sensors

Article
Portable System for Real-Time Detection of
Stress Level
Jesus Minguillon 1,2,3, * ID , Eduardo Perez 2,3 , Miguel Angel Lopez-Gordo 2,3,4 ,
Francisco Pelayo 1,2 and Maria Jose Sanchez-Carrion 5
1 Department of Computer Architecture and Technology, University of Granada, 18014 Granada, Spain;
[email protected]
2 Research Centre for Information and Communications Technologies (CITIC), University of Granada,
18014 Granada, Spain; [email protected] (E.P.); [email protected] (M.A.L.-G.)
3 Department of Signal Theory, Telematics and Communications, University of Granada,
18014 Granada, Spain
4 Nicolo Association, 18194 Churriana de la Vega, Spain
5 School for Special Education San Rafael, 18001 Granada, Spain; [email protected]
* Correspondence: [email protected]; Tel.: +34-958-241-778




Received: 30 June 2018; Accepted: 28 July 2018; Published: 1 August 2018


Abstract: Currently, mental stress is a major problem in our society. It is related to a wide variety
of diseases and is mainly caused by daily-life factors. The use of mobile technology for healthcare
purposes has dramatically increased during the last few years. In particular, for out-of-lab stress
detection, a considerable number of biosignal-based methods and systems have been proposed.
However, these approaches have not matured yet into applications that are reliable and useful
enough to significantly improve people’s quality of life. Further research is needed. In this paper,
we propose a portable system for real-time detection of stress based on multiple biosignals such as
electroencephalography, electrocardiography, electromyography, and galvanic skin response. In order
to validate our system, we conducted a study using a previously published and well-established
methodology. In our study, ten subjects were stressed and then relaxed while their biosignals were
simultaneously recorded with the portable system. The results show that our system can classify
three levels of stress (stress, relax, and neutral) with a resolution of a few seconds and 86% accuracy.
This suggests that the proposed system could have a relevant impact on people’s lives. It can be used
to prevent stress episodes in many situations of everyday life such as work, school, and home.

Keywords: stress; biosignal; EEG; ECG; EMG; GSR; real-time; healthcare; e-Health; m-Health

1. Introduction
Stress is a major concern in our modern society. According to the 2014 report of the
American Psychological Association, most of U.S. population regularly experience physical (77%) or
psychological (73%) symptoms caused by stress, the main ones being fatigue (51%), headache (44%),
and upset stomach (34%). In addition, chronic stress has been proved to facilitate the development
of diseases due to weakening of the immune system [1]. All this adds up to important costs in terms
of people’s quality of life and loss of money (USD 300 billion of annual cost to employers in stress
related health care and missed work). According to the same report, the top causes of stress in the US
are job pressure, money, health, and relationships. Therefore, stress is mainly caused by everyday-life
factors. Thus, it is crucial to develop reliable and usable systems for real-time detection of stress level
in people’s daily life.

Sensors 2018, 18, 2504; doi:10.3390/s18082504 www.mdpi.com/journal/sensors

Sensors 2018, 18, 2504 2 of 15

New technologies have attempted to improve people’s quality of life in the last few years [2].
The development of pervasive and ubiquitous systems and applications has led us into modern
terms such as e-Health and m-Health. These two concepts encompass information, communication,
and mobile technologies for healthcare purposes. e-Health has shown a relevant impact on the quality
and safety of healthcare [3]. For example, facilitating the communications between institutions [4],
incrementing patient engagement to treatment [5], promoting physical activity in older adults [6],
and improving mental health services for trauma survivors [7]. m-Health, for its part, has shown
its effectiveness in multiple scopes, such as monitoring health in elderly people [8], promoting early
diagnosis of cardiovascular diseases [9], differentiating between Parkinson’s disease and essential
tremor diagnosis [10], improving hypertension control in stroke survivors [11], and supporting
recovery from drug addiction [12].
Regarding the stress detection, methods and systems based on biosignal analysis are under
study. These objective approaches are usually more powerful than self-perception of stress level [13].
For example, some patterns extracted from electrocardiography (ECG) such as heart rate or heart
rate variability have been related to mental stress [14–19]. The activity of some muscles such as the
trapezius has been proved to be connected with stress [20–23]. The muscle activity can be measured
by electromyography (EMG). Other studies have demonstrated the relationship between stress and
certain brain rhythms measured by electroencephalography (EEG) [24–31]. The skin conductance has
also been correlated with stress [32–34]. This parameter can be measured using galvanic skin response
(GSR) sensors. All this knowledge has been used by many researchers to propose portable systems
for assessment and detection of mental stress. These systems usually combine multiple biosignals.
Examples include wearable assessment of mental stress of combatants [35], wristband sensor to
measure stress level for people with dementia [36], and stress detection in drivers [37–39]. In short,
much useful work has been done. Nevertheless, beyond the commercial gadgets, ambulatory
stress-monitoring has not matured yet in applications that are reliable and valid enough to convincingly
improve people’s health and quality of life. Further research is needed in this field aimed at tackling
such an important and serious problem.
In this work, we present and validate a portable system for real-time detection of stress level,
based on the RABio w8 (real-time acquisition of biosignals, wireless, eight channels) system. We have
designed and implemented both hardware and software in our laboratory. The hardware is made of
portable, wireless, and low-cost electronics. The software is composed by an application programming
interface (API) and a graphical user interface (GUI). We conducted a study to validate our system using
proven and well-established methodology to induce different levels of stress. Our results demonstrate
the potential application of our system as a useful tool for ubiquitous stress monitoring, detection,
and prevention.

2. Materials and Methods

2.1. Description of the System

As mentioned before, the portable system for real-time detection of stress level presented in this
work is based on the RABio w8 system. RABio w8 is a portable, wireless, low-cost hardware–software
system for the acquisition and processing of multiple biosignals such as EEG, ECG, and EMG. It has
been used in previous works [40].
The electronics of RABio w8 is composed of three blocks (see Figure 1a): acquisition block,
control block, and communication block. The acquisition block uses advanced integrated circuits
for biosignal acquisition from the ADS family of Texas Instruments (Dallas, TX, USA). This block is
in charge of the amplification and the analogue-digital conversion of eight simultaneous channels,
up to 1000 samples per second with 24-bit sample resolution. The gain factor of every single channel
and the sampling rate are configurable. This block interacts with the control block through a serial
peripheral interface (SPI). The control block uses a microcontroller from Microchip Technology
Sensors 2018, 18, 2504 3 of 15

(Chandler, AZ, USA) to receive, synchronize, format, and send the data frames from the first block to
the communication
Sensors blockREVIEW
2018, 18, x FOR PEER through a universal asynchronous receiver–transmitter (UART) port. Finally,3 of 15
the communication
Sensors block
2018, 18, x FOR PEER is responsible for the wireless communication with the software of RABio
REVIEW 3 of 15
RABio
w8 w8 via Bluetooth.
via Bluetooth. All the electronics
All the electronics are powered arebypowered by high-autonomy
high-autonomy lithium
lithium polymer polymer
rechargeable
RABio
batteries w8
andvia
rechargeable Bluetooth.
batteries
contained and All the electronics
in acontained in plastic
3D printed are powered
a 3D printed
casing plastic by1b).
casing
(see Figure high-autonomy
(see Figure 1b). lithium polymer
rechargeable batteries and contained in a 3D printed plastic casing (see Figure 1b).

(a) (b) (c)

(a) (b) (c)
Figure 1. RABio
Figure 1. RABio w8
w8 system:
system: (a)
(a) Diagram
Diagram of
of the
the electronics;
electronics; (b)
(b) Picture
Picture of
of the
the hardware;
hardware; (c)
(c) Screenshot
Screenshot
Figure
of the 1. RABio user
graphical w8 system: (a)(GUI).
interface Diagram of the electronics; (b) Picture of the hardware; (c) Screenshot
of the graphical user interface (GUI).
of the graphical user interface (GUI).
The software of RABio w8 is composed by an application programming interface and a
The software ofofRABio w8w8 is composed by anby application programming interface and a graphical
graphical software
The user interface. RABio
The API isisa dynamic-link
composed an application
library of Windows programming
OS coded in interface and a
C/C++. It allows
user interface.
graphical The
user data API
interface. is a
The dynamic-link
APIthe library
is aelectronics
dynamic-link of Windows
libraryw8, OS
of Windowscoded in C/C++. It allows one to
one to receive frames from of RABio as well asOS to coded
configurein C/C++. It allows
the acquisition
receive
one data
to receive frames
data from the
framesgain electronics
from of
thesampling RABio
electronics w8,
of andas
RABio well as
w8, asto configure
wellmarkers the acquisition
as to configure parameters
the acquisition
parameters (i.e., channels and rate) to send event via Bluetooth. The GUI
(i.e., channels
parameters gain
(i.e., and sampling
channels gain and rate) and torate)
sampling sendandeventto markers
send via
event Bluetooth.
markers via The GUI of
Bluetooth. TheRABio
GUI
of RABio w8 (see Figure 1c) is coded in Matlab from The Mathworks (Natick, MA, USA).
w8
of (see
RABio Figure 1c)
w8 (see Figureis coded in
1c) by Matlab
is coded from
into The
Matlab Mathworks (Natick, MA, USA). The GUI uses the
uses the functions provided the API allow from ThetoMathworks
the user (Natick,and
visualize process MA, USA).
record The
the GUI
signals
functions
uses provided
the functions by the API to
providedinbyreal-time.allow
the API to the user to visualize
allow the user process and record the signals acquired
acquired by the electronics Configuration of to visualize parameters
acquisition process andand record
eventthemarking
signals
by the
acquired electronics in
by the electronicsreal-time. Configuration of acquisition parameters and event marking is also
is also available for the user.in real-time. Configuration of acquisition parameters and event marking
available
is also for the for
available user. thesystem
user. for real-time detection of stress level (see Figure 2) consists of multiple
The full portable
The
The full
full portable
portable system for
system for real-time
real-time detection
detection of of stress
stress level
level (see
(seew8 Figure
Figure 2)
2) consists
consists of
of multiple
multiple
biosignal sensors (EEG, ECG, EMG, and GSR electrodes), the RABio system, a laptop, and the
biosignal
biosignal sensors (EEG,
sensorsplatform
(EEG, ECG,ECG, EMG,
EMG, and and GSR electrodes), the RABio w8 system, a laptop, and the
e-Health sensor of Arduino. TheGSREEG, electrodes),
ECG, and the EMG RABio w8 system,
electrodes a laptop,
are directly and
attached the
to
e-Health sensor
e-Health sensorof platform
platform ofofArduino.
Arduino. TheTheEEG,
EEG,ECG,
ECG,andand EMG EMGelectrodes
electrodesare directly attached to input
input channels the RABio w8 hardware. The GSR electrodes are attachedare directly
to the attached
e-Health to
shield.
channels
input of the RABio
channels of the w8 RABiohardware. The GSRThe electrodes are attached toattached
the e-Health shield. Thisshield.
shield
This shield is powered by anw8 hardware.
Arduino board and GSRprovides
electrodes areconductance
skin to the e-Health
measurements. The
is powered
This by
is an
shieldvalues Arduino
powered boardArduino
and provides skin
andconductance measurements. The measured values
measured are sentby to an
the RABio w8 board
hardware provides
by connecting skin conductance
the analogue measurements.
output of the shieldThe
are sent to values
measured the RABiosent w8 hardware by connecting thebyanalogue output analogue
of the shield (A2)oftothe
anshield
input
(A2) to an inputare channeltooftheRABioRABio w8 hardware
w8. The laptop isconnecting
in chargethe of visualizing, output
processing, and
channel
(A2) to of
an RABio
input w8. The
channel laptop
of RABiois in charge
w8. The of visualizing,
laptop is in processing,
charge of and recording
visualizing, the acquired
processing, and
recording the acquired biosignals using the API and the GUI of RABio w8.
biosignals the
recording using the APIbiosignals
acquired and the GUI usingof the
RABio
APIw8.and the GUI of RABio w8.

Figure 2. Diagram of the full portable system for real-time detection of stress level. The system is
Figure 2.
Figure Diagram
2. Diagram
composed fullmultiple
of thew8,
by the RABio portablebiosignal
system for real-time
sensors detection
placed at head,oftrapezius,
stress
stress level.
level. The
wrist system
and is
fingers,
composed
composed
the Arduinoby the RABio
bye-Health
the RABio w8, multiple
w8, multiple
platform, biosignal sensors placed at head, trapezius, wrist and fingers,
biosignal sensors placed at head, trapezius, wrist and fingers,
and a laptop.
the Arduino
the Arduino e-Health
e-Health platform,
platform, and
and aa laptop.
laptop.
For the purpose of this work (i.e., presentation and validation of our system), a laptop was
used.For the purpose
However, of this
in a final work (i.e.,
version, presentation
we propose and validation ofofour
the cloud-computing system),with
biosignals a laptop was
real-time
used. However,
biofeedback in a final
presented version,devices
in mobile we propose
such asthetablets
cloud-computing of biosignals
or smartphones. with wearable
Also, a more real-time
biofeedback
version of thepresented in mobile devices
EEG cap embedding such
the whole as tabletsis or
electronics smartphones.
feasible and underAlso, a more wearable
development.
version of the EEG cap embedding the whole electronics is feasible and under development.
Sensors 2018, 18, 2504 4 of 15

For the purpose of this work (i.e., presentation and validation of our system), a laptop was used.
However, in a final version, we propose the cloud-computing of biosignals with real-time biofeedback
presented in mobile devices such as tablets or smartphones. Also, a more wearable version of the EEG
Sensors 2018, 18, x FOR PEER REVIEW 4 of 15
cap embedding the whole electronics is feasible and under development.
2.2.Experimental
2.2. ExperimentalProcedure
Procedure
Weconducted
We conducted a study
a study in to
in order order to validate
validate ourfollowing
our system, system, thefollowing the well-established
well-established methodology
methodology of previous published stress studies [28,31]. Ten healthy
of previous published stress studies [28,31]. Ten healthy volunteers were involved volunteers wereininvolved in
the study
the study (five male, five female, age range of 18–23 years, mean age of 20 ± 2 years, all of them
(five male, five female, age range of 18–23 years, mean age of 20 ± 2 years, all of them novice in
novice in stress-related experiments). The recruitment process started one month prior to the
stress-related experiments). The recruitment process started one month prior to the beginning of the
beginning of the study by means of informative emails. The participants were instructed to avoid
study by means of informative emails. The participants were instructed to avoid stimulants or relaxant
stimulants or relaxant substances in the 3 h prior to the experiment. They were not paid for their
substances in the 3 h prior to the experiment. They were not paid for their participation. They were
participation. They were provided with the experiment’s information sheet and the informed
provided with the experiment’s information sheet and the informed consent, both of which were
consent, both of which were approved by the Bioethics Committee of the University of Granada.
approved by the Bioethics Committee of the University of Granada.
The participants were prepared by the research staff after they read, understood, and signed the
The participants were prepared by the research staff after they read, understood, and signed the
informed consent (see Figure 3a). They wore white hospital clothes during the experiment. Four
informed consent (see Figure 3a). They wore white hospital clothes during the experiment. Four EEG
EEG electrodes were placed at Fp1, Fp2, F3, and F4 positions of the 10–20 International System using
electrodes were placed at Fp1, Fp2, F3, and F4 positions of the 10–20 International System using an EEG
an EEG cap. These positions have been successfully used in stress studies [24,25,27,28,31]. One ECG
cap. These positions have been successfully used in stress studies [24,25,27,28,31]. One ECG electrode
electrode was placed on the wrist of the non-dominant hand. Two EMG electrodes were placed on
was placed on the wrist of the non-dominant hand. Two EMG electrodes were placed on the trapezius
the trapezius muscle of the non-dominant-hand side, with an inter-electrode distance of 25 mm. The
muscle of the non-dominant-hand side, with an inter-electrode distance of 25 mm. The activity of the
activity of the trapezius has been related to stress in several published studies [20–23]. Two GSR
trapezius
electrodeshas been
were related
placed on to
thestress
indexinand
several published
the middle studies
fingers [20–23].
of the Two GSRhand
non-dominant electrodes
[32,33].were
All
placed on the index and the middle fingers of the non-dominant hand [32,33]. All
the electrodes were referenced and grounded to the ear lobe of the dominant-hand side. All the the electrodes were
referenced
electrode and grounded
impedances to the
were ear lobe
below of the
30 KΩ. dominant-hand
The EEG, ECG, and side.
EMGAll the electrode
electrodes impedances
were directly
were below 30 KΩ. The EEG, ECG, and EMG electrodes were directly attached to
attached to the input channels 0–6 of RABio w8. The GSR electrodes were attached to the Arduino the input channels
0–6 of RABio
e-Health w8.and
shield Thethe
GSR electrodes
analogue wereA2
output attached to the Arduino
was connected e-Health
to the input shield7 and
channel the analogue
of RABio w8, as
output A2 was connected
described in Section 2.1. to the input channel 7 of RABio w8, as described in Section 2.1.

(a) (b)
Figure 3. (a) Picture of one participant ready for the experiment after preparation; (b) timeline of the
Figure 3. (a) Picture of one participant ready for the experiment after preparation; (b) timeline of the
experiment. Duration of each part is in seconds (s). The total duration was around 30 min, including
experiment. Duration of each part is in seconds (s). The total duration was around 30 min, including the
the transition periods (see text for details). MVC—maximum voluntary contraction; RS—resting
transition periods (see text for details). MVC—maximum voluntary contraction; RS—resting state
state block; MIST—Montreal imaging stress task.
block; MIST—Montreal imaging stress task.

Once the participants were prepared, they were instructed to avoid unnecessary movements
Oncethe
during theexperiment
participantsinwere order prepared,
to preventtheysevere
were instructed
artifacts into recordings.
avoid unnecessary movements
They performed a
during the experiment in order to prevent severe artifacts in recordings. They
maximum voluntary contraction (MVC) of the trapezius during 5 s and a resting state block (RS1) performed a maximum
voluntary
with closedcontraction
eyes during (MVC) of the
2 min. trapezius they
Afterwards, during
were5 s asked
and a resting stateself-perceived
about their block (RS1) with closed
level eyes
of stress
during 2 min. Afterwards, they were asked about their self-perceived level of stress
(T1). The question was posed in Spanish. The English translation is: If 0 is the minimum level and 4 is (T1). The question
was
the posed
maximum in Spanish.
level, whatThe English
is your leveltranslation is: If
of stress? The 0 is the minimum
participants level and
then started 4 is thesession.
a stress maximum level,
In that
what is your level of stress? The participants then started a stress session.
session, they performed the Montreal imaging stress task (MIST), a proven methodology that In that session, they performed
the Montreal
induces imagingstress
psychosocial stressintask
people(MIST), a proven
[41]. Despite methodology
that there are otherthat induces psychosocial
well-described stress
stress methods
insuch
people [41].
as the Despite
variants of that there social
the Trier are other
stresswell-described
task [42], the stress
MIST hasmethods such in
been used as athe variants of
considerable
the Trier social stress task [42], the MIST has been used in a considerable number
number of stress-related works [28,31,41,43–45]. It was classified as well-described stress method by of stress-related
a recent review [46]. The MIST consists of two parts: training and task. In the training part, the
participant is asked to solve arithmetic operations without time limit per operation. The difficulty
level of the operations randomly varies (five levels). In the task part, the participant has to solve
arithmetic operations with time limit. The time limit adapts according to the number of consecutive
wrong and right answers. This enforces a range of 20–45% success ratio, while the participant is
Sensors 2018, 18, 2504 5 of 15

works [28,31,41,43–45]. It was classified as well-described stress method by a recent review [46].
The MIST consists of two parts: training and task. In the training part, the participant is asked to
solve arithmetic operations without time limit per operation. The difficulty level of the operations
randomly varies (five levels). In the task part, the participant has to solve arithmetic operations with
time limit. The time limit adapts according to the number of consecutive wrong and right answers.
This enforces a range of 20–45% success ratio, while the participant is asked to achieve about 80–90%.
The participant is periodically reminded of the relevance of achieving the goal. Detailed information
of this protocol can be found in the literature [41]. In our study, after a training of 3 min, the task lasted
6 min. During that session, the participants were seated on a comfortable chair within a classroom
while they were using the touchpad of a laptop to play a Matlab-based GUI of the MIST. This GUI was
developed by us and further details including screenshots can be found in the literature [28]. After the
stress session, the question about the self-perceived level of stress was asked again (T2).
Immediately after the stress session, the participants started a relaxing session. During that session,
they stayed laid (resting state with opened eyes) down in a blue-lighted room for 10 min. Blue light was
recently proven to accelerate the relaxation process after the MIST in comparison with conventional
white light [31]. In this work, the same room and light were used. Once again, the question about
the self-perceived stress level was asked at the end of the relaxing session (T3). Finally, a new resting
state block (RS2) with closed eyes was performed for 2 min. The timeline of the experiment is shown
in Figure 3b.
All the biosignals (raw data) were recorded during the whole experiment at 1000 samples per
second with amplification gain of 3 for EEG channels and 1 for the others. All the events (e.g., start of
stress session, end of stress session, etc.) were marked in the data. For the aim of this work
(i.e., presentation and validation of our system), the biosignals were processed and analyzed offline.
The real-time capability of our system is discussed in Section 4.2.

2.3. Signal Processing

2.3.1. EEG
EEG data were zero-phase bandpass filtered (1–48 Hz) with a fourth-order Butterworth infinite
impulse response (IIR) filter. Data corresponding to regions of interest (i.e., central minute of each
resting-state block, stress session, and relaxing session) were segmented into two-second epochs
(no overlap of consecutive epochs). Detrending and z-score normalization was applied to each epoch.
The power in theta–alpha (4–13 Hz) and gamma (25–45 Hz) bands was estimated for each channel
and then averaged across channels. The average relative gamma (RG) was computed for every single
epoch as the power ratio between the average gamma power and the average theta–alpha power.
The RG is a stress marker used in emotion and stress studies [28–31]. The following equation defines
the RG:
RG = AvPower (25–45 Hz)/AvPower (4–13 Hz) (1)

2.3.2. ECG
ECG data were zero-phase bandpass filtered (16–24 Hz) with a second-order Butterworth IIR
filter in order to enhance the R-peak of the QRS complex. Data corresponding to parts of interest were
segmented into 10-s epochs (no overlap of consecutive epochs). The average heart rate (HR) in beats
per minute was computed for each epoch by means of the average R–R-interval length. It was not
possible to compute the HR using two-second epochs. The set of HR values corresponding to 10-s
epochs was interpolated using a spline to obtain values corresponding to two-second epochs. The HR
is also a stress marker widely used in stress studies [14–19]. The following equation defines the HR:

HR (bpm) = 60/AvRR (2)

Sensors 2018, 18, 2504 6 of 15

2.3.3. EMG
EMG data were zero-phase bandpass filtered (1–350 Hz) with a second-order Butterworth IIR filter.
In order to obtain differential EMG data, data corresponding to the electrode further from the backbone
was subtracted from data corresponding to the electrode closer to the backbone. Differential data
corresponding to parts of interest were segmented into two-second epochs (no overlap of consecutive
epochs). The average trapezius activity (TA) was computed for each epoch as the ratio between the
root mean square (RMS) value in the epoch and the RMS value in the MVC test. As in the case of RG
and HR, the TA is also a stress marker used in several stress studies [20–23]. The following equation
defines the TA:
TA = RMS (epoch)/RMS (MCV test) (3)

2.3.4. GSR
GSR data corresponding to parts of interest were directly segmented into two-second epochs
(no overlap of consecutive epochs). The average skin conductance (SC) in Siemens was computed for
each epoch by using the equation provided by the Arduino e-Health platform tutorial. The SC is one
of the most used stress markers in literature [32–34]. The following equation defines the SC:

SC = 2 × (AvVoltage − 0.5)/100,000 (4)

2.4. Statistical Analysis

The grand-average across subjects of the time evolution of processed stress markers in the regions
of interests (i.e., set of values of RG, HR, TA, and SC corresponding to two-second epochs) was
computed. For a better visualization, individual data were z-scored and smoothed using a moving
average filter (10 samples) before the computation of the grand-average. The grand-average of the
self-perceived stress level (SPSL) at the three test points (i.e., T1, T2, and T3) was also computed.
A paired-sample t-test was applied in order to assess if the stress markers and the SPSL significantly
differ (p-value < α with α = 0.05) at different time periods. In particular, T1, T2, and T3 were
compared for the SPSL. The last 30 s of the first resting state block, the last 30 s of the stress
session, and the second-to-last 30 s of the relaxing session were compared for the stress markers.
Finally, the Pearson’s correlation coefficient (PCC) between grand-averaged stress markers and the
corresponding 95% confidence interval (CI) was calculated.

2.5. Three-Level Stress Classification

A linear discriminant analysis (LDA) was performed to detect the level of stress using the
processed stress markers (i.e., RG, HR, TA, and SC) as features. Three classes (i.e., levels of stress)
were defined: stress, relax, and neutral. The values corresponding to the two-second epochs of the
minutes 7–8 of the stress session were labeled as stress. These epochs corresponds to the period of
maximum stress. The values corresponding to the two-second epochs of the minutes 2–3 of the relaxing
session were labeled as relax. These epochs corresponds to the period of minimum stress. The values
corresponding to the two-second epochs of the central minute of each resting-state block were labeled
as neutral. These epochs corresponds to the periods of baseline stress level. Therefore, 60 observations
(120 s with two-second-epoch values) per class were used. A leave-one-out cross validation (LOOCV)
was performed for the three-class LDA. That is, for all the observations, 179 out of 180 observations
were used in training to classify the remaining observation. In addition to the leave-one-epoch-out
cross validation, a leave-one-subject-out cross validation was conducted. That is, for all the subjects,
the epochs of one subject were classified using the epochs of the remaining subjects as training data.
The classification accuracy or probability of success (pa ) in stress level detection was computed as the
Sensors 2018, 18, 2504 7 of 15

ratio between the number of successfully classified observations and the total number of observations
(i.e., n = 180). The 95% CI was also estimated as follows:

Sensors 2018, 18, x FOR PEER REVIEW

CI = pa ± 1.96 × sqrt(pa × (1 − pa )/n) 7 of 15(5)

3. Results
3. Results
3.1.3.1.
Time Evolution
Time EvolutionofofBiosignal-Based
Biosignal-BasedMarkers
Markers
Figure
Figure4a–d
4a–dshow
showthethegrand-average
grand-average across subjects of
across subjects of the
thetime
timeevolution
evolutionofofprocessed
processed stress
stress
markers
markers in the regions of interests. Figure 4e also shows the grand-average of the SPSL at the threetest
in the regions of interests. Figure 4e also shows the grand-average of the SPSL at the three
points (i.e., T1,
test points T2T1,
(i.e., andT2T3).
and T3).

(a)

(b)

(c)

Figure 4. Cont.
Sensors 2018, 18, 2504 8 of 15
Sensors 2018, 18, x FOR PEER REVIEW 8 of 15

(d)

(e)
Figure 4. Grand-average across subjects of the time evolution of processed stress markers in the
Figure 4. Grand-average across subjects of the time evolution of processed stress markers in the regions
regions of interests. Base1 and Base2 correspond to the central minutes of resting state blocks RS1
of interests. Base1 and Base2 correspond to the central minutes of resting state blocks RS1 and RS2,
and RS2, respectively. MIST indicates the beginning of the stress session (3 min of training and 6 min
respectively. MIST indicates the beginning of the stress session (3 min of training and 6 min of task).
of task). Relax indicates the beginning of the relaxing session. Shades behind the plots and error bars
Relax indicates the beginning of the relaxing session. Shades behind the plots and error bars indicate
indicate the standard error of the mean (SEM): (a) relative gamma (RG) estimated from
the standard error of the mean (SEM): (a) relative gamma (RG) estimated from electroencephalography
electroencephalography (EEG) data; (b) average heart rate (HR) estimated from electrocardiography
(EEG) data; (b) average heart rate (HR) estimated from electrocardiography (ECG) data; (c) trapezius
(ECG) data; (c) trapezius activity (TA) estimated from electromyography (EMG) data. Asterisk
activity (TA)statistically
indicates estimated significant
from electromyography
difference (p-value(EMG) data.
< 0.05) inAsterisk
average TA indicates
betweenstatistically
the last 30significant
s of the
difference
stress session and the second-to-last 30 s of the relaxing session; (d) skin conductancethe
(p-value < 0.05) in average TA between the last 30 s of the stress session and second-to-last
(SC) estimated
30 from
s of the relaxing
galvanic session;
skin response (d) (GSR)
skin conductance
data. Asterisk (SC) estimated
indicates from galvanic
statistically skindifference
significant response (GSR)
(p-valuedata.
Asterisk indicates statistically significant difference (p-value < 0.05) in average
< 0.05) in average SC between the last 30 s of the stress session and the second-to-last 30 s of the SC between the last 30 s
of relaxing
the stresssession;
session(e) and the second-to-last 30 s of the relaxing session; (e) self-perceived
self-perceived stress level (SPSL) obtained from questions at T1, T2, and T3 stress level
(SPSL) obtained
points. fromcomprises
X-axis only questionsregions
at T1, T2, and T3 points.
of interests and T1,X-axis
T2, andonly comprises
T2 would regions
actually of interests
be located before and
T1,the
T2,stress
and T2 would
session actually
(i.e., be located
just before MIST),before thestress
after the stresssession
session(i.e.,
(i.e.,just
justafter
before MIST),
minute afterafter
9), and the the
stress
session (i.e.,session
relaxing just after
(i.e.,minute 9), and
just after after19),
minute therespectively.
relaxing session (i.e., just
Asterisks after statistically
indicate minute 19), significant
respectively.
Asterisks
differenceindicate
(p-valuestatistically significant
< 0.05) in SPSL between difference
the T1–T2(p-value < 0.05)
and between in SPSL between the T1–T2 and
T2–T3.
between T2–T3.
In addition, the Pearson’s correlation coefficient (PCC) between stress markers and the
corresponding 95% confidence interval is reported in Table 1.
In addition, the Pearson’s correlation coefficient (PCC) between stress markers and the
corresponding 95% confidence interval is reported in Table 1.
Table 1. Pearson’s correlation coefficient (PCC) between processed stress markers and the
corresponding lower (CI low) and upper (CI up) bounds for a 95% confidence interval (CI).
RG—relative gamma; HR—average heart rate; TA—trapezius activity; SC—skin conductance.

Pair PCC CI Low CI Up

RG, HR 0.7296 0.6909 0.7642
RG, TA 0.5753 0.5206 0.6253
RG, SC 0.3293 0.2579 0.3972
Sensors 2018, 18, 2504 9 of 15

Table 1. Pearson’s correlation coefficient (PCC) between processed stress markers and the
corresponding lower (CI low) and upper (CI up) bounds for a 95% confidence interval (CI). RG—relative
gamma; HR—average heart rate; TA—trapezius activity; SC—skin conductance.

Pair PCC CI Low CI Up

RG, HR 0.7296 0.6909 0.7642
RG, TA 0.5753 0.5206 0.6253
RG, SC 0.3293 0.2579 0.3972
HR, TA 0.8338 0.8083 0.8561
HR, SC 0.6327 0.5834 0.6773
TA, SC 0.4632 0.3995 0.5224

3.2. Stress Level Detection

The classification accuracy or probability of success (pa ) in detection of stress level (stress, relax,
and neutral) and the 95% confidence interval are reported in this section. In particular, Table 2 shows
these values when using one stress marker as feature for the LDA classifier. Table 3 shows the same
statistics when using two stress markers as features. Table 4 shows the same when using three or
all the stress markers. Finally, Table 5 shows the same as Table 4, but using a leave-one-subject-out
cross validation instead of leave-one-epoch-out. In these four tables, main values indicate the pa and
error values indicate the 95% CI. Last row indicates the mean and the standard deviation of the mean.
All the values are expressed in percentage.

Table 2. Probability of successful detection of stress level using ones stress marker as feature.

Participant RG HR TA SC
1 72 ± 7 74 ± 6 31 ± 7 49 ± 7
2 61 ± 7 57 ± 7 28 ± 7 69 ± 7
3 61 ± 7 45 ± 7 29 ± 7 84 ± 5
4 51 ± 7 60 ± 7 61 ± 7 51 ± 7
5 28 ± 7 93 ± 4 22 ± 6 69 ± 7
6 44 ± 7 94 ± 3 45 ± 7 61 ± 7
7 47 ± 7 82 ± 6 66 ± 7 60 ± 7
8 33 ± 7 77 ± 6 21 ± 6 61 ± 7
9 67 ± 7 77 ± 6 52 ± 7 18 ± 6
10 33 ± 7 62 ± 7 62 ± 7 76 ± 6
Mean ± Std 50 ± 15 72 ± 16 42 ± 18 60 ± 18

Table 3. Probability of successful detection of stress level using two stress markers as features.

Participant RG, HR RG, TA RG, SC HR, TA HR, SC TA, SC

1 76 ± 6 83 ± 6 69 ± 7 86 ± 5 71 ± 7 64 ± 7
2 73 ± 6 82 ± 6 73 ± 6 61 ± 7 70 ± 7 78 ± 6
3 77 ± 6 60 ± 7 81 ± 6 52 ± 7 92 ± 4 90 ± 4
4 59 ± 7 64 ± 7 72 ± 7 70 ± 7 68 ± 7 87 ± 5
5 92 ± 4 46 ± 7 54 ± 7 93 ± 4 84 ± 5 76 ± 6
6 94 ± 3 69 ± 7 64 ± 7 93 ± 4 96 ± 3 71 ± 7
7 84 ± 5 66 ± 7 64 ± 7 86 ± 5 86 ± 5 66 ± 7
8 74 ± 6 48 ± 7 61 ± 7 76 ± 6 71 ± 7 64 ± 7
9 82 ± 6 72 ± 7 64 ± 7 78 ± 6 73 ± 6 49 ± 7
10 67 ± 7 54 ± 7 67 ± 7 73 ± 6 77 ± 6 81 ± 6
Mean ± Std 78 ± 11 64 ± 13 67 ± 7 77 ± 14 79 ± 10 73 ± 12
Sensors 2018, 18, 2504 10 of 15

Table 4. Probability of successful detection of stress level using three or all the stress markers as features.

Participant RG, HR, TA RG, HR, SC RG, TA, SC HR, TA, SC RG, HR, TA, SC
1 91 ± 4 79 ± 6 84 ± 5 92 ± 4 92 ± 4
2 82 ± 6 78 ± 6 83 ± 6 75 ± 6 82 ± 6
3 77 ± 6 93 ± 4 82 ± 6 92 ± 4 93 ± 4
4 68 ± 7 69 ± 7 78 ± 6 82 ± 6 83 ± 6
5 93 ± 4 84 ± 5 73 ± 7 84 ± 5 84 ± 5
6 93 ± 4 97 ± 3 72 ± 7 98 ± 2 98 ± 2
7 86 ± 5 87 ± 5 67 ± 7 89 ± 4 90 ± 4
8 74 ± 6 75 ± 6 64 ± 7 71 ± 7 74 ± 6
9 81 ± 6 80 ± 6 67 ± 7 76 ± 6 81 ± 6
10 72 ± 7 77 ± 6 79 ± 6 79 ± 6 78 ± 6
Mean ± Std 82 ± 9 82 ± 8 75 ± 7 84 ± 9 86 ± 8

Table 5. Probability of successful detection of stress level using three or all the stress markers as features
for the leave one-subject-out cross validation.

Participant RG, HR, TA RG, HR, SC RG, TA, SC HR, TA, SC RG, HR, TA, SC
1 33 ± 7 33 ± 7 36 ± 7 33 ± 7 33 ± 7
2 67 ± 7 37 ± 7 58 ± 7 64 ± 7 65 ± 7
3 33 ± 7 41 ± 7 36 ± 7 33 ± 7 36 ± 7
4 47 ± 7 36 ± 7 33 ± 7 49 ± 7 34 ± 7
5 66 ± 7 41 ± 7 37 ± 7 64 ± 7 38 ± 7
6 36 ± 7 34 ± 7 33 ± 7 34 ± 7 34 ± 7
7 33 ± 7 33 ± 7 39 ± 7 33 ± 7 33 ± 7
8 34 ± 7 53 ± 7 51 ± 7 54 ± 7 54 ± 7
9 41 ± 7 60 ± 7 56 ± 7 51 ± 7 66 ± 7
10 48 ± 7 48 ± 7 48 ± 7 36 ± 7 42 ± 7
Mean ± Std 44 ± 13 42 ± 9 43 ± 10 45 ± 13 44 ± 13

4. Discussion

4.1. Stress and Biosignals

At first sight, the time evolution of all the stress markers indicates agreement with the
self-perception of stress level. This is partially supported by the statistical tests (see Figure 4). The MIST
causes a significant increase in self-perceived stress level and the relaxing session causes a significant
decrease. However, only two stress markers presented significant differences in stress level at different
time periods. These are the TA and the SC. The significant differences were only between the end of
the stress session (maximum level of stress) and the end of the relax session (minimum or very low
level of stress). The other two markers (i.e., RG and HR) did not present significant differences despite
the noticeable changes. All the markers reflect that the increase in stress level is gradual. This behavior
has been reported in previous literature [28,31]. However, there are some visible differences between
markers. In particular, the RG, the HR, and the TA indicates that the minimum level of stress is quickly
achieved once the relaxing session starts (less than 1 min from the beginning of this session), while the
SC denotes a gradual decrease. This is due to the fact that the sweating process is fast, while the
reabsorption process is slow in comparison with other physiological responses. Accordingly, the SC has
a drawback in terms of time of response. In addition, the RG is the only marker that reflects a gradual
increase in stress level during the relaxing session. This fits with results reported in previous literature
and may be caused by boredom of the participants during this part of the experiment [28,31]. The other
markers indicate a more rapid increase at the end of the relaxing session. In this regard, the RG has an
advantage in terms of response time. In other words, the boredom may have an immediate effect on
EEG, while it may have a delayed effect on the other biosignals.
Sensors 2018, 18, 2504 11 of 15

Regarding the PCC between stress markers, all of them are generally correlated (see Table 1).
The one that correlates the most with the others is the HR (72.96% with RG, 83.38% with TA, and 63.27%
with SC). The SC is the least correlated marker (32.93% with RG, 63.27% and 46.32% with TA). This is
due to the response time discussed in the previous paragraph and to the fact that the GSR is the least
noisy biosignal (see Figure 4). The ECG is the second least noisy biosignal. This suggests that ECG and
GSR are the more appropriate biosignals in the presence of artifacts. Nevertheless, the stress markers
extracted from these two biosignals and from the EMG can be misrepresented by physical activity
(e.g., physical activity may increase the HR even without being stressed). In this respect, the RG
is advantageous.

4.2. Real-Time Detection of Stress Level

The results of the three-class LDA with leave-out-epoch-out cross validation (see Tables 1–4)
indicate that the more biosignals (thus more stress markers) that are combined, the higher probability
of successful detection of stress level (i.e., accuracy). With single markers, the probabilities are 50%,
72%, 42%, and 60% for RG, HR, TA, and SC, respectively. With two markers, probabilities are 78%
(RG–HR), 64% (RG–TA), 67% (RG–SC), 77% (HR–TA), 79% (HR–SC), and 73% (TA–SC). However,
the probabilities increase up to 82% (RG–HR–TA), 82% (RG–HR–SC), 75% (RG–TA–SC) and 84%
(HR–TA–SC). There is no relevant improvement by adding the RG to the trio HR–TA–SC (86%).
By using three or all the markers, the results overcome accuracies reported in previous studies of
biosignal-based systems for stress detection and measurement [32,33,35,39,47]. Nevertheless, this is
not meaningful because the cited studies were neither carried out in the same experimental conditions,
with comparable number of subjects, nor were they conducted with a similar methodology. In reference
to the results of the leave-one-subject-out cross validation (see Table 5), the probabilities of success are
generally close to the chance level (i.e., 33%), taking into account the confidence intervals. This indicates
that the system needs to be calibrated for every single subject. This was expected as stress markers
and thresholds may vary across subjects. Regarding the optimal combination of markers, it depends
on the particular conditions in which the stress has to be detected (e.g., response time and external
factors). For example, in this work, the use of EEG signals does not optimize the results in terms of
accuracy. The EEG has a distinct set of advantages and limitations. Among the advantages, as cited in
Section 4.1, the markers based on brain activity (e.g., the RG) present a shorter response time and are less
susceptible to physical activity. Additionally, the EEG provides powerful endogenous and cognitive
information such as attention [48–50] that can be useful in certain scenarios. Regarding the limitations,
the use of EEG provides a number of technical challenges such as additional sensors (thus less
portability) or higher computational complexity. In order to overcome some of these limitations, we are
developing a more wearable version of the EEG cap embedding the whole electronics and based on
dry electrodes [51]. Our results demonstrate the reliability of our system in the detection of three
levels of stress with a resolution of a few seconds. Still, the results could improve by extracting more
features (i.e., stress markers) from biosignals [38] and by using more powerful classifiers such as
artificial neural networks [52,53]. For the aim of this work, biosignals were processed offline. We used
two-second epochs of data with a low-cost preprocessing, feature extraction, and classification in terms
of computation time. This provides our system with real-time capability. In addition, for the sake of
simplicity, the participants of the study were instructed to avoid unnecessary movements during the
experiment in order to prevent severe artifacts in recordings. This is an unrealistic scenario. For the
use of the proposed system in daily-life scenarios, advanced processing for artifact removal should be
included [2]. Based on the accuracies obtained in this work, we expect that our system can still work
in hostile environments by adding the artifact removal part.

5. Conclusions
In this work, we have proposed a portable system for real-time detection of stress level. We have
presented the methodology and the results of a study aimed at validating the system. In the study,
Sensors 2018, 18, 2504 12 of 15

ten volunteers were stressed and then relaxed using well-established methods, while their biosignals
were recorded. Our portable system can simultaneously record and process four types of biosignals
(i.e., EEG, ECG, EMG, and GSR) in real-time, thereby enabling the detection of three levels of stress
very accurately (86%). The system has some limitations that have been discussed (e.g., portability and
performance under artifacts). In order to overcome them, we are working on a final version in which
the biosignals are cloud-computed, including the needed processing for artifact removal. The real-time
biofeedback (i.e., 2 s plus the computation time) will be presented in mobile devices such as tablets or
smartphones. Moreover, a more wearable version of the EEG cap embedding the whole electronics is
feasible and under development. Having overcome the cited limitations, our system could be used
as a reliable tool for real-time stress monitoring, detection, and prevention in daily life. For example,
prevention of job stress in periods of high level of work intensity, stress monitoring in children at
school, or discovery of new stressors through stress detection in the domestic environment. All of
this has a relevant impact on society as stress is a major problem nowadays and this system could
substantially improve people’s health and quality of life.

Author Contributions: Conceptualization, J.M., M.A.L.-G., F.P., and M.J.S.-C.; Methodology, J.M. and M.A.L.-G.;
Software, J.M. and E.P.; Validation, J.M., M.A.L.-G., and F.P.; Formal Analysis, J.M. and M.A.L.-G.; Investigation,
J.M.; Resources, M.A.L.-G., F.P. and M.J.S.-C.; Data Curation, J.M. and E.P.; Writing—Original Draft Preparation,
J.M.; Writing—Review and Editing, M.A.L.-G. and F.P.; Visualization, J.M., E.P., M.A.L.-G., and F.P.; Supervision,
M.A.L.-G. and F.P.; Project Administration, M.J.S.-C.; Funding Acquisition, M.A.L.-G., F.P., and M.J.S.-C.
Funding: This research was funded by [Ministry of Economy and Competitiveness (Spain)] grant number
[TIN2015-67020P], [Ministry of Economy and Competitiveness (Spain)] grant number [DPI2015-69098-REDT],
[Junta of Andalucia (Spain)] grant number [P11-TIC-7983], [Spanish National Youth Guarantee Implementation
Plan] grant number [Research contract], [Nicolo Association for the R+D in neurotechnologies for disability] grant
number [Research support], and [Orden Hospitalaria San Juan de Dios] grant number [Beca investigacion].
Acknowledgments: The authors would like to thank all the volunteers who participated in the study. The authors
would also like to thank the School for Special Education San Rafael of Granada for their support and the
provided facilities.
Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design of the
study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; and in the decision to
publish the results.

References
1. Cohen, S.; Janicki-Deverts, D.; Doyle, W.J.; Miller, G.E.; Frank, E.; Rabin, B.S.; Turner, R.B. Chronic stress,
glucocorticoid receptor resistance, inflammation, and disease risk. Proc. Natl. Acad. Sci. USA 2012, 109,
5995–5999. [CrossRef] [PubMed]
2. Minguillon, J.; Lopez-Gordo, M.A.; Pelayo, F. Trends in EEG-BCI for daily-life: Requirements for artifact
removal. Biomed. Signal Process. Control 2017, 31, 407–418. [CrossRef]
3. Black, A.D.; Car, J.; Pagliari, C.; Anandan, C.; Cresswell, K.; Bokun, T.; McKinstry, B.; Procter, R.; Majeed, A.;
Sheikh, A. The impact of ehealth on the quality and safety of health care: A systematic overview. PLoS Med.
2011, 8, e1000387. [CrossRef] [PubMed]
4. Blaya, J.A.; Fraser, H.S.F.; Holt, B. E-health technologies show promise in developing countries. Health Aff.
2010, 29, 244–251. [CrossRef] [PubMed]
5. Barello, S.; Triberti, S.; Graffigna, G.; Libreri, C.; Serino, S.; Hibbard, J.; Riva, G. eHealth for patient
engagement: A Systematic Review. Front. Psychol. 2016, 6. [CrossRef] [PubMed]
6. Muellmann, S.; Forberger, S.; Möllers, T.; Bröring, E.; Zeeb, H.; Pischke, C.R. Effectiveness of eHealth
interventions for the promotion of physical activity in older adults: A systematic review. Prev. Med. 2018,
108, 93–110. [CrossRef] [PubMed]
7. Morland, L.A.; Greene, C.J.; Rosen, C.S.; Kuhn, E.; Hoffman, J.; Sloan, D.M. Telehealth and eHealth
interventions for posttraumatic stress disorder. Curr. Opin. Psychol. 2017, 14, 102–108. [CrossRef] [PubMed]
8. Lorenz, A.; Oppermann, R. Mobile health monitoring for the elderly: Designing for diversity.
Pervasive Mob. Comput. 2009, 5, 478–495. [CrossRef]
9. Nguyen, H.H.; Silva, J.N.A. Use of smartphone technology in cardiology. Trends Cardiovasc. Med. 2016, 26,
376–386. [CrossRef] [PubMed]
Sensors 2018, 18, 2504 13 of 15

10. Woods, A.M.; Nowostawski, M.; Franz, E.A.; Purvis, M. Parkinson’s disease and essential tremor
classification on mobile device. Pervasive Mob. Comput. 2014, 13, 1–12. [CrossRef]
11. Lakshminarayan, K.; Westberg, S.; Northuis, C.; Fuller, C.C.; Ikramuddin, F.; Ezzeddine, M.; Scherber, J.;
Speedie, S. A mHealth-based care model for improving hypertension control in stroke survivors: Pilot RCT.
Contemp. Clin. Trials 2018, 70, 24–34. [CrossRef] [PubMed]
12. Liang, D.; Han, H.; Du, J.; Zhao, M.; Hser, Y.I. A pilot study of a smartphone application supporting recovery
from drug addiction. J. Subst. Abuse Treat. 2018, 88, 51–58. [CrossRef] [PubMed]
13. Picard, R.W. Affective Computing: From laughter to IEEE. IEEE Trans. Affect. Comput. 2010, 1, 11–17.
[CrossRef]
14. Dimitriev, D.A.; Saperova, E.V. Heart rate variability and blood pressure during mental stress. Ross Fiziol Zh
Im I M Sechenova 2015, 101, 98–107. [PubMed]
15. Ranganathan, G.; Rangarajan, R.; Bindhu, V. Estimation of heart rate signals for mental stress assessment
using neuro fuzzy technique. Appl. Soft Comput. 2012, 12, 1978–1984. [CrossRef]
16. Chandiramani, S.; Cohorn, L.C.; Chandiramani, S. Heart rate changes during acute mental stress with closed
loop stimulation: Report on two single-blinded, pacemaker studies. Pacing Clin. Electrophysiol. 2007, 30,
976–984. [CrossRef] [PubMed]
17. Regula, M.; Socha, V.; Kutilek, P.; Socha, L.; Hana, K.; Hanakova, L.; Szabo, S. Study of heart rate as
the main stress indicator in aircraft pilots. In Proceedings of the 16th IEEE International Conference on
Mechatronics—Mechatronika, Brno, Czech Republic, 3–5 December 2014; pp. 639–643.
18. Sayette, M.A. Heart rate as an index of stress response in alcohol administration research: A critical review.
Alcohol Clin. Exp. Res. 1993, 17, 802–809. [CrossRef] [PubMed]
19. Michels, N.; Sioen, I.; Clays, E.; De Buyzere, M.; Ahrens, W.; Huybrechts, I.; Vanaelst, B.; De Henauw, S.
Children’s heart rate variability as stress indicator: Association with reported stress and cortisol. Biol. Psychol.
2013, 94, 433–440. [CrossRef] [PubMed]
20. Lundberg, U.; Kadefors, R.; Melin, B.; Palmerud, G.; Hassmén, P.; Engström, M.; Dohns, I.E.
Psychophysiological stress and EMG activity of the trapezius muscle. Int. J. Behav. Med. 1994, 1, 354–370.
[CrossRef] [PubMed]
21. Wijsman, J.; Grundlehner, B.; Penders, J.; Hermens, H. Trapezius muscle EMG as predictor of mental stress.
ACM Trans. Embed. Comput. Syst. 2013, 12. [CrossRef]
22. Larsson, S.E.; Larsson, R.; Zhang, Q.; Cai, H.; Åke Öberg, P. Effects of psychophysiological stress on trapezius
muscles blood flow and electromyography during static load. Eur. J. Appl. Physiol. Occup. Physiol. 1995, 71,
493–498. [CrossRef] [PubMed]
23. Schleifer, L.M.; Spalding, T.W.; Kerick, S.E.; Cram, J.R.; Ley, R.; Hatfield, B.D. Mental stress and
trapezius muscle activation under psychomotor challenge: A focus on EMG gaps during computer work.
Psychophysiology 2008, 45, 356–365. [CrossRef] [PubMed]
24. Papousek, I.; Weiss, E.M.; Schulter, G.; Fink, A.; Reiser, E.M.; Lackner, H.K. Prefrontal EEG alpha asymmetry
changes while observing disaster happening to other people: Cardiac correlates and prediction of emotional
impact. Biol. Psychol. 2014, 103, 184–194. [CrossRef] [PubMed]
25. Seo, S.; Lee, J. Stress and EEG. In Convergence and Hybrid Information Technologies; Crisan, M., Ed.; IntechOpen:
London, UK, 2010; pp. 413–426.
26. Hu, B.; Peng, H.; Zhao, Q.; Hu, B.; Majoe, D.; Zheng, F.; Moore, P. Signal Quality Assessment Model for
Wearable EEG Sensor on Prediction of Mental Stress. IEEE Trans. Nanobiosci. 2015, 14, 553–561. [CrossRef]
27. Brouwer, A.-M.; Neerincx, M.A.; Kallen, V.; van der Leer, L.; ten Brinke, M. EEG alpha asymmetry, heart rate
variability and cortisol in response to Virtual Reality induced stress. J. Cyberther. Rehabil. 2011, 4, 27–40.
28. Minguillon, J.; Lopez-Gordo, M.A.; Pelayo, F. Stress Assessment by Prefrontal Relative Gamma.
Front. Comput. Neurosci. 2016, 10, 1–9. [CrossRef] [PubMed]
29. Steinhubl, S.R.; Wineinger, N.E.; Patel, S.; Boeldt, D.L.; Mackellar, G.; Porter, V.; Redmond, J.T.; Muse, E.D.;
Nicholson, L.; Chopra, D.; et al. Cardiovascular and nervous system changes during meditation.
Front. Hum. Neurosci. 2015, 9, 1–10. [CrossRef] [PubMed]
30. Lutz, A.; Greischar, L.L.; Rawlings, N.B.; Ricard, M.; Davidson, R.J. Long-term meditators self-induce
high-amplitude gamma synchrony during mental practice. Proc. Natl. Acad. Sci. USA 2004, 101, 16369–16373.
[CrossRef] [PubMed]
Sensors 2018, 18, 2504 14 of 15

31. Minguillon, J.; Lopez-Gordo, M.A.; Renedo-Criado, D.A.; Sanchez-Carrion, M.J.; Pelayo, F. Blue lighting
accelerates post-stress relaxation: Results of a preliminary study. PLoS ONE 2017, 12, e0186399. [CrossRef]
[PubMed]
32. Sriramprakash, S.; Prasanna, V.D.; Murthy, O.V.R. Stress Detection in Working People. Procedia Comput. Sci.
2017, 115, 359–366. [CrossRef]
33. Villarejo, M.V.; Zapirain, B.G.; Zorrilla, A.M. A stress sensor based on galvanic skin response (GSR) controlled
by ZigBee. Sensors 2012, 12, 6075–6101. [CrossRef] [PubMed]
34. Sessa, F.; Messina, G.; Valenzano, A.; Messina, A.; Salerno, M.; Marsala, G.; Bertozzi, G.; Daniele, A.;
Monda, V.; Russo, R. Sports training and adaptive changes. Sport Sci. Health 2018, 1–4. [CrossRef]
35. Seoane, F.; Mohino-Herranz, I.; Ferreira, J.; Alvarez, L.; Buendia, R.; Ayllón, D.; Llerena, C.; Gil-Pita, R.
Wearable biomedical measurement systems for assessment of mental stress of combatants in real time.
Sensors 2014, 14, 7120–7141. [CrossRef] [PubMed]
36. Kikhia, B.; Stavropoulos, T.G.; Andreadis, S.; Karvonen, N.; Kompatsiaris, I.; Sävenstedt, S.; Pijl, M.;
Melander, C. Utilizing a wristband sensor to measure the stress level for people with dementia. Sensors
2016, 16, 1989. [CrossRef] [PubMed]
37. Zheng, R.; Yamabe, S.; Nakano, K.; Suda, Y. Biosignal analysis to assess mental stress in automatic driving
of trucks: Palmar perspiration and masseter electromyography. Sensors 2015, 15, 5136–5150. [CrossRef]
[PubMed]
38. Ollander, S.; Godin, C.; Charbonnier, S. Feature and Sensor Selection for Detection of Driver Stress.
In Proceedings of the 3rd International Conference on Physiological Computing Systems, Lisbon, Portugal,
27–28 July 2016; pp. 115–122.
39. Keshan, N.; Parimi, P.V.; Bichindaritz, I. Machine learning for stress detection from ECG signals in automobile
drivers. In Proceedings of the 2015 IEEE International Conference on Big Data, Santa Clara, CA, USA,
29 October–1 November 2015; pp. 2661–2669. [CrossRef]
40. Minguillon, J.; Lopez-Gordo, M.A.; Morillas, C.; Pelayo, F. A Mobile Brain-Computer Interface for Clinical
Applications: From the Lab to the Ubiquity. In Proceedings of the 7th International Work-Conference on
the Interplay between Natural and Artificial Computation; Ferrández Vicente, J.M., Álvarez-Sánchez, J.R.,
De la Paz López, F., Toledo Moreo, J., Adeli, H., Eds.; Springer International Publishing: Corunna, Spain,
2017; pp. 68–76. [CrossRef]
41. Dedovic, K.; Renwick, R.; Mahani, N.K.; Engert, V.; Lupien, S.J.; Pruessner, J.C. The Montreal Imaging Stress
Task: Using functional imaging to investigate the effects of perceiving and processing psychosocial stress in
the human brain. J. Psychiatry Neurosci. 2005, 30, 319–325. [PubMed]
42. Kirschbaum, C.; Pirke, K.-M.; Hellhammer, D.H. The “Trier Social Stress Test”—A Tool for Investigating
Psychobiological Stress Responses in a Laboratory Setting. Neuropsychobiology 1993, 28, 76–81. [CrossRef]
[PubMed]
43. Dedovic, K.; D’Aguiar, C.; Pruessner, J.C. What stress does to your brain: A review of neuroimaging studies.
Can. J. Psychiatry 2009, 54, 6–15. [CrossRef] [PubMed]
44. Zschucke, E.; Renneberg, B.; Dimeo, F.; Wüstenberg, T.; Ströhle, A. The stress-buffering effect of acute
exercise: Evidence for HPA axis negative feedback. Psychoneuroendocrinology 2015, 51, 414–425. [CrossRef]
[PubMed]
45. Nagano-Saito, A.; Dagher, A.; Booij, L.; Gravel, P.; Welfeld, K.; Casey, K.F.; Leyton, M.; Benkelfat, C.
Stress-induced dopamine release in human medial prefrontal cortex—18F-fallypride/PET study in healthy
volunteers. Synapse 2013, 67, 821–830. [CrossRef] [PubMed]
46. Bali, A.; Jaggi, A.S. Clinical experimental stress studies: Methods and assessment. Rev. Neurosci. 2015, 26,
555–579. [CrossRef] [PubMed]
47. Bichindaritz, I.; Breen, C.; Cole, E.; Keshan, N.; Parimi, P. Feature Selection and Machine Learning Based
Multilevel Stress Detection from ECG Signals. In Innovation in Medicine and Healthcare 2017; Chen, Y.-W.,
Tanaka, S., Howlett, R.J., Jain, L.C., Eds.; Springer International Publishing: Cham, Switzerland, 2018;
pp. 202–213.
48. Lopez-Gordo, M.A.; Pelayo, F. A Binary Phase-Shift Keying Receiver for the Detection of Attention to Human
Speech. Int. J. Neural Syst. 2013, 23, 12. [CrossRef] [PubMed]
Sensors 2018, 18, 2504 15 of 15

49. Minguillon, J.; Lopez-Gordo, M.A.; Pelayo, F. Detection of Attention in Multi-Talker Scenarios: A Fuzzy
Approach. Expert Syst. Appl. 2016, 64, 261–268. [CrossRef]
50. Lopez-Gordo, M.A.; Pelayo, F.; Fernandez, E.; Padilla, P. Phase-shift keying of EEG signals: Application to
detect attention in multitalker scenarios. Signal Process. 2015, 117, 165–173. [CrossRef]
51. Lopez-Gordo, M.A.; Sanchez Morillo, D.; Pelayo Valle, F. Dry EEG electrodes. Sensors 2014, 14, 12847–12870.
[CrossRef] [PubMed]
52. Singh, M.; Bin Queyam, A. A Novel Method of Stress Detection Using Physiological Measurements of
Automobile Drivers. Int. J. Electron. Eng. 2013, 5, 13–20.
53. Alić, B.; Sejdinović, D.; Gurbeta, L.; Badnjevic, A. Classification of stress recognition using Artificial Neural
Network. In Proceedings of the 2016 5th Mediterranean Conference on Embedded Computing (MECO),
Bar, Montenegro, 12–16 June 2016; pp. 297–300.

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).