Mmsd: A Multi-modal Dataset for Real-time, Continuous Stress

Detection from Physiological Signals

Mouna Benchekroun

1,2 a

, Dan Istrate

1 b

, Vincent Zalc

1 c

and Dominique Lenne

2 d

Biomechanics and Bioengineering, UMR CNRS 7338, Universit

e de Technologie de Compi

egne, Compi

egne, France

Heudiasyc (Heuristics and Diagnosis of Complex Systems), Universit

e de Technologie de Compi

egne, Compi

egne, France

Keywords:

Multimodal Dataset, Emotion Recognition, Stress Detection, Physiological Data, Affective Computing.

Abstract:

Although chronic stress is proven to be very harmful to physical and mental well being, its diagnosis is

punctual and nontrivial, which calls for reliable, continuous and automated stress monitoring systems that

do not yet exist. Wireless biosensors offer opportunities to remotely detect and monitor mental stress levels,

enabling improved diagnosis and early treatment. There are different algorithms and methods for wearable

stress detection, however, only a few standard and publicly available datasets exist today.

In this paper, we introduce a multi-modal high-quality stress detection dataset with details of the experimental

protocol. The dataset includes physiological, behavioural and motion data from 74 subjects during a lab study.

Different modalities such as electrocardiograms (ECG), photoplethysmograms (PPG), electrodermal activity

(EDA), electromyograms (EMG) as well as three axis gyroscope and accelerometer data were recorded.

In addition, protocol validation was achieved using both subject’s self-reports and cortisol levels which is

considered as gold standard for stress detection.

1 INTRODUCTION

Stress can be deﬁned as a complex reaction pattern

with psychological, cognitive and physiological com-

ponents. The stress response occurs whenever there

is a homeostatic imbalance caused by internal or ex-

ternal factors (Ursin and Eriksen, 2004).

While generally adaptive and safe in the short

term, the presence of stress over the long term can

be harmful to a person’s mental and physical health

(Yaribeygi et al., 2017). For example, chronic stimu-

lation of the cardiovascular system due to stress leads

to sustained increases in blood pressure and vascular

hypertrophy. Stress is also linked to immunosuppres-

sion by directly affecting a variety of hormones in-

volved in immune system function such as cytokines

proﬁles (Schneiderman et al., 2005),(Yaribeygi et al.,

2017).

The gold standard stress measurement modality

today is salivary cortisol levels. This measure remains

however punctual and delayed. It does not allow for

real-time stress monitoring.

https://orcid.org/0000-0001-5833-5776

https://orcid.org/0000-0001-5906-4947

https://orcid.org/0000-0001-5325-6649

https://orcid.org/0000-0002-6225-8854

Since the stress response has psychological de-

terminants as well, stress is among the psychologi-

cal concepts that can be measured through question-

naires. Several psychologists looked into the ques-

tion and developed questionnaires covering a wide

range of psychological symptoms caused by exposure

to chronic stress. State and Trait Anxiety Inventory

(STAI), for example, is the gold standard for measur-

ing preoperative anxiety (Dalal et al., 2015).

The same way as cortisol, questionnaires offer a

punctual measure. Furthermore, they are based on

subjective feedback which is not always reliable.

Given the gap in this area and the value of contin-

uous real-time stress monitoring, scientists have at-

tempted to quantify stress by measuring changes in

physiological parameters such as heart, skin and mus-

cle activity.

Thanks to biosensors that have developed a lot in

recent years, it is possible to collect various physio-

logical data during users daily life and automatically

extract information about their physiology.

Approved medical devices exist today such as

the ”AppleWatch” and the ”WithingDevices” watch

which assesses stress from the heart rate and Heart

Rate Variability (HRV). The Empatica bracelet and

the PIP device are based on skin conductance,

240

Benchekroun, M., Istrate, D., Zalc, V. and Lenne, D.

Mmsd: A Multi-modal Dataset for Real-time, Continuous Stress Detection from Physiological Signals.

DOI: 10.5220/0010985400003123

In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - Volume 5: HEALTHINF, pages 240-248

ISBN: 978-989-758-552-4; ISSN: 2184-4305

whereas the SpireStones device monitors subject’s

breathing frequency to identify stress.

The main drawback common to these commercial de-

vices is the use of a single signal to detect stress,

which reduces their reliability since different psycho-

logical, or even physical states can have similar im-

pacts on the same signal.

Since stress is now recognized as a universal pre-

morbid factor, associated with several risk factors for

various chronic diseases, there is a need to improve

stress monitoring not only in clinical practice, but also

for prevention and early intervention, which is essen-

tial to avoid complications due to cardiovascular dis-

eases (Schneiderman et al., 2005). Affective comput-

ing as an emerging ﬁeld could also beneﬁt from a con-

tinuous real-time stress detection device to improve

human-computer interactions.

2 RELATED WORK

In recent years, many studies have been conducted

on stress detection from physiological parameters.

Besides signals such as electrocardiograms (ECG),

photoplethysmograms (PPG), electrodermal activity

(EDA) . . . , some scientist take interest in outward

characteristics like body posture and facial expres-

sions. In our study, the focus is put on physiologi-

cal signals as they are less sensitive to environmental

variations or noise.

Most of the studies published today are conducted

in laboratory environments using various stressors

such as public speaking, mental arithmetic and stroop

color word test (Adochiei et al., 2019). Others use

physical stressors as a cold pressor for example and

hand grip (Dickerson and Kemeny, 2004).

Although there is intensive research in the ﬁeld of

stress monitoring from wearable devices, there is only

very little publicly available data.

One of the ﬁrst datasets on continuous stress mon-

itoring was published by Healey et al (Healey and

Picard, 2005). The dataset includes ECG, trapezius

EMG, EDA and respiration from 24 subjects during

50min real-world driving. Protocol validation was

achieved by questionnaires and video coding for part

of the subjects. This is one of the rare studies that has

been carried out in an ambulatory environment

Koestra et al also published a dataset for emo-

tion analysis, including stress, using physiological

signals (Koelstra et al., 2011). Electroencephalo-

grams (EEG) as well as ECG, PPG, EDA, trapez-

ius EMG and respiration were recorded from 32 sub-

jects while they watched 40 videos each inducing a

different emotional reaction. Labeling was achieved

by subject’s self assessment after each trial (Koelstra

et al., 2011).

Another publicly available dataset is the

SW ELL − KW published by Koldijk.S et al (Koldijk

et al., 2014). The dataset was collected in an experi-

ment on 25 subjects in their work spaces while they

performed typical knowledge work under stressful

conditions such as time pressure and email inter-

ruptions. Various data including computer logging,

facial expression, body postures, ECG and skin

conductance were recorded in neutral and stressful

conditions. In this study, ground truth was obtained

through subjective rating thanks to various validated

questionnaires such as the Nasa Task Load Index

used to determine task load (Hart and Staveland,

1988) and Self-Assessment-Manikin Scale (SAM)

for emotion response (Bradley and Lang, 1994) and

other questionnaires.

More recently, Schmidt et al. published a multi-

modal dataset of physiological and motion data of 15

subjects during a lab study targeting three different af-

fective states: neutral, stress and amusement (Schmidt

et al., 2018). The recorded data include ECG, PPG,

EDA, trapezius EMG, respiration and three axis ac-

celeration. Subjective feedback was used once again

as ground truth through different established ques-

tionnaires such as the Positive and Negative Affect

Schedule (PANAS), a 6 item STAI, a (SAM) and a

Short Stress State Questionnaire (SSSQ).

One common disadvantage to all the studies cited

above is the small sample size generally below 30 par-

ticipants. Moreover, it is true that self-assessment is

important to have personal feedback from the subject,

but one has to keep in mind the bias of such subjective

assessment, especially when using hand-crafted ques-

tionnaires that haven’t been experimentally tested and

validated. Subjects may indeed have trouble identify-

ing their emotions.

For a more reliable validation, established psy-

chological questionnaires should be backed up by the

gold standard stress measurement which is cortisol

level.

A longitudinal study was carried out, few years

ago in our lab, that aimed to recognize stress dur-

ing users daily routine. Three sensors including an

actigraph, a punctual blood pressure manager and a

PolarH7 belt for heart activity were used to record 53

subjects during 14 days of their daily lives (Tlija et al.,

2019).

The end goal was to study the correlation between

emotional states including stress levels and prognosis

of cardiovascular disease, but it was too ambitious to

start with a long-term, ambulatory study for emotion

recognition for two reasons:

Mmsd: A Multi-modal Dataset for Real-time, Continuous Stress Detection from Physiological Signals

241

1. Study duration was too long for subjects to stay

committed all along and follow all the guidelines.

2. Signal segmentation could not be achieved

reliably because labeling was based only on

subjective journaling which was often not done

carefully by the subjects.

Given the difﬁculty of such longitudinal ambu-

latory user studies, we chose a more constrained

study carried out in a laboratory environment as a

starting point. The exact protocol is detailed in the

next section.

Paper Contribution

In this paper, we present a multi-modal dataset for

stress detection from physiological signals. Our

dataset bridges the gap between uni-modal devices

now available and potentially more reliable multi-

modal stress measurements based on physiological

signals.

To the best of our knowledge, it is one of the

largest datasets available on stress monitoring.

The experimental protocol is described with great

detail so that the potential user can have a global view

and a deep understanding of the data.

Physiological differences and constraints such as

cortisol ﬂuctuation and subjects proﬁle were taken

into consideration in the study design. The experi-

mental protocol is validated with gold-standard stress

measures which makes it suitable for stress detection

through artiﬁcial intelligence algorithms.

3 VARIATIONS IN STRESS

RESPONSES

When studying any physiological reaction of the hu-

man body, one has to be aware of the sources of vari-

ation affecting the response.

Stress is a highly personalized phenomenon that

varies between people depending on individual vul-

nerability and resilience, and between different types

of tasks (Fink, 2016). Parameters affecting the stress

response can be categorized as follows:

• Stressor’s type: Subject’s stress response depends

on the intensity/severity of the stressor, its con-

trollability as well as personal features that deter-

mine the cognitive reactions to each stressor type

(Schneiderman et al., 2005).

• Subject dependant variations: Each individual’s

response to stress is determined by a multitude

of genetic, personal and environmental factors.

Coping skills, personality, psychiatric history and

sociodemographic variables also have an impor-

tant impact on the stress perception and recovery

(Kendler et al., 2003).

• Activity dependant variations: Stress response

characteristics also vary with physical activity

(walking VS running, standing VS sitting . . . )

(Alamudun et al., 2012). Shumm et al. found

that EDA is sensitive to subject’s movement. They

concluded that the faster a person is walking the

more uniformly distributed the skin conductance

(Schumm et al., 2008).

Heart rate is also subject to posture-related

changes. A signiﬁcant increase in heart rate as a

subject transitions from supine to sitting; from sit-

ting to standing, and from standing to walking has

been reported in a study carried out by (Van Stee-

nis and Tulen, 1997).

These variations should be taken into consider-

ation when designing a user study for stress detec-

tion. Efforts should be made to tackle these sources

as much as possible by controlling the stressors, the

environment and subject-related variations by estab-

lishing a subject proﬁle.

It is true that the end goal is ambulatory stress de-

tection in user’s daily life, but this requires to test and

validate continuous stress detection methods in con-

strained environments, where these sources of varia-

tion are kept to a minimum, as a ﬁrst step.

Another very important aspect in such studies is

the use of gold-standards for protocol validation.

4 DATA COLLECTION PROCESS

Once our experimental protocol was approved by The

INSEAD Institutional Review Board (IRB : 202077),

INSEAD Behavioural Lab team took charge of sub-

ject’s recruitment and management. Participants

(aged 18 and older) were invited by email to take an

online survey (the pre-selection questionnaire intro-

duced below) in order to verify their eligibility to the

study.

4.1 Eligibility Criteria Selection

Subjects had to be selected in accordance with ethical

criteria as well as the constraints related to the study

itself.

A pre-selection questionnaire was used as a

ﬁrst step in volunteers selection process with four

objectives:

HEALTHINF 2022 - 15th International Conference on Health Informatics

242

1. Exclude Non Eligible Volunteers:

Subjects were selected in order to have a sta-

tistically representative sample of the french

population in terms of age and gender. Exclusion

criteria included volunteers suffering from: car-

diovascular diseases, chronic diseases (diabetes,

hypertension and mental disorders (depression,

dementia . . . ) since these conditions may have an

impact on the collected physiological data.

2. Build a Subject Proﬁle:

The pre-selection questionnaire was also used

to collect general information about lifestyle

elements that would potentially affect the stress

response such as: the participant’s level of

physical activity, eating habits, sleeping habits,

meditation, . . .

These elements could be used in data interpreta-

tion.

3. Enunciate Guidelines:

Guidelines were given in the pre-selection

questionnaire to check volunteer’s acceptability

to wear the biosensors and their willingness to

respect some instructions. For example, selected

subjects were asked to abstain from alcohol,

caffeine/theine and tobacco, 12 hours, 4 hours

and 2 hours respectively before the experiment.

4. Measure Perceived Stress:

The questionnaire also includes a PSS4 (Per-

ceived Stress Scale) consisting of 4 questions.

This scale assesses the state of perceived stress

and measures the degree to which situations in

one’s life are appraised as stressful (Warttig et al.,

2013).

4.2 Participants

227 participants volunteered to take part of our study.

57 were rejected due to exclusion criteria and 74

healthy subjects were selected from the remaining

volunteers. Our objective was to form a representa-

tive sample of the French population in terms of age

and gender. Figure 1 summarizes the selection pro-

cess. The ﬁnal sample includes 38 women aged 19 to

63 years old (mean age: 33 y.o ± 12.5) and 36 men

aged 21 to 79 years old (mean age: 35 y.o ± 13).

4.3 Sensors

For data collection, we used Shimmer Sensing

biosensors for all physiological data. Four sig-

nals were selected to study their correlation with

stress: electrocardiogram (ECG), photoplethysmo-

gram (PPG), electrodermal activity (EDA) and elec-

tromyogram (EMG). Sensors conﬁguration is de-

picted in ﬁgure 2

ECG is recorded using four self-adhesive elec-

trodes for three leads included in Einthoven triangle.

Two electrodes were placed across the heart, below

the collar-bones, a third one below the chest, and the

grounding electrode in the middle chest.

Pulse is recorded thanks to a transmission Photo-

plethysmograph sensor (PPG) from the right earlobe.

Transmission PPG sensors detect light passed through

the tissue and are therefore commonly used on periph-

eral sites such as ﬁngers and earlobes. We chose the

latter location since it is less sensitive to motion arte-

fact and tissue alterations caused by both voluntary

and involuntary movements as there is no muscle ac-

tivity.

Electrodermal activity is recorded with sticker

electrodes on the participant’s non dominant hand.

One electrode is placed on the palmar surface of the

index medial phalange and the other on the palmar

surface of the middle ﬁngers’ distal phalange.

For trapezius muscle activity, positive and nega-

tive electrodes are placed in parallel with the muscle

ﬁbres, near the centre of the right and left muscle,

2cm apart from each other. The reference electrode

is placed at the elbow as an electrically neutral point

of the body far from the muscle being measured.

ECG, pulse and EDA are recorded at 512Hz

whereas EMG is digitized at 1024Hz.

Recorded data was stored locally in sensors SD

card and transferred at the same time via Bluetooth

to a computer for further processing after the exper-

iment. Shimmer Sensing software (ConsensysPro)

was used to manage data and devices.

All the devices used in the study are equipped with

gyroscope sensors that can measure the orientation

and angular velocity of the device. Three axis gy-

roscope data is available for each of the sensors intro-

duced above.

Computer logging including response time and

task duration were also automatically saved to an SQL

database.

4.4 Experimental Protocol

The goal of this study is to elicit three different affec-

tive states (relaxation, stress and neutral) and identify

correlations between physiological signals and these

states.

To minimize the effect of the different sources of vari-

ation on the stress response, the study is carried out

under laboratory conditions, where the environment is

Mmsd: A Multi-modal Dataset for Real-time, Continuous Stress Detection from Physiological Signals

243

Number of volun-

teers interested in

the study N=227

Waiting list n = 96

Excluded n = 57

Heavy smoking: 9

Mental disorders: 10

Chronic diseases: 5

Medication/drugs: 13

Other diseases: 20

Selected subjects n = 74

Figure 1: Subject selection process.

controlled and movement is reduced. The experimen-

tal protocol is detailed below and depicted in ﬁgure

1. Relaxation/Meditation: Subject is invited to

relax thanks to a 15min guided meditation session

instructed via an audio track. Subjects followed

the instructions with closed eyes, while sitting in

a comfortable position in a dark environment.

2. Stress: Subject performs various stressful tasks

such as the Stroop color word test, mental

arithmetic and other serious games previously

proven to induce mental stress (Dickerson and

Kemeny, 2004). A score as well as a red timer

are used to increase pressure. Subjects were not

aware that the main goal of this step is to induce

stress. Instead, they were told the tasks are used

to compute a QI score.

This phase lasted for around 20min.

3. Recovery: At the end of the protocol, data was

recorded for an additional 10min while subject

is asked to stay seated with a calm music back-

ground.

The purpose of this step is to evaluate how sub-

jects recover from a stressful stimuli.

The study lasts for about 80 minutes. All the trials

took place in the afternoon in order to avoid the wake-

up cortisol peak and because cortisol levels are stead-

ier in the afternoon (Dickerson and Kemeny, 2004).

4.5 Labeling and Protocol Validation

Since the main purpose of our study is to use machine

learning algorithms to classify each state, we needed

to validate our protocol beforehand in order to make

sure subject’s affective state matches the experimen-

tal protocol. Two different measures were used as

labels:

1. Cortisol Samples:

Saliva samples were collected after the ﬁrst and

second phases (baseline and stress respectively) as

a gold standard measure of stress. We expect cor-

tisol levels to be higher in the second sample with

comparison to the ﬁrst one. Since cortisol lev-

els vary throughout the day with the highest peak

reached few hours after waking up, as explained

earlier, all trials took place in the afternoon (after

2pm) to make sure the wake-up peak does not re-

verse the increase in cortisol levels from the ﬁrst

to the second sample.

HEALTHINF 2022 - 15th International Conference on Health Informatics

244

(a) ECG sensor. (b) PPG sensor.

(d) EMG sensor.

Figure 2: Sensors conﬁguration.

2. STAI-S:

State-Trait Anxiety Inventory (STAI) question-

naire comprises separate self-report scales for

measuring two distinct anxiety concepts: trait

anxiety (STAI − T ) and state anxiety (STAI − S).

Each scale is made up of 20 questions.

The ﬁrst one is to assess subjects’ anxiety as

a personality trait and is used, together with

the PSS4, explained in section 4.1, for data

interpretation.

The second one measures anxiety as an emotional

state linked to a particular situation. The STAI − S

is used as a subjective label after each phase to-

gether with cortisol levels to validate our protocol.

Since cortisol is considered to be a stress hor-

mone, we expect cortisol levels to be higher in the

second sample taken after the stress phase with com-

parison to the ﬁrst one. STAI − S scores should also

be higher if the subject was indeed stressed in the sec-

ond phase.

The combination of both physiological and sub-

jective labels makes the validation process more reli-

able.

4.6 Dataset

Our dataset includes different types of parameters

summarized in table 1 from 74 subjects, 38 women

aged 19 to 63 years old (mean age: 33 y.o ± 12.5)

and 36 men aged 21 to 79 years old (mean age: 35

y.o ± 13). Each parameter can be used either for data

interpretation, classiﬁcation or labeling depending on

the end purpose. Some examples are presented but

are not limited to table 1.

Mmsd: A Multi-modal Dataset for Real-time, Continuous Stress Detection from Physiological Signals

245

Figure 3: Study experimental protocol.

Table 1: Types and potential use of collected data.

Type Data Use

Proﬁl Data

Sex,age . . .

Eating and sleeping habits

Exercising

Medidation

Interpretation

Subjective/Phsycological

PSS4 and STAI-T

STAI-S

Interpretation

Labeling

Behavioural

Responses Time

Successes and failures in trials

Task duration

Interpretation

and/or segmentation

Physiological

Electrocardiogram (ECG)

Photoplethysmogram (PPG)

Electrodermal Activity (EDA)

Electromyogram (EMG)

Three axis gyroscope data

Salivary Cortisol Levels

Data Analysis

Classiﬁcation

Interpretation

Labeling

The dataset features physiological signals such as

ECG, PPG, EDA digitalised at 512Hz and EMG at

1000Hz, as well as motion data from 74 subjects for

a total duration of 40 minutes per subject, segmented

in three sessions as explained in section 4.4. The ﬁrst

phase, Relaxation phase, is referred to as session 1,

session 2 is the Stress phase and session 3 is the Re-

covery.

HEALTHINF 2022 - 15th International Conference on Health Informatics

246

5 CONCLUSIONS

The experimental protocol presented in this paper

takes into consideration many of the sources of varia-

tion encountered in previous studies. Its main purpose

was to induce three states: relaxation, stress and neu-

tral/baseline.

The study was carried out in a constrained envi-

ronment and protocol validation was achieved using

both psychological self-reports and ground truth cor-

tisol levels which makes it reliable for machine learn-

ing algorithms. The dataset includes high-quality

physiological modalities commonly used in commer-

cial and medical devices for stress identiﬁcation like

ECG, PPG, EDA, EMG and three axis gyroscope

data. Thanks to the high number of participant as

well as their diversity in terms of age and gender, it

is possible to draw reliable conclusions and statistical

generalisation.

The dataset will be made publicly available once

data cleaning and organisation are complete. It could

be used in many different ways to study the correla-

tion between various physiological signals with stress

and/or stress recovery in a uni-modal or multi-modal

approach. It could also be used to compare chest-

based ECG device to earlobe PPG in terms of signal

quality, prep-rocessing, and classiﬁcation results. The

self-reports could be utilized to create personalised

models able to detect and predict a person’s speciﬁc

affective state.

ACKNOWLEDGEMENTS

Authors would like to thank Idex Sorbonne Univer-

sity for funding this experimental study as part of

french state support for ”Investissements d’Avenir’

program”. Also thanks to all the subjects and to IN-

SEAD lab for their expertise in participant recruit-

ment and management which made the process ex-

tremely easier.

ETHICS DECLARATIONS

All volunteers gave their informed written consent in

accordance with the Declaration of Helsinki and fol-

lowing approval from and in accordance with the IN-

SEAD Institutional Review Board (IRB : 202077).

REFERENCES

Adochiei, I., Adochiei, F., Cepisca, C., Seritan, G., Enache,

B., Argatu, F., and Ciucu, R. (2019). Complex em-

bedded system for stress quantiﬁcation, 2019 (atee).

WOS: 000475904500049.

Alamudun, F., Choi, J., Gutierrez-Osuna, R., Khan, H., and

Ahmed, B. (2012). Removal of subject-dependent and

activity-dependent variation in physiological mea-

sures of stress. In 2012 6th International Conference

on Pervasive Computing Technologies for Healthcare

(PervasiveHealth) and Workshops, pages 115–122.

IEEE.

Bradley, M. M. and Lang, P. J. (1994). Measuring emotion:

the self-assessment manikin and the semantic differ-

ential. Journal of behavior therapy and experimental

psychiatry, 25(1):49–59.

Dalal, K. S., Chellam, S., and Toal, P. (2015). Anaesthesia

information booklet: Is it better than a pre-operative

visit? Indian journal of anaesthesia, 59(8):511.

Dickerson, S. S. and Kemeny, M. E. (2004). Acute stres-

sors and cortisol responses: a theoretical integration

and synthesis of laboratory research. Psychological

bulletin, 130(3):355.

Fink, G. (2016). Chapter 1—stress, deﬁnitions, mech-

anisms, and effects outlined: lessons from anxiety.

Stress: Concepts, Cognition, Emotion, and Behavior;

Fink, G., Ed, pages 3–11.

Hart, S. G. and Staveland, L. E. (1988). Development of

nasa-tlx (task load index): Results of empirical and

theoretical research. In Advances in psychology, vol-

ume 52, pages 139–183. Elsevier.

Healey, J. A. and Picard, R. W. (2005). Detecting stress dur-

ing real-world driving tasks using physiological sen-

sors. IEEE Transactions on intelligent transportation

systems, 6(2):156–166.

Kendler, K. S., Hettema, J. M., Butera, F., Gardner, C. O.,

and Prescott, C. A. (2003). Life event dimensions of

loss, humiliation, entrapment, and danger in the pre-

diction of onsets of major depression and generalized

anxiety. Archives of general psychiatry, 60(8):789–

796.

Koelstra, S., Muhl, C., Soleymani, M., Lee, J.-S., Yazdani,

A., Ebrahimi, T., Pun, T., Nijholt, A., and Patras, I.

(2011). Deap: A database for emotion analysis; using

physiological signals. IEEE transactions on affective

computing, 3(1):18–31.

Koldijk, S., Sappelli, M., Verberne, S., Neerincx, M. A., and

Kraaij, W. (2014). The swell knowledge work dataset

for stress and user modeling research. In Proceedings

of the 16th international conference on multimodal in-

teraction, pages 291–298.

Schmidt, P., Reiss, A., Duerichen, R., Marberger, C., and

Van Laerhoven, K. (2018). Introducing wesad, a

multimodal dataset for wearable stress and affect de-

tection. In Proceedings of the 20th ACM interna-

tional conference on multimodal interaction, pages

400–408.

Schneiderman, N., Ironson, G., and Siegel, S. D. (2005).

Stress and health: psychological, behavioral, and bi-

Mmsd: A Multi-modal Dataset for Real-time, Continuous Stress Detection from Physiological Signals

247

ological determinants. Annu. Rev. Clin. Psychol.,

1:607–628.

Schumm, J., B

achlin, M., Setz, C., Arnrich, B., Roggen, D.,

and Tr

oster, G. (2008). Effect of movements on the

electrodermal response after a startle event. Methods

of Information in Medicine, 47(03):186–191.

Tlija, A., Istrate, D., Gattouﬁ, S., and Bennani, A.-E.

(2019). Stress recognition using connected devices:

experimentation feedback. In Journ

ees d’Etude sur la

eSant

Ursin, H. and Eriksen, H. R. (2004). The cognitive acti-

vation theory of stress. Psychoneuroendocrinology,

29(5):567–592.

Van Steenis, H. and Tulen, J. (1997). The effects of physical

activities on cardiovascular variability in ambulatory

situations [ecg/accelerometry analysis]. In Proceed-

ings of the 19th Annual International Conference of

the IEEE Engineering in Medicine and Biology Soci-

ety.’Magniﬁcent Milestones and Emerging Opportuni-

ties in Medical Engineering’(Cat. No. 97CH36136),

volume 1, pages 105–108. IEEE.

Warttig, S. L., Forshaw, M. J., South, J., and White, A. K.

(2013). New, normative, english-sample data for the

short form perceived stress scale (pss-4). Journal of

health psychology, 18(12):1617–1628.

Yaribeygi, H., Panahi, Y., Sahraei, H., Johnston, T. P., and

Sahebkar, A. (2017). The impact of stress on body

function: A review. EXCLI journal, 16:1057.

HEALTHINF 2022 - 15th International Conference on Health Informatics

248