WORKPLACE STRESS ESTIMATION METHOD BASED ON

MULTIVARIATE ANALYSIS OF PHYSIOLOGICAL INDICES

Hirohito Ide, Guillaume Lopez, Masaki Shuzo, Shunji Mitsuyoshi, Jean-Jacques Delaunay

and Ichiro Yamada

The University of Tokyo, School of Engineering, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan

Keywords:

Stress monitoring, Wearable sensors, Multivariate analysis, Virtual healthcare.

Abstract:

In this research, we have been developing a new integrated analysis method of multiple physiological signals

to estimate stress in daily life, which is important in depression screening and life-style related diseases pre-

vention. Experiments have been carried out on 100 participants, measuring electrocardiogram, pulse wave,

breath rhythm, and skin temperature in four patterns of psychological states; relax state, normal stress state,

monotonous stress state, and nervous state. The newly developed stress state estimation method relies on the

integrated analysis of nine physiological indices related to stress that have been extracted from the four mea-

sured physiological signals. Because variation range of each index is different between individuals and types

of stress, we divided estimation process into three steps. For each step, we performed cross-validation using

various classiﬁcation schemes to select the most relevant set of indices that enable estimation of stress state

with few inﬂuences of individual variations. Through this method we could achieve 87% accuracy for stress

detection, and 63% accuracy for stress type classiﬁcation. Finally a validation study was performed to conﬁrm

this method can be an effective solution to estimate various types of stress state regardless of individuals.

1 INTRODUCTION

Nowadays, most developed countries are facing a

serious problem with the increasing number of dis-

eases caused by excessive stress, not only mental dis-

order diseases (depression, etc.), but also lifestyle-

related diseases (hypertension, metabolic syndrome,

etc.). Indeed, when we are subject to excessive stress,

we tend to overeating, drinking alcohol, smoking and

such lifestyle-related disease risk factors. In this in-

troduction, we deﬁne our ﬁeld of study, describe the

existing approaches and their issues, and present our

approach to address them in the work reported in this

paper.

1.1 Background and Deﬁnition of Stress

Current stress detection methods, when not an after-

wards conclusion, rely on inquiry sheets or interviews

with a medical specialist. Though, because stress is

so pervasive in our social activities, there is an inher-

ent need to be able to monitor stress continuously in

daily life, in a seamless way, and over extended peri-

ods. It is important to propose such new system for

personal continuous stress monitoring, which would

enable prevention of serious stress-related health dis-

orders, through a seamless and regular screening of

stressful experiences an individual is exposed during

his daily life activities. Such system would beneﬁt

both individuals by providing regular feedback about

their stress, as well as physicians by supporting pa-

tient status monitoring and evaluation with objective

and in context information.

We hear a lot about stress, but what is it? Taber’s

Cyclopedic Medical Dictionary deﬁnes stress as ”the

result produced when a structure, system or organism

is acted upon by forces that disrupt equilibrium or

produce strain”. In simpler terms, we will consider

stress as the result of any emotional, physical, so-

cial, economic, or other factors that require a response

or change. It is generally believed that some stress

is okay (sometimes referred to as ”normal stress” or

”positive stress”), but when it occurs in amounts that

cannot be handled, both mental and physical changes

may occur.

In our study, we focus on ”Workplace stress,”

which we deﬁne as the physiological responses that

can happen when there is a conﬂict between job de-

mands on the person and the amount of control this

person has over meeting these demands. In gen-

Ide H., Lopez G., Shuzo M., Mitsuyoshi S., Delaunay J. and Yamada I..

WORKPLACE STRESS ESTIMATION METHOD BASED ON MULTIVARIATE ANALYSIS OF PHYSIOLOGICAL INDICES.

DOI: 10.5220/0003769400530060

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2012), pages 53-60

ISBN: 978-989-8425-88-1

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

eral, the combination of high demands in a job and

a low amount of control over the situation can lead

to stress. Regarding the amount of oxygen consump-

tion, work strongly stimulates heart activity (Carroll

et al., 2009), and both qualitative and quantitative

augmentations are reported to have a strong corre-

lation with dissatisfaction, decline in self-evaluation,

which are risk factors of mental disorders (Araki and

Kawakami, 1993). That is why on both mental dis-

order diseases and lifestyle-related diseases point of

views, there is a considerably strong need in work-

place stress monitoring.

As stated by the Canadian Mental Health Associ-

ation (http://www.cmha.ca/), stress in the workplace

can have many origins: Fear of job redundancy, pres-

sure to perform, increased demands for overtime due

to staff cutbacks, layoffs due to an uncertain econ-

omy act as negative stressors. Though, among these

origins, we chose to remove economical and non-

work-related stressors, to focus on the following three

workplace stress categories as partly proposed by Shi-

mono et al. (Shimono et al., 1998).

• Monotonous stress: stress accompanied by a te-

dious feeling when repeating a work with little

content changes for a long continuous time (job

redundancy, frequent overtimes, etc.).

• Nervous stress: stress accompanied by a feeling

of tension when performing a work that cannot af-

ford any miss (pressure to perform, speech, meet-

ing with hierarchical superiors, etc.).

• Normal stress: stress accompanied by any feeling

different from above described, when performing

a basic work (in other words basic work that does

not generate extra stress).

1.2 Current Technological Solutions

and their Issues

Traditionally, personal medical monitors have been

used only to perform data acquisition. Typical ex-

amples are holter monitors that are routinely used for

electrocardiogram (ECG) and electroencephalogram

(EEG) monitoring. Recently, with the miniaturization

and improved performances of micro-sensors, wear-

able computing, and wireless communication tech-

nologies (Fukuda et al., 2001; Itao, 2007), a new

generation of wearable intelligent sensors have been

developed (Jovanov et al., 2000). Such devices can

signiﬁcantly decrease the number of hospitalizations

and nursing visits (Heidenreich et al., 1999) by act-

ing as a personal quotationvirtual health adviser that

can warn the user of a medical emergency or contact

a specialized medical response service. A wearable

health-monitoring device using a personal area net-

work (PAN) or BAN can be integrated into a user’s

clothing (Park and Jayaraman, 2003), though such

system organization is unsuitable for lengthy, contin-

uous monitoring, particularly during normal activity.

We can classify prior research related to stress

study using wearable physiological sensing into the

following three categories.

1. Studies that demonstrate the causal relationship

between stress and changes in physiological in-

dices (Kim et al., 2008; Ohsuga et al., 2001; Schu-

bert et al., 2009)

2. Studies that evaluate qualitatively and/or quanti-

tatively the stress issued by an external stimulus

(Kotlyar et al., 2008; Watanabe et al., 2008)

3. Studies that estimate the occurrence or not of

stress based on the observation of changes in

physiological indices (Aasa et al., 2006; Fukuda

et al., 2001; Itao et al., 2008)

Aiming at stress monitoring during daily life ac-

tivities, our research corresponds to the third cate-

gory. This category is composed of two groups of

methodologies, being methodologies to retrieve stress

changes based on the observation of long-term evolu-

tion for a single physiological index, and methodolo-

gies that build models for stress status output from

input of physiological indices, based on multivariate

analysis. We consider these groups of methodologies

has four big issues that need to be addressed.

1. As physiological indices are strongly inﬂuenced

by individual differences, their values on stress

occurrence are different depending on each in-

dividual (Miyake, 1997). Therefore, it is neces-

sary to pick-up physiological indices that are less

prone to individual differences when estimating

stress.

2. Depending on the type of stress (in other words

the type of emotion), reacting physiological in-

dices are different, so that it is difﬁcult to esti-

mate stress in detail from a single physiological

index (Miyake, 2001; Ohsuga et al., 2001; Shi-

mono et al., 1998).

3. Stress status output models application is often

limited to only one speciﬁc individual, and cannot

output stress status correctly for a different person

(Ohsuga et al., 2001; Soda and Narumi, 2007).

4. Models are often limited to an output of having

or not stress, and do not estimate stress status in

details (i.e. stress type) (Shin et al., 1998; ?).

According to above statements, the study we

present here aims at addressing the described issues,

HEALTHINF 2012 - International Conference on Health Informatics

by establishing a detailed and high-generality stress

estimation method, in other words a stress type esti-

mation method using physiological indices less prone

to individual differences.

The remainder of the article is structured as fol-

lows. Section 2 described the physiological sig-

nals and indices we selected as an input to our pro-

posed multivariate analysis method and the experi-

mental set-up used to measure these signals. Sec-

tion 3 presents the experimental procedure and the

pre-processing of collected data we executed to build-

up a reliable database of physiological indices corre-

sponding to targeted three types of workplace stress.

Section 4 describes our proposed method, evaluates

its efﬁciency using experimental data, and validates

its reliability regarding the database. Finally, Section

5 sum-up our ﬁndings, raises remaining issues, and

opens a short view on future implementations we plan

to pursue.

2 PHYSIOLOGICAL

INFORMATION USEFUL TO

STRESS ESTIMATION

For monitoring stress, we focus on autonomic nervous

system activity, though we don’t use EEG due to its

difﬁcult processing that makes it difﬁcult to use for a

real-time stress monitoring solution.

It is known that the autonomic nervous system in-

ﬂuences the activity of the heart, the breath, the lung,

and the skin activities. If there is any change on the

autonomic nervous system due to stress, it should be

detectable through the activity of these physiological

elements. Typical studies of the autonomic nervous

system activity monitoring consist in ECG’s heart

beats R peak time interval variations (RRV: R-R inter-

val variations) frequencies analysis, in which strength

of low frequencies zone (LF: 0.04-0.15Hz) reﬂects

sympathetic nerve’s activity, and strength of high

frequencies zone (HF: 0.15-0.4Hz) reﬂects parasym-

pathetic nerve’s activity. Then, LF/HF power ra-

tio is an indicator of activity dominant nervous sys-

tem (large: sympathetic nerve is dominant, small:

parasympathetic nerve is dominant).

Though these studies reported that RRV spectral

analysis was effective to evaluate the physical and

mental loads by quantifying respectively the activity

level of sympathetic and parasympathetic nerves (Ak-

selrod et al., 1981; Itao et al., 2008), this index is

known to be different according to the age, sex and

the individual variation (Miyake, 2001). The physi-

ological indices should meet the following two con-

ditions: the ﬁrst is that they can reﬂect the categories

of stress, and the second is that individual differences

are not large. In this study, we decided to measure

simultaneously electrocardiogram (ECG), pulse wave

by photoplethysmography(PPG), breath, and temper-

ature of ﬁnger’s skin.

From these four physiological signals, we extract

the following nine physiological indices, which we

adopted as the basic information for stress type es-

timation.

• From ECG: HR (Heart rate), RRV, LF/HF

• From PPG: t

PAT

(pulse arrival time)

• From breath: f

(respiratory central frequency),

| f

-f

| (absolute value of difference between

and peak frequency), t

(breath time), stdt

(derivation of breath time)

• From ﬁnger’s skin temperature: T

(average tem-

perature of the ﬁnger’s skin)

To collect above selected physiological signals,

we used multi-channels biological ampliﬁer (Poly-

mate, Digitex lab. Co. ltd.) that is basically composed

of an electrocardiograph (ECG) and an ear clip type

photoplethysmograph (PPG), but to which optional

sensors such as belt-type breath sensor, and temper-

ature sensor needed for our study can be connected.

3 STRESS CORRELATED

PHYSIOLOGICAL INDICES

DATABASE BUILD-UP

For predicting stress by using psycho-physiological

indices, it is necessary to build a database based on

these indices at the situation when people faced tar-

geted different types of stressor and effectively get the

expected stress reaction. The following paragraphs

will present the experimental procedure we deﬁned to

stimulate efﬁciently the three types of stress reactions

targeted, and how we extract a sufﬁcient number of

high quality data sets among the whole measurement.

3.1 Experimental Procedure for

Physiological Indices Data-sets

Measurements

In this study, we focus on normal stress, monotonous

stress, and nervous stress, the most usual stresses that

may occur at a workplace. To collect data that will

populate our database of a person under these differ-

ent types of stress we used the Paced Auditory Serial

WORKPLACE STRESS ESTIMATION METHOD BASED ON MULTIVARIATE ANALYSIS OF PHYSIOLOGICAL

INDICES

Addition Test (PASAT, see Fig. 1), which has an ac-

knowledged authority among scientiﬁc community as

to having high reproducibility (Al’Absi et al., 2005;

?; Willemsen et al., 1998). Based on former research

work about monotonous stress (Yamada and Miyake,

2007) and nervous stress (Al’Absi et al., 1997), we

deﬁned the following three types of PASAT tasks.

• PASAT1: 5 minutes PASAT task to stimulate a

normal stress reaction

• PASAT2: 60 minutes PASAT task to stimulate a

monotonous stress reaction

• PASAT3: 5 minutes PASAT task combined with

reward cutting on miss, to stimulate a nervous

stress reaction

Figure 1: Scheme presenting the principle of PC based

PASAT. PASAT task consists in adding consecutive single

digit numbers presented by voice continuously.

We have built a dedicated black room to avoid any

environmental light and noise disturbance, and a PC

interface to execute PASAT and answer to it in a sim-

ple and quick way. The experimental protocol ﬂow

is described on Figure 2. We have performed above

described experimental for each PASAT task with re-

spectively 48, 18, and 46 participants, whose ages

were from 15 to 47 years old.

However, to be sure that before performing the as-

signed PASAT task each participant was not already

in a psychological status that may inﬂuence expected

stress reaction, due to the lack of sleep, overwork,

alcohol and such, subjective assessments were con-

ducted before and after PASAT task. Subjective as-

sessment is done on dedicated PC interface composed

of a list of short questions to which the participant

can answer using a mouse to set the Visual Ana-

log Scale (VAS) for each question corresponding to

scores from 0% to 100%. To evaluate the effective

reaction to each speciﬁc stress type stimulation, we

used the scores of related questions, among which the

average score of questions about feeling exhausted,

ineffective, and depressed, for monotonousstress, and

questions about feeling strained and palpitating for

nervous stress.

3.2 Data Quality Evaluation Results

To verify if each expected stress type effectively oc-

curred or not, we performed a test to check the sig-

㪠㫅㫊㫋㫉㫌㪺㫋㫀㫆㫅

㪧㫉㪸㪺㫋㫀㪺㪼

㫆㪽㩷㪧㪘㪪㪘㪫

㪩㪼㫃㪸㫏

㪪㫌㪹㫁㪼㪺㫋㫀㫍㪼

㪸㫊㫊㪼㫊㫊㫄㪼㫅㫋

㪧㪘㪪㪘㪫

㪪㫌㪹㫁㪼㪺㫋㫀㫍㪼

㪸㫊㫊㪼㫊㫊㫄㪼㫅㫋

㪈㫄㫀㫅㪌㫄㫀㫅

㪧㪘㪪㪘㪫㪈㪃㩷㪊䋺㪌㫄㫀㫅

㩷㩷㩷㪧㪘㪪㪘㪫㪉䋺㪍㪇㫄㫀㫅䇭

㩿㪠㪽㩷㪸㩷㫇㪸㫉㫋㫀㪺㫀㫇㪸㫅㫋㩷㪾㫀㫍㪼㫊㩷㫌㫇㩷㪧㪘㪪㪘㪫㪃

㪿㪼㩷㪻㫆㪼㫊㩷㫊㫌㪹㫁㪼㪺㫋㫀㫍㪼㩷㪸㫊㫊㪼㫊㫊㫄㪼㫅㫋㩷㫈㫌㫀㪺㫂㫃㫐㪅㪀

㫋㫀㫄㪼

㪜㫏㫇㪼㫉㫀㫄㪼㫅㫋

㫊㫋㪸㫉㫋

㪜㫏㫇㪼㫉㫀㫄㪼㫅㫋

㪼㫅㪻

Figure 2: Experimental procedure ﬂow. A ﬁveminutes-long

relaxing time was set before each test to be able to evaluate

physiological indices in a relax situation.

niﬁcant difference (t-test) among the different PASAT

tasks, using the subjective stress type score variation

between subjective assessment before and after the

PASAT task corresponding to this stress type. Fig-

ure 3 shows the results signiﬁcant differenceexamina-

tion among the three PASAT tasks. Considering only

the group of participants with more than 60% of ac-

curacy rate in PASAT1 and PASAT3 tasks (PASAT1:

19, PASAT2: 18, PASAT3: 21), we could verify high

signiﬁcant differences. From these results, we can

consider that expected stress reaction has effectively

occurred for experimental participants, respectively

corresponding to the stress type stimulated by each

PASAT task.

1 32

PASAT

100

**** **

** : 1% significance

Change of

score of high-tension [-]

Change of

score of boredom [-]

Figure 3: Subjective assessment scores in each PASAT.

Compared to PASAT1, “boredom” score (monotonous

stress) is higher with PASAT2, while “high-tension” score

(nervous stress) is higher with PASAT3.

4 A NEW METHOD FOR

WORKPLACE STRESS TYPE

ESTIMATION

According to the purpose of our study described in in-

troduction, we propose a new method for stress type

estimation applicable with a high-generality. After

describing the points of the method, we will present

the results of its efﬁciency evaluation using the exper-

imental data, and validate its reliability regarding the

physiological indices database.

4.1 Estimation Method

The physiological indices that we should use for a

good estimation should meet the following two con-

HEALTHINF 2012 - International Conference on Health Informatics

ditions: the ﬁrst is that they are able to reﬂect work-

place stress types, and the second is that they are

less prone to individual differences. To specify the

physiological indices set that achieve these condi-

tions, we performed a cross-validation with the avail-

able data-sets, and identiﬁed the combination of phys-

iological that result in the highest estimation accu-

racy. Here we used the leave-one-out cross-validation

(LOOCV) since we don’t need to tackle actually com-

putation performance issues and we still have a rela-

tively small number of samples in the data-set. In our

case, LOOCV involves using a single person data as

sample data for validation, and the remaining persons’

data as the sample data for training, which is repeated

such that each person’s data id used once as the vali-

dation data sample.

Accordingly, we have deﬁned an original multi-

steps logic as shown on Figure 4, to perform

“individual-independent” stress-type estimation with

high accuracy. The ﬁrst step aim at discriminating

with high accuracy stress status from and relax sta-

tus, in other words the presence of any workplace

stress. Once some stress reaction presence has been

detected, the second step consists in discriminating

normal stress and other workplace stress types, which

means identifying the harmfulness of the stress. Fi-

nally, if physiological indices are identiﬁed as reﬂect-

ing some extra stress, the third step consists in dis-

criminating the physiological reaction between ner-

vous stress and monotonous stress.

䋿

Step 1

Step 2

Step 3

Indices selection

for step 1

Indices selection

for step 3

Indices selection

for step 2

Relax

Normal

Monotonous

Nervous

End

Stress estimation start

Calculation of indices

Figure 4: Proposed multi-step estimation procedure with

integrated indices selection. It is composed of three steps to

gradually estimate the psychological status corresponding

to input physiological indices.

According to the results of signiﬁcant difference

study between the three PASAT tasks, which requires

good task performance to guarantee reliable subjec-

tive evaluations, we selected 39 persons among the

whole participants (13 persons by PASAT task type).

For each selected participant, we extracted ﬁve data-

sets for each stress type, which that contain the phys-

iological indices during the ﬁve minutes of PASAT

(the last ﬁve minutes for PASAT2) extracted with a

one minute rolling window. Five data-sets for re-

lax status for each selected participant were also ex-

tracted in the same way. As a result, we obtain a train-

ing database ﬁlled with 390 data-sets, each represent-

ing the physiological reaction to one minute exposure

to relax (195 data-sets), normal stress (65 data-sets),

monotonous stress (65 data-sets), and nervous stress

(65 data-sets) situations.

Then, to evaluate our method in a ﬁrst step inde-

pendently from the classiﬁcation scheme adopted, we

tested 24 classiﬁcation schemes among which Linear

Discriminant Analysis (LDA), Support Vector Ma-

chines (SVM), and so forth. At each step of the esti-

mation procedure and for each adopted classiﬁcation

scheme, we identiﬁed and used the best set of physi-

ological indices (those less prone to individual differ-

ences).

4.2 Validation of the Method

We compared the following four methods combined

with cross validation based on the selected 39 partici-

pants to validate our multi-steps logic combined with

integrated physiological indices selection. The results

of comparison with other conventional methods are

shown on Figure 5, validating the efﬁciency of our

proposed method.

• Method 1 proceeds to only one discrimination

step, using only LF/HF index

• Method 2 proceeds to only one discrimination

step, using all provided physiological indices

• Method 3 proceeds to only one discrimination

step, using identiﬁed best physiological indices

• Method 4 proceeds to multi-step discrimination

with best physiological indices selection at each

step (our method)

4.3 Reliability Study of the Proposed

Stress Estimation Method

In previous paragraph we identiﬁed the physiological

indices that are less prone to individual differences.

However, as a condition to ensure reliability of the es-

timation using these physiological indices, it is neces-

sary to validate its application to stress estimation for

participants whose physiological indices were not in-

cluded into the training database. Precisely, we raised

the following three conditions as an assurance of re-

liability of the proposed workplace stress estimation

method.

WORKPLACE STRESS ESTIMATION METHOD BASED ON MULTIVARIATE ANALYSIS OF PHYSIOLOGICAL

INDICES

0.25 1.5 2.75

Estimation accuracy [%]

SVM

LDA Fuzzy

Gauss

Linear

Method 2Method 1 Method 3 Method 4

σ=

Figure 5: Estimation results for each method using different

classiﬁers. Regardless the classiﬁer we can observe better

estimation accuracy for methods 3 and 4 than methods 1

and 2, validating the introduction of the process of best ﬁt

physiological indices selection, as well as better estimation

accuracy for method 4 compared to method 3, validating the

added value of intelligent multi-step discrimination process.

• The average stress estimation accuracy for the 39

participants composing the database is high

• The standard deviation of the stress estimation

accuracy for the 39 participants composing the

database is low

• The selected set of physiological indices input

into the stress classiﬁcation scheme does not be-

long to the trained database (low dependence on

database)

Accordingly, we selected three sets of physiologi-

cal indices combinations with high accuracy,and used

these three sets with the 24 classiﬁcation schemes, re-

sulting in 72 possible classiﬁcation schemes. Among

these we have investigated the classiﬁcation schemes

that meet the deﬁned reliability conditions, with a par-

ticular focus on the third one, which corresponds to

the dependence on the training database.

In the investigation process to index database de-

pendency criteria, we increased gradually from 9 to

39 the number of participants whose data are included

in the database. At each indentation, the maximum

estimation accuracy is calculated for each classiﬁca-

tion scheme by leave-one-out cross-validation. For

each database size, the smaller the difference between

estimation accuracy using selected scheme and the

maximum accuracy is, the more robust the scheme is.

Indeed, a small difference means that selected scheme

would keep high estimation accuracy even with an

ever-growing database that will reﬂect more and more

human diversity.

In our study, we deﬁned the reliability criterion

to be a difference in accuracy lower than 3%. The

difference between stress estimation accuracy using

below described method (Method A) and maximum

stress estimation accuracy method (recomputed each

time) reaches the required reliability criterion of 3%

when database is composed by data-sets from 29 par-

ticipants above 39 max participants (see Fig. 6).

• Classiﬁcation scheme: SVM with Gaussian Ker-

nel (σ =2.75)

• Physiological indices selected for step 1: f

, | f

|, t

PAT

, T

, RRV

• Physiological indices selected for step 2: stdt

, LF/HF, HR

• Physiological indices selected for step 3: | f

-f

, stdt

, t

PAT

, RRV, LF/HF

Estimation accuracy [%]

Participants number composing DB [-]

Max. accuracy

(SVM σ=2.75)

Method A

(SVM σ=2.75)

Difference

10 20 30 40

Figure 6: Comparison of stress type discrimination ac-

curacy when database’s size changes, between selected

method (Method A) and method with highest accuracy re-

computed each time. It shows the result of reliability study

when using the classiﬁcation scheme and physiological in-

dices set deﬁned in Method A.

Then, we used this 29 participants’ database as the

reference, and calculated for all possible classiﬁcation

schemes a reliability index U using equation 1. In

equation 1, x

, x

, and x

, are respectively the aver-

age accuracy, the standard deviation of the accuracy,

and the database dependency of the adopted stress

classiﬁcation schemes (i=1,2, ,72), which should re-

sult in positive values of U index for reliable classiﬁ-

cation schemes, and small values of U index for non-

reliable schemes.

So, the classiﬁcation scheme corresponding to the

highest U value is the one which characteristics are

described above. This classiﬁcation scheme enables

an accuracy of 64% for the estimation of workplace

stress type reaction , and 89% for occurrence or not

of workplace stress reaction, standard deviation of

28%, and database dependence index of 29 (max 39).

In the opposite, classiﬁcation scheme with lowest U

value, which uses a fuzzy logic algorithm, also en-

ables an accuracy of 64% for the estimation of work-

place stress type reaction, but with a standard devia-

HEALTHINF 2012 - International Conference on Health Informatics

− x

∑

i=1

− x

)

−

− x

∑

i=1

− x

)

−

− x

∑

i=1

− x

)

(1)

Table 1: Best discrimination accuracy for stress presence and type depending on the method used.

Physiological LF/HF only All indices Best Indices Selection Best Indices Selection

Indices Used and Multi-steps Method

Stress Types 26% ±2% 48% ±8% 56% ±3% 63% ±5%

Stress Presence 63% ±1% 83% ±7% 87% ±3% -

tion of 36%, and a database dependence index of 37

(max 39), both very high.

This last point conﬁrms that in our reliability in-

dex U, not only the average accuracy is dominant, but

it is also strongly inﬂuenced by standard deviation and

database dependency index.

5 CONCLUSIONS

In this research study we proposed a new method to

address the problem of inﬂuence of both individual

and stress type on physiological indices values. Based

on the collection of a large amount of physiologi-

cal data-sets under different stress types exposure, we

evaluated the efﬁciency of our method for accurate

stress type estimation by comparing it with other con-

ventional methods. Table 1 presents a detailed result

of efﬁciency analysis, showing that introducing the

process of best ﬁt physiological indices selection, has

a great impact on stress reaction presence estimation

accuracy, while the addition of intelligent multi-step

discrimination process is essential to improve the ac-

curacy of workplace stress type estimation. In addi-

tion, these results were issued with data from 39 dif-

ferent participants in age and sex, demonstrating our

proposed method to be less prone to individual differ-

ences.

However, to achieve our goal of a system as shown

on Figure 7 for personal continuous stress monitoring

in daily life, we still have to tackle many issues among

which the ones we consider in priority are listed be-

low.

• Continue improving stress type discrimination ac-

curacy while limiting at best the number of sen-

sors worn

• Investigate the possibility to discriminate more

types of stress

• Investigate a method that enables to evaluate

quantitatively stress level

• Application of the proposed method to the estima-

tion of chronic stress

Figure 7: Schematic of the virtual stress checker system

that will implement our method. The wearable terminal col-

lecting sensor information communicates through network

connectivity with an online sub-system, at which detailed

processing and various feedback can be performed (Faudot

et al., 2010; Lopez et al., 2009).

ACKNOWLEDGEMENTS

This research was supported by Japan Science and

Technology Agency’s (JST) strategic sector for cre-

ation of advanced integrated sensing technologies for

realizing safe and secure societies: research project

on ”Developmentof a Physiological and Environmen-

tal Information Processing Platform and its Applica-

tion to the Metabolic Syndrome Measures”.

REFERENCES

Aasa, U., Kalezic, N., and Lyskov, E. (2006). Stress

monitoring of ambulance personnel during work and

leisure time. International Archives of Occupational

and Environmental Health, vol. 80, no. 1, pages 51–

59.

Akselrod, S., Gordon, D., Ubel, F., Shannon, C., C.Berger,

and Cohen, J. (1981). Power spectrum analysis of

WORKPLACE STRESS ESTIMATION METHOD BASED ON MULTIVARIATE ANALYSIS OF PHYSIOLOGICAL

INDICES

heart rate ﬂuctuation: a quantitative probe of beat-to-

beat cardiovascular control. Science, vol. 10, pages

220–222.

Al’Absi, M., Bongard, S., and Buchanan, T. (1997). Car-

diovascular and neuroendocrine adjustment to pub-

lic speaking and mental arithmetic stressors. Psy-

chophysiology, vol. 34, no. 3, pages 266–275.

Al’Absi, M., Hatsukami, D., and Davis, G. L. (2005). At-

tenuated adrenocorticotropic responses to psycholog-

ical stress are associated with early smoking relapse.

Psychopharmacology, vol. 181, no. 1, pages 107–117.

Araki, S. and Kawakami, N. (1993). Health care work

stress: a review. Japanese Journal of Industrial

Health, vol. 35, pages 83–97.

Carroll, D., Phillips, A., and Balanos, G. (2009). Metabol-

ically exaggerated cardiac reactions to acute psycho-

logical stress revisited. Psychophysiology, vol. 46, no.

2, pages 270–275.

Faudot, T., Lopez, G., and Yamada, I. (October 11–13

2010). Information system for mapping wearable

sensors into healthcare services: Application to di-

etary habits monitoring. Proceedings of the 2nd In-

ternational Workshop on Web Intelligence and Virtual

Entreprise (WIVE10), held at the 11th IFIP Work-

ing Conference on Virtual Enterprises (PRO-VE’10),

Saint-Etienne, France.

Fukuda, O., Nagata, Y., Homma, K., and Tsuji, T. (2001).

Evaluation of heart rate variability by using wavelet

transform and a recurrent neural network. Proceed-

ings of the 23rd Annual International Conference of

the IEEE Engineering in Medicine and Biology Soci-

ety, vol. 2, pages 1769–1772.

Heidenreich, P., Ruggerio, C., and Massie, B. (1999). Ef-

fect of a home monitoring system on hospitalization

and resource use for patients with heart failure. Amer.

Heart J., vol. 138, no. 4, pages 633–640.

Itao, K. (Oct. 28–31 2007). Wearable sensor network con-

necting artifacts, nature, and human being. Proceed-

ings of the 6th Annual IEEE Conference on Sensors

(IEEE Sensors 2007), Keynote Presentation, pages

1120–1123.

Itao, K., Umeda, T., Lopez, G., and Kinjo, M. (Oct. 26–

29 2008). Human recorder system development for

sensing the autonomic nervous system. Proceedings

of the 7th Annual IEEE Conference on Sensors (Sen-

sors 2008), pages 423–426.

Jovanov, E., Price, J., Raskovic, D., Kavi, K., Martin, T.,

and Adhami, R. (Nov. 2000). Wireless personal area

networks in telemedical environment. Proceedings

of the 3rd Int. Conf. Inf. Technol. Biomed. (ITAB-

ITIS2000), pages 22–27.

Kim, D., Seo, Y., and Salahuddin, L. (2008). Decreased

long term variations of heart rate variability in sub-

jects with higher self reporting stress scores. Proceed-

ings of the 2nd International Conference on Pervasive

Computing Technologies for Healthcare, pages 289–

292.

Kotlyar, M., Donahue, C., and Thuras, P. (2008). Physiolog-

ical response to a speech stressor presented in a virtual

reality environment. Psychophysiology, vol. 45, no. 6,

pages 1034–1037.

Lopez, G., Shuzo, M., and Yamada, I. (7-9 October 2009).

New healthcare society supported by wearable sensors

and information mapping based services. 1st Interna-

tional Workshop on Web Intelligence and Virtual En-

treprise (WIVE’09). CD-Rom proceedings of the 10th

IFIP WG 5.5 Working Conference on Virtual Enter-

prises (PRO-VE 2009).

Miyake, S. (1997). Factors inﬂuencing mental workload

indexes. Journal of University of Occupational and

Environmental Health (J. UOEH). vol. 19, no. 4, pages

313–325.

Miyake, S. (2001). Multivariate workload evaluation com-

bining physiological and subjective measures. Inter-

national Journal of Psychophysiology, vol. 40, no. 3,

pages 233–238.

Ohsuga, M., Shimono, F., and Genno, H. (2001). Assess-

ment of phasic work stress using autonomic indices.

International Journal of Psychophysiology, vol. 40,

no. 3, pages 211–220.

Park, S. and Jayaraman, S. (2003). Enhancing the quality

of life through wearable technology. Engineering in

Medicine and Biology Magazine, IEEE, vol. 22, pages

41–48.

Schubert, C., Lambertz, M., and Nelesen, R. (2009). Effects

of stress on heart rate complexity-a comparison be-

tween short-term and chronic stress. Biological Psy-

chology, vol. 80, no. 3, pages 325–332.

Shimono, F., Ohsuga, M., and Terashita, H. (1998). Method

for assessment of mental stress during higt-tension

and monotonous tasks using heart rate, respiration and

blood pressure. The Japanese Journal of Ergonomics,

vol. 34, no. 3, pages 107–115.

Shin, J., Seongo, H., Cha, D., Yoon, Y., and Yoon, H.

(1998). Estimation of stress status using biosignal

and fuzzy theory. Proceedings of the 20th Annual

International Conference of the IEEE Engineering in

Medicine and Biology Society, vol. 3, pages 1393–

1394.

Soda, T. and Narumi, T. (2007). Assessment of physi-

cal and mental loads using power spectral analysis of

heart rate variability during dressing. Japan society of

Home Economics, pages 91–98.

Watanabe, T., Kokubo, H., Takazawa, K., and Kawano, K.

(2008). Psychophysiological changes during exposure

to natural and urban environments. Journal of Inter-

national Society of Life Information Science, vol. 26,

no. 1, pages 106–111.

Willemsen, G., Ring, C., and et al., D. C. (1998). Secre-

tory immunoglobulin a and cardiovascular reactions

to mental arithmetic and cold pressor. Psychophysiol-

ogy, vol. 35, no. 3, pages 252–259.

Yamada, S. and Miyake, S. (2007). Effects of long term

mental arithmetic on physiological parameters, sub-

jective indices and task performances. Journal of Uni-

versity of Occupational and Environmental Health (J.

UOEH), vol. 29, no. 1, pages 27–38.

HEALTHINF 2012 - International Conference on Health Informatics