2020 Efectos Ansiedad Rodríguez Arce y Colaboradores

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Computer Methods and Programs in Biomedicine 190 (2020) 105408

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine


journal homepage: www.elsevier.com/locate/cmpb

Towards an anxiety and stress recognition system for academic


environments based on physiological features
Jorge Rodríguez-Arce, Ph.D. a,∗, Liliana Lara-Flores, M.Sc. a, Otniel Portillo-Rodríguez, Ph.D. a,
Rigoberto Martínez-Méndez, Ph.D. a
a
Facultad de Ingeniería, Universidad Autónoma del Estado de México, Toluca, Mexico

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

Article history: Background and Objective: Traditional methods to determine stress and anxiety in academic environments
Received 6 April 2019 consist of the application of questionnaires, but the main disadvantage is that the results depend on the
Revised 17 November 2019
students’ self-perception. Being able to detect anxiety-related stress levels in a simple and objective way
Accepted 18 February 2020
contributes greatly to dealing with low performance and school drop-out by students.
Methods: The main contribution of this study is to identify the physiological features that could be used
Keywords:
as predictors of stressful activities and states of anxiety in academic environments using an Arduino
Stress
Anxiety board and low-cost sensors. A test with 21 students was conducted, and a stress-inducing protocol was
Classifiers proposed and 21 physiological features of five signals were analyzed. In addition, the State-Trait Anxiety
Physiological data Inventory (STAI) was used to assess the level of anxiety for each student. Four classifiers were compared
to find the physiological feature subset that provides the best accuracy to identify states of stress and
anxiety.
Results: The stress due to activities performed by students can be identified with an accuracy greater
than 90% (Kappa = 0.84) using the k-Nearest Neighbors classifier, using data from heart rate, skin tem-
perature and oximetry signals and four physiological features. Meanwhile, the identification of anxiety
was achieved with an accuracy greater than 95% (Kappa = 0.90) using the SVM classifier with data from
the galvanic skin response (GSR) signal and three physiological features.
Conclusions: The results provide a clue that anxiety detection in academic environments could be done
using the analysis of physiological signals instead of STAI test scores. Besides, the results suggest that
physiological features could be used to develop stress recognition systems to help teachers to identify the
stressful tasks in an academic environment or to develop anxiety recognition systems to help students to
control their level of anxiety when they are performing either academic tasks or exams.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction piratory rate) [37]. In addition, Berrío and Mazo [5] define aca-
demic stress as the condition that students must endure as a re-
An academic environment is where students experience a vari- sult of a variety of stressors, related to the activities usually carried
ety of stress-related situations such as giving a class presentation, out at school. Kadapatti and Vijayalxmi (Kadapatti & Vijayalaxmi,
solving problems against the clock, and dealing with tests and ex- 2012) point to academic stress as the product of a combination
aminations. The stress that arises in an educational environment of academic demands that exceed the adaptive resources available
is called academic stress [10], and refers to the stress-related sit- to an individual. In general, academic stress refers to the student’s
uations during the educational process in which students usually dealing with demands or stressors in academic environments that
have no control and may cause students’ physiological reactions cause physiological, emotional, cognitive or behavioral discomfort.
(e.g. changes in their blood pressure, skin conductance, and res- In recent years, some studies suggest that academic stress is
related to low performance and dropout, mainly among first-year
undergraduate students. According to García-Ros et al. [10], an

Corresponding author. increase in stress levels of undergraduate students occurs more
E-mail addresses: jrodrigueza@uaemex.mx (J. Rodríguez-Arce), frequently during the year of university admission, because stu-
lililafl@hotmail.com (L. Lara-Flores), oportillor@uaemex.mx (O. Portillo-Rodríguez), dents are dealing with a process in which they must face greater
rmartinezme@uaemex.mx (R. Martínez-Méndez).

https://doi.org/10.1016/j.cmpb.2020.105408
0169-2607/© 2020 Elsevier B.V. All rights reserved.
2 J. Rodríguez-Arce, L. Lara-Flores and O. Portillo-Rodríguez et al. / Computer Methods and Programs in Biomedicine 190 (2020) 105408

demands in school activities, improve their time management, electronic instruments that allow to acquire physiological signals
have greater autonomy, among others. Kitsantas, Winsler, and Huie for further processing and analysis.
[20] report that only 73.6% of newly-enrolled undergraduate stu- It is widely known that biosensors can provide objective mea-
dents return to continue their studies the following year. sures of stress response under specific tasks. For example, Plarre
The early detection of academic stress in students represents et al. [28] exposed 21 subjects to a different kind of stressors, rep-
a challenge for universities. To face this challenge, the institution resentative of those experienced during daily activities. They pro-
must have enough trained personnel to detect the students most posed two models to classify stress using wearable sensors; their
likely to suffer from stress, and to implement preventive actions results showed a significant correlation between their physiological
and consequently reduce the dropout rate caused by academic classifiers and the participants’ self-reported ratings. Wijsman et al.
stress. Anxiety is one of the effects of stress, hindering student [43] used a wearable sensor system to measure physiological sig-
concentration, memory capacity, and other processes causing lower nals such as electrocardiogram (ECG), respiration, galvanic skin re-
academic performance of students [10]. Some authors have studied sponse (GSR), and electromyography (EMG) to detect mental stress.
the relationship between academic stress, anxiety levels and the They presented three different stress conditions to a healthy sub-
academic performance of students [[7],[42]], and the results show ject group and calculated 19 physiological features from all signals.
a negative correlation among them. Their results showed almost 80% of classification accuracy differen-
Traditional methods to determine stress and anxiety consist of tiating stress and non-stress conditions using different classifiers.
the application of questionnaires, self-reports (C. & K., 2016) and Aigrain et al. [1] proposed a multi-assessment methodology to an-
clinical interviews [39]. Beiter et al. [4] used the 21-question ver- alyze stress. They evaluated the predictive power of 101 physiolog-
sion of the Depression Anxiety Stress Scale (DASS-21) to evaluate ical features from EMG, GSR, temperature and heart rate (HR) sig-
374 undergraduate students. Their results show that anxiety, aca- nals. Their results showed that features related to body movement,
demic performance and pressure to succeed are the main stressors blood volume pulse, and heart rate are more useful for the classifi-
among students. The study concludes that problems derived from cation of stress. Vildjiounaite et al. [40] proposed a method to clas-
the state of anxiety tend to hamper the success of university stu- sify stress based on Hidden Markov Models and physiological data.
dents. It is therefore important for universities to continuously as- They tested the method and it achieved an average stress detec-
sess the mental health of their students and implement prevention tion accuracy of 75%. Maaoui and Pruski [22] proposed a compar-
methods to improve their academic performance. ative study of unsupervised approaches to detecting stress using
In their study, DordiNejad et al. [7] discuss the effect of the heart rate signal analysis method. They employed three un-
test anxiety on medical students’ academic performances, us- supervised classification techniques (K-means, the Gaussian mix-
ing the Sarason Test Anxiety (STA) questionnaire to collect data. ture model and self-organizing map). Their results showed that
Their data analysis shows a negative correlation between aca- K-means clustering performs the most accurate stress detection.
demic performance and test anxiety. Vitasari et al. used the Sevil et al. [34] focused on social and competitive stress detection
State-Trait Anxiety Inventory (STAI) to analyze the relation- to discriminate between stress and non-stress states using wrist-
ship between anxiety levels and academic performance [41]. band physiological signals. Their results show that different types
Their experiment shows a significant correlation between high- of stress can be detected with an accuracy of over 87%.
level anxiety and low academic performance among engineering Nevertheless, although previous studies show that the use of
students. some classifiers provides high accuracy of stress detection (in gen-
Other authors such as Jing [16], Peleg [27], and Onyeizugbo eral, accuracy of over 70% [[2],[12],[35]]), it should be noted that
[26] also show that undergraduate students with higher test anx- these studies do not provide conclusive results on the stress in-
iety have low academic performance. Reddy, Menon, and Thattil duction methods and physiological signals that can be useful for
[30] show that academic stress is a devastating problem affecting stress and anxiety identification; it seems that the methods and
the mental health of students and, thus, affecting their academic signals differ from one context to another (for example, between
performance. Their work shows that high stress levels were re- driving activities and academic tasks). Thus, it is necessary to cor-
ported in engineering and medical students, and it emphasizes the rectly define the stressor tasks, signal processing and classification
need for medical interventions during the semester. methodologies for each environment, activity and context. In ad-
Although several authors have shown a significant correlation dition, Giannakakis et al. [12] mention that most of the previous
between anxiety levels and the academic performance of students experiments were performed in well-controlled environments, in
using self-report questionnaires, the main disadvantages of this laboratory, these results could be different if the experiments are
method are the need for an expert (usually a psychologist) to in- executed in real-life environments; then, it is necessary to conduct
terpret the results of questionnaires, and the fact that the results new studies with real-life conditions using induced stressors re-
could be misleading since the answers on the questionnaires de- lated to the task to be evaluated.
pend on students’ perception. As a consequence, this research analyzes the use of Arduino
Stress and anxiety detection methods may also be based on board and low-cost sensors to acquire physiological data in com-
the acquisition and analysis of subjects’ physiological information bination with the STAI self-report questionnaire in an academic
(physiological signals). These methods are based on the fact that environment (this questionnaire is the most widely used by the
the physiological signals present changes when the subject is un- counseling and psychological services office at the School of Engi-
der stressful situations or activities [[12],[35]]. Physiological signals neering, where this research was carried out, in order to evaluate
are obtained from the human body by means of several types of students’ anxiety levels). The advantage of this approach is to avoid
biomedical sensors, providing information of vital functions. These the results of the diagnosis depending on students’ self-perception.
signals can be described in terms of amplitude with respect to The main contribution of this work is to identify the physiological
time and frequency [29]. The physiological features (or character- features that can be used to predict stressful activities and states
istics) are measurable variables related to the physiological signals, of anxiety in students in real academic environments.
which allow to monitor the body function. It is important to un-
derstand that physiological signals are raw signals acquired by sen- 2. Methods
sors, while physiological features are generally “hidden” in the raw
signals and they should be extracted by means of different signal Fig. 1 shows the architecture of the electronics platform used to
processing techniques [17]. Biomedical sensors or biosensors are acquire physiological signals. The platform consists of: (a) a stress-
J. Rodríguez-Arce, L. Lara-Flores and O. Portillo-Rodríguez et al. / Computer Methods and Programs in Biomedicine 190 (2020) 105408 3

Fig. 1. Architecture of the electronics platform used in this study.

inducing protocol, (b) a data acquisition system (DAQ) and sensors


to acquire the psychological signals, (c) a database to store the sig-
nals for further analysis, and (d) a statistical model to identify the
stress and anxiety based on the STAI questionnaire results.

2.1. Stress-inducing protocol (stress conditions)

Determining whether an individual is experiencing stress or not


through physiological variables is not an easy task; it must be
guaranteed that the physiological changes detected at the time of
measurement are caused by stress and not by other factors (e.g.
other emotions, physical activity, and environmental factors). The
methodology used in similar studies to guarantee this was induc-
ing stress by presenting known stressful activities to the subject.
However, the number of stressful activities proposed is lim- Fig. 2. Screenshot of the tablet application used as the stress task (Rising Number
ited in the literature. Moya-Albiol and Salvador [24] mention that application by Wazo for Ipad).

among the activities most commonly used as psychological stres-


sors are: public speaking, arithmetic calculations, “the Stroop task”
simple questions such as “how do you feel?”, “how was your
[18], and other stressful tasks such as reaction time activities,
day?”; in addition, the instructions for the next task are ex-
video-games, and solving cognitive problems.
plained by the experimenter. The objective of this task is for
Sandulescu et al. (Sandulescu, Andrews, Ellis, Bellotto, & Mozos,
participants to rest between tasks.
2015) used a protocol based on the Trier Social Stress Test (TSST) • Non-stress task: this consists of listening to instrumental music
to identify stress in their experiment. The protocol consisted of a
while the subject is seated, as relaxed as possible. Salai et al.
neutral task (known as anticipatory stress phase), followed by the
[32] report that sometimes the stress-task could be more re-
tasks of speaking in public, one cognitive activity and again the
laxing than the non-stress task for some subjects, decreasing
neutral task until completion [19]. The neutral task consisted of
the stress level detection. In order to solve this issue, the sub-
two minutes only answering simple questions. The task of speak-
jects are also asked to close their eyes and to imagine a relaxing
ing in public consisted of a five-minute interview to obtain a job,
landscape, trying to minimize as much as possible any stressful
then each participant was asked to perform the cognitive task,
thoughts in order to enhance the relaxation state. In addition,
which consisted of counting backward in steps of 13 starting from
according to Ogba et al. [25], the fact of listening to music with
1022.
one’s eyes closed reduces levels of academic-related stress in
Salai, Vassányi, and Kósa [32] measured physiological parame-
undergraduate students.
ters of a control group while each individual listened to relaxing • Stress-task: a time-constrained arithmetic task was selected as
music. They recorded the values of the physiological signals and
a stressful task based on the works of Sandulescu [33] and
then measured the same parameters while the participants per-
Salai, Vassányi, and Kósa [32]. Students were asked to solve a
formed the “Stroop color test”. In the last task, the subject was
puzzle in which they had to add two numbers to calculate a
asked to achieve their best score in ten minutes; if an error was
specific value. The puzzle is a tablet-based application game,1
made, the game ended and the score restarted from 0.
based on the principle of addition, in which a set of different
Based on the previous work, in this study a stress-inducing pro-
positive numbers are presented in a random way on the left
tocol for undergraduate students was proposed using three differ-
and right edges of the screen, and a specific value (score) is
ent tasks: a neutral task, a non-stress task, and a stress task. Each
shown in the center of the screen (see Fig. 2). The students had
task is described in detail below:
to choose three numbers from the edges in such ways that,
• Neutral task: this task is similar to that proposed by
Kirschbaum et al. [19] and Sandulescu [33]. The students re- 1
Rising Number App by WAZO. https://appadvice.com/app/rising-number/
main seated without making any movement. They are asked 907455091 (accessed February 2, 2019).
4 J. Rodríguez-Arce, L. Lara-Flores and O. Portillo-Rodríguez et al. / Computer Methods and Programs in Biomedicine 190 (2020) 105408

Fig. 3. On the left, the GSR sensor is placed on the index and middle fingers and the oximeter sensor is placed on the thumbs of the students. On the right, the breathing
sensor was placed near the nose.

when added, the result is the value shown in the center of changes (the nasal temperature air). This sensor was used to
the screen (users must calculate the sum of three numbers). measure the students’ respiratory rate.
Each time the students give the correct answer their score in-
creases and a new value is shown on the screen. In order to
increase the complexity of the task, the new value is greater 2.3. Experimental design
than the previous one. Each student must continue until they
reach a score equal to or higher than 100 points in five minutes. 2.3.1. Experimental procedure
During the task, if a mistake is made, the game ends and the The experiment was conducted at the facilities of the School of
score will restart from zero. According to the participant’s per- Engineering at the Universidad Autónoma del Estado de México.
formance, they could obtain a candy (score equal to or higher Participants followed the protocol (see Fig. 4), as described below:
than 100 points) or an electrocutaneous stimulation (score less
than 100 points). According to Rhudy and Meagher [31], ex- 1 Familiarization session: In this stage, the objective and procedure
posure to electrocutaneous stimulation (or electric shock) in- of this experiment were explained to each participant. They
duces fear, while anticipation of shock (without exposure) in- were also notified about the prize that they will receive ac-
duces anxiety. In this study, the electrocutaneous stimulations cording to their performance during the test. Participants who
consisted of a brief shock delivered with 500 ms pulse dura- agreed to continue filled out an identification sheet and signed
tion and less than 5.0 mA (slight shock felt; not painful [9]). a consent form. Participants were given instructions regard-
ing to the two tasks that they had to complete, and also re-
ceived a brief demonstration of how to perform each task. For
2.2. Devices and sensors example, in the case of the arithmetic task, each participant
had two minutes to practice the task and understand how the
The e-Health V2 (Cooking Hacks, Spain) shield and the Arduino tablet application works. In this session, the participants also
Uno board [3] were used as the DAQ system. The Arduino board received an explanation about the sensors that were used to
has a CPU speed of 16 MHz and a built-in 10-bits ADC (analog-to- collect the physiological data. At the end of this session, all sen-
digital converter) that provides enough resolution to read the sub- sors were placed on the participant and their operation verified
jects’ physiological signals. The e-Health V2 platform is a low-cost (one-minute data acquisition). Participants remained seated for
platform that could be used to acquire raw physiological data in the duration of the experiment; they could look at the screen
real-time from nine sensors. In this study only four of them were showing the time remaining for each activity. All participants
chosen to acquire five physiological signals from students: were right-handed, so the sensors were placed on their left
hand, enabling them to use their right hand to perform the ex-
• Body temperature: this allows skin temperature to be measured perimental activities. Fig. 5 shows the experimental setup.
with a 0.1 °C resolution; the sensor was placed in the axillary 2 Test session: Before starting with the data acquisition, each par-
area of each student. A piece of adhesive tape was used to at- ticipant was encouraged to clarify any concerns because during
tach this sensor to the skin. the experiment they would be not allowed to talk and had to
• Galvanic skin response: this sensor measures the electrical con- remain seated. They were also asked to be as quiet as possible,
ductance of the skin which is affected by the sweat, controlled making only the necessary movements to perform the experi-
by the sympathetic nervous system. As a consequence, in mental tasks. The activities were carried out in the same order
episodes of stress, the electrical resistance of the skin changes. and duration for each participant. The stress-inducing protocol
For this study, the sensor was placed on the index and middle was composed of three main activities: a neutral task, a non-
fingers (see Fig. 3) of the left hand. Adhesive tape was used to stress task, and a stress task. The experiment started with a
attach the sensor to the skin of each finger. neutral task for a period of three minutes. The experiment con-
• Pulse oximeter (for heart rate and oxygen saturation): this is a tinued with the non-stress task. In this case, the selected mu-
sensor capable to measure the amount of oxygen dissolved in sic was a piece of instrumental piano lasting 4:30 min. At the
the blood (oxygen saturation). This sensor can also measure the beginning of the music, the participant was asked to close their
heart rate and provides beats per minute (bpm) values. This eyes and not open them until instructed to do so. Once the mu-
sensor was attached to the thumb of the left hand (see Fig. 3). sic ended, the participant was asked to open their eyes and the
• Breath-rate sensor: this is a thermistor that was attached near neutral task was repeated for three minutes. The next activity
the nose (see Fig. 3) to measure the nasal thermal airflow was the stress-task. As mentioned in the previous section, each
J. Rodríguez-Arce, L. Lara-Flores and O. Portillo-Rodríguez et al. / Computer Methods and Programs in Biomedicine 190 (2020) 105408 5

Fig. 4. Experimental procedure.

Table. 1
Typical spectral characteristics of physiological sig-
nals (adapted from (Mendes, Figueiredo, Fernandes, &
Gama, 2011)).

Physiological signal and label Bandwidth (Hz)

Skin temperature (ST) 0–1


Oximetry (SpO2) 0–30
Breath-flow rate (BR) 0–10
Heart rate (HR) 0–5
Galvanic skin response (GSR) 0–5

2.4. Data acquisition and feature extraction

Five signals of each subject were recorded: heart rate (HR), skin
Fig. 5. Experimental setup: a student is doing the arithmetic task. temperature (ST), galvanic skin response of the hand (GSR), oxy-
gen saturation (SpO2), and breath-flow rate (BR). The sampling rate
was set to 50 Hz. Table 1 shows the typical spectral characteristics
participant had five minutes to achieve a score higher than 100 of the recorded signals; as a consequence, each signal was filtered
points in an arithmetic task. using a low-pass 6th-order Butterworth filter. Fig. 6 shows the raw
3 STAI questionnaire: To collect data about anxiety, the State-Trait and filtered signals of “participant #5 .
Anxiety Inventory was applied to each participant. The ques- MATLAB (2016a, MathworksTM ) software was used for process-
tionnaire consists of 40 questions with a 4-point Likert scale; ing each signal (time-series data) and for feature extraction. The
higher levels of anxiety are positively correlated with higher signal of each subject was divided into individual segments ac-
scores. This psychological inventory measures two types of anx- cording to the experimental tasks (see Fig. 7); two segments be-
iety: anxiety about an event (state-anxiety, 20 items) and anx- long to the neutral task (segments #1 and #3 of Fig. 7), one seg-
iety level as a personal characteristic (trait-anxiety, 20 items). ment belongs to the non-stress task (segment #2 of Fig. 7), and
The experiment ended by removing the sensors attached to the one segment belongs to the stress-task (segment #4 of Fig. 7). The
participant. two neutral task segments were discarded for the data analysis.
The non-stress task and the stress-task segments were divided into
windows-size of 2500 samples intervals; this windows-size was
2.3.2. Participants based on the feature extraction methodology proposed by other
Twenty-one right-handed students participated in this study. authors [[8],[11]].
All participants in this study were undergraduate engineering stu- Four features were extracted for each of the following physio-
dents. There were 9 female students and 12 male students, with logical signals: HR, ST, GSR and SpO2. For the case of the BR signal,
ages ranging from 18 to 21 years. All participants reported normal five features were extracted. Consequently, a total of 21 features
vision and hearing; none of them reported having consumed any were calculated for further analysis; the same feature extraction
medication or being under any medical or psychological treatment. procedures described in [[6],[14],[15],[36],[43]] were used. The fea-
They were chosen by random sampling, only considering the crite- tures for each physiological signal and its feature extraction proce-
rion that they were in the first or second semester at university. dure are summarized in Table 2.
In order to minimize students’ potential concerns about their aca- In order to evaluate the performance among the four classifiers
demic duties and/or exams, all the tests were carried out at the and the four datasets, the accuracy and the Cohen’s Kappa coef-
end of the semester (after the end of the evaluation period). ficient were calculated. In this study, the accuracy is defined as
6 J. Rodríguez-Arce, L. Lara-Flores and O. Portillo-Rodríguez et al. / Computer Methods and Programs in Biomedicine 190 (2020) 105408

Fig. 6. Left side, raw signals; right side, filtered signals of participant #5.

Fig. 7. Each signal was divided into five individual segments: segment numbers 1and 3 belong to the neutral task, segment number 2 belongs to the non-stress task, and
segment number 4 belongs to the stress task.

Table. 2
Feature extraction of each physiological signal (HR – heart rate, ST – skin temperature, GSR – galvanic skin response, SpO2 –
saturation of oxygen, BR- breath flow rate).

Derived from Signal HR Signal ST Signal GSR Signal SpO2 Signal BR

Mean normalization by Healey [14] HR1 ST1 GSR1 SpO2-1 BR1


Root mean square (RMS) [43] HR2 ST2 GSR2 SpO2-2 BR2
Means of differences between adjacent elements [15] HR3 ST3 GSR3 SpO2-3 BR3
Sum of local peak [36] —— —— —— —— BR4
Mean normalization by Iliou [15] HR5 ST5 GSR5 SpO2-5 BR5
J. Rodríguez-Arce, L. Lara-Flores and O. Portillo-Rodríguez et al. / Computer Methods and Programs in Biomedicine 190 (2020) 105408 7

Table. 3 bels were proposed based on the type of task, non-stress task, and
Interpretation of Cohen’s Kappa value (adapted from [[21],[23]]).
stress task.
Kappa value Interpretation % of data that are reliable In addition, a feature selection process was run to reduce the
<0 None or poor agreement 0–4% size of the feature vector that could provide a higher classifica-
0.0 – 0.20 Minimal agreement 4–15% tion value (thus reducing the number of sensors in comparison
0.21 – 0.40 Weak agreement 15–35% with the feature vector of the Full Dataset). Two correlation-based
0.41 – 0.60 Moderate agreement 35–63% attribute subset evaluators (CfsSubsetEval and CorrelationAttributeE-
0.61 – 0.80 Strong agreement 64–81%
val) and one information gain attribute evaluator (InfoGainAttribu-
0.81 – 1.00 Almost perfect agreement 82–100%
teEval) were used to create three new datasets. The CfsSubsetEval
evaluator assesses the worth of a subset of attributes by consider-
ing the individual predictive ability of each feature along with the
the percentage of testing set examples correctly classified; in other degree of redundancy between them [13]. The CorrelationAttribu-
words, it is the fraction of predictions the classifier got right. On teEval evaluator estimates the worth of an attribute by measuring
the other hand, the Cohen’s Kappa value is a statistical measure of the Pearson’s correlation between it and the class. On the other
inter-rater agreement between predicted and actual values, and it hand, the InfoGainAttributeEval evaluator measures each attribute
is useful to handle imbalanced class problems. In this study, the by evaluating their information gain with respect to the class. It
Cohen’s Kappa was used to evaluate the fitness of each classifier, discretizes numeric attributes first using the Minimum Descrip-
and its interpretation is based on the nomenclature proposed by tion Length method [44]. All these attribute evaluators are tools
Landis and Koch [21], where a perfect agreement would equate for data preparation provided by Weka 3.8 software. Table 5 shows
to a value of 1, and a non-agreement would equate to 0. The a summary of the results of each dataset.
nomenclature to describe the strength agreement associated with In Table 5 it is shown that the highest accuracy of 95.98% to
the Kappa’s value is shown in Table 3. McHugh [23] suggested that classify stressful activities in undergraduate students is obtained
80% of data that are reliable when the value Kappa>0.60, in other with the KNN method using the Full dataset (with a feature vector
words, if a classifier has a Kappa value lower than 0.60, it means with 13 features and five signals). On the other hand, the lowest
that the confidence value is low, thus it is not recommended to accuracy of 63.31% is obtained by the LogR algorithm (with a fea-
use it. ture vector with four features and three signals).

3. Results 3.2. Identification of the features influenced by students in states of


anxiety
3.1. Identification of the features influenced by the non-stress and
stress tasks Since one of the main objectives of this study is to define pre-
dictive features to detect anxiety in students in states of stress, the
In order to analyze and identify the features that are influenced participants were divided into two groups according to the score
by the stress and the non-stress tasks, a two-sample t-student test obtained from the STAI questionnaire. The Mexican version of this
was done (N = 21) using Minitab 17 software (Minitab Inc, USA). anxiety test is called IDARE [38], and it reports a mean value of
For this analysis, only the data of the non-stress and the stress 37.68 for male and 38.25 for female in the trait-anxiety question-
segments of each participant were used (see Figs. 4 and 7), and naire, and 36.35 for male and 35.12 for female in the state-anxiety
in total 21 physiological features were calculated from five physi- questionnaire as measures of anxiety reaction in undergraduate
ological signals. Before running the t-student test, the mean value students. In order to separate the students into two groups, the
of each feature in each segment was calculated (the mean of all mean values from the state-anxiety questionnaire of IDARE were
windows for each segment), for example, the mean value of HR1 used. In this way, group A (labeled as non-stress students, N = 8)
for the non-stress task segment was compared with the HR1 mean included students with scores lower than 36.35 for male and 35.12
value of the stress task segment. A confidence interval of 95% was for female; on the other hand, group B (labeled as stress stu-
selected. Table 4 shows the results of the t-student test. dents, N = 13) included students with scores equal to or higher
According to Table 4, the mean differences for 13 out of 21 fea- than 36.35 for male and 35.12 for female. The comparison between
tures were statistically significant (p < 0.05). As a consequence, a group A and B was tested using a t-student test. For this case,
feature vector of 13 characteristics was chosen to classify different the test was run using only the physiological data of the stress-
activities that cause stress in undergraduate students. task segment and the feature vector of 13 features identified in
Four classifiers – Support Vector Machine (SVM), k-Nearest Section 3.1. The results are shown in Table 6.
Neighbors (KNN), Random Forest and Logistic Regression (LogR) – According to Table 6, the mean differences for 6 out of 13 fea-
were used to find a feature subset that provides the best accu- tures were statistically significant (p < 0.05). This means that three
racy to classify stress. Each of these algorithms was tested with out of five physiological signals could be used to identify anxiety
10-fold cross-validation. The classifiers were implemented using according to the score from the STAI questionnaire. Once again, the
Weka 3.8 software (The University of Waikato, New Zealand) with SVM, KNN, Random Forest and LogR classifiers were used to find
the default parameter settings [44]. In this case, two predictive la- a feature subset that provides the best accuracy to classify anx-

Table. 4
Features influenced by the stress and non-stress tasks (the shadow values are the selected features for further analysis).

Heart rate (HR) Oxygen saturation (SpO2) Skin temperature (ST) Breath rate (BR) Galvanic skin response (GSR)

Features p-value Features p-value Features p-value Features p-value Features p-value

HR1 0.001 SpO2-1 0.721 ST1 0.048 BR1 0.094 GSR1 0.001
HR2 0.001 SpO2-2 0.005 ST2 0.538 BR2 0.56 GSR2 0.001
HR3 0.851 SpO2-3 0.006 ST3 0.001 BR3 0.407 GSR3 0.001
—— —— —— —— —— —— BR4 0.001 —— ——
HR5 0.001 SpO2-5 0.005 ST5 0.539 BR5 0.562 GSR5 0.001
8 J. Rodríguez-Arce, L. Lara-Flores and O. Portillo-Rodríguez et al. / Computer Methods and Programs in Biomedicine 190 (2020) 105408

Table. 5

Subsets of features that provide better accuracy to classify stress activities in undergraduate students (note: the symbol near each accuracy value indicates that the
dataset is significantly different from the Full dataset, p < 0.05).

Name Full Dataset Dataset #2 Dataset #3 Dataset #4

Feature
selection
methods None CorrelationAttributeEval CfsSubsetEval InfoGainAttributeEval

HR1, HR2, HR5, SpO2-2, SpO2-3, SpO2-5,


Feature vector ST1, ST3, GSR1, GSR2, GSR3, GSR5, BR4 GSR1, GSR2, GSR5, ST3, BR4 HR1, HR2, SpO2-2, ST3 GSR1, GSR2, GSR5, ST3, HR1

Characteristics 5 signals 13 features 3 signals 5 features 3 signals 4 features 3 signals 5 features

Performance Accuracy (95% CI) Kappa Accuracy (95% Kappa Accuracy Kappa Accuracy (95% Kappa
CI) (95% CI) CI)
KNN (K = 1) 95.98% (±4.73) 0.92 88.90% 0.78 91.85% 0.84 88.23% 0.76
(±7.35)∗ (±6.64) (±7.70)∗
SVM 86.88% (±8.62) 0.74 84.46% (±8.41) 0.69 67.18% 0.33 84.72% (±8.61) 0.69
(±9.00)∗
LogR 87.59% (±8.29) 0.77 83.52% (±9.22) 0.67 63.31% 0.26 85.37% (±8.28) 0.70
(±9.30)∗
RandomForest 94.44% (±5.85) 0.89 90.53% (±7.44) 0.80 84.39% 0.69 87.68% 0.75
(±8.57)∗ (±7.84)∗

Table. 6 from three signals (in this way, the hardware requirements and the
Features that are in-
number of physiological features to analyze are fewer).
fluenced by anxiety in
students (the shadow For the case of the feature vector from Dataset #3, its use to
values are the selected predict stress using either the SVM or LogR classifiers is not rec-
features for further ommended because in both cases the Kappa value is below 0.40,
analysis). meaning that more than 65% of the results are erroneous.
Feature p < 0.05 Vitasari et al. [41] demonstrate that Mathematics causes stress
among some engineering students, causing their academic perfor-
HR1 0.001
HR2 0.546 mance to be lower. Based on the results of Dataset #3 in Table 5,
HR5 0.539 a feature vector of four characteristics of three physiological sig-
SpO2-2 0.001 nals (HR, SpO2 and ST) with the KNN classifier could be used to
SpO2-3 0.437
develop a platform to help teachers to identify the stress tasks in
SpO2-5 0.001
GSR1 0.002 class. Consequently, teachers can help students in handling their
GSR2 0.001 stress due to different kinds of academic tasks to increase the qual-
GSR3 0.387 ity of learning.
GSR5 0.001 On the other hand, to identify the anxiety in undergraduate stu-
ST1 0.442
dents in states of stress the SVM classifier provides the best accu-
ST3 0.574
BR4 0.762 racy using three signals (HR, SpO2 and GRS) and a feature vec-
tor of six characteristics (see Table 7). However, the paired t-test
shows that there is not a significant difference of accuracy between
iety. The predictive labels proposed were without-anxiety (group Dataset #4 and the Full dataset; in this way, anxiety could be iden-
A) and with-anxiety (group B). The default parameter settings in tified with an accuracy of more than 95% and an almost perfect
Weka for each classifier were used with 10-fold cross-validation. agreement (Kappa = 0.90) using only one signal (GSR) and three
Table 7 shows a summary of the results. physiological features.
As shown in Table 7, the maximum accuracy of 98.89% was For the case of the feature vectors from datasets #2 and #3,
achieved using six features and three signals with the SVM clas- although their accuracy values are greater than 70%, there is a sta-
sifier. The lowest accuracy (72.78%) was achieved with the same tistically significant difference (p > 0.05) with the feature vector
classifier using three features and two signals. from the Full Dataset, and their Kappa statistics indicate a moder-
ate to weak agreement (Kappa<0.61). As a result, the use of these
4. Discussion two datasets as features to predict anxiety is not recommended.
Vitasari et al. [41] demonstrate that there is a negative corre-
For the identification of features to classify the non-stress and lation between anxiety levels and academic performance among
stress tasks, according to the results in Table 5, the Full dataset undergraduate engineering students (high-level anxiety results in
provides the best accuracy (95.98%) using the KNN classifier and lower academic performance). Consequently, the results of this re-
a Kappa value of 0.92 (almost perfect agreement). Nevertheless, it search could be used to develop a new ambient intelligence frame-
is necessary to use a feature vector with 13 physiological compo- work to support students effectively while carrying out stress ac-
nents from five signals. In order to determine whether the stress tivities and help them to reduce the level of anxiety to improve
influenced by the non-stress and stress tasks in students could their academic performance. For example, the SVM or KNN clas-
be identified using a feature vector with fewer physiological fea- sifiers with the GSR signal (accuracy of more than 95% to classify
tures, a paired-test with 20 repetitions of 10-fold cross-validation anxiety could be obtained by using three features of the GSR sig-
and a significance value of 0.05 was run. The paired t-test con- nal; see Table 7) could be used to develop a platform to help stu-
firms that the difference in accuracy is not statistically significant dents to control anxiety levels when they are performing academic
(p < 0.05) between Full dataset and Dataset #3 (accuracy=91.85% activities and/or exams. In this way, when the platform detects
and Kappa=0.84) using the KNN classifier; therefore, stress can be anxiety it will trigger an alarm to indicate the student to relax,
identified using a feature vector with four physiological features which means the students can receive a recommendation when
J. Rodríguez-Arce, L. Lara-Flores and O. Portillo-Rodríguez et al. / Computer Methods and Programs in Biomedicine 190 (2020) 105408 9

Table. 7

Subsets of features to identify anxiety according to the STAI score (note: the symbol near each accuracy value indicates that the dataset is significantly different from
the Full dataset, p < 0.05).

Name Full Dataset Dataset #2 Dataset #3 Dataset #4

Feature
selection
method None CorrelationAttributeEval CfsSubsetEval InfoGainAttributeEval

HR1, SpO2-2, SpO2-5, GSR1,


Feature vector GSR2, GSR5 SpO2-2, SpO2-5, GSR2, GSR5 SpO2-2, GSR1, GSR2 GSR1, GSR2, GSR5

Characteristics 3 signals 6 features 2 signals 4 features 2 signals 3 features 1 signal 3 features

Accuracy (95% Accuracy (95%


Performance CI) Kappa CI) Kappa Accuracy (95% CI) Kappa Accuracy (95% CI) Kappa

KNN (K = 1) 95.56% (±6.32) 0.91 85.00% 0.65 82.56% (±13.14)∗ 0.60 95.22% (±7.45) 0.89
(±11.86)∗
SVM 98.89% (±3.35) 0.97 74.56% 0.34 72.78% (±12.27)∗ 0.33 95.56% (±6.51) 0.90
(±12.16)∗

LogR 95.22% (±7.28) 0.89 75.11% 0.41 74.33% (±14.96) 0.43 87.56% (±6.23) 0.84
(±12.25)∗
RandomForest 88.22% 0.74 76.00% 0.38 74.22% (±13.01)∗ 0.35 83.78% (±7.79) 0.85
(±10.09) (±10.32)∗

they experience anxiety, so they could continue with the learning Selection - SFS) to compare with the statistical methods used in
activity to increase their learning engagement. Finally, Zhu, Ober, this study to select the most relevant features.
and Jafari [45] propose the use of physiological signals to assist
professors with the assessment of the students’ attention levels. Compliance with ethical standards
Their results show an accuracy of more than 98% for predictions
of interest level in a reading activity. For future work, the authors Ethical Approval: All procedures performed in studies involving
propose to run new experiments in order to evaluate and identify human participants were in accordance with the ethical standards
the main physiological features that could be used as predictors of the institutional and/or national research committee and with
of students’ attention levels using another kind of academic activ- the 1964 Declaration of Helsinki and its subsequent amendments
ity (i.e. mathematical computations, solving engineering problems, or comparable ethical standards.
answering a test, etc.). Informed Consent: Informed consent was obtained from all indi-
vidual participants included in the study.
5. Conclusions

Academic stress is a cause of poor academic performance in un- Declaration of Competing Interest
dergraduate students and results in dropout from university. Self-
reports are a tool to identify students prone to anxiety. However, The authors declare no potential conflicts of interest with re-
the interpretation of the test requires an expert and the diagnosis spect to the research, authorship, and/or publication of this article.
depends on the veracity of the answers and perception of the sub-
jects. In this study, a stress-inducing protocol was proposed, and 21 Acknowledgments
features of five signals were analyzed from time-series physiolog-
ical data from 21 undergraduate students. In addition, the anxiety The authors would like to thank the Autonomous University of
due to an arithmetic task was corroborated through the application the State of Mexico (UAEM) for the financial support of this project
of the STAI questionnaire for each student. (4313/2017/CI). Liliana Lara-Flores acknowledges the Mexican Na-
In order to identify stress due to activities performed by the tional Council for Science and Technology (CONACYT) for the schol-
student, accuracy in classification greater than 90% and a Kappa arship CONACYT CVU 786012 to carry out this research.
value > 0.80 were achieved by using the KNN classifier and only
three sensors (HR, SpO2 and ST) with four physiological features. References
Meanwhile, the classification of anxiety with an accuracy of more
than 95% and Kappa = 0.90 can be done using only three char- [1] J. Aigrain, M. Spodenkiewicz, S. Dubuisson, M. Detyniecki, D. Cohen,
M. Chetouani, Multimodal stress detection from multiple assessments, IEEE
acteristics from the GSR signal and the SVM classifier. The main Trans. Affective Comput. (2016), doi:10.1109/taffc.2016.2631594.
advantage of the use of the GSR signal as a predictor of anxiety is [2] A Alberdi, A Aztiria, A. Basarab, Towards an automatic early stress recognition
to avoid noise or artifacts due to subjects’ body part movements system for office environments based on multimodal measurements: a review,
J. Biomed. Inform. 59 (2016) 49–75, doi:10.1016/j.jbi.2015.11.007.
(other signals could be affected by this kind of artifacts, i.e. EMG, [3] S.F. Barrett, Arduino, Microcontroller processing for everyone!, Synth. Lect.
ECG, HR [12]). Thus, the results provide a clue that anxiety detec- Digit. Circuits Syst. 8 (4) (2013), doi:10.220 0/s0 0522ed1v01y201307dcs043.
tion could be done using the analysis of the physiological signals [4] R Beiter, R Nash, M McCrady, D Rhoades, M Linscomb, M. Clarahan, S Sammut,
The prevalence and correlates of depression, anxiety, and stress in a sample of
instead of the STAI test score.
college students, J. Affect. Disord. 173 (2015) 90–96, doi:10.1016/j.jad.2014.10.
One limitation of this study could be the experimental proce- 054.
dure because only one type of stress task was tested. Nevertheless, [5] N Berrío García, R Mazo Zea, Estrés académico, Revista de Psicología Universi-
dad de Antioquia 3 (2) (2011).
this research proposes to use another kind of academic task, to run
[6] Y Deng, Z Wu, C.-.H. Chu, Q Zhang, D.F Hsu, Sensor feature selection and com-
future experiments, and to evaluate the level of anxiety according bination for stress identification using combinatorial fusion, Int. J. Adv. Rob.
to the complexity of the task and the academic performance of the Syst. 10 (8) (2013) 306, doi:10.5772/56344.
students. In addition, future work derived from this study includes [7] F.G DordiNejad, H Hakimi, M Ashouri, M Dehghani, Z Zeinali, M.S Daghighi,
N Bahrami, On the relationship between test anxiety and academic perfor-
the use of automatic feature selection methods (e.g. Minimum Re- mance, Proc. Soc. Behav. Sci. 15 (2011) 3774–3778, doi:10.1016/j.sbspro.2011.
dundancy Maximum Relevance - mRMR and Sequential Forward 04.372.
10 J. Rodríguez-Arce, L. Lara-Flores and O. Portillo-Rodríguez et al. / Computer Methods and Programs in Biomedicine 190 (2020) 105408

[8] B Egilmez, E Poyraz, W Zhou, G Memik, P Dinda, N Alshurafa, UStress: under- [28] K Plarre, A Raij, S.M Hossain, A.A Ali, M Nakajima, M Al’absi, . . . L.E Wittmers,
standing college student subjective stress using wrist-based passive sensing, Continuous inference of psychological stress from sensory measurements col-
in: 2017 IEEE International Conference On Pervasive Computing and Commu- lected in the natural environment, in: Proceedings of the 10th ACM/IEEE In-
nications Workshops, IEEE, 2017, doi:10.1109/percomw.2017.7917644. ternational Conference on Information Processing in Sensor Networks, 2011,
[9] R.M. Fish, L.A. Geddes, Conduction of electrical current to and through the hu- pp. 97–108.
man body: a review, Open Access J. Plastic Surg. 9 (2009) 1–24. [29] M.B.I Reaz, M.S Hussain, F Mohd-Yasin, Techniques of {EMG} signal analysis:
[10] R García-Ros, F Pérez-González, J Pérez-Blasco, L.A Natividad, Academic stress detection, processing, classification and applications, Biol. Proc. Online 8 (1)
in first-year college students, Revista Latinoamericana de Psicología 44 (2) (2006) 11–35, doi:10.1251/bpo115.
(2012) 143–154. [30] K.J. Reddy, K.R. Menon, A. Thattil, Academic stress and its sources among uni-
[11] A Ghaderi, J Frounchi, A Farnam, Machine learning-based signal processing us- versity students, Biomed. Pharmacol. J. 11 (1) (2018) .10.13005/bpj/1404.
ing physiological signals for stress detection, in: 2015 22nd Iranian Confer- [31] J.L. Rhudy, M.W. Meagher, Fear and anxiety: divergent effects on human pain
ence On Biomedical Engineering {ICBME}, IEEE, 2015, doi:10.1109/icbme.2015. thresholds, Pain 84 (1) (20 0 0) 65–75, doi:10.1016/s0304-3959(99)00183-9.
7404123. [32] M. Salai, I. Vassányi, I. Kósa, Stress detection using low cost heart rate sensors,
[12] G. Giannakakis, D Grigoriadis, K Giannakaki, O Simantiraki, A Roniotis, M Tsik- J. Healthcare Eng. (2016) 1–13, doi:10.1155/2016/5136705.
nakis, Review on psychological stress detection using biosignals, IEEE Trans. [33] V. Sandulescu, S. Andrews, D. Ellis, N. Bellotto, O.M. Mozos, Stress detection
Affective Comput. 1 (2019), doi:10.1109/taffc.2019.2927337. using wearable physiological sensors, in: Artificial Computation in Biology and
[13] M.A Hall, Correlation-based Feature Subset Selection for Machine Learning, Medicine, Springer International Publishing, 2015, pp. 526–532, doi:10.1007/
University of Waikato, Hamilton, New Zealand, 1998. 978- 3- 319- 18914- 7_55.
[14] J.A Healey, R.W Picard, Detecting stress during real-world driving tasks using [34] M. Sevil, I. Hajizadeh, S. Samadi, J. Feng, C. Lazaro, N. Frantz, . . . A. Cinar, Social
physiological sensors, IEEE Trans. Intell. Transp. Syst. 6 (2) (2005) 156–166, and competition stress detection with wristband physiological signals, in: 2017
doi:10.1109/tits.2005.848368. IEEE14th International Conference On Wearable and Implantable Body Sensor
[15] T Iliou, G Konstantopoulou, M Ntekouli, C Lymperopoulou, K Assimakopou- Network, IEEE, 2017, doi:10.1109/bsn.2017.7936002.
los, D Galiatsatos, G Anastassopoulos, ILIOU machine learning preprocess- [35] N Sharma, T. Gedeon, Objective measures, sensors and computational tech-
ing method for depression type prediction, Evol. Syst. (2017), doi:10.1007/ niques for stress recognition and classification: a survey, Comput. Methods
s12530- 017- 9205- 9. Programs Biomed. 108 (3) (2012) 1287–1301, doi:10.1016/j.cmpb.2012.07.003.
[16] H. Jing, Analysis on the relationship among test anxiety, self-concept and aca- [36] K. Soman, V. Alex, C Srinivas, Analysis of physiological signals in response to
demic competency, J. US-China Foreign Lang. 5 (1) (2007) 48–51. stress using ecg and respiratory signals of automobile drivers, in: 2013 Interna-
[17] E Kaniusas, E. Kaniusas (Ed.), Springer-Verlag, 2012, doi:10.1007/ tional Mutli-Conference on Automation, Computing, Communication, Control
978- 3- 642- 24843- 6 1. and Compressed Sensing, IEEE, 2013, doi:10.1109/imac4s.2013.6526476.
[18] P Karthikeyan, M Murugappan, S Yaacob, Analysis of stroop color word [37] G. Spangler, R. Pekrun, K. Kramer, H. Hofmann, Students emotions, physiolog-
test-based human stress detection using electrocardiography and heart rate ical reactions, and coping in academic exams, Anxiety Stress Coping 15 (4)
variability signals, Arab. J. Sci. Eng. 39 (3) (2013) 1835–1847, doi:10.1007/ (2002) 413–432, doi:10.1080/1061580021000056555.
s13369-013-0786-8. [38] C.D. Spielberger, F. Gonzalez-Reigosa, A. Martinez-Urrutia, L. Natalicio, D. Na-
[19] C Kirschbaum, K.-M Pirke, D.H Hellhammer, The trier social stress test a tool talicio, Development of the spanish edition of the state-trait anxiety inventory,
for investigating psychobiological stress responses in a laboratory setting, Neu- Int. J. Psychol. (1971).
ropsychobiology 28 (1–2) (1993) 76–81, doi:10.1159/0 0 01190 04. [39] R.L. Spitzer, The structured clinical interview for DSM-III-R (SCID), Arch. Gen.
[20] A Kitsantas, A Winsler, F Huie, Self-Regulation and ability predictors of aca- Psychiatry 49 (8) (1992) 624, doi:10.1001/archpsyc.1992.01820 080 0320 05.
demic success during college: a predictive validity study, J. Adv. Acad. 20 (1) [40] E. Vildjiounaite, J. Kallio, J. Mäntyjärvi, V. Kyllönen, M. Lindholm, G. Gimel’farb,
(2008) 42–68, doi:10.4219/jaa- 2008- 867. Unsupervised stress detection algorithm and experiments with real life data,
[21] J.R. Landis, G.G Koch, The measurement of observer agreement for categorical in: Progress in Artificial Intelligence, Springer International Publishing, 2017,
data, Biometrics 33 (1) (1977) 159, doi:10.2307/2529310. pp. 95–107, doi:10.1007/978- 3- 319- 65340- 2_9.
[22] C Maaoui, A Pruski, Unsupervised stress detection from remote physiological [41] P. Vitasari, T. Herawan, M.N.A. Wahab, A. Othman, S.K Sinnadurai, Explor-
signal, in: 2018 IEEE International Conference on Industrial Technology, IEEE, ing mathematics anxiety among engineering students, Proc. Soc. Behav. Sci. 8
2018, doi:10.1109/icit.2018.8352409. (2010) 482–489, doi:10.1016/j.sbspro.2010.12.066.
[23] M.L McHugh, Interrater reliability: the kappa statistic, Biochemia Medica [42] P. Vitasari, M.N.A. Wahab, A. Othman, T. Herawan, S.K. Sinnadurai, The rela-
(2012) 276–282, doi:10.11613/bm.2012.031. tionship between study anxiety and academic performance among engineer-
[24] L Moya-Albiol, A Salvador, Empleo de estresores psicológicos de laboratorio en ing students, Proc. Soc. Behav. Sci. 8 (2010) 490–497, doi:10.1016/j.sbspro.2010.
el estudio de la respuesta psicofisiológica al estrés, Anales de Psicología 17 (1) 12.067.
(2001) 69–81. [43] J Wijsman, B. Grundlehner, H. Liu, H. Hermens, J. Penders, Towards mental
[25] F.N Ogba, M.O Ede, C.N Onyishi, P.U Agu, A.B Ikechukwu-Ilomuanya, J.N Igbo, stress detection using wearable physiological sensors, in: 2011 Annual Interna-
. . . S.C Ugwoke, Effectiveness of music therapy with relaxation technique on tional Conference of the {IEEE} Engineering in Medicine and Biology Society,
stress management as measured by perceived stress scale, Medicine 98 (15) IEEE, 2011, doi:10.1109/iembs.2011.6090512.
(2019) e15107, doi:10.1097/md.0 0 0 0 0 0 0 0 0 0 015107. [44] I.H. Witten, E. Frank, M.A. Hall, Data Mining: Practical Machine Learning Tools
[26] E.U Onyeizugbo, Self-Efficacy and test anxiety as correlates of academic per- and Techniques, Morgan Kaufmann, 2011.
formance, J. Educ. Res. (2010). [45] Z. Zhu, S. Ober, R. Jafari,, Modeling and detecting student attention and interest
[27] O Peleg, Test anxiety, academic achievement, and self-esteem among Arab ado- level using wearable computers, in: 2017 IEEE14th International Conference On
lescents with and without learning disabilities, Learn. Disabil. Q. 32 (1) (2009) Wearable and Implantable Body Sensor Networks, IEEE, 2017, doi:10.1109/bsn.
11–20, doi:10.2307/25474659. 2017.7935996.

You might also like