Stress State Evaluation by an Improved Support Vector Machine
Effective methods of evaluation of the psychological pressure can detect and assess realtime stress states, warning people to pay necessary attention to their health. This study is focused on the stress assessment issue using an improved support vector machine (SVM) algorithm on the base of surfac...
Збережено в:
| Дата: | 2016 |
|---|---|
| Автори: | , , , , , |
| Формат: | Стаття |
| Мова: | English |
| Опубліковано: |
Інститут фізіології ім. О.О. Богомольця НАН України
2016
|
| Назва видання: | Нейрофизиология |
| Онлайн доступ: | http://dspace.nbuv.gov.ua/handle/123456789/148339 |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
| Назва журналу: | Digital Library of Periodicals of National Academy of Sciences of Ukraine |
| Цитувати: | Stress State Evaluation by an Improved Support Vector Machine / L. Xin, Ch. Zetao, Zh. Yunpeng, X. Jiali, W. Shuicai, Z. Yanjun // Нейрофизиология. — 2016. — Т. 48, № 2. — С. 96-102. — Бібліогр.: 15 назв. — англ. |
Репозитарії
Digital Library of Periodicals of National Academy of Sciences of Ukraine| id |
irk-123456789-148339 |
|---|---|
| record_format |
dspace |
| spelling |
irk-123456789-1483392019-02-19T01:25:25Z Stress State Evaluation by an Improved Support Vector Machine Xin, L. Zetao, Ch. Yunpeng, Zh. Jiali, X. Shuicai, W. Yanjun, Z. Effective methods of evaluation of the psychological pressure can detect and assess realtime stress states, warning people to pay necessary attention to their health. This study is focused on the stress assessment issue using an improved support vector machine (SVM) algorithm on the base of surface electromyographic signals. After the samples were clustered, the cluster results were given to the loss function of the SVM to screen training samples. With the imbalance amongst the training samples after screening, a weight was given to the loss function to reduce the prediction tendentiousness of the classifier and, therefore, to decrease the error of the training sample and make up for the influence of the unbalanced samples. This improved the algorithm, increased the classification accuracy from 73.79% to 81.38%, and reduced the running time from 1973.1 to 540.2 sec. Experimental results show that this algorithm can help to effectively avoid the influence of individual differences on a stress appraisal effect and to reduce the computational complexity during the training phase of the classifier Ефективні методи визначення ступеня психологічного тиску можуть забезпечувати виявлення та оцінку стресових станів у реальному часі, примушуючи людей приділяти необхідну увагу їх здоров’ю. Метою нашого дослідження було оцінити стан стресу з використанням покращеного методу опорних векторів (SVM), базуючись на відведенні поверхневих електроміограм. Після того, як зразки даних були кластеризовані, результати передавалися до функції розділення SVM для того, щоб представити тренувальні зразки. Після встановлення дисбалансу між тренувальними зразками після скринінга для функції розділення надавався параметр ваги для зменшення тенденційності прогнозування класифікатора і, таким чином, зменшення похибки тренувального зразка і впливу незбалансованих зразків. Це покращувало алгоритм, підвищувало точність класифікації від 73.79 до 81.38 % та зменшувало час обробки від 1973.1 до 540.2 с. Результати експериментів показали, що даний алгоритм може допомогти ефективно уникнути впливу індивідуальних відмінностей на оцінювання стресу та зменшити складність комп’ютерних розрахунків у перебігу тренувальної фази діяльності класифікатора. 2016 Article Stress State Evaluation by an Improved Support Vector Machine / L. Xin, Ch. Zetao, Zh. Yunpeng, X. Jiali, W. Shuicai, Z. Yanjun // Нейрофизиология. — 2016. — Т. 48, № 2. — С. 96-102. — Бібліогр.: 15 назв. — англ. 0028-2561 http://dspace.nbuv.gov.ua/handle/123456789/148339 616.891 en Нейрофизиология Інститут фізіології ім. О.О. Богомольця НАН України |
| institution |
Digital Library of Periodicals of National Academy of Sciences of Ukraine |
| collection |
DSpace DC |
| language |
English |
| description |
Effective methods of evaluation of the psychological pressure can detect and assess realtime stress states, warning people to pay necessary attention to their health. This study is
focused on the stress assessment issue using an improved support vector machine (SVM)
algorithm on the base of surface electromyographic signals. After the samples were clustered,
the cluster results were given to the loss function of the SVM to screen training samples. With
the imbalance amongst the training samples after screening, a weight was given to the loss
function to reduce the prediction tendentiousness of the classifier and, therefore, to decrease
the error of the training sample and make up for the influence of the unbalanced samples.
This improved the algorithm, increased the classification accuracy from 73.79% to 81.38%,
and reduced the running time from 1973.1 to 540.2 sec. Experimental results show that this
algorithm can help to effectively avoid the influence of individual differences on a stress
appraisal effect and to reduce the computational complexity during the training phase of the
classifier |
| format |
Article |
| author |
Xin, L. Zetao, Ch. Yunpeng, Zh. Jiali, X. Shuicai, W. Yanjun, Z. |
| spellingShingle |
Xin, L. Zetao, Ch. Yunpeng, Zh. Jiali, X. Shuicai, W. Yanjun, Z. Stress State Evaluation by an Improved Support Vector Machine Нейрофизиология |
| author_facet |
Xin, L. Zetao, Ch. Yunpeng, Zh. Jiali, X. Shuicai, W. Yanjun, Z. |
| author_sort |
Xin, L. |
| title |
Stress State Evaluation by an Improved Support Vector Machine |
| title_short |
Stress State Evaluation by an Improved Support Vector Machine |
| title_full |
Stress State Evaluation by an Improved Support Vector Machine |
| title_fullStr |
Stress State Evaluation by an Improved Support Vector Machine |
| title_full_unstemmed |
Stress State Evaluation by an Improved Support Vector Machine |
| title_sort |
stress state evaluation by an improved support vector machine |
| publisher |
Інститут фізіології ім. О.О. Богомольця НАН України |
| publishDate |
2016 |
| url |
http://dspace.nbuv.gov.ua/handle/123456789/148339 |
| citation_txt |
Stress State Evaluation by an Improved Support Vector Machine / L. Xin, Ch. Zetao, Zh. Yunpeng, X. Jiali, W. Shuicai, Z. Yanjun // Нейрофизиология. — 2016. — Т. 48, № 2. — С. 96-102. — Бібліогр.: 15 назв. — англ. |
| series |
Нейрофизиология |
| work_keys_str_mv |
AT xinl stressstateevaluationbyanimprovedsupportvectormachine AT zetaoch stressstateevaluationbyanimprovedsupportvectormachine AT yunpengzh stressstateevaluationbyanimprovedsupportvectormachine AT jialix stressstateevaluationbyanimprovedsupportvectormachine AT shuicaiw stressstateevaluationbyanimprovedsupportvectormachine AT yanjunz stressstateevaluationbyanimprovedsupportvectormachine |
| first_indexed |
2025-07-12T18:45:41Z |
| last_indexed |
2025-07-12T18:45:41Z |
| _version_ |
1837467909111152640 |
| fulltext |
NEUROPHYSIOLOGY / НЕЙРОФИЗИОЛОГИЯ.—2016.—T. 48, № 296
UDC 616.891
L. XIN1,2,3, Ch. ZETAO1,2, Zh. YUNPENG1,2, X. JIALI1,2, W. SHUICAI3，
and Z. YANJUN3
STRESS STATE EVALUATION BY AN IMPROVED SUPPORT VECTOR MACHINE
Received March 26, 2014
Effective methods of evaluation of the psychological pressure can detect and assess real-
time stress states, warning people to pay necessary attention to their health. This study is
focused on the stress assessment issue using an improved support vector machine (SVM)
algorithm on the base of surface electromyographic signals. After the samples were clustered,
the cluster results were given to the loss function of the SVM to screen training samples. With
the imbalance amongst the training samples after screening, a weight was given to the loss
function to reduce the prediction tendentiousness of the classifier and, therefore, to decrease
the error of the training sample and make up for the influence of the unbalanced samples.
This improved the algorithm, increased the classification accuracy from 73.79% to 81.38%,
and reduced the running time from 1973.1 to 540.2 sec. Experimental results show that this
algorithm can help to effectively avoid the influence of individual differences on a stress
appraisal effect and to reduce the computational complexity during the training phase of the
classifier.
Keywords: surface electromyographic signals, stress state evaluation, support vector
machine, clustering, weight.
1 Institute of Biomedical Engineering, Yanshan University, Qinghuangdao,
China.
2 Measurement Technology and Instrumentation Key Laboratory of the Hebei
Province, Qinghuangdao, China.
3 College of Life Science and Bio-Engineering, Beijing University of
Technology, Beijing, China.
Correspondence should be addressed to L. Xin (e-mail: yddylixin@ysu.edu.cn),
or Z. Yanjun (e-mail: yjzeng@bjut.edu.cn).
INTRODUCTION
Psychological or mental stress is a psychosomatic
tension state that tends to be manifested in various
types of psychological and physiological reactions
when a person becomes aware that the confronting
environment is important but rather difficult to deal
with [1]. Moderate stress can make subjects to produce
positive energy and may be changed into a motive
power to improve the work efficiency. However, stress
that exceeds the person’s limit may cause negative
effects or even significantly affect one’s normal life
[2]. The recognition of the cause of stress before it
becomes chronic is a key step in managing it.
Methods that can be used for stress appraisal in
the psychology usually require a significant response
of and positive cooperation with the subject. The
application of stress appraisal will be wider and
beneficial for the research on stress and health if it
can operate without self-evaluation.
In engineering, affective computing is used to
assess the intensity of stress. Stress states of drivers
were evaluated based on the speech and physiological
parameters [3, 4]. A traditional game, Tetris, could be
used as the stress source, and this allowed researchers
to collect respiratory and electromyographic (EMG)
signals in a group of 129 subjects; this study reached
an average recognition rate of above 80% by using
linear discriminant analysis and r to analyse the
results [5]. In other study [6], a stress detection
method based on physiological measurements of 22
subjects was proposed. Each subject in this study was
exposed to a protocol containing four stressors and six
rest periods. A simple wireless device was designed
[7] for detection and daily evaluate of routine stress
automatically and permanently.
In the next study [8], nine call centre employees
as examined subjects were asked to wear a skin
conductance sensor on their wrists for a week at work,
to record the stress levels for each call. Individual
differences were addressed by either modifying the
loss function of a support vector machine (SVM)
NEUROPHYSIOLOGY / НЕЙРОФИЗИОЛОГИЯ.—2016.—T. 48, № 2 97
STRESS STATE EVALUATION BY AN IMPROVED SUPPORT VECTOR MACHINE
to adapt to the varying priors or by placing more
significance on training samples from the most similar
people in terms of the skin conductance lability to
automatically recognise classes of the stressful/non-
stressful calls [8].
Physiological measurements can detect stress with
a minimal discomfort for the subjects and are useful
in reflecting emotions [9]. In our study, we selected
four stress sources to stimulate the participants. One
hundred forty-four groups of EMG signals were
collected from each of the 16 subjects. The existing
SVM algorithm was improved to solve the stress
evaluation problem with increased classification
accuracy and to reduce the classifier’s processing
time. The feasibility of the experimental scheme and
the algorithm were proven by using real data.
METHODS
Improved SVM. A kernel-based machine, SVM,
learns a family of methods used to accurately classify
both linearly separable and linearly inseparable data
[10]. The SVM has been used in numerous machine
learning fields, such as classification, regression
estimation, and kernel principal component analysis,
because it is capable of providing a good generalisation
capacity.
Contrast ive analysis showed that different
participants have different perception degrees of
the same stimulus, and that the same participant
often produces different responses to different stress
sources at the same time. Actual verification showed
that solving the diversity problem between samples is
difficult even after normalisation. Making predictions
for a single participant causes a certain amount of
interference and affects on the classifier performance
if all sample information is included in the training
set. Furthermore, too many training samples increase
the computational complexity and influence the
classification accuracy. Therefore, making a targeted
selection for the training set is obviously necessary.
Filtration of the training sets results in an imbalance
between positive and negative samples. Thus, different
weights are given to positive and negative samples,
and sample information is added to the loss function
to weaken this imbalance.
Optimised Algorithm. The overall flow chart of
the algorithm is shown in Fig. 1.The improved SVM
algorithm can reduce the amount of training samples
by selecting training sets and addressing stress
evaluation. Thus, only the most beneficial message for
the classifier’s model building is put in. Otherwise,
the adjustment of the weight of positive and negative
samples can amend the imbalance in training samples.
Therefore, this algorithm can reduce the computation
complexity at the training stage and, in a parallel
manner, to improve the classification accuracy.
Selection of the Training Samples. This algorithm
aimed at seeking the correlation degree between each
training sample and the testing sample by clustering.
The correlation factor that describes the correlation
degree to the loss function of the SVM’s is then
provided to combine the classifier with the practical
issue and make it different from the classifier that
fits all of the common classification problems.
The algorithm specifically improves the evaluation
accuracy.
The standard expression of the SVM is
,
s.t. and εi ≥ i = 1,2, … n (1),
where C is the misclassification cost, and … is the
slack variable for the sample….
The SVM’s loss function can be expressed as
loss function = . (2).
The improved algorithm’s loss function expression
is the following:
loss function = ) (3),
where n is the number of training samples, and νi de-
fines the similarity between sample i and the testing
sample for classification, to solve the sample differ-
ences problem.
Each group of the data centre section of 10 sec is
truncated, and K = 2 is set to divide all samples into
two categories based on the k-means clustering algo-
rithm, and νi = 1 is set when the training sample be-
longs to the same category as the testing one. Other-
wise, a weak correlation is considered to exist between
the training and testing samples, and νi = 0 is set.
The leave-one-out method is used for classification
evaluation.
Classification Weight . An imbalance occurs
between the almost equal original positive and negative
samples are used after making selection to the training
samples with the improved SVM algorithm. This
NEUROPHYSIOLOGY / НЕЙРОФИЗИОЛОГИЯ.—2016.—T. 48, № 298
L. XIN, Ch. ZETAO, Zh. YUNPENG, et al.
F i g. 1. Flow chart of an improved SVM algorithm.
Р и с. 1. Блок-схема вдосконаленого алгоритму SVM.
imbalance leads to the error sum of the positive class
smaller than that of the negative class. An example of
this imbalance is when the amount of positive samples
is significantly smaller than that of negative samples.
Thus, a larger penalty weight exists with respect to the
negative class, causing the separation plane to move
toward the positive class.
The weights of the positive are set to δ_,and
negative samples are set to δ_, to solve the data
imbalance problem. This problem is addressed by
changing the classified weight in the loss function.
The loss function of the improved SVM is expressed
as
loss function = (4).
Let be the training set, where Xi
represents the feature vector of the sample i, and yi is
the class label, where yi = {–1, 1}. Let the class priors
of this set be
Two ways have been tried to improve the classified
weight. One way is to set ; then, the
probabilities that the classifier groups divide the
samples into positive and negative classes tend to be
similar to each other.
2
Figure 1 - Flow chart of improved SVM algorithm
1v i0v i
)(C oss
-N
-1}{y
-
N
1}{y
j
j
i
in
FunctionL
O
NEUROPHYSIOLOGY / НЕЙРОФИЗИОЛОГИЯ.—2016.—T. 48, № 2 99
STRESS STATE EVALUATION BY AN IMPROVED SUPPORT VECTOR MACHINE
The other way is to optimise the weights.
The error sums of the positive and negative classes
are better to be equal to the balance of positive and
negative errors [11], as follows:
(5).
The from the formula is uncertain; thus,
determining the exact relationship between δ+ and δ–
is difficult. However, an approximation relationship
can be obtained by assuming that the mathematical
expectations of the positive and negative class errors
are the same:
N+ δ+
2 = N– δ–
2 (6),
where N+ is the number of positive samples, and N- is
the number of negative samples; δ– is set to be 1 for
calculation convenience; thus, δ+ tends to be .
A more compromise way adopted is to avoid an over
adjustment. Optimisation for δ+ is made between 1 and
by consulting an SVM (c, g) optimisation
approach, which is a method to optimize the SVM
parameters.
Experiment Design. Subjects. The selected subjects
were 16 students (8 men and 8 women), students and
postgraduate students of the Yanshan University. All of
the subjects were healthy and right-handed. The EMG
signals under conditions of stress stimulation were
collected using the method described below.
Questionnaire. Before data collection, the subjects
were asked to fill in a questionnaire, to evaluate
their recent psychological states. Distributing of the
questionnaires is a convenient method for identification
of the one’s real situation and evaluation of the effect
of the experiment. If the recent physiological state
was out of the requirements, we should change the
volunteer. The subjects were also asked to fill another
questionnaire PSTRI (Psychosomatic Tension/related
Inventory) after the data collection, to evaluate
their present psychological state and, therefore, to
test whether the experiment raised the stress level
of the subjects. The results of application of these
questionnaires were treated as references to the stress
level. The questionnaires were used for data analysis.
Data col lect ion. An MP150 mult i -channel
physiological recorder (Biopac Company, ………..)
was used to record surface EMG (sEMG) signals from
the subjects (Fig. 2). The experiment was conducted
within four consecutive days.
Day 1. Studies showed that office noises even of
a low intensity cause emotional stress to the workers
[12, 13]. To obtain a better effect of stimulation, five
of the top 10 intranquil voices from a survey result
(Prof. T. Cox, Salford University, Great Britain) were
chosen to be the experimental environment of stress
stimulation. The subjects were asked to remember
a large group of numbers within a limited time and
against a noisy background. The stimulation mode is
shown in Table 1.
T a b l e 1. Experimental process of number memorisation
Т а б л и ц я 1. Експериментальний процес
запам’ятовування чисел
Stimulus Duration Function
Light music +
scenery pictures
2 min Place subjects in a state
of relaxation
Slide (first time) 50 sec Place subjects in a state
of stress
Slide(second time) 35 sec Place subjects in a state
of stress
Slide(third time) 15 sec Place subjects in a state
of stress
Light music +
scenery pictures
1 min Place subjects in a state
of relaxation
Day 2. College students were facing a growing
pressure of employment with the growing numbers
of undergraduates and the increasingly fierce social
competition. A video describing the employment
outlook of university students was used as the stimulus
to arouse potential stress among the participants. The
process is presented in detail in Table 2.
F i g. 2. EMG measurements with an MP150 multi-channel physio-
logical recorder.
Р и с. 2. Вимірювання електроміограм із використанням багато-
канального реєстратора фізіологічних даних MP150.
NEUROPHYSIOLOGY / НЕЙРОФИЗИОЛОГИЯ.—2016.—T. 48, № 2100
L. XIN, Ch. ZETAO, Zh. YUNPENG, et al.
T a b l e 2. Experimental process of video stimulation
Т а б л и ц я 2. Експериментальний процес відеостимуляції
Video material Duration Function
Light music +
scenery pictures
2 min Place subjects in a state of
relaxation
Employment
outlook video
15 min Place subjects in a state of
stress
Light music +
scenery pictures
1 min Place subjects in a state of
relaxation
Day 3. A word memorisation task was conducted as
the experiment. Participants were asked to remember
a group of relatively difficult English words within
a 3-min-long interval. The sound of a stopwatch was
played within the last minute, to give the participants a
sense of urgency. At the same time, all the words were
designed in negative meanings, to produce a certain
psychological hint and place the subjects in a negative
emotional state, producing an environment conducive
to stress [14]. The detailed processes are shown in
Table 3.
T a b l e 3. Experimental process of word memorisation
Т а б л и ц я 3. Експериментальний процес
запам’ятовування слів
Stimulus Duration Function
Light music +
scenery pictures
2 min Place subjects in a state of
relaxation
Word
memorisation
3 min Place subjects in a state of
stress
Light music +
scenery pictures
1 min Place subjects in a state of
relaxation
Day 4. Adopting the stress stimulation method
applied in the Augsburg University (Germany), we
used the traditional game Tetris as the stress source.
The game difficulty (falling speed of the blocks) was
increased during the experiment to arouse higher stress
among the subjects. The participants were given only 5
min to relax before the game. The data of the subjects
under the highest speed were collected to determine
the stress level.
RESULTS
Data Pre-Processing. One hundred forty-four groups
of the data, including 72 groups of the stress data
and 72 groups of the non-stress data, were obtained
after the processing. Finally, the data were saved and
labelled.
Surface EMG is characterized by weak signals,
strong noises within a low-frequency range, and strong
randomicity due of the influence of numerous factors.
Thus, the first step was to denoise the collected sig-
nals. The db9 wavelet was used to conduct wavelet de-
composition of the noisy sEMG. Finally, 28 statistical
characteristics were obtained by making feature ex-
traction from the noise-cancelled sEMG signal.
Stress Evaluation . The improved algorithm
showed a better classification performance, compared
with the ordinary (c, g) optimisation SVM and (c, g)
optimisation using the genetic algorithm SVM [15].
The 10-sec-long data were intercepted to determine
the effect of the data length on the classification
accuracy. Specific results are reported in Tables 4
and 5.
T a b l e 4. Truncated 20-sec data
Т а б л и ц я 4. Скорочені 20-секундні дані
Algorithm Accuracy of the training sets Accuracy of the testing sets Time
Optimised (c, g) by grid search 97.29% 70.34% 895.3 sec
SVM based on genetic algorithm
optimisation 98.27% 68.28% 2 0 2 6 . 5
sec
Improved SVM by screening training samples 98.09% 77.93% 569.3 sec
Improved SVM by screening training samples and
searching the SVM weight 98.03% 79.31% 541.3 sec
NEUROPHYSIOLOGY / НЕЙРОФИЗИОЛОГИЯ.—2016.—T. 48, № 2 101
STRESS STATE EVALUATION BY AN IMPROVED SUPPORT VECTOR MACHINE
DISCUSSION
Experimental results showed that this program can
arouse the subjects’ stress state to a certain degree.
Changes in the stress state were reflected in sEMG
signals. Different stressors were selected to stimulate
the participants within four consecutive days. Thus,
the “adaptability” problem was avoided, and the
feasibility of the laboratory-induced stress project
was improved. Subsequently, the collected data can be
used to evaluate and analyse the stress state.
According to the research findings in [7], the
SVM model was optimised directly at the individual
difference problem. The testing samples included the
skin conductance data of all subjects in one day or the
skin conductance data of one subject for four days.
The rest of the samples were included in the training
set for the prediction. The imbalance between samples
was amended by making S+ = P–+/ and S– = P––/, where
and are the corresponding weights of the two groups
of samples, and are the class priors of the different
samples. This algorithm joined the information of the
testing set into the building of SVM model whereas
it is impossible to know the proportion of stressful/
non-stressful events for the subjects in the actual
stress evaluation work. Obviously, this approach has
a limitation because the values of and are unknown.
For this limitation, two approaches were proposed
to modify the class weights that only relied on the
known sample information rather than introduce the
testing set details. The classification model based
on this theory was built. This model can be used for
stress evaluation in the real environment. In addition,
only one sample was selected for the test each time,
and the samples associated with it were selected for
training. The cycle was repeated until every sample
was assessed as the testing set. That is a classification
model for each sample that conformed to its own
characteristics was built for each cycle. Using the
real leave-one-out method to validate the theoretical
results was more aligned with the research purposes.
The SVM optimisation based on the genetic
algorithm had the highest training accuracy but the
lowest testing accuracy, which indicated that the
classifier came from an overfitting problem.
The classification time showed that the SVM
optimisation based on the genetic algorithm expended
the amount of time and memory overhead. The SVM
optimisation based on normal (c, g) searching required
a long training time. By contrast, the improved
algorithm greatly reduced the computation complexity
in the training stage and, therefore, shortened the
classifier’s training time. This advantage is even more
significant when the training sample size is large.
Comparative analysis showed that selection of the
improved algorithm for the training samples avoided
the interference of the redundant information and
reduced the classifier’s computation complexity in
the training phase, thereby improving the overall
classification accuracy, while decreasing the training
time. The classified accuracy was generally improved
when the truncated data length was equal to 10 sec.
Our study aimed to solve the stress evaluation
problem using sEMG as the study object. The stress-
inducing stimulation, analysis, and evaluation system
with the problem of individual differences were
established. This study proposed an improved SVM
classification assessment algorithm by classifying after
clustering, to solve the individual difference problem
during evaluation of stress/non-stress reactions
automatically and giving the clustering results to SVM
loss function, to improve the quality of classification
evaluation.
Experimental results showed that the accuracy of
stress evaluation classification based on SVM was
68%, which was noticeably lower than the accuracy
of the improved algorithm (79%). Thus, the improved
algorithm showed higher classification accuracy
T a b l e 5. Truncated 10-sec data
Т а б л и ц я 5. Скорочені 10-секундні дані
Algorithm Accuracy of the training sets Accuracy of the testing sets Time
Optimised (c, g) by grid search 99.76% 73.79% 702.4 sec
SVM based on genetic algorithm
optimisation 99.12% 77.93% 1973.1 sec
Improved SVM by screening training samples 98.65% 79.31% 545.2 sec
Improved SVM by screening training samples
and searching the SVM weight 98.92% 81.38% 540.2 sec
NEUROPHYSIOLOGY / НЕЙРОФИЗИОЛОГИЯ.—2016.—T. 48, № 2102
L. XIN, Ch. ZETAO, Zh. YUNPENG, et al.
for stress evaluation with individual differences.
The improved algorithm could decrease the stress
evaluation time by selecting training samples. The
classification time was decreased from 2026.5 to
541.3 sec after the improvement. The improved
SVM algorithm can address the effects of individual
differences during the stress state assessment.
Acknowledgement. This work was supported by the funding
project for outstanding experts go abroad in the Hebei province
and the key project in department of education of the Hebei
province.
All participants were informed about the experimental
process and subsequently signed and provided the informed
consent. The ethical protocol of this study was based on the
Declaration of Helsinki and existing international ethical
standards.
The authors, L. Xin, Ch. Zetao, Zh.Yunpeng, X. Jiali,
W. Shuicai,and Z. Yanjun, confirm that there were no conflicts
of any kind relating to commercial or financial relations,
relations with organizations or persons, which could in any way
be associated with the investigation, and with the relationship
of the co-authors of the article.
Л. Ксін1,2,3, Ч. Зетао1,2, Ж. Юнпен1,2, Кс. Джіалі1,2,
В. Шуїкай3, З. Янчжун3
ОЦІНКА СТРЕСОВОГО СТАНУ ЗА ДОПОМОГОЮ
ПОКРАЩЕНОГО МЕТОДУ ОПОРНИХ ВЕКТОРІВ
1 Інститут біомедичного інженірінгу Університету
Яншань, Циньхуанда́о (Китай).
2 Ключова лабораторія технології вимірювань та
інструментів провінції Хебей, Циньхуандáо (Китай).
3 Коледж наук про життя та біоінженірінгу Пекінського
технологічного університету (Китай).
Р е з ю м е
Ефективні методи визначення ступеня психологічного
тиску можуть забезпечувати виявлення та оцінку стресових
станів у реальному часі, примушуючи людей приділяти
необхідну увагу їх здоров’ю. Метою нашого дослідження
було оцінити стан стресу з використанням покращеного
методу опорних векторів (SVM), базуючись на відведен-
ні поверхневих електроміограм. Після того, як зразки да-
них були кластеризовані, результати передавалися до функ-
ції розділення SVM для того, щоб представити тренувальні
зразки. Після встановлення дисбалансу між тренувальними
зразками після скринінга для функції розділення надавався
параметр ваги для зменшення тенденційності прогнозування
класифікатора і, таким чином, зменшення похибки
тренувального зразка і впливу незбалансованих зразків. Це
покращувало алгоритм, підвищувало точність класифікації
від 73.79 до 81.38 % та зменшувало час обробки від 1973.1
до 540.2 с. Результати експериментів показали, що даний
алгоритм може допомогти ефективно уникнути впливу
індивідуальних відмінностей на оцінювання стресу та
зменшити складність комп’ютерних розрахунків у перебігу
тренувальної фази діяльності класифікатора.
REFERENCES
1. S. J. Lupien, F. Maheu, M. Tu, et al., “The effects of stress
and stress hormones on human cognition: Implications for the
field of brain and cognition,” Brain Cognit., 65, No. 3, 209-
237 (2007).
2. D. Giakoumis, D. Tzovaras, and G. Hassapis, “Subject-
dependent biosignal features for increased accuracy in
psychological stress detection,” Human comp. studies, 71,
No. 4, 425-439 (2013).
3. R. Fernandez and R. W. Picard, “Modeling drivers’ speech
under stress,” Speech Commun., 40,145-159 (2003).
4. J. A. Healey and R. W. Picard, “Detecting stress during
real-world driving tasks using physiological sensors,” IEEE
Transact. Intel. Transport. Syst., 6, No. 2, 156-166 (2005).
5. W. Johannes, Augsburg Biosignal Toolbox (AuBT). User Guide.
[http://emotion-research.net/].
6. G. Tartarisco, G. Baldus, D. Corda, et al., “Personal health
system architecture for stress monitoring and support to
clinical decisions,” Comp. Commun., 35, No. 11, 1296-1305
(2012).
7. J. Hernandez, R. R. Morris, and R. W. Picard, “Call center
stress recognition with person-specific models,” Affect.
Comput. Intel. Interact., Part 1, 125-134 (2011).
8. N. Sharma and T. Gedeon, “Objective measures, sensors
and computational techniques for stress recognition and
classification: a survey,” Comput. Methods Progr. Biomed.,
108, No. 3, 1287-1301 (2012).
9. S. Cheng, Q. Mo, and H. He, “Effects of office noise on staff
health,” J. Envir. Health, 17, 154-155 (2000).
10. H. Jahncke, S. Hygge, N. Halin, et al., “Open-plan office noise:
cognitive performance and restoration,” J. Envir. Psychol., 31,
No. 4, 373-382 (2011).
11. http://www. sound101.com/.
12. A. Hogh, A. M. Hansen, E. G. Mikkelsen, and R. Persson,
“Exposure to negative acts at work, psychological stress
reactions and physiological stress response,” J. Psychosom.
Res., 73, No. 1, 47-52 (2012).
13. V. N. Vapnik, The Nature of Statistical Learning Theory.
Second Ed., Spinger Verlag, NewYork (2000).
14. O. L. Mangasarian and D. R. Musicant, “Successive
overrelaxation for support vector machines,” IEEE Trans.
Neural Netw., 10, No. 5, 1032-1037 (1999).
15. X. Li, Y. Wang, J. Song, and J. Shan, “Research on
classification method of combining support vector machine and
genetic algorithm for motor imagery EEG,” J. Computation.
Inform. Syst., 17, No. 12, 4351-4358 (2011).
|