Enhancement of Physiological Stress Classiﬁcation using Psychometric

Features

Karl Magtibay

1 a

, Xavier Fernando

2 b

and Karthikeyan Umapathy

Faculty of Engineering and Architecture Science, Ryerson University, 350 Victoria St., Toronto, Ontario, Canada

Department of Electrical, Computer, and Biomedical Engineering, Ryerson University,

Keywords:

Stress, Well-being, Feature Extraction, Joint Data Analysis.

Abstract:

Psychological data features are underutilized in many acute stress studies since they are challenging to repli-

cate and validate due to their inherent subjectivity. However, psychology and perception play essential roles

in stress research according to the well-established allostatic load model. Therefore, we demonstrate the

importance of accounting for psychological data in acute stress research in an ambulatory setting through a

joint analysis. We enhanced stress classiﬁcation by combining psychometric features with standard phys-

iological signal features. We used the publicly available Wearable Stress and Affect Database (WESAD),

from which we obtained physiological signals and psychological self-assessments from 15 participants. For

each participant, a set of physiologically relevant features were extracted from each signal type. In paral-

lel, we adapted psychometric features, positive emotion (PE

score

) and negative emotion (NE

scores

) scores, by

calculating the weighted average of self-evaluation scores. Using a stepwise feature selection and a linear-

discriminant-analysis-based classiﬁer, we found that PE

scores

, along with select physiological signal features,

could enhance cross-validated stress classiﬁcation accuracy by 8%, higher than a previous benchmark study

using the same dataset. More importantly, we found that such a classiﬁcation accuracy could be achieved with

signiﬁcantly fewer physiological signal features (by 20 times) with the aid of a psychometric feature. Finally,

we found that psychometric features could indicate the type of perceived stress relating to an individual’s mood

descriptor scores. Thus, a combination of psychometric and physiological data could be beneﬁcial towards

improving the detection and management of stress and support the development of holistic stress models.

1 INTRODUCTION

Stress is a complex physiological, and psychological

response of the human body towards perceived or ac-

tual threats to its well-being (O’Connor et al., 2021).

Stress has been reported to cause many medical and

mental health issues experienced by workers in highly

demanding jobs, such as personal support workers

and healthcare practitioners (Pappa et al., 2020). In

addition, prolonged stress could induce many long-

term health issues like cardiovascular disease and de-

pression (Legault et al., 2017; O’Connor et al., 2021).

Therefore, persistent or chronic stress without proper

and regular intervention could be detrimental to an in-

dividual’s well-being.

Some stress studies on workers in harsh envi-

ronments primarily focused on using commercial-

https://orcid.org/0000-0001-5846-5327

https://orcid.org/0000-0001-7120-528x

grade wearables(Choi et al., 2011; Runkle et al.,

2019) or building wireless body area networks

(WBANs)(de Fazio et al., 2020; Wu et al., 2019) to

monitor and detect physiological and environmental

indicators of stress. Studies cited above are crucial

for developing stress studies standards, such as ex-

perimental protocols and devices, in a dynamic set-

ting such as the workplace. However, the psychologi-

cal correlates of stress, such as affect, perception, and

past experiences have been under-represented.

Evidence of the dynamic interaction of human

physiology and psychology in response to stres-

sors has been long established (Guidi et al., 2021;

McEwen and Rasgon, 2018). The interaction of dif-

ferent physiological and psychological systems in the

human body is the building block of the allostatic

model(Guidi et al., 2021; Fava et al., 2019; McEwen

and Rasgon, 2018). The allostatic model provides a

framework to elucidate the physiological and psycho-

430

Magtibay, K., Fernando, X. and Umapathy, K.

Enhancement of Physiological Stress Classiﬁcation using Psychometric Features.

DOI: 10.5220/0010823300003123

In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - Volume 5: HEALTHINF, pages 430-437

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

logical mechanisms that contribute to the overall wear

and tear of the body (allostatic load) as a result of

prolonged exposure to stressors. Although adminis-

tering psychological evaluations in a dynamic envi-

ronment is challenging, accounting for psychological

biomarkers along with physiological biomarkers of a

stress response could be beneﬁcial to improve stress

classiﬁcation tasks.

The disjunct in methods of measurement or as-

sessment of the stress response is apparent in some

studies. While some have focused on characterizing

stress through physiological signal features(Schmidt

et al., 2018; Choi et al., 2019), others solely relied

on evaluating its psychological effects through obser-

vational studies and self-reports analysis(Vitale et al.,

2021; Edmondson et al., 2014). A previous study by

(Schmidt et al., 2018) called for accounting psycho-

logical data from self-reports to improve physiolog-

ical characterization of stress and affective states for

individuals. The importance of psychological data to

stress modeling using ambulatory data has been pre-

viously shown (Hovsepian et al., 2015; Sarker et al.,

2016; Plarre et al., 2011) however, a joint quantitative

analysis of raw psychological and physiological stress

factors has been underutilized.

Our work highlights the importance of accounting

for psychological correlates of stress, such as affect

and perception, along with physiological signals from

wearable devices. We jointly analyze the physiologi-

cal signal and psychometric features to enhance stress

classiﬁcation. In addition, we introduce new psycho-

metric features, the emotion scores, translated from

an established psychological questionnaire, and as-

sess their relevance in improving stress classiﬁcation,

consistent with the allostatic load model. Outlined in

Figure 1 is the overall organization of our work pre-

sented in this paper.

2 METHODS

2.1 Database

Data used in this work was obtained from the Wear-

able Stress and Affect Database (WESAD) as part

of the study done by (Schmidt et al., 2018), made

available through the University of California Irvine’s

machine learning repository (Asuncion and Newman,

2007). Below, we brieﬂy describe WESAD and sum-

marize its authors’ data collection methods relevant

to our proposed work. Then, we direct the reader to

(Schmidt et al., 2018) for a detailed description of the

data acquisition and validation of WESAD.

Figure 1: Overall organization of proposed study to improve

stress classiﬁcation accuracy using both physiological and

psychometric data features.

WESAD provides a standard multi-modal dataset

for stress and affect detection in an ambulatory set-

ting. In addition, the authors of WESAD offered

benchmark classiﬁcation results using standard data

features and machine learning methods for compari-

son against future stress and affect studies.

WESAD consists of physiological and psycholog-

ical data collected from 15 participants, 12 males and

3 females, with an average age of 27.5 ± 2.4 years

old. Physiological data from each participant were

obtained from two wearable devices, one strapped

around the chest (RespiBAN professional) and the

other around the wrist (Empatica E4). In this work,

we limited our data source to the chest device due

to its superior signal quality compared to the wrist

device. Therefore, from the chest device, only ac-

celerometer (ACC), plethysmograph (RESP), elec-

trocardiograph (ECG), electrodermal activity (EDA),

and temperature (TEMP) data were included for fea-

ture extraction.

Moreover, from all of the questionnaires used in

the Schmidt study, we chose only the Positive Affect

and Negative Affect Schedule (PANAS)(Crawford

and Henry, 2004) as the source of our psychometric

features since it was consistently used throughout that

study. The authors of WESAD added affect state de-

scriptors (i.e., Stressed, Frustrated, Happy, and Sad)

to the original PANAS questionnaire to suit the goals

of the original study. We used the ‘Stressed’ descrip-

tor as ground truth for later classiﬁcation. Finally, we

Enhancement of Physiological Stress Classiﬁcation using Psychometric Features

431

included data only from neutral/relaxed and stressed

states to perform binary classiﬁcation.

2.2 Physiological Signal Feature

Extraction

Due to the length of both baseline (20 mins) and stress

(6.5 mins) conditions, we calculated signal features

within a sliding one-minute window on increments

of 10 seconds for all signal types. The set of sig-

nal features for each signal type were ensured to have

been proven in the literature to be physiologically rel-

evant in stress research. The median of sliding win-

dow increments for each physiological signal feature

was calculated for baseline and stress conditions. Ta-

ble 1 summarizes physiological signal features used

in this study.

2.2.1 Electrocardiograph (ECG)

Standard beats per minute (BPM), heart rate variabil-

ities (HRV), and its root mean squared (HRV

rms

) were

calculated from the ECG data. Moreover, features

of the HRV spectrum were calculated. HRV features

are critical because they can relate psychological pro-

cesses to physiological processes (Grossman and Tay-

lor, 2007). Therefore, to indicate physiological re-

sponse to stressful situations, we calculated the ratio

(HR

ratio

) of the area under the curve of low-frequency

bands (0.05 – 0.15 Hz) over high-frequency bands

(0.15 – 0.5 Hz), the HRVratio (Healey and Picard,

2005).

2.2.2 Plethysmograph (RESP)

Several relevant features were calculated from RESP

data. Respiration cycle features could provide

clues on blood-oxygen saturation, which is essen-

tial in maintaining proper organ function (Schmidt

et al., 2018). We calculated inhalation and ex-

halation duration (RESP

indur

& RESP

exdur

) and ra-

tio (RESP

ratio

), respiration duration (RESP

dur

), chest

stretch (RESP

stretch

), inhalation volume (RESP

invol

and respiration rate (RESP

rate

2.2.3 Respiratory Sinus Arrhythmia (RSA)

Respiratory sinus arrhythmia (RSA) describes irreg-

ularity in heart rate due to cardiac vagal efferent dis-

charge and time-alignment to breath cycles (Gross-

man and Taylor, 2007). It has been observed that

during inhalation, the heart beats faster and slower

during exhalation. In a highly stressful situation, hy-

perventilation could occur therefore increasing the oc-

currence of RSA (Campbell and Wisco, 2021; Tavel,

2021). Therefore, RSA is a multimodal feature, de-

pendent on both the ECG and RESP signals. The

RSA feature was calculated using the peak-valley

method(Grossman and Taylor, 2007) such that the

shortest beat interval during inhalation was subtracted

from the longest beat interval during exhalation. We

also calculated the beat number ratio between exhala-

tion over inhalation segments (RSA

ratio

). The respira-

tion window was extended 750 ms forward to account

for phase shifts between respiration and heart rates

in sync with respiratory rates (Grossman and Taylor,

2007).

2.2.4 Electrodermal Activity (EDA)

Skin conductance level (SCL) and skin conductance

response (SCR) are two components of the EDA sig-

nal. While SCL reﬂects general changes in autonomic

arousal, SCR indicates autonomic responses speciﬁc

to external stimuli (Boucsein, 2012). We separated

these components via regularized least-squares de-

trending method used by (Choi et al., 2011) in a previ-

ous study. However, no particular startle events were

noted in the Schmidt study; therefore, we included

only the statistical features from EDA’s SCL compo-

nents in our analysis.

2.2.5 Accelerometer (ACC) and Skin

Temperature (TEMP)

From ACC, we calculated an approximation of the

energy expenditure of each subject through the inte-

gral of the modulus of acceleration (IMA)(Karantonis

et al., 2006). Finally, we calculated the average skin

temperature (TEMP) for each subject within each

condition.

2.3 Psychometric Features

We calculated psychometric features, positive

(PE

score

) and negative (NE

score

), from each subject

based on their self-reported perceptions of emotional

descriptors within the PANAS questionnaires. Each

item in the questionnaire was scored using a 5-point

scale. A score of 1 indicates descriptor perception

as ‘Very Slightly’ or ‘Not at All’ while 5 as ‘Ex-

tremely.’ First, we grouped the questionnaire items

into 10 positive and 10 negative adjectives associated

with positive and negative emotions. Second, we

normalized the participants’ scores for each adjective

using min-max normalization, where 0 is the lowest

and 1 is the highest. We implemented a weighted

average (ρ

i j

, η

i j

) on the normalized scores (s

i j

) for

each subject (i) to calculate their PE

score

and NE

score

features such that

HEALTHINF 2022 - 15th International Conference on Health Informatics

432

Table 1: Summary of physiological signal features calculated from wearable data (Schmidt et al., 2018) for stress classiﬁca-

tion.

Summary of physiological signal features

Signal type Feature Description

ECG

BPM Beats per minute

HRV Heart rate variability

HRV

rms

Root mean squared of HRV

ratio

Ratio of low and high frequency bands of HRV spectrum

RESP

indur

Inspiration duration

RESP

exdur

Expiration duration

RESP

ratio

Ratio of RESP

indur

and RESP

exdur

RESP

dur

Overall respiration duration

RESP

stretch

Chest stretch due to respiration

RESP

invol

Inspiration volume

RESP

rate

Respiration rate

ECG & RESP RSA Respiratory sinus arrhythmia

EDA

SCL

mean

Average skin conductance level

SCL

std

Standard deviation of SCL

SCL

var

Variance of SCL

ACC IMA Integral of modulus of acceleration (energy expenditure)

TEMP TEMP Skin temperature

score,i

∑

j=1

i j

(1)

and

score,i

∑

j=1

i j

. (2)

The weights were calculated according to the ad-

jective groupings from a mood checklist(Crawford

and Henry, 2004). Based on the mood checklist,

negative emotions are weighed equally while positive

emotions have different subgroups hence have their

unique weights, and so a distinction between ρ

i j

and

i j

was made.

2.4 Feature Selection and Classiﬁcation

Using SPSS®(George and Mallery, 2019), we per-

formed stepwise feature selection, with a 95% conﬁ-

dence interval (α = 0.05), to determine a set of inputs

to a linear discriminant analysis (LDA) based classi-

ﬁer that will yield the best cross-validated (CV) clas-

siﬁcation accuracy.

Selected features were served as inputs to the

LDA-based classiﬁer to separate stressed individu-

als from those relaxed. Subject data was labeled as

‘stressed’ if they graded the corresponding descriptor

in the PANAS questionnaire between 2 and 5. Sub-

jects with a score of 1 for the same descriptor were la-

beled as ‘relaxed’ or ‘baseline.’ In total, 7 LDA mod-

els were generated, i.e., physiological only (Φ), psy-

chometric PE

score

(Ψ

), psychometric NEscore (Ψ

psychometric only (Ψ), physiological and psychome-

tric PE

score

(Φ, Ψ

), physiological and psychometric

score

(Φ, Ψ

), and ﬁnally, physiological and psy-

chometric (Φ, Ψ). Due to sample size, prior prob-

abilities were adjusted based on the number of par-

ticipants in each group. We validated each classi-

ﬁer model using the leave-one-out cross-validation

(LOOCV) method (George and Mallery, 2019).

3 RESULTS

Eight (8) participants experienced some degree of

stress during a controlled stress test, while 7 partici-

pants minimally experience stress. Participant reports

served as the ground truth for classiﬁcation. Both

physiological and psychometric feature types were

fed to the LDA-based classiﬁer. Table 2 shows the

detailed classiﬁcation results and the confusion tables

for only 5 LDA-based models since only the combi-

nations of PE

score

with physiological signal features

improved classiﬁcation accuracy.

Classiﬁcation accuracy of 93.3% can be achieved

with only 3 physiological features, while 86.7% can

be achieved using the PE

score

feature. Furthermore,

when stepwise feature selection was applied to the

combined physiological and psychometric features,

3 physiological features and PE

scores

improved to a

perfect classiﬁcation (100%), improving the previous

results presented by the benchmark study (Schmidt

et al., 2018) using the same binary classiﬁcation

Enhancement of Physiological Stress Classiﬁcation using Psychometric Features

433

Table 2: Evaluation of stress classiﬁcation performance

via leave-one-out cross validation (LOOCV) of linear dis-

criminant analysis-based classiﬁer models. O: original, P:

predicted, Φ: physiological only, Ψ: psychometric only,

: psychometric PE

score

, Ψ

: psychometric NE

score

, and

score

(Φ, Ψ

): physiological and psychometric PE

score

Model

Confusion Matrix

LOOCV (%)

R S

R 7 0

93.3

S 1 7

R 6 1

86.7

S 1 7

R 6 1

86.7

S 1 7

R 6 1

80.0

S 2 6

Φ, Ψ

R 7 0

100.0

S 0 8

method by 8%

. In addition, our results show that

such classiﬁcation accuracy could be achieved with

20 times fewer physiological signal features com-

pared to the same study above. Although the results

obtained are from a small dataset, our results demon-

strate potential in the joint analysis of physiological

and psychometric features to enhance stress classiﬁ-

cation.

Table 3 shows the descriptive statistics of the se-

lected features, RSA

ratio

, HRV

ratio

, IMA, and PE

score

We also show in Figure 2 the comparison of the distri-

bution of values of a physiologically dominant signal

feature, RSA

ratio

, and the dominant psychometric fea-

ture PE

score

We demonstrate that stressful conditions increase

an individual’s RSA

ratio

values. Speciﬁcally, the av-

erage number of beats during exhalation was higher

than during inhalation in stressed individuals, which

explains their higher RSA

ratio

values.

Furthermore, we observed a signiﬁcant increase

in HRV

ratio

values in stressed individuals, which re-

ﬂects the degree of sympathetic over parasympa-

thetic nervous activity, respectively. A closer ex-

Table 3: Selected features for stress classiﬁcation and fea-

tures descriptive statistics.

Features Relaxed (n = 7) Stressed (n = 8)

RSA

ratio

0.59 ± 0.06 0.67 ± 0.04

HRV

ratio

2.77 ± 1.77 5.28 ± 4.43

IMA 53, 575 ± 8, 625 50, 484 ± 5, 428

score

0.28 ± 0.11 0.56 ± 0.16

(Φ, Ψ

) and (Φ, Ψ) are not shown in table 2 since they

do not improve LOOCV results beyond (Φ, Ψ

) provided,

therefore are redundant.

Figure 2: Distribution comparison between relaxed and

stressed states of a dominant physiological signal feature,

RSA

ratio

(A), and a psychometric feature, PE

score

(B), for

stress classiﬁcation.

amination of the area under the low-frequency HRV

spectrum reveals that, while it remained higher than

its high-frequency component for both states, there

was a signiﬁcant increase in its median values in

stressed individuals compared to those who were re-

laxed (stressed: 74% vs. relaxed: 56%). Such an ob-

servation indicates that individuals under stress have

a greater sympathetic nervous response (i.e., ﬁght-or-

ﬂight) than those in a neutral state, as corroborated by

earlier studies(Choi et al., 2011; Healey and Picard,

2005).

Energy expenditure, as approximated by IMA, de-

creased in individuals under stress. Due to the nature

of IMA calculation (i.e., the sum of integrals of each

accelerometer axis channel), a decrease in an indi-

vidual’s IMA indicates a reduction in an individual’s

overall movement during stress, decreasing their en-

ergy expenditure. This observation could be due to an

evolutionary tendency to conserve energy when faced

with challenges (Pontzer, 2015); however, this war-

rants further investigation within the scope of stress

research in an ambulatory setting as well as an exam-

ination of the stress protocol of a study.

Interestingly, one of our psychometric features,

score

, is also an important feature for stress classiﬁ-

cation. Although PE

score

and NE

score

increased for in-

dividuals during stress, 78% and 50% more than base-

line values, respectively, PE

score

remained higher than

score

in both states suggesting participants experi-

enced eustress (positive stress) through a stress pro-

tocol indicating an overall positive engagement. Nat-

urally, as they were subjected to stressful situations,

individuals’ distress (negative stress) became more

prominent. Different stress types have been previ-

ously described (Le Fevre et al., 2003), and the trend

shown from our psychometric features is concordant

to their observations.

HEALTHINF 2022 - 15th International Conference on Health Informatics

434

It is important to note that PE

score

are based on

the weighted average of the positive mood descriptors

as assessed by each individual. While we achieved a

perfect classiﬁcation with the inclusion of this psy-

chometric feature, we did not include in its calcu-

lation a participant’s perceived stress score. Many

versions of the PANAS questionnaire have been pre-

viously introduced (i.e., PANAS-C(Laurent et al.,

1999), I-PANAS-SF(Thompson, 2007), and PANAS-

X(Watson and Clark, 1999)) however no version of

the PANAS questionnaire include a ’Stressed’ de-

scriptor. A ’Stressed’ descriptor was added by the au-

thors of WESAD in their PANAS questionnaire to suit

the goals of the original classiﬁcation study. There-

fore, the perceived stress score of a subject was not

accounted in the calculation of our psychometric fea-

tures and was used only as a label for binary classiﬁ-

cation as described above.

4 DISCUSSIONS

We introduced new psychometric features to enhance

stress classiﬁcation using physiological data from

wearable technologies obtained from WESAD. Our

psychometric features, derived from traditional ques-

tionnaires from psychological evaluations, were used

to perform a joint analysis with physiologically rele-

vant features for stress research. More importantly,

we demonstrated that stress classiﬁcation could be

improved when both the physiological and psycho-

logical components of stress are considered.

In comparison, the work by Nkurikiyeyezu et al

(Nkurikiyeyezu et al., 2019) provided a simple frame-

work for creating person-speciﬁc models to predict

stress using only physiological signals. Their pro-

posed framework achieved a classiﬁcation accuracy

of 95.2%, comparable to the results of the original

WESAD study. Another WESAD-based study by Lai

et al (Lai et al., 2021) introduces a stress monitoring

assistant that demonstrates a 96.7% binary stress clas-

siﬁcation accuracy using only features of chest-based

physiological signals. However, the above examples

do not account for the available psychological data as

part of the feature pool for later classiﬁcation and pre-

diction tasks.

While our work does not introduce any stress

models or frameworks, we emphasize the importance

of a joint analysis of physiological and psychologi-

cal data to improve stress classiﬁcation tasks. As pre-

viously described through the allostatic load model,

physiological and psychological systems are closely

intertwined, if not cascaded. A physiological re-

sponse could be a result of psychological stressors

and vice versa. Moreover, physiological and psycho-

logical systems could provide concerted reactions to

external stressors. While the relationship of physi-

ological and psychological systems has been previ-

ously established in stress research, the contribution

of psychological data to stress classiﬁcation tasks is

under-recognized.

Self-reports in a dynamic environment are chal-

lenging to validate and replicate due to their inherent

subjectivity. Person-speciﬁc models and frameworks

that consider psychological data also require regu-

lar calibration and updates as both physiology and

psychology change over time(Nkurikiyeyezu et al.,

2019), especially in the context of stress and allostatic

load since they are only observed after extended ex-

posure to stressors. However, as we demonstrated in

this work, when physiological and psychological data

are jointly analyzed, stress classiﬁcation tasks could

be improved, and the number of physiological signal

features needed could be greatly reduced. Evidence

and recommendations from previous works(Schmidt

et al., 2018; Hovsepian et al., 2015; Sarker et al.,

2016; Plarre et al., 2011) provide further motivation

for the inclusion of psychological data in stress clas-

siﬁcation tasks.

While we present good stress classiﬁcation accu-

racy, we acknowledge that our work has limitations.

We recognize that our stress classiﬁcation task could

be affected by choice of features and pre-processing

methods. Recommendations for analysis window du-

rations for different physiological signals are in the

literature (i.e., HRV (Shaffer and Meehan, 2020) and

EDA (Boucsein, 2012)). Our choice of analysis win-

dow is in keeping with the WESAD benchmark study

(Schmidt et al., 2018), which was based on previ-

ous work by (Kreibig, 2010). On the other hand,

our choice of features was based on existing literature

cited presented in our Methods section. We were care-

ful to select our signal features so that they could ad-

equately represent a stress response. While this strat-

egy naturally decreases the amount of features fed in

to an classiﬁcation algorithm, it ensures sufﬁcient rep-

resentation of an individual’s autonomic activity re-

lated to external stressors.

We also recognize that our choice of database lim-

its the sample size on which we could test our new

psychometric features. Indeed, a 15-subject cohort is

a small sample size. Subject recruitment for human-

centric studies in biomedical or biomedical-adjacent

ﬁelds remains a signiﬁcant challenge especially for

short-term studies such as acute stress. Despite the

limited number of subjects, we ensured that our meth-

ods for classiﬁcation and cross-validation were suit-

able for small sample sizes, such as LOOCV.

Enhancement of Physiological Stress Classiﬁcation using Psychometric Features

435

Future work aims to collect and analyse data from

a larger cohort of subjects to further test and validate

our psychometric features. In addition, we aim to use

our validated psychometric features to create a holis-

tic predictive model concordant with the established

allostatic load model(McEwen and Rasgon, 2018),

giving greater focus on chronic stress and its effect

on a person’s overall well-being.

5 CONCLUSIONS

We demonstrated the importance of including psy-

chological data in an acute stress study. Through a

joint analysis of physiological and psychological fea-

tures, we showed that stress classiﬁcation could be

enhanced. Furthermore, accounting for psychomet-

ric data reduces the number of physiological signal

features needed stress classiﬁcation. We also found

that our psychometric features could aid in identifying

the type of stress (eustress or distress) an individual

perceives, as indicated by a self-assessment question-

naire’s independent contributions of each mood de-

scriptor (affect). Our work provides an incremental

step towards translating affect linked to stress to suit-

able quantitative measurements similar to those of-

fered by physiological sensors. Joint analysis of psy-

chological and physiological data could be beneﬁcial

towards the detection and management of stress. Fur-

thermore, our work could support the future devel-

opment of holistic stress models consistent with the

well-established allostatic load model. Such models

could be beneﬁcial for workers in harsh environments

like healthcare and personal support workers.

REFERENCES

Asuncion, A. and Newman, D. (2007). Uci machine learn-

ing repository.

Boucsein, W. (2012). Electrodermal activity. Springer Sci-

ence & Business Media.

Campbell, A. A. and Wisco, B. E. (2021). Respiratory si-

nus arrhythmia reactivity in anxiety and posttraumatic

stress disorder: A review of literature. Clinical Psy-

chology Review, page 102034.

Choi, B., Jebelli, H., and Lee, S. (2019). Feasibility analy-

sis of electrodermal activity (eda) acquired from wear-

able sensors to assess construction workers’ perceived

risk. Safety science, 115:110–120.

Choi, J., Ahmed, B., and Gutierrez-Osuna, R. (2011). De-

velopment and evaluation of an ambulatory stress

monitor based on wearable sensors. IEEE trans-

actions on information technology in biomedicine,

16(2):279–286.

Crawford, J. R. and Henry, J. D. (2004). The positive and

negative affect schedule (panas): Construct validity,

measurement properties and normative data in a large

non-clinical sample. British journal of clinical psy-

chology, 43(3):245–265.

de Fazio, R., Cafagna, D., Marcuccio, G., Minerba, A., and

Visconti, P. (2020). A multi-source harvesting system

applied to sensor-based smart garments for monitor-

ing workers’ bio-physical parameters in harsh envi-

ronments. Energies, 13(9):2161.

Edmondson, D., Green, P., Ye, S., Halazun, H. J., and

Davidson, K. W. (2014). Psychological stress and 30-

day all-cause hospital readmission in acute coronary

syndrome patients: an observational cohort study.

PloS one, 9(3):e91477.

Fava, G. A., McEwen, B. S., Guidi, J., Gostoli, S., Ofﬁ-

dani, E., and Sonino, N. (2019). Clinical characteri-

zation of allostatic overload. Psychoneuroendocrinol-

ogy, 108:94–101.

George, D. and Mallery, P. (2019). IBM SPSS statistics 26

step by step: A simple guide and reference. Routledge.

Grossman, P. and Taylor, E. W. (2007). Toward understand-

ing respiratory sinus arrhythmia: Relations to cardiac

vagal tone, evolution and biobehavioral functions. Bi-

ological psychology, 74(2):263–285.

Guidi, J., Lucente, M., Sonino, N., and Fava, G. A.

(2021). Allostatic load and its impact on health: a

systematic review. Psychotherapy and psychosomat-

ics, 90(1):11–27.

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.

Hovsepian, K., Al’Absi, M., Ertin, E., Kamarck, T., Naka-

jima, M., and Kumar, S. (2015). cstress: towards a

gold standard for continuous stress assessment in the

mobile environment. In Proceedings of the 2015 ACM

international joint conference on pervasive and ubiq-

uitous computing, pages 493–504.

Karantonis, D. M., Narayanan, M. R., Mathie, M., Lovell,

N. H., and Celler, B. G. (2006). Implementa-

tion of a real-time human movement classiﬁer us-

ing a triaxial accelerometer for ambulatory monitor-

ing. IEEE transactions on information technology in

biomedicine, 10(1):156–167.

Kreibig, S. D. (2010). Autonomic nervous system activ-

ity in emotion: A review. Biological psychology,

84(3):394–421.

Lai, K., Yanushkevich, S. N., and Shmerko, V. P. (2021).

Intelligent stress monitoring assistant for ﬁrst respon-

ders. IEEE Access, 9:25314–25329.

Laurent, J., Catanzaro, S. J., Joiner Jr, T. E., Rudolph, K. D.,

Potter, K. I., Lambert, S., Osborne, L., and Gathright,

T. (1999). A measure of positive and negative affect

for children: Scale development and preliminary vali-

dation. Psychological assessment, 11(3):326.

Le Fevre, M., Matheny, J., and Kolt, G. S. (2003). Eu-

stress, distress, and interpretation in occupational

stress. Journal of managerial psychology.

HEALTHINF 2022 - 15th International Conference on Health Informatics

436

Legault, G., Clement, A., Kenny, G. P., Hardcastle, S., and

Keller, N. (2017). Cognitive consequences of sleep

deprivation, shiftwork, and heat exposure for under-

ground miners. Applied ergonomics, 58:144–150.

McEwen, B. S. and Rasgon, N. L. (2018). The brain and

body on stress: Allostatic load and mechanisms for

depression and dementia.

Nkurikiyeyezu, K., Yokokubo, A., and Lopez, G. (2019).

The effect of person-speciﬁc biometrics in improv-

ing generic stress predictive models. arXiv preprint

arXiv:1910.01770.

O’Connor, D. B., Thayer, J. F., and Vedhara, K. (2021).

Stress and health: A review of psychobiological pro-

cesses. Annual review of psychology, 72:663–688.

Pappa, S., Ntella, V., Giannakas, T., Giannakoulis, V. G.,

Papoutsi, E., and Katsaounou, P. (2020). Prevalence

of depression, anxiety, and insomnia among health-

care workers during the covid-19 pandemic: A sys-

tematic review and meta-analysis. Brain, behavior,

and immunity, 88:901–907.

Plarre, K., Raij, A., Hossain, S. M., Ali, A. A., Naka-

jima, M., Al’Absi, M., Ertin, E., Kamarck, T., Kumar,

S., Scott, M., et al. (2011). Continuous inference of

psychological stress from sensory measurements col-

lected in the natural environment. In Proceedings of

the 10th ACM/IEEE international conference on in-

formation processing in sensor networks, pages 97–

108. IEEE.

Pontzer, H. (2015). Energy expenditure in humans and other

primates: a new synthesis. Annual Review of Anthro-

pology, 44:169–187.

Runkle, J. D., Cui, C., Fuhrmann, C., Stevens, S., Del Pinal,

J., and Sugg, M. M. (2019). Evaluation of wearable

sensors for physiologic monitoring of individually ex-

perienced temperatures in outdoor workers in south-

eastern us. Environment international, 129:229–238.

Sarker, H., Tyburski, M., Rahman, M. M., Hovsepian, K.,

Sharmin, M., Epstein, D. H., Preston, K. L., Furr-

Holden, C. D., Milam, A., Nahum-Shani, I., et al.

(2016). Finding signiﬁcant stress episodes in a dis-

continuous time series of rapidly varying mobile sen-

sor data. In Proceedings of the 2016 CHI conference

on human factors in computing systems, pages 4489–

4501.

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.

Shaffer, F. and Meehan, Z. M. (2020). A practical

guide to resonance frequency assessment for heart rate

variability biofeedback. Frontiers in neuroscience,

14:1055.

Tavel, M. E. (2021). Hyperventilation syndrome: Why is

it regularly overlooked? The American Journal of

Medicine, 134(1):13–15.

Thompson, E. R. (2007). Development and validation of

an internationally reliable short-form of the positive

and negative affect schedule (panas). Journal of cross-

cultural psychology, 38(2):227–242.

Vitale, E., Galatola, V., and Mea, R. (2021). Observational

study on the potential psychological factors that af-

fected italian nurses involved in the covid-19 health

emergency. Acta Bio Medica: Atenei Parmensis,

92(Suppl 2).

Watson, D. and Clark, L. A. (1999). The panas-x: Manual

for the positive and negative affect schedule-expanded

form.

Wu, F., Wu, T., and Yuce, M. R. (2019). Design and im-

plementation of a wearable sensor network system for

iot-connected safety and health applications. In 2019

IEEE 5th World Forum on Internet of Things (WF-

IoT), pages 87–90. IEEE.

Enhancement of Physiological Stress Classiﬁcation using Psychometric Features

437