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

net/publication/331608230

Edge Computing For Smart Health: Context-aware Approaches,

Opportunities, and Challenges

Article  in  IEEE Network · March 2019

DOI: 10.1109/MNET.2019.1800083

CITATIONS READS

139 780

5 authors, including:

Alaa Awad Abdellatif Amr Mohamed

Politecnico di Torino 290 PUBLICATIONS   4,106 CITATIONS   
67 PUBLICATIONS   866 CITATIONS   
SEE PROFILE
SEE PROFILE

Carla-Fabiana Chiasserini Aiman Erbad

Politecnico di Torino Hamad bin Khalifa University
459 PUBLICATIONS   8,697 CITATIONS    215 PUBLICATIONS   3,201 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Intrusion Detection System for IoTs View project

5Growth: 5G-enabled Growth in Vertical Industries View project

All content following this page was uploaded by Alaa Awad Abdellatif on 08 March 2019.

The user has requested enhancement of the downloaded file.

 Edge Computing For Smart Health:
Context-aware Approaches, Opportunities, and
Challenges
Alaa Awad Abdellatif∗† , Amr Mohamed∗ , Carla Fabiana Chiasserini† , Mounira Tlili+ , and Aiman Erbad∗
∗ Department of Computer Science and Engineering, Qatar University
† Department of Electronics and Telecommunications, Politecnico di Torino
+ Department of Computer Science, Carnegie Mellon University

E-mail: {aawad, amrm, aerbad}@qu.edu.qa, [email protected], [email protected]

Abstract—Improving efficiency of healthcare systems is detection, heart failure, etc.). All these things will report
a top national interest worldwide. However, the need of an impressive amount of data that need to be transported,
delivering scalable healthcare services to the patients while swiftly processed, and stored, while ensuring privacy pro-
reducing costs is a challenging issue. Among the most
promising approaches for enabling smart healthcare (s- tection. Given these requirements, the conventional cloud
health) are edge-computing capabilities and next-generation computing paradigm becomes unsuitable for s-health, since
wireless networking technologies that can provide real-time a centralized approach cannot provide a sufficiently high
and cost-effective patient remote monitoring. In this paper, we level of scalability and responsiveness, and it will impose
present our vision of exploiting multi-access edge computing while an exceedingly heavy traffic load to communication
(MEC) for s-health applications. We envision a MEC-based
architecture and discuss the benefits that it can bring to networks. A new approach has therefore emerged, known
realize in-network and context-aware processing so that the as Multi-access Edge Computing (MEC), defined as the
s-health requirements are met. We then present two main ability to process and store data at the edge of the network,
functionalities that can be implemented leveraging such an i.e., in the proximity of the data sources. The advantage
architecture to provide efficient data delivery, namely, mul- of MEC in a smart heath environment is multifold as it
timodal data compression and edge-based feature extraction
for event detection. The former allows efficient and low dis- can provide short response time, decreased energy con-
tortion compression, while the latter ensures high-reliability sumption for battery operated devices, network bandwidth
and fast response in case of emergency applications. Finally, saving, as well as secure transmission and data privacy [1].
we discuss the main challenges and opportunities that edge Furthermore, it can be applied to various network scenar-
computing could provide and possible directions for future ios, including cellular, WiFi and fixed access technologies.
research.
Index Terms—Edge computing, smart health, Internet This paper paves the way for MEC usage in smart heath
of Medical Things (IoMT), context-aware processing, deep environment through answering the following questions:
learning. • What are the motivations and main expected benefits
of leveraging the MEC architecture in s-health sys-
I. I NTRODUCTION
tems?
The evolution of computational intelligence and Internet • What are the s-health requirements, solutions of
of Medical Things (IoMT), along with the advances of MEC, and open challenges?
next-generation wireless technologies, has boosted the In what follows, Section II introduces a MEC-based
development of traditional healthcare processes into smart- system architecture that meets the s-health requirements,
healthcare services. Smart-health (s-health) can be consid- highlighting the benefits of pushing data processing and
ered as the context-aware evolution of mobile-health, lever- storage toward the data sources. Section III presents
aging wireless communication technologies to provide context-aware solutions for implementing multimodal data
healthcare stakeholders with innovative tools and solutions compression, in-network processing, and event-detection at
that can revolutionize service provisioning. In particular, s- the edge. Section IV then discusses some challenges that
health enables remote monitoring, where patients and care- MEC poses and further opportunities that such a paradigm
givers leverage mobile technologies to provide information offers. Finally, Section V concludes the paper.
about their health remotely – a service that is expected to
reduce hospitalization considerably and allow for timely II. MEC- BASED A RCHITECTURE FOR S MART H EALTH
delivery of healthcare services to remote communities at We now give a brief description of the proposed MEC-
low costs. based architecture for e-health applications, and discuss
S-health systems will also leverage various wireless the benefits that it offers to s-health systems.
sensors, cameras, and controllers, which permit patients’
automatic identification and tracking, correct drug–patient A. MEC-based S-Health Architecture
associations, and intensive real-time vital signs monitoring The proposed system architecture, shown in Figure 1,
for early detection of clinical deterioration (e.g., seizure stretches from the data sources located on or around
 Edge
Cloud
e
ns
po
Smart Home
R es
n cy
ge
er
Em
MEN

Hybrid Sensing Edge Cloud Monitoring and Services

Mobile/Infrastructure Edge Node (MEN)
Sources Provider

Data
PDA Raw data Doctor
Data Knowledge
Data Feature Event Adaptive MEC

Data
EEG Compressed Reconstruction
Data

Discovery
ECG Raw data Extraction Detection Compression Optimizer
data

Acquisition Classification
Emergency Response
Features

Data
SpO2 Pulse

Accelerometer

Patients' Relatives
and other Followers

Fig. 1. Proposed smart health system architecture.

patients to the service providers. It contains the following patient’s state) and wireless network conditions.
major components: Edge Cloud: It is a local edge cloud where data storage,
Hybrid sensing sources: A combination of sensing de- sophisticated data analysis methods for pattern detection,
vices attached/near to the patients represent the set of data trend discovery, and population health management can be
sources. Examples include: body area sensor networks (in- enabled. An example of the edge cloud can be a hospital,
cluding implantable or wearable medical and non-medical which monitors and records patients’ state while providing
sensors), IP cameras, smartphones, and external medical required help if needed.
devices. All such devices are leveraged for monitoring pa- Monitoring and services provider: A health service
tients’ state within the smart assisted environment, which provider can be a doctor, an intelligence ambulance, or
facilitates continuous-remote monitoring and automatic even a patient’s relative, who provides preventive, curative,
detection of emergency conditions. These hybrid sources emergency, or rehabilitative healthcare services to the
of information are attached to a mobile/infrastructure edge patients.
node to be locally processed and analyzed before sending
it to the cloud (see Figure 1). B. Benefits for s-health
Patient Data Aggregator (PDA): Typically, the wire- Given the characteristics and requirements of e-health
less Body Area Network (BAN) consists of several sensor applications, Table I summarizes some of the e-health
nodes that measure different vital signs, and a PDA which systems that can benefit from the above architecture. It
aggregates the data collected by a BAN and transmits it to is not the objective of this paper to provide an in depth
the network infrastructure. Thus, the PDA is working as a technical comparison on the different proposed e-health
communication hub that is deployed near to the patient to systems. However, we investigate the practical benefits of
transfer the gathered medical data to the infrastructure. leveraging MEC in such systems. In what follows, we will
Mobile/Infrastructure Edge Node (MEN): Herein, discuss the advantages of the proposed MEC architecture
a MEN implements intermediate processing and storage in the light of these systems.
functions between the data sources and the cloud. The 1) Monitoring systems using wearable devices: Heart
MEN fuses the medical and non-medical data from dif- monitoring applications are the most common type of
ferent sources, performs in-network processing on the remote monitoring applications. Monitoring vital signs
gathered data, classification and emergency notification, related to the heart reveals many types of diseases, e.g.,
extracts information of interest, and forwards the processed Cardiac arrhythmia, chronic heart failure, Ischemia and
data or extracted information to the cloud. Importantly, Myocardial Infarction [2][3][4]. In [2], authors present
various healthcare-related applications (apps) can be im- a real-time heart monitoring system, where the extract
plemented in the MEN, e.g., for long-term chronic disease medical data of the patients are transmitted to an Android
management. Such apps can help patients to actively based listening port via Bluetooth. Then, this listening
participate in their treatment and to ubiquitously interact port forwards these data to a web server for processing.
with their doctors anytime and anywhere. Furthermore, Also, [3] exploits Android smartphone to gather patient’s
with a MEN running specialized context-aware processing, information from wearable sensors and forward it to a web
various data sources can be connected and managed easily portal in order to facilitate the remote cardiac monitoring.
near the patient, while optimizing data delivery based on However, in these systems, the smartphone is used only
the context (i.e., data type, supported application, and as a communication hub to forward collected data to the
 TABLE I
S UMMARY OF THE E - HEALTH RELEVANT SYSTEMS .
Application Collected Data Description Limitations MEC Benefits
Cardiac disorder Electrocardiography Heart monitoring system is developed All data processing Data reduction
detection [2] (ECG) for detecting status of the patient tasks are performed and BW saving
and sending an alert message at a web server
in case of abnormalities
Requirements: long lifetime for
the battery-operated devices
Remote Cardiac Heart rate A location based real-time Cardiac Fewer number of Location
monitoring [3] blood pressure monitoring system is developed subjects participated Awareness and
body temperature Requirements: long lifetime for in the experiments energy saving
the battery-operated devices
Detection of Ischemia ECG and Electronic Presenting different methods leveraged Majority of the reviewed Data reduction
and Myocardial Health Records (EHR) ECG signal with EHR information literature did not exploit and BW saving
Infarction [4] to detect Ischemia and MI contextual information
Requirements: low computational
complexity
Parkinson’s disease (PD) Voice signal A PD monitoring system over the cloud All data processing Data reduction
detection [5] is proposed using feature selection and tasks are performed and BW saving
classification of a voice signal at the cloud
Requirements: Reliability and high
classification accuracy
Contactless heart rate Heart rate Heart rate measurement from facial Illumination variance, Data reduction
measurement [6] videos using digital camera sensor motion variance, and and BW saving
Requirements: Reliability and high motion artifacts
measurement accuracy
Prediction of ECG Leveraging point process analysis of Using single-channel Low latency
Bradycardia in preterm the heartbeat time series to predict ECG data to predict
infants [7] infant Bradycardia prior to onset Bradycardia
Requirements: Fast prediction of
emergency situations
Real-time epileptic Electroencephalography Automatic epileptic seizure detection Requiring large amount Low latency
seizure detection [8] (EEG) system using wavelet decomposition of data for training
is proposed to improve specificity
Requirements: Fast seizure detection of the detector
ECG change ECG A centralized approach for the detection Using one type of data Low latency
detection [9] of abnormalities and intrusions in for detecting abnormality
the ECG data is developed and emergency situations
Requirements: Fast detection of
abnormalities
Remote monitoring of Pulmonary Function Real-time tracking system of chronic Relying on one type Low latency
chronic obstructive Test (PFT) pulmonary patients comfortable in of data
pulmonary [10] their home environment is developed
Requirements: Fast detection of
abnormalities

cloud. Hence, continuous data transmission is not viable patient’s activities [6]. However, transmitting large vol-
due to the high energy toll it implies. The advantages of umes of data generated from these camera sensors using
implementing the proposed MEC architecture in such sys- conventional cloud-based architecture is not advisable and
tems are twofold. First, energy saving can be significantly may deem some of these applications impractical given the
increased by carefully managing the devices operational limited bandwidth availability. For instance, the amount
state and their data transfer at the MEN. In addition, data of digital data generated from a single-standard camera
compression as well as the proximity between sensors can reach to 40 GB per day. Accordingly, processing,
and MEN further reduce the energy consumption due to compressing, and extracting most important information
data transmission [11][12]. Second, the network edge can from the gathered data at the MEN greatly reduce the
be fruitfully exploited to extract context information and amount of data to be transferred toward the cloud, hence
apply localization techniques, which allows matching the the bandwidth consumption, and even makes it possible to
patient’s geographical position with the nearest appropriate store the data locally.
caregivers (e.g., hospital or ambulance).
3) Disorder prediction/detection systems: One of the
2) Contactless monitoring systems: Along with the promising applications of s-health, is the predictive mon-
evaluation of remote sensing, contactless monitoring has itoring of high-risk patients. The aim of these techniques
attained much focus recently. The main motivation of is improving prediction/detection of the emergency to
using contactless sensors is enabling ordinary life as much implement preventative strategies for reducing morbid-
comfortable as possible to all patients, since the patients ity and mortality associated with high-risk patients. For
are required only to be present within a few meters from instance, [7] presented a simplistic framework for near-
the sensors [5]. Heart rate measurement from facial videos term prediction of Bradycardia in preterm infants using
using digital camera sensors is one of the rapidly growing statistical features extracted from ECG signal. Also, [8]
directions to extract physiological signals without affecting proposed a quick seizure detection algorithm using fast
wavelet decomposition method. In such real-time predic- Our solution leverages deep learning, which is a good
tion/detection systems, the swift delivery of data to the candidate for multimodal data compression due to its abil-
server is a necessity. In many cases, this requires that data ity to efficiently exploit, not only the intra-modality cor-
are analyzed and even a diagnosis is made as close as relation, but also inter-correlation among different modal-
possible to the patient. However, detecting the changes of ities. Specifically, we use Stacked Auto-Encoders (SAE),
the physiological signals (e.g., changing in ECG values) in i.e., a special type of neural networks allowing for the
continuous health monitoring systems is not an easy task. hierarchical extraction of data representation [13]. SAE
It can be an indication for an emergency situation (e.g., consists of two main layers: (i) the encoding layers where
occurrence of a heart attack) [9][10]. This abnormality the data features are extracted, and (ii) the decoding layers
detection task becomes even more challenging during where the signal is reconstructed from the data coming
wireless communication transfer of patient’s data to the from the encoding layers. In our case, we implement the
cloud due to the erroneous communication and security encoding layers at the MEN, while the decoding layers
attacks that could introduce errors or makes changes in the are placed at a server in the cloud. Our key idea is to
patient’s data. Hence, quick detection of the changes in the progressively reduce the number of neurons in each of the
gathered medical data at the MEN is essential for real-time encoding layers at the MEN, and make the neural network
abnormal event detection. In a nutshell, the implementation learn from compressed version of the data. As a result,
of MEC architecture addresses all these issues, and the through the last encoding layer at the MEN (i.e., the one
ability of the MEN to perform event detection/prediction with the least number of neurons), we obtain a set of
fulfills these requirements even in the case of emergency features that are a compressed representation of the initial
applications. data. In summary, at the MEN, our SAE encoder converts
the input data x into the compressed data z, provided by
III. I MPLEMENTING THE E DGE N ODE F UNCTIONS the last encoding layer. At the server side (in the cloud),
The ultimate goal of our MEC architecture is to fulfill the SAE decoding layers will obtain the reconstructed data
the different requirements of e-health applications men- x̃, using the compressed representation z. The compressed
tioned above and enable s-health services through imple- and reconstructed signals can be written as:
menting the following main functionality at the network
edge: z = Wx + b (1)
• data compression, in order to reduce energy and band-
x̃ = W̃ z + b̃. (2)
width consumption in the case of health monitoring
systems; where W and W̃ are the encoder and decoder weights
• feature extraction and classification, in order to ensure matrices, respectively, while b and b̃ are the bias vectors.
high-reliability and fast response time in disorder The objective of SAE is to find the optimal configuration
prediction and detection. of the weight matrices and bias vectors that minimize
Below, we present how the above functionality can be the reconstruction error. In our case, instead, we first set
implemented at the MEN and highlight the benefits that the number of neurons in the last layer at the encoder,
the MEC architecture can bring. according to the desired compression ratio. Then we
optimize the number of neurons to be placed in each of
A. Multimodal data compression using deep learning the other encoding layers. Finally, by training the neural
The conventional approach used for health monitoring, network, we determine the optimal weight matrices and
i.e., transmitting the entire medical data wirelessly to bias vectors. We remark that, although the network training
the cloud, implies the transfer of a massive amount of is a computational expensive task, it can be conducted
data. For instance, in brain disorder monitoring systems, offline at the server side. Then such a configuration can be
EEG, Electromyography (EMG), and Electrooculography sent and used at the MEN for on-line data compression,
(EOG) data need to be stored and accessed remotely, thus leading to low-complexity, on-line data compression
along with video recording patient’s activities, in order to and transfer.
correlate the patient’s activities with her EEG pattern. This The advantages of multimodal data over single modality
would result in generating 8-10 GB per patient per day. A compression are twofold. First, we can account for inter-
promising methodology to deal with this issue in s-health modality correlation during compression, which results in
systems is to perform local in-network and data-specific a lower distortion while reconstructing the signal. Sec-
compression on the gathered data before transmission, ond, by concatenating the different modalities (i.e., EEG
while taking into account the applications’ requirements and EOG signals), it enables encoding the modalities in
and the characteristics of the data. a single-joint representation (i.e., the single stream z).
Here, we consider the EEG-EOG monitoring system as Figure 2 compares the proposed multimodal SAE (M-
a case study and present an efficient technique that deals SAE) with the Single Modality (SM) compression scheme,
with multimodal data, as required by s-health applications. which compresses each signal separately using SAE. As
In particular, we use the multimodal dataset in [13], the compression ratio varies, we observe that multimodal
which contains EEG and EOG signals of 32 people, who SAE allows for up to 50% reduction in EEG distortion
volunteered for this experiment, watching to 40 music with respect to SM, while EOG distortion increases by
videos. just 2%.
 50
M-SAE EEG PDA
M SAE EOG Transmit

Data
45
SM EEG Raw Data
40 SM EOG Data Feature Swift
Acquisition Extraction Classification
Transmit

Data

EEG Features
35

Raw Data
Features
Distortion %

30
Cloud
25 Classification

20 Knowledge Feature Extraction

Storage
Discovery & Classification
15

10 Fig. 3. Efficient class-based data transmission for s-health systems.

5

0
10 20 30 40 50 60 70 80 90 EEG data into the frequency domain, the normal/abnormal
Compression Ratio %
EEG classes under study exhibit different mean, median,
Fig. 2. Signal distortion as a function of the compression ratio for EEG and amplitude variations. Also, Root Mean Square (RMS)
and EOG signals, using M-SAE and SM compression schemes.
and Signal Energy (SE) are good signal strength estimators
in different frequency bands. Hence, to distinguish between
seizure and non-seizure events we select the following
B. Edge-based feature extraction and classification
five Frequency Features (FF): mean (µ), median (M), peak
Many neurodegenerative diseases detection methods, amplitude (P), RMS, and SE.
such as Parkinson’s, Epilepsy, Alzheimer’s, and Hunt- 2) Event-detection at the edge : The second step in
ington’s, have been reported in the literature based on our procedure consists in developing a reliable, edge-based
extracting some features from the patients’ vital signs, classification technique for seizure detection leveraging the
voice, or captured videos. Such features are used to dif- extracted features [15]. A number of machine learning
ferentiate a potential patient from a healthy person, or to techniques, including supervised, unsupervised and rein-
identify emergency situations. For instance, [5] proposes forcement learning, have been investigated for the purpose
a method to detect Parkinson’s disease (PD) leveraging of classification, for a variety of applications. In a nutshell,
certain features of the voice signal using cloud computing. supervised learning algorithms leverage a labeled training
Specifically, at the cloud server, the voice signal features data set to learn the relation between inputs and outputs.
are extracted and used for classification; the results are In contrast, unsupervised learning algorithms classify the
then sent to registered doctors for proper action. provided data sets into different clusters by discovering
In our study, we focus on epileptic seizure detection and the correlation between input samples. The third category
show the advantages of implementing feature extraction includes reinforcement learning algorithms and exploits
and classification at the MEN for efficient transmission and online learning, which involves the exploration of the
fast detection of abnormalities. We assume that the MEN environment and the exploitation of current knowledge,
gathers EEG data from the patient using an EEG Headset, in order to classify the data [16]. However, some im-
processes the data, and forwards them to the cloud. We portant issues arise when machine learning techniques
now use the EEG dataset in [14], which comprises three are applied to s-health: (i) an optimal trade-off between
classes of data, in the following denoted by A, B, and algorithms computational complexity and classification ac-
E, respectively. Each class contains 100 EEG records curacy should be established, (ii) sufficiently large datasets
corresponding to different patients. Each record includes have to be considered, in order to ensure high accuracy,
samples collected for 23.6 seconds at a 173.61 Hz rate. Sets (iii) a mathematical formulation of the learned model, as
A and B represent seizure-free subjects with eyes opened well as full control over the knowledge discovery process,
(A) and closed (B), respectively, while set E contains data is hard to obtain.
related to epileptic patients. In the considered case study, we define an IF-THEN
Using such data, we first perform feature extraction and classification rule using generated FF to detect abnormal
classification at the MEN. Then, depending on whether variations in sensed EEG data due to seizure. Thus, the
a seizure event was detected through classification, the status of the patient, S, is given by:
system sends to the cloud server the all data, or only
(
Normal if µ+M +P +RM S+SE ≤ γ
the computed features. Figure 3 summarizes the proposed S=
key concept. Below, we describe an efficient technique for Seizure if µ+M +P +RM S+SE > γ
feature extraction in Sec. III-B1, then we address event where γ is the classification threshold obtained during
detection and classification in Sec. III-B2. the offline training phase. Thus, leveraging the proposed
1) Feature extraction : In order to carry out the first low-complexity classifier, a quick emergency notification
step of our procedure, two approaches can be imple- system can be implemented at the edge to notify patient’s
mented: time-domain and frequency-domain feature ex- caregivers in case of emergency, as well as doctors at the
traction. Herein, we consider the frequency-domain ap- remote site.
proach due to its insensitivity to signal variations resultant In Figure 4, we compare the accuracy of the proposed
from electrode placement. By transforming the gathered Frequency Features Classifier (FFC) against that of differ-
 100
ent machine learning approaches, including random deci- CBS
90
sion forests (RandomForest), Naive Bayes (NaiveBayes), FFC

k-Nearest Neighbors (IBk), and classification/regression 80

trees (REPTree). Each of these classifiers is run using the 70

Battery Level (%)

default algorithm configuration in WEKA explorer with 60
5-fold cross-validation [17]. In FFC, when γ is small,
50
most of the obtained statuses will be Seizure, while at
high values of γ, most of the obtained statuses will be 40

Normal. In both cases, our classifier cannot accurately 30

differentiate between the patients’ classes. In the middle 20
region, when γ ranges between 0.5 and 0.8, our classifier
10
can discriminate between different classes yielding a high
0
accuracy. Notably, for γ ranging between 0.5 and 0.8, 0 5 10 15 20 25 30 35 40 45
our classifier outperforms other solutions, achieving 98.3% Time (hours)
accuracy for seizure detection for γ = 0.7. Fig. 5. Comparison of the proposed FFC technique with respect to CBS
Figure 5 assesses the performance of the proposed scheme, in terms of battery lifetime.
class-based data transmission (CDT) scheme, in terms
of MEN’s battery lifetime, comparing with cloud-based
system (CBS). CDT refers to the proposed scheme where privacy and security is not straightforward. Wireless med-
the MEN locally classifies the acquired data to decide ical devices are typically susceptible to various types of
whether the all data, or only the computed features, should threats, such as patient tracking and relaying, as well as
be sent to the cloud, while CBS refers to a traditional denial of service attacks, which violate confidentiality and
system where the MEN is used only as a communication integrity of the devices. Data processing algorithms and
hub to forward all acquired data to the cloud. Herein, we data storage may also be subject to attacks. Below, we
used a Samsung Galaxy S4 smartphone as a MEN, which discuss some challenges and opportunities that MEC poses
is connected to both a data emulator and server via WiFi. in this respect.
Interestingly, Figure 5 shows that CDT can improve the First is the ownership of the collected data from the
MEN’s battery lifetime by 50% with respect to CBS. patients. Storing the data at the patients’ proximity, where
it is collected, and enabling the patients to fully own the
data is a better solution for privacy protection. Also, the
patient will be able to control if the data should be stored
at the edge or transmitted to the cloud after removing or
hiding some of the private information from the data.
Second is the trade-off between increasing security level
and QoS. Increased security through strong cryptographic
algorithms or effective key management schemes [18],
adds more processing and additional overhead at the edge,
which may have a significantly adverse impact on QoS,
especially for real-time applications with strict delay and
throughput requirements. This imposes an essential need
to design joint QoS and security mechanisms for s-health
applications that maximize QoS while meeting the appli-
cation security requirements.

Fig. 4. Comparison of the proposed FFC technique with respect to B. Collaborative edge
RandomForest, NaiveBayes, IBk, and REPTree algorithms, in terms of
classification accuracy, with varying γ.
Healthcare requires data sharing and collaboration
among different stakeholders in multiple domains. How-
ever, sharing of data owned by a stakeholder rarely hap-
IV. C HALLENGES AND O PPORTUNITIES pens due to privacy concerns and the high cost of data
transfer. In this context, collaborative edge, which con-
In this section, we discuss three main challenges and op- nects the edges of multiple stakeholders that are geographi-
portunities that characterize MEC-based s-health systems cally distributed (such as hospitals, centers for disease con-
and represent interesting lines for future research. trol and prevention, pharmacies, and insurance companies),
is beneficial in threefold. First, it provides distributed
A. privacy and security data sharing among different stakeholders at low cost,
Great potential of s-health system can only be achieved thanks to computation and processing at the participant
if individuals are confident about the privacy of their edges. Second, in the case of remote monitoring, it enables
health-related information and providers are confident patients to forward their medical data to the cloud through
about the security of gathered data. However, ensuring other users/edge nodes. This also improves spectrum and
energy efficiency and allows data transferring even in ge- acquired data, and high level requirements of the consid-
ographically remote areas by exploiting D2D data transfer ered application should be integrated in order to provide
[19]. Third, it enables a patient’s edge node to directly sustainable and high-quality services for s-health systems.
connect to the nearest hospital’s edge in the proximity for In this context, we identified some computing tasks that
continuous monitoring, without the need of going through can be implemented at the edge and presented effective
the cloud. This helps to increase monitoring efficiency, approaches to implement them, so as to ensure short
reduce energy consumption and operational cost, as well response time, efficient processing and minimal energy
as improve high-quality services. and bandwidth consumption. Finally, we highlighted some
challenges and opportunities of edge computing in the s-
C. Combining heterogeneous sources of information health field that are worth further research.
Various sources of information are used in S-health
systems for efficient monitoring, hence, leveraging ad- ACKNOWLEDGMENT
vanced multimodal data processing techniques for com-
This work was made possible by GSRA grant # GSRA2-
bining these sources of information at the edge is a
1-0609-14026 from the Qatar National Research Fund (a
promising trend toward automating supervision and remote
member of Qatar Foundation). The work of Amr Mohamed
monitoring tasks. However, several challenges remain open
and Alaa Awad Abdellatif is partially supported by NPRP
when it comes to the s-health systems with hybrid sensing
grant # 8-408-2-172. The findings achieved herein are
sources. First, in terms of multiple modalities, it is not
solely the responsibility of the authors.
straightforward to incorporate and transmit multiple data
streams in s-health systems, where power consumption is a
R EFERENCES
limiting factor; indeed, transmission of highly informative
biosignals (e.g., EEG, EMG, and electrocardiogram) is [1] A. A. Abdellatif, M. G. Khafagy, A. Mohamed, and C. Chiasserini,
an energy hungry process for battery-operated devices. “Eeg-based transceiver design with data decomposition for health-
care iot applications,” IEEE Internet of Things Journal, pp. 1–1,
Second, signals artifacts arise from internal sources, e.g., 2018.
muscle activities and movements, as well as from external [2] A. Bansal, S. Kumar, A. Bajpai, V. N. Tiwari, M. Nayak,
sources related to noise, interference, and signals offset, S. Venkatesan, and R. Narayanan, “Remote health monitoring
system for detecting cardiac disorders,” IET Systems Biology, vol. 9,
which have critical implications on data quality [20]. no. 6, pp. 309–314, Dec 2015.
In this context, adopting a MEC-based s-health system [3] P. Kakria, N. K. Tripathi, and P. Kitipawang, “A real-time health
architecture would be beneficial in two ways. First, it monitoring system for remote cardiac patients using smartphone
and wearable sensors,” International Journal of Telemedicine and
permits to address system complexity associated with such Applications, vol. 2015, 2015.
heterogeneous and variable data-stream inputs. This is [4] S. Ansari, N. Farzaneh, M. Duda, K. Horan, H. B. Andersson,
done through implementing multimodal in-network pro- Z. D. Goldberger, B. K. Nallamothu, and K. Najarian, “A review
of automated methods for detection of myocardial ischemia and
cessing techniques that yield the correlation between dif- infarction using electrocardiogram and electronic health records,”
ferent modalities, in addition to the temporal correlation IEEE Reviews in Biomedical Engineering, vol. 10, pp. 264–298,
within each modality. Moreover, a MEC-based architecture 2017.
enables extracting high level application-based features at [5] M. Alhussein, “Monitoring Parkinson’s disease in smart cities,”
IEEE Access, vol. 5, pp. 19 835–19 841, 2017.
the edge rather than the cloud. By doing so, a MEN [6] M. Hassan, A. Malik, D. Fofi, N. Saad, B. Karasfi, Y. Ali, and
can send a limited number of the extracted features, or F. Meriaudeau, “Heart rate estimation using facial video: A review,”
the obtained correlations, instead of transmitting either Biomedical Signal Processing and Control, vol. 38, pp. 346–360,
2017.
the original or the compressed data. Second, advanced [7] A. H. Gee, R. Barbieri, D. Paydarfar, and P. Indic, “Predicting
signal processing for artifact removal can be incorporated Bradycardia in preterm infants using point process analysis of heart
at the edge, in order to improve signals quality before rate,” IEEE Transactions on Biomedical Engineering, vol. 64, no. 9,
pp. 2300–2308, Sept 2017.
transmission. [8] L. S. Vidyaratne and K. M. Iftekharuddin, “Real-time epileptic
seizure detection using EEG,” IEEE Transactions on Neural systems
V. C ONCLUSION and Rehabilitation Engineering, vol. 25, no. 11, Nov. 2017.
[9] F. A. Khan, N. A. H. Haldar, A. Ali, M. Iftikhar, T. A. Zia, and A. Y.
In this paper, we presented our vision of an s-health sys- Zomaya, “A continuous change detection mechanism to identify
tem leveraging the multi-access edge computing paradigm. anomalies in ecg signals for wban-based healthcare environments,”
IEEE Access, vol. 5, pp. 13 531–13 544, 2017.
Such an approach can indeed boost the system perfor- [10] A. H. Isik, I. Guler, and M. U. Sener, “A low-cost mobile adaptive
mance by efficiently handling the enormous amount of data tracking system for chronic pulmonary patients in home environ-
generated by sensors and personal as well as medical de- ment,” Telemed J E Health, vol. 19, no. 1, Jan 2017.
vices at the edge of the network, and addressing the limited [11] A. Awad, A. Mohamed, C.-F. Chiasserini, and T. Elfouly, “Dis-
tributed in-network processing and resource optimization over
energy capabilities of such devices. In particular, edge- mobile-health systems,” Journal of Network and Computer Appli-
based processing like compression and event detection cations, vol. 82, pp. 65–76, March 2017.
can greatly reduce the amount of data transferred toward [12] A. Awad, A. Saad, A. Jaoua, A. Mohamed, and C. F. Chiasserini,
“In-network data reduction approach based on smart sensing,”
the cloud, thus removing one of the major bottlenecks in IEEE Global Communications Conference (GLOBECOM), Dec
in s-health systems. Furthermore, processing data at the 2016, pp. 1–7.
edge will ensure better user privacy than when raw data [13] A. B. said, M. F. Al-Sa’D, M. Tlili, A. A. Abdellatif, A. Mohamed,
T. Elfouly, K. Harras, and M. D. O’Connor, “A deep learning
is uploaded to the cloud. Accordingly, we recommend approach for vital signs compression and energy efficient delivery
that wireless network components, characteristics of the in mhealth systems,” IEEE Access, vol. 6, pp. 33 727–33 739, 2018.
 [14] A. A. Abdellatif, A. Mohamed, and C. Chiasserini, “Automated and Computer Science California State University.
class-based compression for real-time epileptic seizure detection,” [18] X. Du, M. Guizani, Y. Xiao, and H. H. Chen, “A routing-driven
in Wireless Telecommunications Symposium (WTS), April 2018, pp. elliptic curve cryptography based key management scheme for
1–6. heterogeneous sensor networks,” IEEE Transactions on Wireless
[15] A. A. Abdellatif, A. Emam, C. Chiasserini, A. Mohamed, A. Jaoua, Communications, vol. 8, no. 3, pp. 1223–1229, March 2009.
and R. Ward, “Edge-based compression and classification for smart [19] A. Awad, A. Mohamed, C. F. Chiasserini, and T. Elfouly, “Network
healthcare systems: Concept, implementation and evaluation,” Ex- association with dynamic pricing over D2D-enabled heterogeneous
pert Systems with Applications, 2018. networks,” in IEEE Wireless Communications and Networking
[16] M. A. Alsheikh, S. Lin, D. Niyato, and H. P. Tan, “Machine Conference (WCNC), March 2017, pp. 1–6.
learning in wireless sensor networks: Algorithms, strategies, and [20] K. T. Sweeney, T. E. Ward, and S. F. McLoone, “Artifact removal
applications,” IEEE Communications Surveys Tutorials, vol. 16, in physiological signals: Practices and possibilities,” IEEE Trans-
no. 4, pp. 1996–2018, Fourthquarter 2014. actions on Information Technology in Biomedicine, vol. 16, no. 3,
[17] S. S. Aksenova, “Weka explorer tutorial,” School of Engineering pp. 488–500, May 2012.

View publication stats