IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 58, NO. 10, OCTOBER 2011 4857
A Stress-Detection System Based on Physiological
Signals and Fuzzy Logic
Alberto de Santos Sierra, Carmen Sánchez Ávila, Javier Guerra Casanova, and Gonzalo Bailador del Pozo
Abstract—A stress-detection system is proposed based on phys-
iological signals. Concretely, galvanic skin response (GSR) and
heart rate (HR) are proposed to provide information on the
state of mind of an individual, due to their nonintrusiveness and
noninvasiveness. Furthermore, specific psychological experiments
were designed to induce properly stress on individuals in order
to acquire a database for training, validating, and testing the
proposed system. Such system is based on fuzzy logic, and it
described the behavior of an individual under stressing stimuli in
terms of HR and GSR. The stress-detection accuracy obtained is
99.5% by acquiring HR and GSR during a period of 10 s, and
what is more, rates over 90% of success are achieved by decreasing
that acquisition period to 3–5 s. Finally, this paper comes up
with a proposal that an accurate stress detection only requires
two physiological signals, namely, HR and GSR, and the fact
that the proposed stress-detection system is suitable for real-time
applications.
Index Terms—Biometrics, fuzzy logic, galvanic skin response
(GSR), heart rate (HR), physiological signals, stress detection,
stress template.
I. INTRODUCTION
C
URRENT biometric systems attempt to identify/
authenticate an individual in a unique and precise
manner. Biometrics focus on such specific aim mainly under
the assumption that registered or identified user will not act
maliciously, and therefore, overall security will be ensured.
However, a biometric system (whatever its complexity is) fails
in one single aspect: What would happen if a registered user is
forced to be utilized as a key to enter the system?
Nowadays, biometrics cannot provide any solution to that
scenario. Nonetheless, those previous scenarios where a person
is forced to access a system or entrance have a well-known
common denominator: stress. Human body will react by
increasing the blood volume pressure and the quantity of
hormones. Furthermore, some basic functions will be avoided
like hunger, sleepiness, and the like, focusing on the scenario
and the malicious agent. In other words, the human body
prepares to fight [1], [2].
This stress response is almost impossible to disguise
and, therefore, is an accurate indicator about the security
compromise.
Manuscript received June 11, 2010; revised September 28, 2010; accepted
October 7, 2010. Date of publication January 6, 2011; date of current version
August 30, 2011.
The authors are with the Group of Biometrics, Biosignals, and Security,
Centro de Domótica Integral, Polytechnical University of Madrid, Cam-
pus de Montegancedo, 28223 Madrid, Spain (e-mail: alberto@cedint.upm.es;
csa@cedint.upm.es; jguerra@cedint.upm.es; gbailador@cedint.upm.es).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIE.2010.2103538
A stress-detection system offers a solution for the previous
situation where biometrics failed. This system will come up
with an understanding of the state of mind of the user who
attempts to access the system, so that if such a person is
abnormally out of her or his normal values for stress/relax
relation, then something strange is happening, and therefore,
security could be compromised.
This paper proposes a stress-detection system based on two
physiological parameters, namely, heart rate (HR) and galvanic
skin response (GSR), together with fuzzy expert systems to
elucidate to what extent an individual is under stress. This
system is able to provide a very fast decision on the state of
mind of the user, which is very suitable for real applications.
Furthermore, its simplicity and noninvasiveness make this ap-
proach a possible system to be easily embedded not only in
current biometric systems but also in general accessing systems.
Results highlight the fact that stress detection can be
achieved with an accuracy of 99.5% by measuring HR and
GSR during 10 s. Furthermore, rates higher than 90% can be
achieved by decreasing that acquisition time.
The layout of this paper consists of a state-of-the-art in
relation to stress detection (Section II), a description of the
physiological signals involved in this system (Section III), and
which experiments where carried out to validate the system
(Section IV). The system is described in detail in Section V,
and its implementation is presented in Section VI. An eval-
uation of the whole system is provided within Section VII.
Finally, conclusions and future work (Section VIII) end this
paper.
II. L
ITERATURE REVIEW
Stress detection has been considered from different points
of view and approaches. The work presented by Andren and
Funk [3] provides a system that is able to compute the stress
level of an individual by the manner and rhythm in which a
person types characters on a keyboard or keypad. Furthermore,
Dinges et al. [4] provide a study of stress detection based on
facial recognition.
However, the method proposed in this paper focuses on stress
detection based only on physiological signals. Considering this,
there exist many previous works related to this topic. The
essay presented by Begum et al. [5] presents a study of stress
detection based only on finger temperature (FT), together with
fuzzy logic [6] and case-based reasoning [3].
HR variability has been also considered as a stress marker
in human body. Due to this reason, HR has been widely
studied and analyzed. Several authors consider this signal in
0278-0046/$26.00 © 2011 IEEE
4858 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 58, NO. 10, OCTOBER 2011
their reports: Jovanov et al. [7] presented a stress monitoring
system based on a distributed wireless architecture imple-
mented on intelligent sensors. HR was recorded along differ-
ent locations in individual body by means of sensors located
beneath clothes.
Nonetheless, it is not common to focus only on one certain
physiological feature but to focus on many of them in order to
obtain further and more precise information about the state of
mind. Considering this multimodal approach, there are several
articles which study a variety of parameters and signals, as well
as the combination of them.
Continuing on this research line, the research provided in [8],
[9] proposes a system gathering FT, GSR, and blood volume
pulse (BVP). The main characteristic of this system lies on
the fact that signals are acquired in a nonintrusive manner,
and furthermore, these previous physiological signals provide
a predictable relation with stress variation.
However, not all physiological signals involve an electrical
component. Pupil dilation (PD) and eyetracking (ET) provide
very precise information about frame stress. When an individ-
ual is under stress, PD is wider, and the eye movement is faster.
The article presented in [10] not only considers PD and ET but
also GSR, BVP, and FT. The main purpose of this approach
is to recognize emotions, interest, and attention from emotion
recognition. Moreover, it is possible to deduce the intention of
the individual from these results, a very remarkable conclusion
for future computer applications and for the sake of a better
human–computer interaction (HCI) [11], [12].
The work presented by Sarkar [11] proposes fuzzy logic (as
Jiang and Wang [13]) to elucidate up to what extent a user is
under stress. Furthermore, it introduces an approach oriented
to improve HCI. Moreover, the work presented by Healey and
Picard [2] deserves special mention, since they are considered
to be pioneers on stress-detection field.
The research by Lisetti and Nasoz [14] provides a complete
study on emotion recognition, including a deep literature review
on the experiments carried out to provoke emotions considering
populations, algorithms, approaches, and so forth.
III. P
HYSIOLOGICAL SIGNALS
Several indicators have been considered within the literature
to detect stress (Section II). However, this paper proposes the
use of only two signals: GSR, also known as skin conductance
(SC), and HR. These two signals were selected based on their
properties regarding noninvasivity when being acquired and
because their variation is strongly related to stress stimuli [2],
[10], [15].
GSR, known also as electrodermal activity, is an indicator of
SC [15], [16]. In details, glands in the skin produce ionic sweat,
provoking alterations of electric conductivity. First experiment
on this nature dates back to 1907, when Jung first described
some relations between emotions and the response of this
parameter [8], [9].
GSR can be obtained by different methods, but the device
proposed to acquire signals (Section IV-A) is based on an
exosomatic acquisition. In other words, extracting skin conduc-
tivity requires a small current passing through the skin. GSR is
typically acquired in hand fingers, and its units of measure are
microsiemens (micromhos) [8].
The main parameters of GSR, such as basis threshold, peaks,
or frequency variation, vary enormously among different indi-
viduals, and thus, no general features can be extracted from
GSR signals. Therefore, the parameters extracted from GSR
signals are strongly related to each individual. However, despite
the differences among individuals, GSR signal is not distinctive
enough to identify an individual in terms of biometrics.
On the other hand, HR measures the number of heartbeats
per unit of time. HR can be obtained at any place on the human
body, being an accessible parameter that can be acquired easily
[7], [17].
HR describes the heart activity when the autonomic nervous
system (ANS) attempts to tackle with the human-body demands
depending on the stimuli received [18]. ANS orders to increase
the blood volume within the veins so that the rest of the body
can react properly, if required. Concretely, a stressing stimulus
provokes a similar reaction, increasing the number of heartbeats
when facing these situations.
Among the wide number of methods to extract HR, the
most common methods consider to measure the frequency of
the well-known QRS complex in an electrocardiogram (ECG)
signal [19], [20]. In contrast to ECG biometric properties [21],
HR is not distinctive enough to identify an individual. Never-
theless, HR signal behaves in a different manner depending on
the stimulus and the individual.
Summarizing, both HR and GSR behave differently for each
individual, and therefore, posterior stress template must gather
properly these unique responses in order to obtain an accurate
result in stress detection.
IV. D
ATABASE ACQUISITION
Every biometric system requires a set of different biometric
acquisitions to train, validate, and assess the whole system [22].
This section provides an overview of how the data set was built
considering the experimental setup and the characteristics of the
database.
A. Overview
The experiments were carried out in the Faraday room in
the Human Psychology Laboratory, Psychology Faculty, Com-
plutense University of Madrid (UCM), endowed with electro-
magnetic, thermal, and acoustic insulation.
The main aim of this step is to collect HR and GSR signals
from each participant. The device proposed to carry out these
experiments is I-330-C2 PHYSIOLAB (J &J Engineering),
which is able to process and store six channels, including
electromyography, ECG, respiration rate, HR, and GSR. Sen-
sors were attached to the hand’s right (or left, but not both)
fingers [23], wrist, and ankle in order to acquire both HR
and GSR, avoiding sensor detachments, unplugged connectors
to the analog-to-digital converter, and/or software acquisition
errors. Moreover, the sample acquisition rate is made with one
sample per second for both HR and GSR.
DE SANTOS SIERRA et al.: STRESS-DETECTION SYSTEM BASED ON PHYSIOLOGICAL SIGNALS 4859
B. Participants
The participants were mainly students from the Psychol-
ogy Faculty (UCM) and Social Work (UCM), with a total of
80 female individuals of ages from 19 to 32 years old, with
an average of 21.8 years old and a standard deviation of 2.15.
The lack of male individuals is due to the Faculty where the
experiment took place, since the percentage of male to female
students is almost negligible. Therefore, no male candidates
were included during collection of the database.
C. Task Justification
Provoking stress on an individual requires a specific exper-
imental design in order to obtain a proper arousal in terms of
physiological signal [2], [4]. Concretely, this paper proposes to
induce stress by using hyperventilation (HV) and talk prepara-
tion (TP) [24].
HV is defined as a certain kind of breath, which exceeds
standard metabolic demands as a result of excess in respiratory
rhythm.
As a consequence, several physiological changes emerge,
such as arterial-pressure diminution in blood until a certain
level, so-called hypocapnea [24], [25], and blood pH increment,
known as alkalosis.
However, voluntary HV does not always produce an actual
anxiety reaction [24], and therefore, an additional anxiogenic
task is required to ensure that a positive valence in terms of
stress response is provoked. Such a task is TP.
Results provided by Cano-Vindel et al. [24] and Zvolensky
and Eifert [25] highlight that HV produces a physiological
reaction (in terms of physiological registration, HR, and GSR)
similar to the reaction induced by a threatening task of prepar-
ing a talk.
As a conclusion, TP and HV provoke both an alteration
in physiological parameters together with different emotional
experiences. These previous tasks have been widely studied and
evaluated with positive results, and they are very suitable to
induce stressing stimuli on individuals.
D. Procedure
Two groups (namely, Group 1 and Group 2) were created,
ensuring that the distribution of their respective anxiety levels,
measured by psychological tests [26], [27], were similar. In
other words, this selection seeks to avoid one group containing
people which barely react against stress and another group with
people which overreact under stressing conditions. Therefore,
both groups must be well balanced in terms of anxiety levels in
order to validate the experiments.
Participants from Group 1 underwent an experimental
session using physiological and subjective signals under the
following conditions: calm state (baseline, namely, BL1), stim-
ulating task (HV), threatening task (TP), and baseline poststress
(BL2). On the other hand, the order of tasks was swapped for
participants from Group 2: calm state (baseline), threatening
task (TP), stimulating task (HV) and baseline poststress. The
main reason for altering the order is to make the task order
independent from the results obtained [24], [26].
Obviously, BL1 implies no stressing stimuli on the individual
in contrast to HV and TP. However, nothing can be assured
in relation to BL2, since it cannot be considered either as a
stressing or as a relaxing state [24], [26], [27].
E. Database Discussion
Two questions arise from this database acquisition: Do these
experiments assure that the final system is able to detect
stress/relax states when performing real applications by being
trained with this database? To what extent could the results
obtained with these experiments be generalized to a wider
population, including males and females with a larger age
range?
The former question has an affirmative answer. The physio-
logical response to a stressing agent is strongly related to each
individual, and such a response is similar, independent of the
time during the stressing stimulus provoked the response [28].
Moreover, stress mechanism extracts some information from
the stimuli so that if such stressing agent appears again, the
human body is able to react faster and better compared with
the first time [1], [29]. This characteristic makes useless the
repetition of the same tasks after a certain period of time and
furthermore makes unnecessary a third session with different
tasks, since the response will not be the same as the stimuli
provided by different task provoke different responses.
Finally, the latter question is hard to answer since it is
difficult, even for expert psychologists, to state whether the
response among female and male individuals differs, inasmuch
as the previous response varies within female individuals [30].
Several researchers support the idea that male and female
individuals suffer different responses when stress agent endures
through time, (e.g., a great amount of work at job, a bad eco-
nomical situation, and so forth), but they have similar responses
when the stress stimuli consist of specific actions in a very short
period of time, e.g., an accident, an armed robbery, and the
like [29].
Thereby, it is justified to extend the results obtained with this
database to a wider population.
V. S
TRESS-DETECTION SYSTEM
A stress-detection system inherits several characteristics
from biometric systems. First of all, a template extraction is
required so that the system could create a profile in order to con-
trast, in future accesses, whether a user is actually under stress.
This template is based on specific characteristics extracted
from individual concerning parameters from the physiological
signals HR and GSR.
On the other hand, once the user is associated to a template,
the individual is able to access the system, and therefore, a
template comparison is required. Both steps are described in
following sections.
A. Template Extraction
Mathematically, both HR and GSR are considered as sto-
chastic signals. Therefore, H represents the space of HR possi-
ble signals, and G represents the space of GSR possible signals.
4860 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 58, NO. 10, OCTOBER 2011
Fig. 1. Graphic representation of γ. Notice how the relation between HR and
GSR varies depending on the stressing stimuli (BL1, TP, HV, and BL2).
Each stage will come up with a pair of signals h ∈H
and g ∈G according to the experimental task conducted in
each situation. Thus, a template extraction requires four pair
of signals, namely, γ =[(h
1
,g
1
), (h
2
,g
2
), (h
3
,g
3
), (h
4
,g
4
)] ∈
H×G corresponding to how the individual behaves under
different states. Notice that signals h
i
and g
i
are not normalized,
in contrast to previous approaches [2], [8]. The decision to
avoid normalization was done based on experience, since data
without normalization provide more accurate results in terms of
stress detection.
Once γ is obtained, for each pair of signals (h
i
,g
i
), i =
{1, 2, 3, 4}, a mean vector is obtained together with the devia-
tion for each pair. In other words, four parameters are obtained:
ζ
h
i
=
¯
h
i
, and ζ
g
i
=¯g
i
, which represent the mean of signals h
i
and g
i
in addition to σ
h
i
and σ
g
i
, related to the dispersion for
each pair. Finally, the stress template, namely, T is described
by T =(ζ
h
i
,ζ
g
i
,σ
h
i
,σ
g
i
), i = {1, 2, 3, 4}.
Fig. 1 shows a visual example of a scattering representation
of each pair of signals γ. Notice how nonstressing stimuli pro-
vokes a low excitation in GSR (Fig. 1, ◦) and, on the contrary,
the evidence of an arousal when undergoing an stressing tasks,
such as TP (Fig. 1, ) and HV (Fig. 1, ).
The aim of this action is to described the information in HR
and GSR by four Gaussian distributions, centered in (ζ
h
i
,ζ
g
i
)
and with standard deviations σ
h
i
and σ
g
i
. This approach will
facilitate the implementation of fuzzy antecedent membership
functions by Gaussian distributions in a posterior fuzzy deci-
sion algorithm.
Let t
T
be the time used to acquire both signals in order
to extract the stress template. Evidently, the performance of
the system depends on this parameter since the longer t
T
is,
the more information the system obtains, and therefore, the
stress template may be more accurate. A study regarding this
relation between t
T
and system performance is presented in
Section VII-D.
Finally, after template extraction, the template must be
stored. The template T requires 16 × 32 b since each template
element (whatever ζ
h
i
, ζ
g
i
, σ
h
i
or σ
g
i
be), is represented by a
float element or, in other words, 512 b, i.e., 64 B.
B. Template Update
As a matter of fact, T is different for each individual, and
therefore, it must be stored as a whole template. However,
these values do not remain unalterable but change along time.
Consequently, this template must be refreshed every time the
user accesses the system so that the template can adapt to the
variation of the specific individual. Such template update must
be performed each time the user uses the system. This template-
update scheme remains as future work (Section VIII).
VI. I
MPLEMENTATION
This paper proposes two possible system implementations:
a manual implementation, where the system parameters are set
without any learning procedure, and an automatic implementa-
tion, where the system adapts its internal parameters to relate
input with output during training stage.
The motivation for designing two systems is to compare
an understandable system based on expert knowledge (manual
implementation), containing very simple and intuitive rules,
with an automatic system, which learns the rules according to
the data, providing a more complex system.
In detail, the main difference between them relies on the
fuzzy decision algorithm: First, manual implementation is
based on a Mamdani [31] fuzzy decision algorithm using
template T to describe antecedent membership functions and
implementing the output with triangular distributions. Second,
the automatic implementation involves an adaptive-network-
based fuzzy inference system [31], [32] fuzzy algorithm carried
out to provide a fuzzy decision system adapted to the specific
data (i.e., HR and GSR signals in different tasks—BL1, TP, HV,
and BL2).
A. Manual Implementation
As introduced previously, the manual implementation is de-
signed with a Mamdani fuzzy system since it has been widely
used in expert decision systems due to their comprehensible
rules [31], [33]. This implementation provides a fuzzy output
on the interval [0, 1] based on template T .
This fuzzy system is described by Gaussian-based antecedent
functions whose parameters coincide with centroids ζ
h
i
and ζ
g
i
and deviations σ
h
i
and σ
g
i
. In other words, the variables are
represented by HR and GSR, considering linguistic labels for
each task in the experiments (BL1, HV, TP, and BL2). There-
fore, these functions describe the behavior of HR and GSR,
respectively, under stressing stimuli provoked by experimental
tasks BL1, TP, HV, and BL2. Gaussian-function selection is
justified since they are very suitable for the data provided by
HR and GSR signals.
Furthermore, two triangular functions describing each output
possibility (stress or nonstress) characterize the consequent
membership functions. The triangular functions were selected
according to their adequate properties to provide an accurate
output [33].
The proposed set of rules considers four rules, one for each
stressing/nonstressing situation (BL1, HV, TP, and BL2). These
DE SANTOS SIERRA et al.: STRESS-DETECTION SYSTEM BASED ON PHYSIOLOGICAL SIGNALS 4861
rules are very intuitive, and they are provided by an expert,
involving the previously described parameters ζ
h
i
, ζ
g
i
, σ
h
i
,
and σ
g
i
.
Before describing the rules in detail, some nomenclature is
required. Let h
k
i
and g
k
i
be the physiological signals HR and
GSR, respectively, for a given k subject (k ∈{1,...,80}) in
the state i, i ∈{1, 2, 3, 4}. This state i represents the four situ-
ations corresponding to the experiments to induce stress: BL1,
HV, TP, and BL2. Let s
k
m
be the output of the system, given a
k subject and m representing the two possible aforementioned
situations: stress (m = S) or nonstress (m = NS). Therefore,
the rules are mathematically described as follows:
h
k
1
∩ g
k
1
→ s
k
m=NS
h
k
2
∩ g
k
2
→ s
k
m=S
h
k
3
∩ g
k
3
→ s
k
m=S
h
k
4
∩ g
k
4
→ s
k
m=S or NS
.
The latter rule makes reference to the poststress state, which
is difficult to be classified as stress or nonstress, as stated
previously. This duality will come up with two possible imple-
mentation schemes with each of them considering this output
as stress and nonstress (Section VII-B).
Finally, the defuzzification method is carried out by a cen-
troid method [6].
B. Automatic Implementation
A different approach is used to improve the previous manual
implementation since the algorithm learns which membership
functions are to be used and which parameters must be selected
in order to obtain a given output. This automatic implemen-
tation based on adaptive neuro-fuzzy inference system, which
provides a Sugeno-type fuzzy inference system.
In other words, Sugeno implementation considers as ini-
tial conditions the membership functions from the previous
Mamdani implementation. Consequently, this implementation
needs a training stage to obtain the knowledge that is enough to
properly detect stress [32].
C. Differences Between Implementations
The main difference between the previous implementations
(manual versus automatic) is related to the manner by which
both systems are modeled. The Mamdani system models from
the beginning both the output and the input and requires stress-
template parameters (centroids and standard deviations). The
output was modeled previously to the data, and that model was
established manually based on a trial-and-error strategy.
On the other hand, the Sugeno implementation is able to
come up with a more accurate and precise output starting
from an initial condition based on T . Furthermore, no previous
knowledge about the output distribution is required since this
system is able to provide an optimum output, by adapting its
rules and membership function to the training data.
Regarding these rules, both implementations were eventually
forced to have as many rules as the stages involved during
the training. For instance, if the training stages require only
data from BL1 and TP, then there will be only two rules, one
corresponding to each stage. Finally, there is a difference in
terms of performance, since the Sugeno implementation offers
a higher success rate in differentiating stress from relax states.
D. Stress Measurement
Previous systems provide an output for each single pair of
values (h
i
,g
i
) ∈H×G. In other words, these systems can
provide an output each second.
However, the system provides an output each t
acq
seconds
in order to obtain a more representative result. Therefore, the
output consists of a vector of t
acq
points. The final value will
result as a metric from this previous vector.
Two measurements are extracted from that output, namely,
media μ and median μ
1/2
. The result provided by these indica-
tors is also verified to be in the interval [0,1].
These proposed measurements provide a result each t
acq
seconds, and, as discussed posteriorly in Section VII, measure-
ments based on median μ
1/2
usually produces a more accurate
result when compared with μ measurements.
Finally, the system must decide which output provided
by previous metrics (μ or μ
1/2
) corresponds to a stress or
a calm state. Although the output is provided within the
interval [0, 1], there must be a threshold on that interval in-
dicating the boundary between values belonging to stress state
or calm state. This threshold, namely, ρ
th
will be obtained so
that the performance of the system is maximized. Furthermore,
threshold ρ
th
is fixed and unique for the whole database.
The idea of obtaining one threshold for each individual could
improve the system performance. The study of this proposal
remains as future work (Section VIII).
VII. R
ESULTS
This section attempts to gather a study of the system perfor-
mance in relation to those parameters proposed in the previous
sections: threshold ρ
th
and temporal parameters (t
T
and t
acq
),
implementations (Mamdani or Sugeno), and stress measure-
ments (μ and μ
1/2
).
A. Database: Training, Validation, and Testing Data
In order to obtain valid results, the database must be divided
into three groups.
1) Training data: Used to extract the template, i.e., T =
(ζ
h
i
,ζ
g
i
,σ
h
i
,σ
g
i
).
2) Validation data: Used to fixed threshold ρ
th
and temporal
parameters (t
T
and t
acq
) in order to maximize the perfor-
mance of the system.
3) Testing data: Used to obtain which implementation and
metric are most suitable, and therefore, what is the per-
formance of the whole system.
For each individual, a vector containing t
T
seconds of γ (for
each task BL1, HV, TP, and BL2) was used for training data;
a vector of t
acq
seconds for each task was used to validate the
system, and the rest of the data were used as testing data. These