenergy are indirectly proportional to er . In Fig. 2b and Fig. 2c, we see that as er increases by
0.2, both mctime and txtime decrease by 41 ms and 1.1 ms respectively. Decrease in both mctime and txtime lead to an overall
decrease in energy consumed in both processing and communication subsystems. This in turn results in a reduction in total
system energy given by totenergy = mcenergy + txenergy . Note that the sensor energy (senenergy ) is assumed constant and also
turned out to be very small as seen from our energy consumption data and hence, has not been mentioned explicitly in the
total energy calculation for this paper. Thus, from Fig. 2a - Fig. 2d, a tradeoff between the amount of tolerable error and
the energy consumption can be observed. To achieve this system level energy improvement with negligible impact of signal
quality, the following algorithms have been proposed in this paper:

• Encoding Algorithm: that can directly control the sparsity level and hence, the amount of error introduced in the recon-
structed signal with respect to the original signal.
• Optimization Heuristic: that aims to minimize the the total amount of energy consumption of the WBSN while incurring
negligble impact on application-level quality. The tolerable small quality/PRD degradation is stated as an input design
parameter.

This paper presents one example of IoT based ECG system that we use to demonstrate how approximate computing can
be used to reduce the energy consumption drastically of these types of IoT-based medical systems. There will be available
other similar systems too, and our generic approach is to be applied to those systems as the approximations are performed
both at the system (hardware) and the application (software) levels. As stated earlier, although it is mandatory to reduce
PRD, occasionally with a very small increase in it, we can get significant levels of energy savings. Sometimes, very small PRD
may not be even required to get correct ECG analysis, and in those cases the PRD can be traded-off for much higher amounts
of system-level energy savings as performed in a systematic way in this paper. We introduce our proposed error-resilient
algorithms for demonstrating the usefulness of the approximation which would lead to energy savings in the designed
system. These values can be set initially based on inputs provided by medical and signal processing experts and then these
values can be adapted accordingly based on the ECG signal characteristics and the specified error bounds given by the
clinical experts. In a nutshell, the major contributions of this paper are listed as follows:

1. Design of an end-to-end WBSN system consisting of sensing, processing, and communication subsystems.
2. Identification of window size and threshold as approximation knobs that also play the role of quality vs energy (Q-E)
control knobs for the current encoding scheme that can trade-off energy for quality in this ECG acquiring WBSN.
3. Proposal of a heuristic that can tune the Q-E approximation knobs in O(n) time to find the optimal settings (degree of
approximations) that will result in the least energy consumption for a specified quality degradation (in terms of PRD).
This heuristic can generate the overall system-level Q-E relationship that determines the window size and threshold
value for each quality loss bound in the set of specified bounds of the quality loss.

The rest of the paper is organized as follows. Section 2 provides a detailed account of the related work and how they
differ from the current work. Section 3 outlines the theory behind the data acquisition through noise filters and subse-
quently the details of the Geometry-Based Method (GBM) and the Wavelet Transform Based Iterative Thresholding (WTIT)
methodologies used to condition the ECG signals before the process of sparse signal encoding. Section 4 shows the system
design of the WBSN and the approximation mechanism to trade-off energy for quality in the system. Finally, Section 5 and
Section 6 describe the experimental methodology and the experimental results, respectively. Section 7 concludes the paper.

2. Related Work and Our Contributions

With the increase in health awareness among people, new demand for real time remote health monitoring had been
observed recently [20,21]. Information and Communications Technology (ICT) had been able to cater for this demand signif-
icantly [22]. Over the years, ICT improved the efficiency of healthcare sector at a great extent [23,24]. This led to the growth
of e-health devices: devices which were used in ICT for autonomous monitoring and controlling of healthcare related issues
[25–28].However, there are certain concerns related to the application of ICT in healthcare service [29,30]. The major issue
is that e-health devices consumed enormous energy, much of which was wasted [2,17,31]. These devices were connected
with a large number of sensors and hardware components for information processing and signal transmission [32,33]. Real
time health monitoring demanded these e-health systems remained to be powered on, to collect data in a continuous man-
ner, for longer periods of time. However, not all collected data are clinically useful, and only a fraction of the information
collected was utilized by the concerned authorities for making decisions [34]. Thus, the wastage of energy was enormous
in these devices [35,36]. Another issue with these devices was to mention that they generated a high volume of data which
often required larger storage [37,38]. This demanded for efficient real time compression scheme [39,40]. Note that most of
the existing worked on signal compression either achieved low PRD between the original and reconstructed signals or high
compression ratio (CR) [41–45]. Compression Ratio is the ratio of number of bits required to encode the original signal (Bo )
to the number of bits required to encode the compressed signal (Bc ). But, the main problem lay in achieving both low PRD
and high CR simultaneously. Some previous works on ECG signal compression assumed arbitrary threshold values for sparse
encoding of signals [46–49]. The real life situations showed that the changes of threshold values of ECG were depending
 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166 5

Fig. 3. IoT-based health monitoring system.

upon heart conditions of persons under clinical observations. Thus, hard thresholding might result into loss of clinically
important information [50].
The work in [51] proposed two algorithms 1) GBM (that removes baseline drift of the ECG signal) and 2) WTIT (which
encodes the signal in the wavelet domain). The WTIT algorithm extracted the high frequency components of ECG signal by
setting a high threshold value for approximate (low-level) coefficients and a low threshold value for detailed coefficients
(high frequency components). In order to achieve this, a window function of fixed length in WTIT algorithm slid over each
of the approximate and detailed signals of the original ECG signal. In each of the window segments, a fixed threshold value
set the value of the signal coefficients below a certain threshold value equal to zero. This promoted sparsity of the ECG
signal in the wavelet domain. Though this earlier work [51] achieved a good CR for a given allowable PRD, the following
had not been addressed in [51]:

• The implementation of the ECG compression using WTIT and GBM on a real hardware platform was unexplored.
• The work [51] did not address the realization of an end-to-end system that consisted of sensing, processing and commu-
nication subsystems.
• As the physical realization of the system was unexplored, the effect of variation of the window-size and threshold pa-
rameter on the energy consumed by the system was not addressed. Hence, the WTIT algorithm was performed with a
fixed value of window-size and threshold parameter in the work [51].

In this work, we design a real time IoT based health monitoring system with optimized process parameters using the
principles of approximate computing [52–55]. The optimization heuristic provides solution scheme that allows to reduce
power consumption, and improves data compression while maintaining significantly low PRD. Since our primary objective
is to optimize the energy contribution of the ECG-acquiring WBSN system, first the energy contribution of the different
components of the overall system are deduced. Subsequently, it is figured out how the energy consumption is impacted by
the approximation knobs. Using this knowledge, the heuristic to acquire the best approximation settings for window size
and threshold are developed that results in the least energy consumption for a specified quality loss (PRD).

3. Theory

In the three tiered architecture of an IoT-based health monitoring system illustrated in Fig. 3, Wireless Body Area Network
(WBAN) forms the first tier IoT platform of WBSNs. WBAN comprises miniature biosensors called WBSNs that can be both
wearable and implantable. These biosensors collect physiological signal, in this context of ECG signal, and perform sampling,
filtering and processing operations on the data. WBSNs used in this work are smart sensors containing user modifiable sec-
tion of signal processing. All the biosensors connected through short range communication form an ad-hoc sensor network.
The second tier is constituted of the gateway devices like smartphones and laptops. The function of the gateway system
is to provide the connection between ad-hoc sensor networks and infrastructure based Wide Area Networks (WANs). Ultra
Low-power WiFi is used in this work to launch the transmission of the processed and compressed data from the WBAN to
the Gateway Devices. It transmits information to one or more heterogeneous networks depending upon the application. The
third tier comprises WAN and the number of end-user healthcare application providers as illustrated in Fig. 3. The end-user
healthcare application provider has a processing unit and an user interface namely graphical user interface (GUI). Vital signs
appearing in GUI are monitored real-time by healthcare professionals to keep track the health status of a patient under ob-
servation. Note that these WBSNs are highly energy constrained as they are embedded or implanted on or within the body
of the human being.
6 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166

Fig. 4. WBSN circuit diagram.

The IoT-based health monitoring system shown in Fig. 3, constructs Tier 1 nodes in WBAN that is segregated into three
subsystems: i) Sensing, ii) Processing and iii) Communication as shown already in Fig. 1. The power subsystem is abstracted
in this system. The processing and the communication subsystems are the areas in which we develop energy efficient tech-
niques in this work. The digital signal processing stage is user modifiable, and subsequently an inexact signal processing is
implemented in it for achieving energy efficiency. In this work, we use a commercial-off-the-shelf (COTS) microcontroller
for designing the sensing, processing and communication subsystems. The circuit design of a typical COTS microcontroller,
namely WBSN in this paper, is shown in Fig. 4.

• The Sensing subsystem comprises AD8232 ECG monitor with a lead setup of three electrodes. It is a dedicated front end
integrated circuit heart rate monitor. It is designed to extract, amplify, filter small biosignals in the presence of noisy
conditions, such as those created by motion artifacts or remote electrode placement. The monitor contains an ultra low
power Analog to Digital Converter (ADC). The output signal is acquired on an embedded microcontroller. Input signals
are taken from Right Arm, Left Arm and Right Leg forming an Einthoven triangle.
• To develop the Processing subsystem, we use Arduino UNO R3, which removes noise from acquired biosignals, adjusts
baselines, encodes signals using sparse coefficients, and optimizes system parameters to consume minimum energy. Ar-
duino UNO comprises a low power unit, an efficient microcontroller with built-in ADC and an USART communication
feature.
• The WiFi module in Arduino transmits real time data to a server located typically in public/private cloud. Those data
will be retrieved by physicians or hospitals’ personnel. This Arduino based WiFi shield, ESP8266, is used in this system
to design the Communication subsystem.

While considering the DSP stage of WBSN, a practical scenario of embedded bio-signal analysis is assumed here for our
case study. Embedded bio-signal analysis is a domain in which energy efficiency is the key, while the applications are in-
herently error-tolerant to a certain degree. In this work, we utilize this intrinsic error resilience of algorithm to achieve
unprecedented level of saving of the energy of the system. We develop the algorithm in a way that it requires minimal
hardware and a negligible energy overhead with real-time transmission. The bio-signal, considered in this work, is an elec-
trocardiogram (or ECG signal). In ECG signal, the deterministic and stochastic feature are exhibited, and the repertoire of the
methods applied for its analysis include (i) the extractions of morphological features, (ii) the analysis in time and frequency
domain, (iii) the statistical methods and classification procedures. In order to study them, we develop, in this work, the pa-
rameter extraction method namely geometry based method (GBM) [51], and wavelet transform based iterative thresholding
(WTIT) [60]. Utilizing the quasi-periodic nature of ECG signal, GBM increases the sparsity of the signal in the time domain.
Due to localization properties in time and frequency domains, WTIT further approximates the signal in wavelet domain
with a few coefficients creating a sparser signal, which is finally transmitted after having undergone Huffman encoding.
Therefore, the proposed framework is divided into three sub-algorithms: 1) Noise filtering, 2) GBM, and 3) WTIT.

3.0.1. Data Acquisition Through Noise Filters

The processing section of WBSN begins with the wavelet filtering of ambulatory sampled ECG signals. For proper ECG
signal acquisition, the first step is noise suppression performed by noise filters, and then the enhancement of ECG signal
amplitudes through an amplifier circuit built of operational amplifiers. The filtering techniques used for denoising ECG sig-
nal involve linear and non-linear methods; either model based or model free approach. In this paper, noise filters employ
wavelet based denoising (model-free thresholding method), brought it to completion by decomposing the signal using band
pass filters corresponding to a mother wavelet basis function. The mother wavelet chosen should be maximally compatible
with the ECG signal structure.
Wavelet based denoising involves the use of Donoho’s universal thresholding. We set decomposition level dependent
threshold value Tl per Donoho’s universal thresholding [61]:

Tl = σ 2 loge l (3)
 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166 7

Fig. 5. Typical ECG signal showing P, Q, R, S and T waveforms.

where, σ is noise variance of the signal at each level of wavelet decomposing and l is the length of the analyzing signal. For
a given signal u[n], let there exist an orthogonal matrix W such that the discrete wavelet transformed matrix U[n] is given
by:
U = Wu (4)
Let, the additive noise to signal x[n] be y[n], which has noise variance σ 2. Hence, the resultant signal z[n] is given by:
z[n] = x[n] + y[n] (5)
The inverse wavelet transform is computed as:
u = W TU (6)
The wavelet transformed matrix U[n] is modified to U˜ [n] by applying thresholding and as a result u[n] becomes u˜[n]
u˜ = W T U˜ (7)
Reduction of wavelet transformed coefficients produces significantly noise free estimate, which is done either by hard or
soft thresholding.
Thard (U˜ ; Tl ) = U O(|U | > Tl ) (8)

Tso f t (U˜ ; Tl ) = sgn(U )(U − Tl )O(|U | > Tl ) (9)

O is the usual indicator function and sgn represents signum function. In order to eliminate noise, we consider soft thresh-
olding of wavelet coefficients because hard thresholding results loss of information from the original signal. The noise free
ECG signal is free from zero mean band limited Gaussian noise i.e., the minute interference presents in the signal but it
is not entirely free from the periodic baseline drifts. The noiseless ECG signal is processed using our proposed encoding
algorithm.

3.0.2. Geometry Based Method (GBM)

Clinical assessment of ECG (Fig. 5) relies on the evaluation of its morphological features in time domain, those features
are position, duration, amplitude and slope of its QRS complex and segment. The first step is the detection of the QRS
complex which serves as a marker for averaging of heart cycles. To determine feature extraction from ECG beat, differ-
ent approaches using Hermite Functions, wavelet transforms etc., were proposed. In this paper, we develop a time-domain
based method GBM for the same purpose. GBM belongs to the preprocessing stage of Sensing subsystem that is shown in
Fig. 1. This pre-processing technique is only applicable when ECG signal shows likely a periodic variation. In an ECG signal
shown in Fig. 5, only P, Q, R, S and T waveforms and their relative separations are clinically important features. From preset
threshold values and the rate of change of the amplitudes of these waveforms, GBM automatically identifies them. It then
diminishes all the points, between T waveform and next period of P waveform, to 0. This process facilitates elimination of
noise and adjusting the baseline of ECG signal to 0.

3.0.3. Wavelet Transform Based Iterative Thresholding (WTIT)

After preprocessing of ECG signal having completed through GBM, it is decomposed by discrete wavelet transformation
(DWT), which involves passing this signal through high and low pass filters. The wavelet coefficients are obtained by the
convolution process on the input signal with the impulse responses of high (g) and low (h) pass filters. Eqn.10 provides
the output of high pass filter after performing downsampling on convolved signals to give detailed coefficients (CD), while
Eqn.11 obtains the output of low pass filter that gives approximate coefficients (CA) after performing downsampling.
∞

CD[k] = x[n].g[2k − n] (10)
n=−∞

∞

CA[k] = x[n].h[2k − n] (11)
n=−∞
8 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166

Fig. 6. Variation of log10 [totenergy ] with Th and Win obtained from experimental results.

This WTIT algorithm is mainly dependent upon two system parameters: (i) Window Size (Win ) and (ii) Threshold (Th ). How
these parameters directly affect the performance of the energy aware sparse encoding scheme, discussed in the next few
sections. A window is traversed along the detailed and approximate wavelet coefficients. The number of coefficients inside
the window is referred to as the window size. If the largest coefficient inside a window is M, then all elements with value
lesser than Th ∗ M are approximated to zero. Here, Th is referred to as the normalized threshold. Its ranges vary from 0 to 1.
In particular, we use WTIT to extract the high frequency components of ECG signal. This WTIT sets high approximated value
for lower level coefficient, and lower approximation to higher level coefficient. This is continued to reach higher sparsity in
ECG signal, keeping the higher frequency components intact. These coefficient matrices obtained from WTIT are transmitted
after having undergone Huffman encoding. Huffman encoding is a lossless data compression technique which encodes sparse
matrices in terms of only non-zero values. In the receiver end, the received signal has undergone Huffman decoding and
inverse wavelet transformation for time domain representation of the signal to be analyzed by the person under concern.

4. System Design

4.1. Interaction between System Parameters

This energy efficient IoT-health monitoring system requires a mathematical trade-off between quality of the reconstructed
signal and the energy spent in the Processing and Communication subsystems that is shown in Fig. 1. As wavelet coefficients
provide a sparse representation of a signal, wavelet based thresholding is applied on the ECG signal that is preprocessed by
applying GBM. Furthermore, WTIT provides two independent parameters, threshold (Th ) and window size (Win ). Th and Win
are used as approximation knobs for the entire system. The selection of sparse wavelet coefficients for representation of
a signal is supervised by Th . Moreover, Win can be adjusted according to the need of study of ECG transients for clinical
purposes. So, Th and Win are the independent approximation parameters which control the IoT-based health monitoring
system.
Fig. 6 and Fig. 7 have been drawn from experimental results. They represent the relationships between error bound and
system energy consumption to the threshold parameter and the window size. When the error bound er is the deviation
between original and reconstructed signals, Fig. 6 represents how er varies with Th for various Win of lengths 128, 256, 512,
1024, 2048 and 4096 respectively. If Th increases, the ECG signal encoded by WTIT becomes more sparse, and hence, er
increases. With an increase of Win , fewer coefficients are thresholded to zero, thus er decreases. If mcenergy is the energy
consumed by the microcontroller during the operation and txenergy is the transmission energy then totenergy = mcenergy +
txenergy . and Fig. 7 represents the variation of totenergy with Th for various values of Win . To decrease the range of variation
of totenergy w.r.t Win and Th , logarithm scale is used. While mcenergy and microcontroller execution time (mctime ) are nearly
independent of Th , txenergy changes inversely related to Th . With an increase in the number of sparse coefficients, fewer bits
are transmitted. Then, transmission time (txtime ) decreases, which results in a decrease in txenergy . Thus, totenergy decreases
marginally with an increase in Th . However, as Win increases, the sparsity of encoded coefficients decreases and hence
txenergy increases.
Fig. 6 and Fig. 7 are obtained from experimental results that represent the nature of relationship between system param-
eters, thus driving us to develop an energy optimization heuristic. Therefore, provided a maximum tolerable error bound,
we need to select that coordinate of window size and threshold parameter that result in the least total energy consumption.

4.1.1. Effect of window-size variation

Upon variation of approximation parameters Win and Th , the change of values for er , mctime and txtime are observed. These
observed values are used to generate polynomial relations among the system parameters that are given in Table 2. These
 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166 9

Fig. 7. Variation of er with Th and Win obtained from experimental results.

Table 1
Comparison of our proposed algorithm with previous algorithms.

Algorithm PRD (%) CR

Simultaneous OMP [56] 2.57 7.23

Basis Pursuit(1) [57] 1.66 4
Basis Pursuit(2) [58] 9 3.44
Basis Pursuit(3) [59] 9 2.5
Our Encoding Algorithm 2.27 9.4

Table 2
Mathematical relations among various system parameters (c and a being
positive real number variables).

Parameters Curve fitting expression

er vs Win c + ax−0.5
mctime vs Win c + ax1.024 ≈ c + ax
txtime vs Win c + ax
er vs Th c + ax
mctime vs Th c
txtime vs Th c − ax

relations are generated by polynomial regression of the values of system parameters collected from actual experimentation
executed in this work. The types of the polynomials clearly indicate that er decreases as Win increases. This effect can also be
observed from Fig. 6. When the length of window increases in WTIT, the less number of coefficients are thresholded to zero.
This results the decrement of er . The time complexity of WTIT is in the order of O (n ) and hence mctime increases linearly
with the signal length under consideration. Increasing the Win results in improvement of quality of the reconstructed signal
at the expense of increased mcenergy and txenergy . These are illustrated in both Fig. 6 and Fig. 7 and implied from Table 2.

4.1.2. Effect of threshold variation

To study the effect of variation of Th , er is varied with Th for a fixed Win . With increase in the value of Th , the number of
sparse wavelet coefficients increases, thus introducing subtle differences between original and reconstructed signals. From
Fig. 6, we see that er varies linearly with Th . As Th increases, fewer coefficients are transmitted as rest of the coefficients
below the threshold level are set to zero. These zero values, discarded via Huffman encoding, are not transmitted. The
non-zero coefficients are transmitted in the form of packets. Higher the number of non-zero coefficients the larger is the
packet-size. Thus, txtime , and hence txenergy , are directly proportional to the packet-size and inversely proportional to Th .
However, Th does not affect mctime and mcenergy . It is observed that Th has negligible influence when totenergy is calculated. In
Fig. 7, totenergy has a small negative slope, when varied with Th for a given Win because txenergy decreases with the increase
of Th .
Thus, the IoT-based ECG monitoring system is designed in such a way that for a pre-specified maximum tolerable error
bound (K), the totenergy is to be minimized by varying the independent approximation parameters Th and Win . This require-
ment of IoT-based ECG monitoring system demands to solve an optimization problem.
10 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166

Fig. 8. Dependencies of various subparameters on independent metrics obtained from experimental results: (a) Variation of microcontroller execution time
and window size (b) Variation of normalised error with window size (c) Variation of normalised error with threshold.

Table 3
Notation of system parameters.

Notation Sub-parameter variable

Win window-size
Th threshold parameter
txtime module transmission time (ms)
txenergy module transmission energy
mctime microcontroller system time
mcenergy module microcontroller energy
totenergy total energy (txenergy + mcenergy )
er error between original and reconstructed signal

4.2. Optimization Heuristic

In order to discuss the optimization heuristic, the notations for sub-parameter variables are given in Table 3. Variables
er , totenergy , mcenergy , mctime , txenergy , txtime are dependent on the independent variables Win and Th . Both the variables er and
totenergy can be represented as a function of independent variables Win and Th . From simulation data, the observed values
are modelled to form 3-dimensional curves by using curve-fitting toolbox in MATLAB 2017a. Thin plate spline is chosen as
the interpolant because it gives the best fitting model with least sum of squared errors (SSE) as compared to other best
fitting models like linear, polynomial and power fitting. The ideal fitting will have a SSE tending to zero. The goodness of fit
using thin plate spline is given in Table 4.
 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166 11

Table 4
Sum of squared errors (SSE) using thin plate spline interpolant.

Modelled function Interpolant type Goodness of fit

er (Win , Th ) Thin plate spline SSE = 2.78e-27

totenergy (Win , Th ) Thin plate spline SSE = 1.625e-08

To develop an energy efficient system, we minimize the totenergy for a given maximum tolerable error bound (K). So, the
optimization problem can be stated as:

Minimize totenergy (Win , Th );

subject to er (Win , Th ) ≤ K

The error bound K is set as preset requirement for studying transients of ECG signal. The optimization heuristic namely
Algorithm 1 chooses the optimum (Winopt , Thopt ) coordinate from the set of coordinates (Wini , Thi ) where i ∈ Z. Algorithm 1 is
a optimum parameter search algorithm that does not take into account the parametric relationship between the independent
variables (Win and Th ) and the system parameters (er and totenergy ). In Fig-10, the variation of Total Energy with Window-size
is much higher than the variation of Total Energy with Threshold Parameter. Exploiting this fact Algorithm 2 is developed
which reduces the algorithm complexity from O (n2 ) to O (n ).

Algorithm 1: Finding optimized values of Win and Th for energy minimization (Winopt , Thopt ) using a generic search
method.
Input: Error Bound K
Output: Optimum window-size and threshold (Winopt , Thopt )
1 G=∅ ; // G is a null set
2 for (i = 0; i ≤ n; i = i + 1) do
3 while er (Wini , Thi ) ≤ K do
4 G = G ∪ (Wini , Thi ); // include the new coordinate into the existing set
5 end
6 end
7 g = |G|; // cardinality of the set G
8 for (i = 0; i ≤ g; i = i + 1) do
9 (Wini , Thi ) = (∞, ∞ )
10 if totenergy (Wini , Thi ) ≤ totenergy (Winopt , Thopt );
11 (Winopt , Thopt ) = (Wini , Thi ); // select the optimum coordinates from the set G
12 end
13 return (Winopt , Thopt )

4.2.1. Selection of coordinates that satisfy inequality constraint

The set of coordinates (Wini , Thi ) that satisfy the inequality constraint of the optimization problem is denoted by set
G. All the operating points which belong to the set G has value of Normalized Error less than the error bound K. Hence,
(Wini , Thi ) ∈ G implies (Wini , Thi ) ≤ K. Selection of operating points in set G will be shown in Fig. 9 in Section 6.

4.2.2. Selection of optimum coordinate from set G

The search space for the optimum coordinate is narrowed down by selecting operational points from set G. The coordi-
nate (Wini , Thi ) that minimizes the value of totenergy is the optimum coordinate. Algorithm 1 applies brute search method on
set G to obtain the optimum operating point (Winopt , Thopt ). Selection of the optimum operating point from the set G will be
shown in Fig. 10 in Section 6.

4.2.3. Deriving the Quality-Energy (QE) relation

The previous two steps are repeated for various values of error bound K, and consequently, the optimum coordinate is
obtained for each value of K. This optimum coordinate minimizes total energy and is chosen for energy-efficient operation of
IoT based health monitor system. Whenever this IoT based health monitoring system is in operation for some given values
of (K), the system operates at (Winopt , Thopt ), which is the optimal operating point for the minimization of the total energy.
In this case, Fig. 11 will show the Quality-Energy (Q-E) relationship of the system that denotes the minimum achievable
energy for a given error bound.
12 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166

Algorithm 2: Finding optimized values of Win and Th for energy minimization (Winopt , Thopt ) exploiting parametric
relationship.
Input: Error Bound K
Output: Optimum window-size and threshold (Winopt , Thopt )
1 G=∅ ; // G is a null set
2 for (i = 0; i ≤ n; i = i + 1) do
3 while er (Wini , Thi ) ≤ K do
4 G = G ∪ (Wini , Thi ); // include the new coordinate into the existing set
5 end
6 end
7 g = |G| ; // cardinality of the set G
8 for (i = 0; i ≤ g; i = i + 1) do
9 (Winopt ) = (∞ );
10 if totenergy (Wini ) ≤ totenergy (Winopt );
11 Winopt = Wini ; // select the optimum window-size coordinate from the set G
12 end
13 for (i = 0; i ≤ g; i = i + 1) do
14 (Thopt ) = (∞ );
15 if totenergy (Winopt , Thi ) ≤ totenergy (Winopt , Thopt );
16 Thopt = Thi ; // select the optimum Threshold Parameter coordinate
17 end
18 return (Winopt , Thopt )

Fig. 9. Variation of normalized error with window size and threshold obtained from experimental results. A red plane indicating constant error of K is
drawn which cuts the surface. All coordinates lying below this red plane belong to the set G.

Fig. 10. Variation of total energy with threshold and window size obtained from experimental results. The coordinates belonging to the set G are shown
by the arrowed region. The optimum coordinate is the one that simultaneously belongs to this region and has the least value of total energy.
 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166 13

Fig. 11. Q-E relationship diagram denoting the minimum achievable normalized energy for a pre-specified error bound K. Normalized energy is Total
Energy consumed in our compression-transmission scheme divided by the Total Energy without any prior compression. Using the Elbow method, the
optimum Normalized Energy is found to be 0.04 indicating decrease in energy consumption of about 96%. Q-E plot has been obtained for three different
ECG subjects.

Algorithm 1 searches for all the coordinate points in set G making it of complexity O (n2 ). However, we have already
obtained a relation of Total Energy with both Threshold Parameter and Window-Size in Fig. 10. Exploiting this fact that vari-
ation of totenergy with respect to Win is greater than, we first perform a linear search on window size and obtain the Winopt .
Then keeping Winopt fixed, we perform a second linear search on the Threshold Parameter to obtain Thopt . This reduces/prunes
the search space significantly as Algorithm 2 reduces to a O (n ) search problem as when compared to Algorithm 1 which
is a O (n2 ) search problem. Hence, by exploiting the relationship of independent parameters with Total Energy, we avoid a
generic search algorithm. This allows us to extract the final design parameters of Winopt and Thopt quickly and are the same
as the ones shown (emboldened) in Table 5.

5. Experimental Methodology

In order to describe the hardware experimental set-up, we refer to Fig. 4. We use AD8232 ECG heart rate monitor with
an Arduino UNO R3 for ECG signal acquisition from a subject (patient) body. The set of pins of AD8232 are connected to
the corresponding pins of Arduino; those pins are namely ground pin, 3.3V power supply pin, output signal pin, positive
leads-off detect pin, negative leads-off detect pin of AD8232 which are connected with ground pin, 3.3V pin, A0 pin, pin
number 10 and pin number 11 of the Arduino respectively. AD8232 has three coloured cables: black, blue and red, which are
placed on the right arm, left arm and right leg of the patient respectively. Arduino C processing IDE v2.2.1 is used to write
sketch to read and process ECG signals directly. Arduino IDE shows a list of available communication ports, any of which
can be used for ECG signal recording. From the Arduino, ECG signal is transmitted to a server located in private/public cloud
using ESP826 module which is used to access the Internet via WiFi connection. The energy consumption of the system is
measured using a fluke high precision multimeter.
We apply static thresholding, i.e., it is assumed that during the transmission period, the heart condition of the patient
remains unchanged or the threshold level is constant. However, the physician can change the normalized error value on
ECG monitor as and when it is required to study the ’Quality’ of the signal (refer to Fig. 11). We perform the experiment
by taking a generalised value of Th , which will be further changed. In our simulation, MATLAB R2017a is used to evaluate
the mathematical relations among various system parameters that are given in Table 2. We use the Curve Fitting Toolbox in
Matlab R2017a, which would interpolate the sample observation points to generate a well-defined function in order to study
the dependencies among various system parameters. The interpolating function is chosen in such a way that it minimizes
SSE of the observed points and the interpolating curve.

6. Experimental Results

In this work, we have developed a real time encoding scheme which performs iterative thresholding and approximation
of wavelet coefficients for sparse encoding of ECG signals. We have optimized various process parameters to model a low
power hardware prototype for the implementation of our algorithm on a real time microcontroller based IoT platform that
exhibits an end-to-end WBSN system.
In order to describe the performances of these algorithms, various system parameters and finally the implementation of
hardware prototype, we start the discussion the results from Fig. 8. The nature of variations of mctime with Win is evident in
Fig. 8a. With the increment of Win , mctime increases almost linearly. This graph depicts that when Win increases from sample
size window of 16 to sample size window of 4096, mctime increases from 0 ms to 73 ms. This is because of the fact that
14 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166

Table 5
Variation of optimum system variables for different values of K.

K (Wini , Thi ) mctime (ms) mcenergy (mJ) txtime (ms) txenergy (mJ) totenergy (mJ)

0.015 (2048, 0.0056) 191.98 0.37 55.07 0.2297 0.6069

(4096,0.0071) 823.60 1.61 742.23 3.096 4.71

0.02 (1024, 0.0054) 41.71 0.08 3.21 0.0134 0.0953

(2048,0.0076) 191.99 0.3772 54.99 0.2294 0.6066
(4096,0.0096) 823.66 1.618 741.26 3.092 4.71

0.03 (256, 0.0053) 1.42 0.0028 2.68 0.011 0.014

(512,0.0134) 7.57 0.0156 2.61 0.0109 0.0265
(1024,0.0176) 41.71 0.082 3.20 0.0134 0.0954
(2048,0.0116) 192.01 0.3772 54.83 0.228 0.605
(4096,0.0149) 823.7 1.61 739.21 3.08 4.70

0.04 (128, 0.0057) 0.745 0.00146 2.069 0.00863 0.0100

(256,0.0071) 1.424 0.0028 2.679 0.0117 0.0145
(512,0.0087) 7.95 0.0156 2.605 0.0108 0.026
(1024,0.0114) 41.71 0.081 3.195 0.0133 0.095
(2048,0.001575) 192.03 0.377 54.67 0.228 0.605

0.045 (128, 0.0064) 0.746 0.00147 2.065 0.00861 0.01008

(256,0.0089) 1.42 0.0028 2.67 0.011 0.0139
(512,0.0098) 7.95 0.00156 2.60 0.010 0.0264
(1024,0.0129) 41.71 0.081 3.19 0.0133 0.0952
(2048,0.018) 192.04 0.377 54.58 0.227 0.6055

0.05 (128, 0.0072) 0.745 0.0015 2.069 0.0086 0.0101

(256,0.0089) 1.42 0.0029 2.67 0.011 0.014
(512,0.010) 7.95 0.0156 2.60 0.010 0.0264
(1024,0.0144) 41.72 0.082 3.18 0.013 0.095

0.06 (64, 0.0059) 1.05 0.0021 1.50 0.0062 0.008

(128,0.0087) 0.7449 0.0015 2.05 0.0085 0.010
(256,0.010) 1.42 0.0028 2.66 0.011 0.013
(512,0.0134) 4.95 0.0097 2.58 0.0108 0.0205
(1024,0.0176) 41.722 0.082 3.17 0.0132 0.095

the asymptotic time complexity of WTIT algorithm is in the order of O(n). As Win increases, the more number of discrete
samples of ECG signals under consideration are processed together by WTIT, and hence mctime increases. However, mctime
remains constant with respect to Th . Irrespective of the variations of Th , the number of discrete samples processed by WTIT
remain unchanged. Hence, mctime does not vary with the variations of Th .
Fig. 8 b and Fig. 8c describe the variations of er with sub-optimal values of Win and Th respectively. For a fixed value of
Th , as Win increases, er decreases due to the fact that the WTIT algorithm approximates fewer wavelet coefficients to zero.
When Win increases from 512 to 4096 sample sizes, er decreases from 0.043 to 0.02. An increase in Th results in higher
number of sparse wavelet coefficients. Normalized er increases from 0.1 to 0.15, when Th increases from 0.01 to 0.015. Thus,
er and Th are directly proportional to each other. Similar nature of variations of er with Th and Win was illustrated in Fig. 6.
Fig. 9 demonstrates the relationship of er with Win and Th using a 3-D surface generated by thin plate spline interpolation
having used on the experimental data. On the variation of both Win from 16 to 4096 sample sizes, and Th from 0.005 to
0.02, normalized er varies from 0.05 to 0.2. To explain Fig. 10, it is imperative to discuss that txtime increases and decreases
linearly with respect to Win and Th , respectively. With the increment of sample sizes in Win , WTIT has processed these
number of samples of ECG signal and subsequently transmitted to server located in private/public cloud.This results in the
increase of txtime . As Th increases, the number of sparse wavelet coefficients increases. Thus, after having undergone Huffman
encoding, both the amount of transmitted data and txtime decrease. Note that txtime and mctime are directly related to txenergy
and mcenergy respectively. As defined earlier, totenergy is the sum of mcenergy and txenergy . It is observed that while mcenergy does
not vary with Th , txenergy decreases with an increase of Th . This results in the negative slope of totenergy versus Th that was
illustrated in Fig. 7. Since both mcenergy and txenergy increases with Win , totenergy increases with Win . In Fig. 10, generated with
experimental data, similar variation of totenergy with different values of Th and Win is observed in the 3-D surface generated
by thin plate spline interpolant. With the variation of both Win from 16 to 4096 window sample sizes, and Th from 0.005
to 0.02, totenergy varies from 0.008 mJ to 4.71 mJ. It is evident in Fig. 10.
Algorithm 1 and Algorithm 2 search for (Winopt , Thopt ) for all values of (Wini , Thi ). From Fig. 9, given an error bound K
marked by the red plane, the points (Wini , Thi ) are chosen in such a way that they lie below this plane. For a given (Wini , Thi )
pair, totenergy is determined that is shown in Fig. 10. The pair (Wini , Thi ) for which totenergy becomes minimum, is the optimal
operational point (Winopt , Thopt ) for the design of IoT based monitoring system.
 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166 15

Table 5 enlists all (Wini , Thi ) for various values of normalized error bound K. For each (Wini , Thi ), totenergy is calculated from
the sum of mcenergy and txenergy . The optimal operation point (Winopt , Thopt ) minimizes totenergy (marked boldfaced in Table 5)
for each K. For a normalized error bound K = 0.03, (Winopt , Thopt ) = (256, 0.053 ) and the totenergy consumed is 0.014 mJ. Sim-
ilarly, for K = 0.05, (Winopt , Thopt ) = (128, 0.0072 ) and the totenergy consumed is 0.0101 mJ. For various values of K (marked
boldfaced in Table 5), the optimal operational point (Winopt , Thopt ) is derived for all values of (Wini , Thi ) using Algorithm 1 and
Algorithm 2. Accordingly, the totenergy consumed is calculated and given in Table 5. Fig. 11 shows the final Quality-Energy
(QE) relationship of quality and energy parameters and it demonstrates a significant decrease in optimized totenergy con-
sumption for a negligible degradation of the quality of signal. Through the application of Algorithm 1 and Algorithm 2,
energy consumption is reduced to 96% with a deterioration of 2% in signal quality. Also when the deterioration of the sig-
nal quality is 1%, the energy consumption is reduced to only 22% as shown in Fig. 11. This abrupt change in the energy
consumption due to change in error level leads to the selection of the optimum Normalized Error bound by elbow method
shown in Fig. 11. Thus, energy consumption decreases with a increase in error bound K. With an increment in tolerable
error, sparsity of the encoded vector increases resulting in a significant amount of the decrease of txenergy .
We introduce the quality/error-levels (degree of tolerable error) just for demonstrating the benefits of approximation of
the ECG signal. These quality/error levels can be adjusted by the physicians as per the quality of ECG signal originated from
patients and then these values can be adapted accordingly. To demonstrate that Energy-Efficient compression-transmission
scheme can be generalised to any standard ECG signal, we derived the Q-E plot for three different ECG signals namely ECG-
1, ECG-2 and ECG-3 respectively obtained from three different patients each having a different health condition. This leads
to the conclusion that the irrespective of the health conditions of the patients, the compression-transmission scheme for
those ECG signals shows approximately 96% energy savings with negligible deterioration in signal quality (2%).

7. Conclusion

In this work, an energy efficient hardware prototype developed on a IoT-based microcontroller platform has been de-
ployed to operate as an end-to-end WBSN system in real time. This WBSN has the capability of the compression of real
time ECG signal via sparse encoding of these signals in time-frequency domain using wavelet coefficients. The signal com-
pression scheme has been developed a two step approach: GBM has encoded the signal in time domain and WTIT has
worked in time-frequency domain to perform approximation of wavelet coefficients. However, excessive approximations of
these coefficients result into an appreciable degradation of signal quality and the loss of clinically important information.
On the other hand, sub-optimal amount of approximations has generated large volumes of transmission data that consumes
higher energy and higher bandwidth. In this work, we have proposed a methodology to perform optimum trade-off be-
tween the quality of signal, the consumption of energy and the usage of bandwidth (shown in Fig. 11). Algorithm 1 and
Algorithm 2 have optimized the various design parameters by introducing optimal amount of approximations that have
led to develop low power IoT-based ECG health monitoring hardware prototype. With these proposed algorithms, energy
consumption of the IoT system has been reduced by 96% with only 2% deterioration in signal quality.

Declaration of Competing Interest

This work has not been published previously and is not under consideration for publication elsewhere. This publication
is approved by all authors and if accepted it will not be published elsewhere in the same form, in English or in any other
language, including electronically without the written consent of the copyright-holder. There is no conflict of interest.

References

[1] B. Farahani, F. Firouzi, V. Chang, M. Badaroglu, N. Constant, K. Mankodiya, Towards fog-driven iot ehealth: Promises and challenges of iot in medicine
and healthcare, Future Generation Computer Systems 78 (2018) 659–676.
[2] G. Sebestyen, A. Hangan, S. Oniga, Z. Gál, ehealth solutions in the context of internet of things, in: 2014 IEEE International Conference on Automation,
Quality and Testing, Robotics, IEEE, 2014, pp. 1–6.
[3] M. Rapin, F. Braun, A. Adler, J. Wacker, I. Frerichs, B. Vogt, O. Chételat, Wearable sensors for frequency-multiplexed eit and multilead ecg data acqui-
sition, IEEE Transactions on Biomedical Engineering 66 (3) (2019) 810–820.
[4] L. Greco, P. Ritrovato, F. Xhafa, An edge-stream computing infrastructure for real-time analysis of wearable sensors data, Future Generation Computer
Systems 93 (2019) 515–528.
[5] F. Firouzi, B. Farahani, M. Ibrahim, K. Chakrabarty, Keynote paper: From eda to iot ehealth: Promises, challenges, and solutions, IEEE Transactions on
Computer-Aided Design of Integrated Circuits and Systems 37 (12) (2018) 2965–2978.
[6] H. Mamaghanian, N. Khaled, D. Atienza, P. Vandergheynst, Compressed sensing for real-time energy-efficient ecg compression on wireless body sensor
nodes, IEEE Transactions on Biomedical Engineering 58 (9) (2011) 2456–2466.
[7] A. Khanna, V. Chaudhary, S.H. Gupta, Design and analysis of energy efficient wireless body area network (wban) for health monitoring, in: Transactions
on Computational Science XXXIII, Springer, 2018, pp. 25–39.
[8] S. Pirbhulal, H. Zhang, W. Wu, S.C. Mukhopadhyay, Y.-T. Zhang, Heartbeats based biometric random binary sequences generation to secure wireless
body sensor networks, IEEE Transactions on Biomedical Engineering 65 (12) (2018) 2751–2759.
[9] J. Zhang, Z.L. Yu, Z. Gu, Y. Li, Z. Lin, Multichannel electrocardiogram reconstruction in wireless body sensor networks through weighted minimization,
IEEE Transactions on Instrumentation and Measurement 67 (9) (2018) 2024–2034.
[10] A.H. Sodhro, L. Chen, A. Sekhari, Y. Ouzrout, W. Wu, Energy efficiency comparison between data rate control and transmission power control algorithms
for wireless body sensor networks, International Journal of Distributed Sensor Networks 14 (1) (2018). 1550147717750030.
[11] A. Galli, G. Frigo, G. Giorgi, Robust ecg denoising for ehealth applications, in: 2018 IEEE International Symposium on Medical Measurements and
Applications (MeMeA), IEEE, 2018, pp. 1–6.
16 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166

[12] Z. Ning, X. Wang, F. Xia, J.J. Rodrigues, Joint computation offloading, power allocation, and channel assignment for 5g-enabled traffic management
systems, IEEE Transactions on Industrial Informatics (2019).
[13] K.S. Kim, D.K. Kim, C.-B. Chae, S. Choi, Y.-C. Ko, J. Kim, Y.-G. Lim, M. Yang, S. Kim, B. Lim, et al., Ultrareliable and low-latency communication techniques
for tactile internet services, Proceedings of the IEEE 107 (2) (2019) 376–393.
[14] R. Braojos, D. Bortolotti, A. Bartolini, G. Ansaloni, L. Benini, D. Atienza, A synchronization-based hybrid-memory multi-core architecture for energy-ef-
ficient biomedical signal processing, IEEE Transactions on Computers 66 (4) (2017) 575–585.
[15] S. Basu, L. Duch, R. Braojos, G. Ansaloni, L. Pozzi, D. Atienza, An inexact ultra-low power bio-signal processing architecture with lightweight error
recovery, ACM Transactions on Embedded Computing Systems (TECS) 16 (5s) (2017) 159.
[16] F. Nasri, A. Mtibaa, Smart mobile healthcare system based on wbsn and 5g, IJACSA) International Journal of Advanced Computer Science and Applica-
tions 8 (10) (2017).
[17] J. Pagan, R. Fallahzadeh, M. Pedram, J.L. Risco-Martin, J.M. Moya, J.L. Ayala, H. Ghasemzadeh, Toward ultra-low-power remote health monitoring: An
optimal and adaptive compressed sensing framework for activity recognition, IEEE Transactions on Mobile Computing 18 (3) (2019) 658–673.
[18] S. Popli, R.K. Jha, S. Jain, A survey on energy efficient narrowband internet of things (nbiot): Architecture, application and challenges, IEEE Access 7
(2019) 16739–16776.
[19] M. García-Valls, C. Calva-Urrego, A. García-Fornes, Accelerating smart ehealth services execution at the fog computing infrastructure, Future Generation
Computer Systems (2018).
[20] O. Albahri, A.A. Zaidan, B. Zaidan, M. Hashim, A. Albahri, M. Alsalem, A. Mohsin, K. Mohammed, A. Alamoodi, O. Enaizan, et al., Based multiple
heterogeneous wearable sensors: A smart real-time health-monitoring structured for hospitals distributor, IEEE Access (2019).
[21] T. Wu, J.-M. Redouté, M. Yuce, A wearable, low-power, real-time ecg monitor for smart t-shirt and iot healthcare applications, in: Advances in Body
Area Networks I, Springer, 2019, pp. 165–173.
[22] S. Ravindran, H.R. Sheikh, M. Polley, Y. Yoo, Wearable augmented reality head mounted display device for phone content display and health monitoring,
2019. US Patent App. 15/646,048.
[23] S. Soumya, S. Kumar, Healthcare monitoring using internet of things, in: First International Conference on Artificial Intelligence and Cognitive Com-
puting, Springer, 2019, pp. 485–494.
[24] I. Budhiraja, S. Tyagi, S. Tanwar, N. Kumar, J.J. Rodrigues, Tactile internet for smart communities in 5g: An insight for noma-based solutions, IEEE
Transactions on Industrial Informatics (2019).
[25] J. Martín-Vaquero, A. Hernández Encinas, A. Queiruga-Dios, J. José Bullón, A. Martínez-Nova, J. Torreblanca González, C. Bullón-Carbajo, Review on
wearables to monitor foot temperature in diabetic patients, Sensors 19 (4) (2019) 776.
[26] T.A.H. Rocha, N.C. da Silva, A.C.Q. Barbosa, C. Elahi, J.R.N. Vissoci, mhealth: Smart wearable devices and the challenges of a refractory context, in: The
Internet and Health in Brazil, Springer, 2019, pp. 347–367.
[27] H.F. Badawi, F. Laamarti, A. El Saddik, Iso/ieee 11073 personal health device (x73-phd) standards compliant systems: A systematic literature review,
IEEE Access 7 (2019) 3062–3073.
[28] S.-H. Fang, C.-C. Li, W.-C. Lu, Z. Xu, Y.-R. Chien, Enhanced device-free human detection: Efficient learning from phase and amplitude of channel state
information, IEEE Transactions on Vehicular Technology (2019).
[29] S.G. Alonso, J. Arambarri, M. López-Coronado, I. de la Torre Díez, Proposing new blockchain challenges in ehealth, Journal of medical systems 43 (3)
(2019) 64.
[30] M.U. Hassan, M.H. Rehmani, J. Chen, Privacy preservation in blockchain based iot systems: Integration issues, prospects, challenges, and future research
directions, Future Generation Computer Systems (2019).
[31] W.H. Organization, et al., Who compendium of innovative health technologies for low-resource settings: 2016-2017: medical devices, ehealth(2018).
[32] F. Firouzi, A.M. Rahmani, K. Mankodiya, M. Badaroglu, G.V. Merrett, P. Wong, B. Farahani, Internet-of-things and big data for smarter healthcare: from
device to architecture, applications and analytics, 2018.
[33] A.A. Abdellatif, M.G. Khafagy, A. Mohamed, C.-F. Chiasserini, Eeg-based transceiver design with data decomposition for healthcare iot applications,
IEEE Internet of Things Journal 5 (5) (2018) 3569–3579.
[34] N. Genes, S. Violante, C. Cetrangol, L. Rogers, E.E. Schadt, Y.-F.Y. Chan, From smartphone to ehr: a case report on integrating patient-generated health
data, npj Digital Medicine 1 (1) (2018) 23.
[35] M. El Tanab, W. Hamouda, Machine-to-machine communications with massive access: Congestion control, IEEE Internet of Things Journal (2018).
[36] O. Salem, A. Serhrouchni, A. Mehaoua, R. Boutaba, Event detection in wireless body area networks using kalman filter and power divergence, IEEE
Transactions on Network and Service Management 15 (3) (2018) 1018–1034.
[37] J. Grundy, M. Abdelrazek, M.K. Curumsing, Vision: Improved development of mobile ehealth applications, in: 2018 IEEE/ACM 5th International Confer-
ence on Mobile Software Engineering and Systems (MOBILESoft), IEEE, 2018, pp. 219–223.
[38] K. Wang, Y. Shao, L. Shu, C. Zhu, Y. Zhang, Mobile big data fault-tolerant processing for ehealth networks, IEEE Network 30 (1) (2016) 36–42.
[39] B. Liu, Z. Zhang, The distortion of data compression via compressed sensing in eeg telemonitoring for the epileptic, in: 2016 IEEE Biomedical Circuits
and Systems Conference (BioCAS), IEEE, 2016, pp. 512–515.
[40] J. Ma, T. Zhang, M. Dong, A novel ecg data compression method using adaptive fourier decomposition with security guarantee in e-health applications,
IEEE journal of biomedical and health informatics 19 (3) (2015) 986–994.
[41] M. Elgendi, A. Mohamed, R. Ward, Efficient ecg compression and qrs detection for e-health applications, Scientific reports 7 (1) (2017) 459.
[42] D. Craven, B. McGinley, L. Kilmartin, M. Glavin, E. Jones, Adaptive dictionary reconstruction for compressed sensing of ecg signals, IEEE journal of
biomedical and health informatics 21 (3) (2017) 645–654.
[43] O. Yildirim, R. San Tan, U.R. Acharya, An efficient compression of ecg signals using deep convolutional autoencoders, Cognitive Systems Research 52
(2018) 198–211.
[44] T.Y. Liu, K.J. Lin, H.C. Wu, Ecg data encryption then compression using singular value decomposition, IEEE journal of biomedical and health informatics
22 (3) (2018) 707–713.
[45] C. Tan, L. Zhang, H.-t. Wu, A novel blaschke unwinding adaptive-fourier-decomposition-based signal compression algorithm with application on ecg
signals, IEEE journal of biomedical and health informatics 23 (2) (2019) 672–682.
[46] H.A. Akkar, W.A. Hadi, I.H. Al-Dosari, A squared-chebyshev wavelet thresholding based 1d signal compression, Defence Technology (2018).
[47] H. Fu, Y. Chi, Quantized spectral compressed sensing: Cramer–rao bounds and recovery algorithms, IEEE Transactions on Signal Processing 66 (12)
(2018) 3268–3279.
[48] Q. Wang, G. Qu, G. Han, A thresholding algorithm for sparse recovery via laplace norm, Signal, Image and Video Processing 13 (2) (2019) 389–395.
[49] A. Esmaeili, E.A. Kangarshahi, F. Marvasti, Iterative null space projection method with adaptive thresholding in sparse signal recovery, IET Signal
Processing 12 (5) (2018) 605–612.
[50] J. Maly, L. Palzer, Analysis of hard-thresholding for distributed compressed sensing with one-bit measurements, arXiv preprint arXiv:1805.03486
(2018).
[51] R.B. Roy, A. Roy, A. Mukherjee, A. Ghosh, S. Bhattacharyya, M.K. Naskar, Sparse encoding algorithm for real-time ecg compression, in: Recent Trends
in Signal and Image Processing, Springer, 2019, pp. 31–38.
[52] S. Mittal, A survey of techniques for approximate computing, ACM Computing Surveys 48 (4) (2016).
[53] A. Raha, V. Raghunathan, QLUT: Input-Aware Quantized Table Lookup for Energy-Efficient Approximate Accelerators, ACM Transactions on Embedded
Computing Systems 16 (5s) (2017).
[54] A. Raha, S. Venkataramani, V. Raghunathan, A. Raghunathan, Energy-efficient reduce-and-rank using input-adaptive approximations, IEEE Transactions
on Very Large Scale Integration (VLSI) Systems 25 (2) (2017) 462–475.
 A. Ghosh, A. Raha and A. Mukherjee / Internet of Things 9 (2020) 100166 17

[55] A. Raha, V. Raghunathan, Approximating beyond the processor: Exploring full-system energy-accuracy tradeoffs in a smart camera system, IEEE Trans-
actions on Very Large Scale Integration (VLSI) Systems 26 (12) (2018) 2884–2897.
[56] H. Mamaghanian, Chebyshev functional link artificial neural networks for denoising of image corrupted by saltand pepper noise, in: IEEE Trans.
Biomed. Eng, 58, IEEE, 2011, pp. 2456–2466.
[57] L. Polania, Compressed sensing based method for ecg compression, Proc. IEEE Int. Conf. Acoust., Speech Signal Process (2011) 761–764.
[58] S.M. S. K. Mishra G. Panda, Chebyshev functional link artificial neural networks for denoising of image corrupted by saltand pepper noise, International
Journal of Recent Trends in Engineering 1 (2009) 413–417.
[59] D. Chae, Performance study of compressive sampling for ecg signal compression in noisy and varying sparsity acquisition, IEEE Int. Conf. on Acoust.,
Speech and Signal Process (2009). 1306–109
[60] A. Ghosh, R.B. Roy, S. Bhattacharyya, A. Roy, M.K. Naskar, A. Mukherjee, Model of real time biomedical signal transmission using wavelet based
adaptive thresholding, in: 2017 IEEE Calcutta Conference (CALCON), IEEE, 2017, pp. 336–340.
[61] D.L. Donoho, I.M. Johnstone, Threshold selection for wavelet shrinkage of noisy data, in: Proceedings of 16th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society, 1, IEEE, 1994, pp. A24–A25.