Edge-Assisted Control for Healthcare Internet-of-Things: A Case Study on PPG-based Early Warning Score

Abstract

Recent advances in pervasive Internet of Things (IoT) technologies and edge computing have opened new avenues for development of ubiquitous health monitoring applications. Delivering an acceptable level of usability and accuracy for these healthcare IoT applications requires optimization of both system-driven and data-driven aspects which are typically done in a disjoint manner. While decoupled optimization of these processes yields local optima at each level, synergistic coupling of the system and data levels can lead to a holistic solution opening new opportunities for optimization. In this paper, we present an edge-assisted resource manager that dynamically controls the delity and duration of sensing w.r.t. changes in the patient's activity and health state, thus ne-tuning the trade-o between energy-e ciency and measurement accuracy. The cornerstone of our proposed solution is an intelligent low-latency real-time controller implemented at the edge layer that detects abnormalities in the patient's condition and accordingly adjusts the sensing parameters of a recon gurable wireless sensor node. We assess the e ciency of our proposed system via a case study of PPG-based medical Early Warning Score (EWS) system. Our experiments on a real full hardware-software EWS system reveal up to 49% power savings while maintaining the accuracy of the sensory data.

Full text

Edge-Assisted Control for Healthcare Internet-of-Things:
A Case Study on PPG-based Early Warning Score
ARMAN ANZANPOUR, University of Turku, Finland
DELARAM AMIRI, University of California, Irvine, USA
IMAN AZIMI, University of Turku, Finland
MARCO LEVORATO, University of California, Irvine, USA
NIKIL DUTT, University of California, Irvine, USA
PASI LILJEBERG, University of Turku, Finland
AMIR M. RAHMANI, University of California, Irvine, USA
Recent advances in pervasive Internet of Things (IoT) technologies and edge computing have opened new
avenues for development of ubiquitous health monitoring applications. Delivering an acceptable level of
usability and accuracy for these healthcare IoT applications requires optimization of both system-driven and
data-driven aspects, which are typically done in a disjoint manner. While decoupled optimization of these
processes yields local optima at each level, synergistic coupling of the system and data levels can lead to
a holistic solution opening new opportunities for optimization. In this paper, we present an edge-assisted
resource manager that dynamically controls the delity and duration of sensing w.r.t. changes in the patient’s
activity and health state, thus ne-tuning the trade-o between energy-eciency and measurement accuracy.
The cornerstone of our proposed solution is an intelligent low-latency real-time controller implemented at the
edge layer that detects abnormalities in the patient’s condition and accordingly adjusts the sensing parameters
of a recongurable wireless sensor node. We assess the eciency of our proposed system via a case study of
PPG-based medical Early Warning Score (EWS) system. Our experiments on a real full hardware-software
EWS system reveal up to 49% power savings while maintaining the accuracy of the sensory data.
Additional Key Words and Phrases: Health Monitoring, Wearable Electronics, Early Warning Score, Internet
of Things, Edge Computing, Edge-Assisted Control
1 INTRODUCTION
Remote health monitoring is expected to fundamentally transform many healthcare applications
to be used in everyday settings. Beyond the use of health monitoring systems for general health
reports, these systems can predict and prevent deterioration and death in chronic patients [
77
].
Such applications demand a satisfactory level of quality, which poses tremendous challenges in the
face of real-time and personalized health monitoring systems. A key mechanism to manage and
implement such applications is leveraging pervasive Internet of Things (IoT) technologies [
53
,
64
].
In the most prevalent IoT architecture, smart objects interchange their data with remote servers
through a gateway so that the processing resources are mostly concentrated on the server side. In
healthcare applications, utilizing such fully centralized processing methods puts the continuity of
remote patient monitoring services at the risk of failure. Edge (i.e., Fog) computing is a method for
extending the compute, storage, and network facilities to the edge of the network by outsourcing
the tasks needing low-latency and real-time support [5].
Edge computing not only provides local notication, processing, and storage but also reduces the
data trac through data fusion. Moreover, the lower radio signal strength required for wireless data
transmission in shorter ranges reduces the sensor node power consumption signicantly [56,57].
2020. XXXX-XXXX/2020/7-ART1 $15.00
DOI: 10.1145/nnnnnnn.nnnnnnn
1

1 A. Anzanpour et al.
Fig. 1. Coupling system-driven and data-driven ap-
proaches Fig. 2. ODA system control loop
There is a high demand to enhance the quality of experience in healthcare IoT applications
adaptively, generally from two dierent often disjoint aspects: system-driven and data-driven
(see Figure 1). There are several studies [
63
,
69
,
76
] on optimizing system-driven aspects by
leveraging cognition to enhance single or multiple system characteristics; however, these eorts
are in fact agnostic to the content and semantics of the transmitted data. For instance, consider
a remote healthcare monitoring application, where the health condition of an at-risk patient is
transitioned between “normal” to “abnormal”. Here, these semantics (from the data-driven aspects)
allow us to properly adjust the sensing settings and sensing duration (in the system-driven aspects).
Therefore, the system-driven parameters (e.g., energy and bandwidth) can be optimized considering
the requirements of “normal” and “abnormal” states without loss of signicant information.
With the prevalence of IoT-base remote patient monitoring, a vast number of medical conditions
are trackable, and most of their complications are preventable. Specically, in chronic diseases,
which are the leading cause of death and disability worldwide, the emergency situation precedes
with a sudden deterioration. It has been shown that the early signs of such deterioration are ob-
servable in the patient vital signs several hours earlier [
39
,
41
,
42
], and preventing the deterioration
reduces the risk of death [
75
]. Monitoring the vital signs of a chronic patient remotely enables the
detection of a potential deterioration and prevention of death or disability. However, the main issue
here is due to patient daily activities out of the hospital, which aects the quality of the recorded
signals [3].
In this paper, we propose an edge-assisted solution that makes the sensing, processing, and
conguration of the IoT system cognitive and self-aware. The system is capable of adapting its
sensing energy and signal quality parameters to the patient’s activity and health status, which
are dened as system context. The context-aware solution we propose to preserve the quality of
biosignals for an active subject is set based on an ODA loop (Observe,Decide,Act). In an ODA
system control loop (Figure 2) the process of Observation examines the current status of the system,
the Decision process decides how to control the system according to observations, and the Action
process applies the results of decision making on the system. The Decision that comes to Action
according to Observations on the involved system layers provides the system with a level of context-
awareness. In our solution, the observation and act processes happen on the sensing layer and the
decision-making which is the main contribution of our work happens on the edge. Observation
includes the assessment of the patient activity and health status, the Decision is the lowest power
state of the signal recording device while targeting an acceptable level of quality, and the Action
is a conguration command for sensing device dened in the Decision process. We chose edge
computing because it is the most ecient IoT architecture to implement our idea without imposing
any computation overhead to the sensor layer while oering rapid response at the edge. In the
healthcare domain, highly reliable services are needed which can be served more consistently and
promptly when the decision-making core is closer to the patient. The Action process is faster with
edge computing, and the risk of network failure is minimum.
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

Edge-Assisted Control for Healthcare IoT 1
The main novelty of our solution is i) a decision-making method which leverages both the
physical activity and health status of the patient (via Early Warning Score method) to decide
the required actuation on the sensor node, and ii) a cross-layer control mechanism in which the
application/data layer informs the system layer when opportunistic energy-savings would be
feasible while preserving the accuracy of sensing.
In this work, we also present a real use case of our system via a Photoplethysmograph (PPG)
sensor-based case study on Early Warning Score (EWS). EWS is a risk scoring method designed for
early detection of possible health deterioration using the patient’s vital signs [45].
The contributions of this paper are the following:
•
We design an edge-based remote health monitoring architecture and a recongurable
sensor node for Early Warning Score (EWS) assessment in out-of-hospital settings.
•
We assess the error/noise resiliency of feature extraction in the bio-signals for varying
sensing energy consumption and a person’s activity.
•
We formulate an optimization problem to minimize the long-term energy consumption of
a sensor node for a given target accuracy of extracted parameters.
•
We develop a run-time control algorithm state machine residing at the edge layer to enable
the real-time adaptation of the sensing parameters to patients’ states.
•
We evaluate our proposed system in terms of accuracy and energy consumption of the
sensing layer for an EWS case-study where a user engages in various physical activities
during daily routine.
The rest of the paper is organized as follow: In Section 2, we introduce essential concepts as a
background. Section 3outlines related works. The proposed healthcare IoT system is presented in
Section 4. Section 5demonstrates and evaluates the system performance, and Section 6concludes
the work.
2 BACKGROUND
The development of an edge-assisted remote health monitoring system through a context-based
optimization solution needs knowledge about patient health status [
4
,
46
]. Therefore, a reliable
health indicator is required to nd abnormalities. Here, we use the Early Warning Score (EWS) for
this purpose, which similar to most health indicators, relays on the changes in the vital signs. To
make a compact and wearable sensing device, we choose the PPG sensor as a source for most of the
vital signs. PPG signal is a rich source of medical information but very sensitive to patient activities.
In this section, we briey introduce the Early Warning Score (EWS) and Photoplethysmography
(PPG) methods. Moreover, we outline the limitations of PPG and our methods for vital signs and
activity detection.
2.1 Early Warning Score (EWS)
Studies show that health deterioration detection within a short time frame plays a crucial role
in patients’ survival [
52
,
68
]. There is solid evidence that such deterioration is distinguishable in
patients’ vital signs several hours prior to adverse medical events.[
65
] Thus far, several risk score
methods have been proposed to track the vital signs manually and notify such events earlier.[
15
]
Early Warning Score (EWS) proposed by Morgan et al. [
45
] is a tool for detecting the early signs of
health deterioration in order to prevent ICU admission. This method is an instruction for recording
and comparing patient vital signs (i.e., heart rate, blood pressure, respiration rate, body temperature,
blood oxygen saturation, and level of consciousness) to determine the severity of sickness. Based
on the observation of certain vital signs, EWS calculates a score for each vital sign where the sum
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

1 A. Anzanpour et al.
Table 1. A conventional Early Warning Scores (EWS) chart [73]
Score 3 2 1 0 1 2 3
Heart rate10-39 40-50 51-59 60-100 101-110 111-129 130+
Systolic BP20-69 70-80 81-100 101-149 150-169 170-179 180+
Respiratory rate30-8 9-14 15-20 21-29 30+
Body temperature40-35 35.1-38 38.1-39.5 39.6+
SpO2 (%) 0-84 85-89 90-94 95-100
Level of consciousness Alert Reacting to voice Reacting to pain Unresponsive
1beats per minute, 2mmHg, 3breaths per minute, 4◦C
of all individual scores shows the degree of abnormality. Table 1indicates a conventional early
warning score [45].
EWS is used in hospitals as a guideline for the track-and-trigger system, which updates the
number of patient’s observations per hour when the medical team is alerted in case of high scores.
The standard EWS scoring ranges for each vital sign are dened for a patient in a hospital setting.
However, in everyday settings, the activity and environment aect these limits and scores. In this
paper, we use a modied self-aware early warning score, which provides a scale for patient health
status considering the patient’s activity and environment [
9
,
11
,
25
]. For this purpose, a patient’s
activities should also be recorded together with standard vital signs.
2.2 Photoplethysmograph (PPG)
Photoplethysmograph (PPG) is an optical technique representing blood volume variations in the
microvascular [
28
] from which vital signs, including heart rate, respiration rate, and blood oxygen
saturation can be obtained. PPG measurement can be performed using a non-invasive and low-cost
miniaturized sensor. For this reason, PPG sensors are widely integrated into portable medical and
wearable sensors (e.g., tness trackers, wristbands, smart watches) to continuously capture vital
signs.
A PPG sensor consists of 1-2 single-wavelength light sources together with a light sensor in
contact with a body organ containing microvascular end-points such as ngertip (Figure 4(a)).
This sensor measures the amount of light reection or absorption by blood which depends on blood
volume, light wavelength and the amount of oxygenated and deoxygenated hemoglobin molecules.
Figure 4(b) shows the dierences in light absorption intensity for fully saturated (red line) and
fully desaturated (blue line) hemoglobin molecules in two selected wavelengths. The ratio of light
absorbance in these two sample wavelengths provides the value of blood oxygen saturation.
A PPG waveform reects volumetric changes in the microvascular bed of tissues. Such a signal
comprises two major components as alternating current (AC) and direct current (DC), whose
changes are attributed to synchronous variations in the blood volume oscillating with heartbeat
and respiration, respectively [
2
]. The heart beat and respiratory oscillations in a PPG signal are
indicated in Figure 5.
Recent works show that in addition to the above data, several other procedures, such as con-
tinuous measurement of nger arterial blood pressure, cardiac output measurement, and vascular
assessments are also possible via continuous PPG signal monitoring [12,32].
2.3 PPG Limitations
Despite many benets of using the PPG method for collecting vital signs, there are also some
limitations utilizing this method. The main drawback of the PPG method is its high power con-
sumption compared to other medical sensors. While the power consumption of ECG, respiration,
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

Edge-Assisted Control for Healthcare IoT 1
Fig. 3. PPG signal quality samples collected from finger tip: (a) high-current, sleeping (b) low-current, sleeping
(c) high-current, running
Fig. 4. (a) PPG sensors including two light sources and two light sensors (b) Light absorption intensity for
fully saturated and fully desaturated hemoglobin molecules [23,34]
and temperature sensors are in the range of microWatt, the power consumption of most PPG
sensors are in the range of milliWatt and without power management policies, the application of
this sensor in wearables would be minimal. PPG sensor has one or two light-emitting diodes (LED)
which turning them on draws 1 to 10 mA current. While applying higher currents makes the LEDs
brighter and leads to higher signal quality, in lower currents, the ambient light interferes with the
sensing process and makes a noisy signal.
The other drawback of the PPG signal is its vulnerability to body movements. Body movements
change the shape of blood vessels and surrounding tissues and make rapid changes in the DC part
of the signal. The noisy signals due to activities and low power measurements lead to low-accuracy
vital signs. [
47
]As mentioned earlier in section 1, the main contribution of this study is an edge-assisted
control that manages the power consumption of a PPG-based sensor node while keeping the signal
noise (and therefore the accuracy of vital signs) in an acceptable range. Figure 3shows the eect of
applying current and activity on the PPG signal.
2.4 Bio-signals Extraction
Thus far, several methods have been proposed to extract respiration rate and heart rate signals
from the raw PPG waveform [
17
,
54
]. For the respiration rate, there are two major methods to
obtain the signal named as feature-based and lter-based techniques.
In the feature-based techniques, specic features (e.g., maximum intensity of the pulse and
baseline variations) are extracted to provide respiration rate values [
33
]. However, these methods
are insucient for our system because the features cannot be obtained due to the eect of ambient
and motion noise on the signal. Such sources of noise are inevitable in ubiquitous health monitoring,
as users might engage in various physical activities in dierent environments.
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

1 A. Anzanpour et al.
Fig. 5. Oscillations reflected in a PPG signal. (a) heart rate. (b) respiration rate.
Fig. 6. Modified EWS calculation using PPG, temperature and acceleration signals
On the other hand, the respiration rate is obtained via lter-based techniques, where certain
band-pass lters are designed to discard non-respiratory frequency domain [
21
]. Correspondingly,
such lters can be designed for heart rate extraction.
In our system, a lter-based technique is tailored to mitigate the impact of motion artifacts on
the outcome. In this regard, two band-pass lters are designed according to the respiration rate and
heart rate frequency domains. The respiration and heartbeat frequency ranges are generally limited
to 0.1-1 Hz (6-60 breath rate/minute) and 0.5-3.0 Hz (30-180 heart rate/minute). The boundaries can
be chosen for the lters’ cuto frequency, although such a selection method might be inappropriate
when the user is moving, because the pass frequency is too wide. Hence, the lter’s pass frequency
needs to be precisely selected w.r.t. the current situation.
In this regard, a lter’s cut o frequency selection is extemporaneously performed exploiting
the peak values in the power spectral density (PSD) of the incoming PPG signal [
38
]. Figure 7
clearly illustrates frequency peaks corresponding to respiration and heartbeat oscillations in the
PSD of a one-minute PPG signal collected from a healthy person while he is sleeping. Heartbeat
signal is extracted using the peak value in the heart rate frequency range. As shown in the gure,
respiration frequency range might hold the heart rate peak as well. Since the heartbeat signal is rst
discarded, the peak value in the respiration frequency range reects the respiration signal. Note
that a high noise level nevertheless alters the structure of the PSD of the signal and subsequently
hinders the bio-signals extraction. Therefore, an adequate signal-to-noise ratio (SNR) is required in
this approach. Using this property, we can opportunistically adjust the sensing energy (i.e., the
power used by the sensor to emit/capture IR and red signals) depending on the current state of the
user,—e.g., “jogging” state leads to a low SNR, therefore a high sensing energy is needed to enhance
the signal, while “Sleeping” state results in a high SNR which demands a low sensing energy.
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

Edge-Assisted Control for Healthcare IoT 1
Fig. 7. The power spectral density of a one-minute PPG signal
Fig. 8. The alternative currents and direct currents of the IR and red signals to retrieve 𝑆𝑝𝑂2
2.5 Vital Signs Detection
In our system, we exploit a peak detection method in which both heart rate and respiration rate
are dierentiated. The average time interval between consecutive local maxima in the derivative
signals indicates the heart rate and respiration rate values.
Despite the aforementioned detection algorithm, we determine the
𝑆𝑝𝑂2
value via a feature-based
method, where four features are extracted: the alternative currents and direct currents of the IR
and red signals (i.e.,
𝐴𝐶𝐼𝑅
,
𝐴𝐶𝑅𝐸𝐷
,
𝐷𝐶𝐼 𝑅
and
𝐷𝐶𝑅𝐸 𝐷
) (see Figure 8). The
𝑆𝑝𝑂2
is determined using
the following equations:
𝑅=
𝐴𝐶𝑅𝐸𝐷 .𝐷𝐶𝐼𝑅
𝐴𝐶𝐼𝑅 .𝐷𝐶𝑅𝐸𝐷
(1)
𝑆𝑝𝑂2=𝛼𝑅2+𝛽𝑅 +𝛾(2)
where 𝛼,𝛽and 𝛾are constants retrieved from the sensor’s specication [31].
As indicated in Table 1, various vital signs are required to provide the warning score for users.
We use PPG signals to extract three of the vital signs. A temperature sensor is also used to measure
body temperature. Eventually, a modied warning score is calculated leveraging the four vital
signs along with physical activity data. Since patient deterioration is mainly caused by respiration
ineciency and lack of enough oxygen intake [
40
], and due to non-continuous nature of blood
pressure monitoring, the blood pressure is excluded from this setup. Figure 6illustrates the schema
of the score calculation process.
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

1 A. Anzanpour et al.
Fig. 9. The three-layer architecture of the proposed system
2.6 Physical Activity Recognition
In our work, the 3D acceleration data is utilized to determine the user’s physical activity continu-
ously. As the accelerometer sensor is placed at the user’s hand, we use hand movements to extract
the user’s steps. In this regard, rst, the acceleration data is ltered, mitigating the ambient noise.
Then, walking cycles are extracted from the data via a peak detection algorithm. Subsequently,
the steps are counted at each time interval. When the user is still (i.e., no step is detected) the
orientation of the user is leveraged to specify whether the current activity is sleeping or sitting.
In our setup, we use the steps and orientation data to determine a physical activity for each
time interval. The physical activities are “sleeping”, “sitting”, “walking”, “jogging” (i.e., moderate
activity) and “running” (i.e., vigorous activity) [71,72].
2.7 Modified EWS Calculation
The extracted vital signs are used in EWS calculation. As mentioned in Section 2.1, the conventional
EWS method is inappropriate for out-of-hospital setting, therefore we leverage a modied EWS
method proposed in [
9
,
11
] where context data (e.g., physical activities) are also considered in the
calculations. In such a method, the score ranges (Table 1) are dynamically ne-tuned according
to the user’s situations. This adjustment is performed via a pre-dened rule-based algorithm, in
which meta-data (e.g., user’s feedback) could also be added throughout the monitoring to deliver a
personalized decision making.
3 RELATED WORKS
The development of this research benets from several concepts and technologies. The main
objective of this study is to design an architecture for remote health monitoring, which satises
users’ needs with a minimum level of quality in every situation. Since an increase in physical
activity level decreases the signal-to-noise ratio for most medical signals, we recongure the
parameters of the sensing device according to the subject’s activity level and health status to keep
the quality of signal in an acceptable range. Such recongurations are applicable for a signal which:
a) is prone to noise during the activity b) has an adjustable knob which feeds more resources
to the system to provide a less noisy signal. Among medical signals, ECG, EMG, GSR, and PPG
signals have these characteristics. Since the patient’s health status is also an essential factor in
reconguration, we use the EWS method for health assessment and because the EWS requires the
vital signs, we choose PPG signal which has a compact sensor and can alone provide three vital
signs. To make the monitoring remotely possible, we leverage an IoT-based architecture for data
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

Edge-Assisted Control for Healthcare IoT 1
collection and to be able to respond to patient state fast, we process the data on the fog layer. To
analyze the incoming data and make the most right conguration for the sensor, we optimize the
activity/accuracy problem during design-time and run-time. To the best of our knowledge, the
proposed architecture is not described in the literature yet, but in each of the mentioned sections it
has several related works that we summarize them here.
Edge computing is dened as methods for relocating the processing load and network resources
from traditional cloud servers to local gateways [
13
,
61
]. The aim of these methods is to ooad
compute-intensive tasks to the edge devices to save energy [
18
,
19
,
37
,
55
], reduce the response
latency by avoiding massive data transmission to the cloud [
14
,
26
,
30
,
59
], and reduce the band-
width by data preprocessing [
30
]. Several works also present the benets of fog-assisted remote
reconguration of sensor nodes [
29
,
36
,
48
,
50
]. However, these solutions solely focus on the data
processing aspect, without an in-depth exploitation of edge devices to perform system optimization
through a closed-loop low-latency local control.
Some recent eorts have attempted to provide a degree of adaptivity to enable ecient IoT
services. Several frameworks have been proposed for bringing intelligence to IoT-enabled solutions
via architectural designs [
76
], energy management [
69
], object virtualization [
60
], and protocol
design [
1
]. Most of these works develop intelligence to optimize and control the device layer
at a user-state agnostic level. Our work presents a cognitive model that provides a system-user
tightly-coupled control of the data sensing and transmission processes by considering the state of
the user obtained from the same process.
Power consumption reduction techniques in sensor nodes have been proposed earlier using
dierent methods such as powering o sensors and radio transmission modules to reduce the power
consumed by those parts [
22
,
74
], reducing local processings [
6
], reducing data transmission[
67
],
and turning processing units to sleep or deep sleep modes[44].
Dierent schemes have been introduced to optimize the energy eciency of sensor network while
detecting the activities of an individual. Partially observable Markov decision process (POMDP)
frameworks are proposed in [
78
] and [
80
] to detect one’s activity based on noisy measurements of
mean and variance values of an accelerometer sensor and period of an ECG sensor using a Kalman
lter estimator. Optimization of an individual’s physical activity detection while minimizing the
energy consumption of wireless networks is proposed in [
79
]. However, Zois et al. [
79
] assume
that the energy bottleneck is at the gateway level rather than the sensor nodes, and reducing the
signal quality while maintaining a minimum level of accuracy has not been considered in these
solutions. These works have applied their power consumption reduction solution on the sensing or
the gateway device completely independent from the user’s health, while in our proposed method,
not only we set sensor node congurations according to user activity status, but also we pay
attention to his/her health status.
Most of the proposed sensor control solutions focus either on the quality of data (which we
consider it as data-driven aspects) or the eciency of the power consumption in the system (which
we name it system-driven aspects) and those few works that managed to couple these two aspects,
the optimization target has not been in the sensing layer. Our formulation in this paper is based
on the assumption that sensors are energy-constrained and their accuracy is a function of the
user-state (i.e., activity) and can be controlled at run-time.
Moreover, there are several solutions for IoT-based health monitoring [
7
,
8
], measuring the
accuracy of wearable health monitors [
27
,
49
,
58
], investigating the eect of activity on the accuracy
of health signals [
43
,
78
–
80
], assessment the eect of environmental parameters on the quality of
PPG signal [
16
,
20
,
70
], and adjusting the power consumption of a wearable medical device via
self-awareness [
9
], context-awareness [
24
], and goal management methods [
10
], but none of them
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

1 A. Anzanpour et al.
Fig. 10. Sensor device functions flowchart
has proposed a solution for conguring a sensing device in a fog-assisted control loop to optimize
the power consumption while preserving the quality of data in an acceptable level.
4 THE PROPOSED HEALTHCARE IOT SYSTEM
The proposed system architecture consists of three layers: the sensing layer, edge layer, and cloud
layer.
In the rst layer, recongurable sensing devices collect medical and activity data and send them
to smart gateways in the edge layer. In reply, the gateways send a new conguration instruction
to the sensors after evaluating the state of the user as well as the system. They also forward
the aggregated user data to the cloud. The cloud stores the patient information in a database for
long-term evaluations. Based on the stored data, the cloud updates the baseline parameters used in
our edge controller at regular intervals. Figure 9shows the three-layer architecture of the proposed
system.
4.1 Reconfigurable Sensor Node
To leverage the local control feature provided by the edge layer, we propose a recongurable sensor
node whose sensing delity and hibernate duration can be changed at run-time. The recongurable
sensor node is a remotely programmable data collector that receives a command from the smart
gateway regarding its new task in every iteration. In this work, our sensor node is a wireless
activity and medical vital signs data recorder. It uses a PPG sensor for recording the light reection
intensities from the ngertip, a temperature sensor for measuring the skin temperature, and a 3D
accelerometer for recording the patient’s activities. A micro-controller communicates with these
sensors and records collected data to a ash memory. After recording, the sensor node connects to
the gateway using a low-power Wi-Fi module, transmits the recorded data to the edge layer, asks
in reply for a new conguration command, and then goes to the hibernate mode for a time interval
specied in the conguration. After hibernation duration, the sensor node wakes up and starts
recording again according to the latest instruction. A web service in the gateway device receives,
stores, and analyzes data and denes a new command for the sensor node according to the current
patient’s state. Figure 10 shows the details of the sensor device functions owchart.
The conguration command contains instructions for dening record/hibernate duration and
adjusting the power consumption of active sensor modules. The power consumption states of the
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

Edge-Assisted Control for Healthcare IoT 1
Fig. 11. High level Edge-assisted control architecture
sensor node are changed by turning o the radio communication during data recording and also
by micro-controller hibernation. The power consumption of PPG sensor LEDs is also congurable
and denes the recording sub-states.
4.2 Personalized Edge-assisted Controller
One of our design goals is to provide a high Quality of Experience (QoE) to IoT services and
technologies. To this aim, we implement an edge-assisted controller interfaced with the sensing
and processing layers of the system. We contend that the edge processor is in a unique position to
bridge dierent scales of the system, enabling adaptation to local availability of resources and the
state of the monitored person based on global information.
In the proposed architecture, the system includes a set of cognitive algorithms that observe and
monitor states of the user to control sensing accuracy and duration and adapt the data acquisition
process to the current context. For instance, when a patient is in the “Sleeping” state, the magnitude
of motion artifacts is signicantly lower than the “Jogging” state. Therefore, it demands a lower
sensing energy to provide the same level of sensing accuracy (e.g., RMSE). Thus, activity-based
sensing control can save a considerable amount of energy.
Figure 11 depicts an overview of the proposed edge-assisted control architecture for our target
application. The self-aware cognitive engine at the edge layer optimizes the parameters of sensing
and hibernation in response to the person’s state. The objective is to maximize system’s lifetime
while maintaining a target detection accuracy from the PPG signal. The proposed optimization
involves two interconnected aspects: Sensing Fidelity and Sensing Duration.
4.2.1 Sensing Fidelity.
The system cognition consists of two phases, during which the edge
determines the sensor current level based on the Root Mean Squared Error (RMSE) requirements:
i) Sensing Fidelity Design-time:
In this phase, the edge processor builds a model connecting the
control parameters to the overall accuracy. The PPG signal sent from the sensor is pre-processed
and dierent features including,
𝑆𝑝𝑂2
, heart rate and respiration rate are extracted. Then, a set of
reference signals (
𝑆𝑝𝑂2
, heart rate and respiration rate) are compared with the calculated features.
Specically, we use an ECG sensor as a reference for heart rate, an airow sensor for respiration
reference, and another PPG sensor with higher overall quality as a reference for 𝑆 𝑝𝑂2.
We compute the RMSE of the calculated features in the edge layer to be a function of both the
current level, 𝑈and the activity state of the monitored person:
𝑋∈ {Sleeping, Sitting, Walking, Jogging, Running}.
Varying the PPG sensor’s current level
𝑈
, we calculate the total RMSE as the weighted sum of
the RMSE of the three individual features. The total error, then, is
𝜖(𝑈 , 𝑋 )={𝛾1𝜖1(𝑈 , 𝑋 ) + 𝛾2𝜖2(𝑈 , 𝑋 ) + 𝛾3𝜖3(𝑈 , 𝑋 )}
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

1 A. Anzanpour et al.
Fig. 12. System design-time block diagram
where
𝜖1(𝑈 , 𝑋 )
,
𝜖2(𝑈 , 𝑋 )
and
𝜖3(𝑈 , 𝑋 )
are associated with heart rate,
𝑆𝑝𝑂2
and respiration rate,
respectively. Please note that RMSE for each feature is calculated as the dierence between the
ground truth and the measured feature using dierent current levels,
𝑈
and activity state
𝑋
and
each value is squared. We then square each dierence to get a positive value, nd the summation
over each time period, divide the result by the length of the calculated feature, and calculate the
root of it. The variables
𝛾1,𝛾2, 𝛾3
are normalized weights corresponding to each feature, with
0
<𝛾1≤𝛾2≤𝛾3<
1. The control parameter
𝑈
species the PPG sensor’s current level. The block
diagram of system design-time is shown in Figure 12.
Optimization Problem:
In this method, the energy consumption in the sensor
𝐸(𝑈 , 𝑋 )
is a
function of the patient’s activity and sensor’s current level. Note that error measurements of the
sensor
𝜖(𝑈 , 𝑋 )
as a function of the patient’s activity and sensor’s current level should satisfy a
desired threshold
𝜏
. Therefore, the sensing delity is selected to minimize energy expense under a
constraint on the minimum accuracy in terms of overall RMSE. Thus,
min
𝑈𝐸(𝑈 , 𝑋 )subject to 𝜖(𝑈 , 𝑋 ) ≤ 𝜏(3)
The dened formulation is a convex optimization problem where one optimal solution can be
obtained. Lagrangian function
L(𝑈 , 𝜆)
is dened as a function of Lagrangian multiplier
𝜆
and
current level 𝑈to solve the optimization problem,
L(𝑈 , 𝜆)=𝐸(𝑈 , 𝑋 ) + 𝜆(𝜖(𝑈 , 𝑋 ) − 𝜏)(4)
where,
𝜆
is the trade-o parameter determining the relative importance of RMSE and the Energy
cost. The goal is to select the variable
𝑈
for each activity level
𝑋
such that the total cost will
be minimized. Therefore, we take the derivative of the Lagrangian function with respect to the
sensor’s current level,
𝜕𝐸 (𝑈 , 𝑋 )
𝜕𝑈 +𝜆𝜕𝜖 (𝑈 , 𝑋 )
𝜕𝑈
=0(5)
Considering a linear relation between energy consumption and sensor’s current level, derivative of
𝐸(𝑈 , 𝑋 )
with respect to
𝑈
is a constant value
𝑎𝑈
. The total RMSE
𝜖(𝑈 , 𝑋 )
can be interpolated as a
linear combination of the current level
𝑈
. The derivative of
𝜖(𝑈 , 𝑋 )
with respect to
𝑈
will be a
constant value
𝑏𝑈
. Therefore, the Lagrangian multiplier can be calculated as
𝜆=−𝑏𝑈
𝑎𝑈
. Solving this
optimization problem will obtain the optimal current level
𝑈∗
. Using this optimization solution, we
developed the Sensing Fidelity Run-time to determine the optimal current level in the monitoring.
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

Edge-Assisted Control for Healthcare IoT 1
Fig. 13. The readjustment algorithm
ii) Sensing Fidelity Run-time:
In this phase, the edge processor periodically determines the
lowest sensor’s current level in real-time such that the total RMSE of the extracted features is less
than a desired threshold.
4.2.2 Sensing and Hibernation Duration.
To adjust the sensing device parameters, we
consider a state machine which consists of four states for the patient and ve states for the device.
We dene two states for patient health status according to the EWS score [
35
], where any score
below 3 is considered as normal conditions (H1), and scores equal or above 3 as abnormal condition
(H2). Similarly, we consider sleeping and walking as low-intensity activities (A1) and walking,
jogging, and running as high-intensity activities (A2). Recent work has shown that ve minutes
is a reasonable interval to capture changes in one’s activity [
51
]. Therefore, we set the sensing
duration to ve minutes and ve dierent duration for hibernation from m1 to m5 minutes. Since
we aim to capture the states changing moment, the hibernation time is preferred to be elongated as
long as no change happens to the patient health or activity status. The sensing device follows the
state machine shown in Figure 14. It starts with state S1, which has the longest hibernation time
and stays there as long as no adverse event happens. With any abnormal or intense activity event,
it switches to the state S5, which has the lowest hibernation time (i.e., zero hibernation time or
continuous monitoring) and stays there while the abnormality is there. When the patient faces the
normal condition again, it increases the hibernation time step by step by switching to other states
from S4 to S1. The device does not stay in S4 to S2 states and just passes them one by one to reach
the initial state. It is a reasonable decision since the abnormality may return again, and the health
re-assessment of the patient would be possible with the shortest delay.
5 DEMONSTRATION AND EVALUATION
A remote recongurable sensor node was implemented together with the required edge services
to collect vital signs, assess the accuracy of sensors in dierent modes, and evaluate the energy
eciency of the system. A system for collecting the baseline for each parameter of interest was
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

1 A. Anzanpour et al.
Fig. 14. Sensing device readjustment state machine
also designed. We use ESP8266-12E WiFi module as a core for the sensor node. This module has an
80MHz 32-Bits RISC microprocessor with 96KB RAM and 4MB QSPI ash memory. Three sensor
modules are connected to this core, all of which are digital sensors congurable by internal registers
via
I2C
communication. The primary sensor module is MAX30102 PPG sensor placed on the left
hand index ngertip. This module has two LEDs and a light sensor working in 880nm wavelength
(infrared) and 660nm wavelength (red). The sampling rate and the current level of LEDs (and thus
their brightness) are congurable. The current level for driving the LEDs in the PPG sensor can
be congured to be one of the 0.8mA, 3.5mA, 6.3mA, 9.2mA or 12mA values. Other sensors are
TEMP102 temperature sensor module and MMA8451 3D accelerometer sensor module attached to
the left hand palm.
Figure 15 shows the sensor node. Since the ngertip is the most common site for PPG measure-
ment [
66
], we collected our reference signals from the ngertip and designed the sensing device
to record the PPG signal from this site on left hand index nger. The wearable device is in the
form of a glove which contains the electronic circuit and covers the sensor to prevent ambient light
from interfering with it. The user is expected to wear the glove during the monitoring. Compare
to other potential solutions for measuring the EWS, which requires a chest strap for heart rate, a
respiration sensor on the nostril, and a grip on nger or ear for
𝑆𝑝𝑂2
, this glove-shaped sensing
device is very compact and easy-to-wear. Since the sensor node uses Wi-Fi communication for data
transmission, the gateway should be a portable device. Therefore, we use an Android phone as a
gateway in our experiments. We implement the readjustment algorithm on the phone using the
Python programming language. A Samsung S6 Android phone with 1.5GHz octa-core processor
and 3GB RAM acts as a hotspot and creates a wireless network. It uses QPython script engine [
62
]
to create a web server and run the Python readjustment code. The cloud server is a virtual private
server (VPS) running an Apache web server and Python on an Ubuntu OS to perform periodic
RMSE updates.
The web service on the gateway device responds to conguration requests and performs the
readjustment algorithm for setting the sensor node conguration. The sensor node has three
states: recording state, transmission state, and hibernate state. According to the conguration, the
recording state has its own sub-states. We measure the power consumption of each state using a
power monitor device. Table 2shows sensor node power consumption in each state.
5.1 Accuracy Assessment
In this section, the performance of the proposed system is evaluated in terms of accuracy. We
conduct 12.5 hours of experiments including 150 (6
×
5
×
5) dierent conditions where six healthy
users engage ve dierent activities (i.e., Sleeping, Sitting, Walking, Jogging and Running) and the
PPG sensor varies as a function of current (i.e., 0.8mA, 3.5mA, 6.3mA, 9.2mA and 12mA).
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

Edge-Assisted Control for Healthcare IoT 1
Fig. 15. Sensor node and its components
Fig. 16. Reference signals wireless data logger
For each activity type, data collection was performed for 5 minutes. 30-second interval time was
selected as a computation window to extract vital signs from the PPG signal. Therefore, 10 values
are obtained for each activity. To collect the assessment reference signals, we develop a wireless
vital sign data logger which consists of four sensor nodes. The sensor nodes are developed using
Adafriut HUZZAH which is a development board for ESP8266-12E WiFi module. The recording
device is a Raspberry Pi 3 board equipped with a 3.5in display hat that records data and displays the
signals in real-time. A NodeJS server on the Raspberry Pi 3 board is responsible for receiving data
samples via UDP packets and recording them on the MicroSD card. Figure 16 shows the sensor
nodes and the Raspberry Pi 3 data logger. In this setup:
- A wristband xes two nodes on the left wrist that record the main PPG signal from the index
nger’s tip and reference PPG signal from the middle nger’s tip.
- A head strap xes a node on the forehead and a very accurate and sensitive temperature sensor
(MCP9808) sits in front of the nostril to sense the real respiration signal. The sensor gets warm
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

1 A. Anzanpour et al.
State Sensor node power consumption
R1: Recording, LEDs setting: 0.8mA 69.30 mW
R2: Recording, LEDs setting: 3.5mA 73.26 mW
R3: Recording, LEDs setting: 6.3mA 79.86 mW
R4: Recording, LEDs setting: 9.2mA 84.15 mW
R5: Recording, LEDs setting: 12mA 89.43 mW
T: Transferring 244.2 mW
H: Hibernate 5.94 mW
Table 2. Sensor node states power consumption
Hibernation Time
(Minute) Total Recording Time
(% of the day)
Total Missing Events
(%)
Average Power Consumption
(mW)
m1 m2 m3 m4 m5
Scenario 1 16 12 8 4 0 28.7% 30% 39.2mW
Scenario 2 8 6 4 2 0 41.1% 20% 53.8mW
Scenario 3 8 4 2 1 0 41.5% 10% 54.2mW
Scenario 4 7 5 3 1 0 44.6% 0 58mW
Scenario 5 4 3 2 1 0 57.8% 0 69.5mW
Baseline 0 0 0 0 0 83.3% 0 115.2mW
Table 3. Total monitoring time, average power consumption, and missing event due to dierent scenarios for
sensing device hibernation duration
during the exhale and gets cold during the inhale. The result is an oscillating signal which we use
to calculate the respiration rate reference.
- The 4th sensor node utilizes an ECG measurement chest strap (Polar T31) as a reference for
the real heart rate. It also contains a body temperature sensor (TMP102) and a 3D-accelerometer
sensor (MPU-9250).
To assess the accuracy, the extracted values through the PPG sensor are compared with the
true values. The total RMSE is calculated (see Section 4.2.1) to determine errors in the monitoring
with dierent user’s activities and dierent PPG sensor’s current level. As illustrated in Figure
17, the error increases when (a) the the amount of activity increases, and (b) the sensing current
decreases. This is more signicant in case of vigorous activities. 3.5mA is the boundary (i.e., lowest
current level) where an acceptable result could be obtained in any physical activity. However,
this boundary is 6.3mA for running and jogging activities. We tailor such information for our
cognitive edge controller design to deliver a satisfactory outcome, ensuring that the RMSE does not
exceed the threshold through the monitoring process. In the proposed architecture, the procedure
of calculating RMSE happens on the cloud server and its periodic recalculation requires collecting
new reference data that can be performed when the patient is ready to repeat the reference signals
data collection procedure.
5.2 Energy Evaluation
Energy management in the edge layer is an important factor to control the limited resources in the
sensor layer. To evaluate the proposed state machine in Sections 4.2 and 4.2.2, a healthy individual
was monitored during a 24-hour activity using the proposed sensor layer and controller in the edge
with dierent sets of timings for hibernation. We dened these timings in our state machine as
𝑚1
,
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

Edge-Assisted Control for Healthcare IoT 1
Fig. 17. (a) The eect of activity type on total RMSE, (b) The average of total RMSE for heart rate, respiration
rate and 𝑆𝑝𝑂2in dierent activities
Fig. 18. 24-hours health monitoring of a healthy person. (a) user’s activity level. (b) calculated modified EWS.
(c) sensor’s current level. (d) total RMSE expected regarding the user’s activity. (e) device states including
hibernate, dierent recordings and data transferring. (f) device power consumption through the monitoring.
The red line and green line indicate the baseline and our proposed system power consumption, respectively.
𝑚2
,
𝑚3
,
𝑚4
, and
𝑚5
. Denitely, using higher values for hibernation duration leads to lower energy
consumption, but with longer hibernation, some of the important health or activity events that
happen during the hibernation may be missed. To nd an optimal set of timings we dene ve
dierent scenarios and assess the energy eciency of the sensing device in each scenario. Table 3
shows the timing of each scenario, the total daily monitoring time, the percentage of missing events,
and the average power consumption of the sensing device. We choose the 4th scenario which is the
most ecient one without missing events. Recording the PPG signal and body temperature in the
sensor layer along with the activity level in the edge layer, we calculated the modied EWS score
from the extracted vital signs. Figures 18 (a) and 18 (b) demonstrate the activity level and calculated
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

1 A. Anzanpour et al.
EWS score. By extracting the EWS score and activity level, the readjusting algorithm is execucted.
At each given time, based on the EWS score and activity level, the threshold of RMSE will be
determined. High-level activities and/or high EWS score necessitate accurate monitoring leading
to choosing lower RMSE threshold values. In contrast, low-level activities including, ”Sitting” and
”Sleeping” with normal EWS score can tolerate some higher threshold of total RMSE.
In this experiment, we considered high-level activities and/or high EWS score that can tolerate
at maximum 1 in error in HR, maximum of 1 error in RR and maximum of 0.5 error in
𝑆𝑝𝑂2
. The
parameters dening the total RMSE are chosen based on the sensitivity of each vital signs during
dierent activities. Oxygen saturation has high sensitivity compared to other features in lower
current levels and can hardly be detected in low current levels due to mitigation of noise in the
signal. Respiratory rate is less sensitive to noise compared to the oxygen saturation. However, the
respiration rate is more prone to noise compared to the heart rate. Therefore, the three parameters
𝛾1,𝛾2, 𝛾3
are chosen based on the variations of the vital signs, so features with more sensitivity have
a slightly higher impact on calculating
𝑅𝑀𝑆𝐸
to take the sensitivity into account. Therefore, the
total RMSE threshold will be
𝜏=𝛾1𝜖1+𝛾2𝜖2+𝛾3𝜖3
for
𝛾1<𝛾2<𝛾3
and
𝛾1+𝛾2+𝛾3=
1(described in
Section 4.2). This leads to a threshold of
𝜏=
0
.
25
×
1
+
0
.
35
×
1
+
0
.
4
×
0
.
5
=
0
.
8. Low level activities
with normal EWS score can tolerate an error of
𝜖1=
2,
𝜖2=
2and
𝜖3=
1. The threshold for RMSE
will be
𝜏=
0
.
25
×
2
+
0
.
35
×
2
+
0
.
4
×
1
=
1
.
6. At each given time, the edge layer chooses the lowest
sensor’s current level to satisfy the threshold
𝜏
given the activity and EWS score calculation of an
individual. Figure 18 (c) determines the total RMSE that is expected by choosing each current level.
Note that the lowest RMSE values that can be obtained during ”Running” and ”Jogging” states are
2.931 and 1.431, respectively. In other words, due to motion artifacts, even the maximum current
level cannot improve the accuracy over a certain threshold. The average RMSE observed during
the experiment was 1
.
3276. Figure 18 (d) illustrates the implementation of the state machine while
running the scenario 4 for hibernation timings, determining the time the sensor toggles between
the on/transmit/o states. At each given time, when the sensor is on, detection of high-level activity
and/or high EWS score resets the hibernating time of the sensor to zero. Lower level activities
with normal EWS score enables the sensor node hibernation duration to increase state by state.
Figure 18 (e) shows the total power consumption of the sensor taking account the system state
and sensing duration. The average total power consumption in this experiment was recorded to
be 58 mW. Considering monitoring an individual with constant 12 mA sensor’s current level for
24 hours in 5-minutes intervals (i.e., the baseline) 115
.
2mW power consumption in the sensor
node can be obtained. Consequently, we observe that the edge controller is able to save 49% of
the battery power for this case study, demonstrating the ecacy of our edge-assisted intelligent
control scheme.
6 CONCLUSIONS
IoT technology and edge computing create new avenues for health monitoring and enable the
delivery of desirable Quality of Experience (QoE) to patients and healthcare providers in the
face of resource constraints and dynamic user behavior. In this context, traditional schemes
have investigated the optimization of system-driven and data-driven aspects separately, resulting
in missed opportunities for better optimization and system responsiveness to meet dynamically
changing healthcare application scenarios. In this paper, we outlined a scheme that jointly combines
system-driven and data-driven aspects using cognitive edge-assisted control, to deliver desirable
QoE while saving precious energy resources.
We presented a case study that deploys an edge-based health monitoring system that adjusts itself
to the most energy-ecient setting while retaining a desirable level of accuracy, considering patient
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

Edge-Assisted Control for Healthcare IoT 1
medical and activity state. We implemented our optimization method and run-time algorithm on a
recongurable sensor node and assessed the eciency of our proposed system through a PPG-based
Early Warning Score system case study. We demonstrated that our proposed edge-assisted system
reduced 49% of the overall power consumption by adapting to patient health state and activity
level while assuring a minimum level of accuracy. This scale provides a vision that a context-aware
system will consume approximately half of the energy that a non-context-aware system uses to
provide the same service. In the future work, we will optimize data transfer bandwidth considering
sensor selection and sensing sampling rate w.r.t. the patient’s conditions during the monitoring.
ACKNOWLEDGEMENTS
This work was partially supported by the US National Science Foundation (NSF) WiFiUS grant
CNS-1702950 and Academy of Finland grants 311764 and 311765.
REFERENCES
[1]
A. Aijaz et al. Cognitive machine-to-machine communications for internet of things: A protocol stack perspective.
IEEE Internet of Things, 2:103–112, 2015.
[2]
J. Allen. Photoplethysmography and its application in clinical physiological measurement. Physiol Meas., 28(3):1–39,
2007.
[3] D. Amiri et al. Edge-assisted sensor control in healthcare iot. In GLOBECOM’18 Conf., pages 1–6, 2018.
[4]
D. Amiri et al. Context-aware sensing via dynamic programming for edge-assisted wearable systems. ACM Trans.
Comput. Healthcare, 1(2), 2020.
[5] D. Amiri et al. Optimizing Energy Eciency of Wearable Sensors Using Fog-assisted Control, chapter 9, pages 245–268.
John Wiley and Sons, Ltd, 2020.
[6]
D. Anguita et al. Energy ecient smartphone-based activity recognition using xed-point arithmetic. J. UCS,
19(9):1295–1314, 2013.
[7]
A. Anzanpour et al. Internet of things enabled in-home health monitoring system using early warning score. In
MobiHealth’15 Conf., 2015.
[8]
A. Anzanpour et al. Context-aware early warning system for in-home healthcare using internet-of-things. In
HealthyIoT’16 Conf., 2016.
[9]
A. Anzanpour et al. Self-awareness in remote health monitoring systems using wearable electronics. In DATE’17
Conf., 2017.
[10]
A. Anzanpour et al. Energy-ecient and reliable wearable internet-of-things through fog-assisted dynamic goal
management. In ANT’19 Conf., 2019.
[11]
I. Azimi et al. Self-aware early warning score system for iot-based personalized healthcare. In LNICST’16 Conf., 2016.
[12]
K. Bartels et al. Advances in photoplethysmography: beyond arterial oxygen saturation. Canadian Journal of
Anaesthesia, 62(12):1313–1328, 2015.
[13]
F. Bonomi et al. Fog computing and its role in the internet of things. In First Edition of the MCC Workshop on Mobile
Cloud Computing, pages 13–16, 2012.
[14]
F. Bonomi et al. Fog computing: A platform for internet of things and analytics, in big data and internet of things: A
roadmap for smart environments. In Big Data and Internet of Things, 2014.
[15]
M. Brabrand et al. Risk scoring systems for adults admitted to the emergency department: a systematic review.
SJTREM, 18:8, 2010.
[16] P. C.-P. Chao et al. A portable, wireless photoplethysomography sensor for assessing health of arteriovenous stula
using class-weighted support vector machine. Sensors (Basel), 18, 2018.
[17]
P. Charlton et al. An assessment of algorithms to estimate respiratory rate from the electrocardiogram and photo-
plethysmogram. Physiol Meas., 37(4):610–26, 2016.
[18] B.-G. Chun et al. Clonecloud: Elastic execution between mobile device and cloud. In EuroSys’11 Conf., 2011.
[19] E. Cuervo et al. Maui: Making smartphones last longer with code ooad. In MobiSys’10 Conf., pages 49–62, 2010.
[20]
O. Dur et al. Design rationale and performance evaluation of the wavelet health wristband: Benchtop validation of a
wrist-worn physiological signal recorder. JMIR Mhealth Uhealth, 6, 2018.
[21]
A. Garde et al. Estimating respiratory and heart rates from the correntropy spectral density of the photoplethysmogram.
PLoS One, 9(1), 2014.
[22] D. Gordon et al. Energy-ecient activity recognition using prediction. In ISWC’12 Conf., 2012.
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

1 A. Anzanpour et al.
[23] W. B. Gratzer. . Med. Res. Council Labs, Holly Hill, London.
[24]
M. Götzinger et al. Enhancing the self-aware early warning score system through fuzzied data reliability assessment.
In MobiHealth’17 Conf., 2017.
[25]
M. Götzinger et al. Condence-enhanced early warning score based on fuzzy logic. ACM/Springer Mobile Networks
and Application, 2019.
[26] K. Ha et al. Towards wearable cognitive assistance. In MobiSys’14 Conf., pages 68–81, 2014.
[27] J. Hendrikx et al. Clinical evaluation of the measurement performance of the philips health watch: A within-person
comparative study. JMIR Mhealth Uhealth, 2(5), 2017.
[28]
A. Hertzman. The blood supply of various skin areas as estimated by the photoelectric plethysmograph. American
Journal of Physiology, 124:328–340, 1938.
[29]
C. M. Hsieh et al. Dance: Distributed application-aware node conguration engine in shared recongurable sensor
networks. In DATE’13 Conf., 2013.
[30]
P. Hu et al. Fog computing based face identication and resolution scheme in internet of things. IEEE Trans. Ind.
Inform., 13(4):1910–1920, 2017.
[31] M. Integrated. , (accessed 2018-08-01). https://www.maximintegrated.com/en/products/sensors/MAX30102.html.
[32]
B. M. Jayadevappa et al. Photoplethysmography: Design, development, analysis and applications in clinical and
physiological measurement – a review. IJIRSET, 5(3):3519–3531, 2016.
[33]
W. Karlen et al. Multiparameter respiratory rate estimation from the photoplethysmogram. IEEE Trans. on Biomedical
Engineering, 60(7):1946–53, 2013.
[34] N. Kollias. . Wellman Lab, Harvard Medical School, Boston.
[35]
U. Kyriacos et al. Monitoring vital signs: development of a modied early warning scoring (mews) system for general
wards in a developing country. PloS one, 9(1), 2014.
[36]
X. S. Le et al. Cardin: An agile environment for edge computing on recongurable sensor networks. In CSCI’16 Conf.,
2016.
[37]
R. LiKamWa et al. Energy characterization and optimization of image sensing toward continuous mobile vision. In
MobiSys’13 Conf., pages 69–82, 2013.
[38]
L. G. Lindberg et al. Monitoring of respiratory and heart rates using a bre-optic sensor. Med. Biol. Eng. Comput.,
30(5):533–7, 1992.
[39]
J. Ludikhuize et al. Identication of deteriorating patients on general wards; measurement of vital parameters and
potential eectiveness of the modied early warning score. J. Crit. Care., 27(4), 2012.
[40]
L. A. Lynn and J. P. Curry. Patterns of unexpected in-hospital deaths: a root cause analysis. Patient Safety in Surgery,
5(1), 2011.
[41]
I. D. Mapp et al. Predicting clinical deterioration in the hospital: the impact of outcome selection. Resuscitation, 84(5),
2013.
[42]
I. D. Mapp et al. Prevention of unplanned intensive care unit admissions and hospital mortality by early warning
systems. Dimens Crit Care Nurs., 32(6), 2013.
[43]
Y. O. Mohammed and U. A. Baroudi. Partially observable markov decision processes (pomdps) and wireless body area
networks (wban): A survey. KSII Trans. on Internet and Information Systems, 7, 2013.
[44] J. Moody et al. A -76dbm 7.4nw wakeup radio with automatic oset compensation. In ISSCC’18 Conf., 2018.
[45]
R. Morgan et al. An early warning scoring system for detecting developing critical illness. Clin Intensive Care, 8:100,
1997.
[46]
E. K. Naeini et al. An edge-assisted and smart system for real-time pain monitoring. In CHASE’19 Conf., pages 47–52,
2019.
[47]
E. K. Naeini et al. A real-time ppg quality assessment approach for healthcare internet-of-things. Procedia Computer
Science, 151:551–558, 2019.
[48] S. Natheswaran and G. Athisha. Remote recongurable wireless sensor node design for wireless sensor network. In
ICCSP’14 Conf., 2014.
[49]
B. Nelson and N. Allen. Accuracy of consumer wearable heart rate measurement during an ecologically valid 24-hour
period: Intraindividual validation study. JMIR Mhealth Uhealth, 7(3), 2019.
[50] T. Nyländen et al. Recongurable miniature sensor nodes for condition monitoring. In SAMOS’12 Conf., 2012.
[51] H. Oh and R. Jain. From multimedia logs to personal chronicles. In MM’17 Conf., 2017.
[52]
C. O’Keee et al. Role of ambulance response times in the survival of patients with out-of-hospital cardiac arrest.
Emerg Med J., 28(8):703–6, 2011.
[53] P. Patel et al. On using the intelligent edge for iot analytics. IEEE Intelligent Systems, 32(5):64–69, 2017.
[54]
M. A. F. Pimentel et al. Probabilistic Estimation of Respiratory Rate from Wearable Sensors, pages 241–62. Springer, 2015.
[55]
M. R. Ra et al. Odessa: Enabling interactive perception applications on mobile devices. In MobiSys’11 Conf., pages
43–56, 2011.
, Vol. 1, No. 1, Article 1. Publication date: July 2020.

Edge-Assisted Control for Healthcare IoT 1
[56] A. M. Rahmani et al. Fog Computing in the Internet of Things - Intelligence at the Edge. Springer, 2017.
[57]
A. M. Rahmani et al. Exploiting smart e-health gateways at the edge of healthcare internet-of-things: a fog computing
approach. Future Generation Computer Systems, 78(2):641–58, 2018.
[58]
R. K. Reddy et al. Accuracy of wrist-worn activity monitors during common daily physical activities and types of
structured exercise: Evaluation study. JMIR Mhealth Uhealth, 6(12), 2018.
[59]
S. Sarkar and S. Misra. Theoretical modelling of fog computing: a green computing paradigm to support iot applications.
IET Netw., 5(2):23–29, 2016.
[60]
S. Sasidharan et al. Cognitive management framework for internet of things: x2014; a prototype implementation. In
WF-IoT’14 Forum, pages 538–543, 2014.
[61] M. Satyanarayanan et al. Edge analytics in the internet of things. IEEE Pervasive Computing, 14:24–31, 2015.
[62] Q. scripting engine. , (accessed 2018-08-01). https://www.qpython.com/.
[63]
A. Sheth. Internet of things to smart iot through semantic, cognitive, and perceptual computing. IEEE Intelligent
Systems, 31(2):108–112, 2016.
[64] D.-H. Shin. Conceptualizing and measuring quality of experience of the internet of things: Exploring how quality is
perceived by users. Information &Management, 54:998–1011, 2017.
[65]
A. Taenzer et al. A review of current and emerging approaches to address failure-to-rescue. Anesthesiology, 115(2):421–
431, 2011.
[66] T. Tamura et al. Wearable photoplethysmographic sensors—past and present. Electronics, 3:282–302, 2014.
[67]
G. B. Tayeh et al. A new autonomous data transmission reduction method for wireless sensors networks. In
MENACOMM’18 Conf., 2018.
[68] TeleTracking. The value of time in healthcare. White paper, 2004.
[69]
J. Tervonen et al. Cognitive internet-of-things solutions enabled by wireless sensor and actuator networks. In
CogInfoCom’14 Conf., pages 97–102, 2014.
[70]
V. Toral et al. Recongurable ppg: Wearable system for biosignal acquisition and monitoring based on recongurable
technologies. Sensors (Basel), 19, 2019.
[71]
C. Tudor-Locke et al. Pedometer-determined step count guidelines for classifying walking intensity in a young
ostensibly healthy population. Can J Appl Physiol., 30(6):666–76, 2005.
[72] C. Tudor-Locke et al. Patterns of adult stepping cadence in the 2005-2006 nhanes. Prev Med., 53(3):178–81, 2011.
[73]
R. W. Urban et al. Modied early warning system as a predictor for hospital admissions and previous visits in
emergency departments. Advanced emergency nursing journal, 37(4):281–289, 2015.
[74]
Y. Wang et al. A framework of energy ecient mobile sensing for automatic user state recognition. In Mobisys’09
Conf., 2009.
[75]
M. Weenk et al. Continuous monitoring of vital signs using wearable devices on the general ward: Pilot study. JMIR
Mhealth Uhealth., 5(7), 2017.
[76]
Q. Wu et al. Cognitive internet of things: A new paradigm beyond connection. IEEE Internet of Things Journal,
1:129–143, 2014.
[77]
M. P. Young et al. Inpatient transfers to the intensive care unit: delays are associated with increased mortality and
morbidity. J Gen Intern Med., 18:77–83, 2003.
[78]
D. S. Zois et al. A pomdp framework for heterogeneous sensor selection in wireless body area networks. In INFOCOM’12
Conf., pages 2611–5, 2012.
[79]
D. S. Zois et al. Energy-ecient, heterogeneous sensor selection for physical activity detection in wireless body area
networks. IEEE Trans. signal processing, 61(7):1581–94, 2013.
[80]
D. S. Zois et al. Active classication for pomdps: A kalman-like state estimator. IEEE Trans. on Signal Processing,
62(23):6209–24, 2014.
, Vol. 1, No. 1, Article 1. Publication date: July 2020.