Thermal Comfort and Stress Recognition in Ofﬁce Environment

Kizito Nkurikiyeyezu, Kana Shoji, Anna Yokokubo and Guillaume Lopez

Graduate School of Science and Engineering, Aoyama Gakuin University,

Keywords:

Personalized Thermal Comfort, Work Stress, Heart Rate Variability, Smart Buildings, Humanized Computing,

Affective Computing.

Abstract:

Work stress and thermal discomfort are some of the hurdles that ofﬁce workers face every day. Ofﬁce workers

experience a periodic work stress because work is long and mentally challenging. At the same time, current

thermal comfort provision technologies are inefﬁcient and consume a large amount of energy. In our previ-

ous works, we proposed an efﬁcient thermal comfort provision system that is based on a person

s heart rate

variability (HRV). However, because work stress can also affect the person

s HRV, this paper investigates the

possibility to distinguish HRV changes that are due to thermal discomfort from changes that emanate from

work stress. We conducted experiments on subjects taking Advanced Trail Making Test (ATMT) and observed

that stress alters HRV and that it is possible to distinguish stressed and non-stressed subjects with a 100%

accuracy. We validated our method on the multimodal SWELL knowledge work (SWELL-KW) stress dataset

and achieved similar results (99.25% accuracy and 99.75% average recall). Further analysis suggests that,

although both thermal comfort and work stress affect HRV, their effect is perhaps non-overlapping, and that

the two can be distinguished with a near-perfect accuracy. These results indicate that it could be possible to

design an automatic and unobtrusive system that delivers thermal comfort and predicts work stress based on

people's HRV

1 INTRODUCTION

Stress and the dissatisfaction in thermal comfort are

among the biggest challenges in modern workplaces.

In Europe, it is estimated that stress affects half of

the workers, is responsible for half of all lost work-

ing days, is the second reason for most work-related

health problems and causes signiﬁcant losses to busi-

nesses due to absenteeism and high employee turnover

(EU-OSHA, 2017). As for thermal comfort dissatisfac-

tion, although thermal comfort has been actively stud-

ied for almost a century, currently, its provision mech-

anisms are based on fundamentally ﬂawed assump-

tions, achieve a lackluster performance and require an

excessive energy to operate (Nicol and Roaf, 2017;

Brager et al., 2015). Indeed, thermal comfort is de-

ﬁned as “the condition of mind that expresses satisfac-

tion with the thermal environment and is assessed by

subjective evaluation” (ANSI/ASHRAE, 2013). Thus,

it is a personal psychological sensation. On the con-

trary, HVAC systems are based on heat and energy

transfer principles; thus, can only inﬂuence the per-

son

s thermal comfort. However, they fail to account

for the complexity and the dynamics of people

s ther-

moregulation and cannot assess the satisfaction of the

provided thermal comfort. Further, they ignore some

important precursors to thermal comfort such as peo-

ple

s psychophysics, their genders, their thermal adap-

tation, their physiological makeup, and their age dif-

ferences. Consequently, prominent researchers have

urged a paradigm shift in the way thermal comfort is

provided (Nicol and Roaf, 2017; Brager et al., 2015).

Unfortunately, in the future, if nothing is done to

deal with the above issues, work-related stress and

thermal discomfort can only increase. On one hand,

modern workplaces require stressful, high workload

and long working hours. On the other hand, recently,

there are policies to curtail agents of anthropogenic cli-

mate change that impose, inter-alia, strict reductions

in energy use in buildings. These policies, given the

limitations of current thermal comfort provision tech-

nologies, can only aggravate the level of thermal dis-

comfort in ofﬁces. Hence, this dilemma entails alterna-

tive thermal comfort delivery mechanisms that provide

higher quality thermal comfort at lower energy.

In our previous works (Nkurikiyeyezu et al., 2017),

we proposed to provide thermal comfort based on a

person

s physiological changes due to their surround-

256

Nkurikiyeyezu, K., Shoji, K., Yokokubo, A. and Lopez, G.

Thermal Comfort and Stress Recognition in Ofﬁce Environment.

DOI: 10.5220/0007368802560263

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 256-263

ISBN: 978-989-758-353-7

ing thermal environment and discussed the beneﬁts

and approaches to designing a personalized, real-time

and energy efﬁcient thermal comfort delivery appara-

tus that is based on the person’s estimated thermal

comfort sensation (Nkurikiyeyezu et al., 2018). Our

system (Lopez et al., 2018; Nkurikiyeyezu and Lopez,

2018) consists of an unobtrusive wrist-wearable de-

vice that records a person’s photoplethysmogram

(PPG) signal. The person’s heart rate variability (HRV)

is calculated from the PPG signal and is used to pre-

dict, in real-time, the person’s thermal comfort with an

accuracy greater than 90%. There were, however, con-

cerns about the system

s performance in real-world set-

tings. Certainly, HRV can change dues to other reasons

including physiological, pathological, neuropsycho-

logical, lifestyle factors, and, most importantly, stress

(Kim et al., 2018). This is because both thermal com-

fort and stress are regulated via a complementary ac-

tion of the sympathetic nervous system (SNS) and the

parasympathetic nervous system (PSN). Succinctly,

when the brain perceives a danger, the SNS ushers

a release of hormones to prepare the body to swiftly

react to life-threatening situations (for example, faster

heartbeat and an increase in airways for easier breath-

ing) and inhibits some non-urgent physiological needs

(e.g., digestion and sexual arousal). When the per-

ceived danger ﬁnally recedes, the PSN restores the

normal hormonal balance (Harvard Health, 2011). In

the case of thermal comfort, the brain’s hypothala-

mus serves as a “thermostat” of the body (Figure 1).

In a nutshell, the hypothalamus receives sensory in-

puts from thermo-receptors located in the skin, liver

and skeletal muscles and initiate appropriate processes

to keep constant the body

s core temperature. For

instance, when it is hot, the hypothalamus activates

heat dissipating and body cooling mechanisms such as

sweating and vasodilation. Conversely, when it is cold,

the hypothalamus activates thermogenesis mechanism

(e.g., shivering in skeletal muscle and heat generation

in brown adipose tissues) and other mechanisms to re-

duce heat dissipation (e.g.,cutaneous vasoconstriction

and piloerection) (Morrison, 2011). The human ther-

moregulation can be indirectly monitored via e.g., the

person’s heart rate variability (HRV). There is indeed

evidence that the very-low-frequency (VLF) band of

the HRV power spectra mirrors thermoregulatory vaso-

motor control activities (Thayer et al., 1997) and that

thermal discomfort generally alters heartbeat patterns

(Liu et al., 2008). Likewise, when in a stressful situa-

tion, the brain’s amygdala decodes the stress and sends

an appropriate message to the hypothalamus. The hy-

pothalamus, in turn, facilitates the release of hormones

such as adrenaline in the blood; thus, increases the per-

son’s heart rhythms, his respiration rate, his blood pres-

sure, and his pulse rate. Due to this hormonal changes,

the person becomes alert, his hearing and sight be-

come sharper and his blood sugar increases (Harvard

Health, 2011). The PNS restores this hormonal imbal-

ance when the stressor abates. However, if the stressor

persists for a long time, the hypothalamus keeps the

person on an enduring high alert by activating the cen-

tral stress response system —the hypothalamic pitu-

itary adrenal (HPA) axis —which releases cortisol, a

well-known biomarker of chronic stress (Schulz et al.,

1998).

Hypothalamus

Central

thermoreceptors

Peripheral

thermoreceptors

[Cold][Hot]

Thermal comfort decision

Temperature reduction

mechanisms

Temperature generation

mechanisms

Sweat glands

Vascular smooth

muscles

Brown adipose

tissues

Somatic

muscles

Transpiration VasodilatationVasoconstriction Heat production Shiver

Cholinergic neurons of the

sympathetic nervous system

Noradrenergic neurons of the

sympathetic nervous system

Somatic motor

neurons

Figure 1: A simpliﬁed process of human thermal regulation

—the hypothalamus checks the body’s core temperature and

starts necessary thermogenesis or heat dissipating processes

to maintain the core temperature [adapted from (Wikilivres,

2018)].

Over the past decade, there have been numerous

attempts to automate stress recognition. (Lopez et al.,

2014) developed an adaptive method that uses phys-

iological indices integrated into an intelligent multi-

steps discrimination process to predict stress occur-

rence and other workplace stress type. Their meth-

ods predicts workplace stress type with an accuracy

of 64%. In their review article, Alberdi and his coau-

thors (Alberdi et al., 2016) discuss various modalities

for stress recognition. They further propose a frame-

work for an automatic stress recognition system that

records a person’s physiological signals (e.g., elec-

troencephalogram (EEG), electrocardiography (ECG),

blood volume pulse(BVP), pupil diameter, eye gaze,

blinking, electrodermal activity (EDA), electromyo-

gram (EMG), respiration, etc..), behavioral (keystroke,

mouse usage, posture, facial expression, speech, mo-

bile phone usage, etc), environmental and contextual

(calendar events, location and ambiance) data to pre-

dict the likelihood of the person’s stress. Koldijk

(Koldijk et al., 2018) analogously used an amalgam

of sensor data (e.g., computer logging, posture, facial

expression and physiological features) and developed

a stress classiﬁer that achieves a 90% accuracy. This is

an important achievement—especially since the stress

prediction was based on data recorded in realistic of-

ﬁce settings. Nevertheless, in their paper, they failed

Thermal Comfort and Stress Recognition in Ofﬁce Environment

257

to use physiological signal to reliably predict stress;

thus, concluded that physiological reaction to ofﬁce

stress is not a strong and reliable indicator of stress in

an ofﬁce setting. However, while it could be possible

to predict stress using a fusion of many sensors data,

this approach raises both practical challenges (e.g.,

multimodal data acquisition, data fusion and, data in-

tegration) and user privacy concerns (e.g., the impli-

cation of user’s computer keystroke logging, video

and speech recording), and, may not be used in real-

world settings because of company-wide computer se-

curity policies or due to international workplace pri-

vacy laws.

This paper aims to address two research questions:

Is it possible to distinguish stressful from non-

stressful situations in a typical workplace using only

a single unobtrusive sensor? and

Is it possible to

distinguish HRV due to thermal distress and that due

to work stress? Concretely, with this paper, we make

the following contributions:

We achieve a 99.25% stress prediction accuracy.

We believe this is the ﬁrst published work that

achieves such a high stress prediction accuracy.

Further, our approach is based on only the person’s

HRV; thus, it could be possible to detect stress in

workplace using a single unobtrusively sensor (for

example by using a commercial photoplethysmo-

gram (PPG) wrist-worn smart-watch)

2. We show that, although both thermal comfort and

work stress affect HRV, their effect on a person

HRV is perhaps non-overlapping and that the two

can be distinguished with a 99.25% accuracy,

99.27% precision, and a 99.60% micro-average re-

call.

2 METHODS

2.1 HRV Indices Calculations

HRV indices listed in Table 1 were computed as fol-

lows: First, we extracted an inter-beat interval (IBI)

signal from the peaks of the Electrocardiography

(ECG) of each subject. Then, each HRV index is com-

puted on a 5 minutes IBI array. A new IBI sample is

appended to the IBI array while the oldest IBI sample

is removed from its beginning. The new resulting IBI

array is used to compute the next HRV index. This pro-

cess is repeated until the end of the entire IBI signal.

Unlike other HRV computation methods proposed by

other researchers —which mostly consist of comput-

ing HRV on the whole signal— in our previous studies,

we noticed that this approach allows a more granular,

detailed and accurate study of how each heartbeat in-

ﬂuences the person’s HRV.

2.2 Datasets

The study is based on three datasets described below:

2.2.1 PHRENIC Dataset

The PHRENIC (so named because it aims at under-

standing the mind) is based on an experiment we con-

ducted on 20 subjects. The aim of the experiment

was to study the subjects’ mental fatigue. The sub-

jects were randomly divided into two groups and as-

signed the same number of tasks. Subjects in group A

were told the number of tasks they have to complete,

while subjects in group B were not aware of the exact

number of tasks they will complete. During the ex-

periment, each subject took an advanced trail making

test (ATMT), a neuropsychological test of visual atten-

tion and task switching (Mizuno and Watanabe, 2008).

The objective was to evaluate their performance dur-

ing mental fatigue. We developed an application that

runs on a tablet with a touchscreen. The application

generates a random set of number on the subject’s

tablet’s screen, and the subject is requested to connect

them in sequence as fast as he can. Additionally, the

app logs the subject’s performance on the test. Each

experiment ended when the subject completed all the

assigned tasks. During the experiment, each subject

wore myBeat (Union Tool Co., Ltd), a small (41mm x

35.5mm x 10mm), light (25g) and unobtrusive wear-

able sensor to record the subject’s heartbeats at 1000

Hz. The study reported in this paper was approved by

the Aoyama Gakuin University’s Institutional Ethic

Review Board (No. M1-15, 2015/7/9).

2.2.2 SWELL Dataset

The SWELL knowledge work (SWELL-KW) dataset

was collected by researchers at the Institute for Com-

puting and Information Sciences at Radboud Univer-

sity and is described in details in (Koldijk et al., 2014).

It is a result of experiments conducted on 25 subjects

doing typical ofﬁce work (for example writing reports,

making presentations, reading e-mail and searching

for information). The subject went through typical

working stressors such as receiving unexpected emails

interruptions and pressure to complete their work on a

tight schedule. The experiment recorded various data

including computer logging, facial expression, body

postures, ECG signal and skin conductance. Each par-

ticipant went through three different working condi-

tions:

HEALTHINF 2019 - 12th International Conference on Health Informatics

258

Table 1: Description of the selected HRV indices.

HRV index Short description Equation/Reference

MEAN_RR Mean of all RR intervals

MEDIAN_RR Median of all RR intervals

SDRR Standard deviation of all interval

RMSSD

Square root of the mean of the sum of the squares of the difference

between adjacent RR intervals

√

∑

N−1

i=1

(RR

i+1

−RR

)

N−1

SDSD

Standard deviation of all interval of differences between adjacent

RR intervals

SDRR_RMSSD Ratio of SDRR over RMSSD

HR Heart Rate (beats per minute)

pNN25 % of adjacent RR intervals differing by more than 25 ms

∑

i=1



−R

i+1

>25ms



N−1

pNN50 % of adjacent RR intervals differing by more than 50 ms

∑

i=1



−R

i+1

>50ms



N−1

SD1 Poincaré plot descriptor of the short-term HRV

variance(

−RR

i+1

√

)

SD2 Poincaré plot descriptor of the long-term HRV

variance(

+RR

i+1

√

)

KURT Kurtosis of all RR intervals

SKEW Skewness of all RR intervals

MEAN_REL_RR Mean of all relative RR intervals see. (Vollmer, 2015)

MEDIAN_REL_RR Median of all relative RR intervals see. (Vollmer, 2015)

SDRR_REL_RR Standard deviation of all relative RR interval see. (Vollmer, 2015)

RMSSD_REL_RR

Square root of the mean of the sum of the squares of the difference

between adjacent relative RR intervals

see. (Vollmer, 2015)

SDSD_REL_RR

Standard deviation of all interval of differences between adjacent

relative RR intervals

see. (Vollmer, 2015)

SDRR_RMSSD_REL Ratio of SDRR_REL over RMSSD_REL

KURT_REL_RR Kurtosis of all relative RR intervals see. (Vollmer, 2015)

SKEW_REL_RR Skewness of all relative RR intervals see. (Vollmer, 2015)

VLF

Very low (0.003Hz - 0.04Hz) frequency band of the HRV power

spectrum

see (Malik et al., 1996)

Low (0.04Hz - 0.15Hz) frequency band of the HRV power spec-

trum

see (Malik et al., 1996)

High (0.15Hz - 0.4Hz) frequency band of the HRV power spec-

trum

see (Malik et al., 1996)

TP Total HRV power spectrum see (Malik et al., 1996)

LF/HF Ratio of LF to HF see (Malik et al., 1996)

HF/LF Ratio of HF to LF see (Malik et al., 1996)

sampen Sample entropy of the RR signal see (Sassi et al., 2015)

higuci Higuchi Fractal Dimension see(Gomes et al., 2017)

no stress: the subjects are allowed to work on the

tasks as long as they needed for a maximum of 45

minutes but they are not aware of the maximum

duration of their tasks.

time pressure: during this time, the time to ﬁnish

the task was reduced to 2/3 of the time the partici-

pant took in the “no stress” condition.

interruption: the participants received eight emails

in the middle of their assigned tasks. Some emails

were relevant to their tasks —and the participant

was requested to take speciﬁc actions—while oth-

ers were just irrelevant to their tasks.

The experiment lasted for about 3 hours for each sub-

ject. We make available the computed HRV dataset

and it can be used by other researchers to validate our

ﬁndings

2.2.3 COMFORT Dataset

This dataset is a result of our previous work on ther-

mal comfort prediction(Nkurikiyeyezu et al., 2017). In

summary, we recorded ECG signals of 17 subjects in

three different thermal chambers whose thermal set-

tings correspond to cold, a neutral, and a hot sensation

on the American Society of Heating, Refrigerating and

Air-Conditioning Engineers (ASHRAE) thermal sen-

sation scale. The experiment lasted for about 30 min-

utes. From the recorded ECG, we extracted HRV in-

The dataset available at https://bit.ly/2RZWJ3Z

Thermal Comfort and Stress Recognition in Ofﬁce Environment

259

dices that were used to predict the thermal comfort.

2.2.4 Datasets Merging

One of the objectives of the study is to investigate

the possibility to distinguish variation in HRV due to

thermal comfort from that due to work stress. There-

fore, we merged together the three datasets (SWELL,

PHRENIC, COMFORT) to see if the resulting HRV

datasets would allow distinguishing stress from ther-

mal comfort. Generally, there are pitfalls in merging

two datasets that were recorded in different environ-

ments because there could be unknown bias. In our

case, however, because the functionality of the heart

does not depend much on the physical location (as-

suming other variables stay the same) of where the ex-

periment was conducted, we surmised that, unless the

person is both stressed and thermally dis-comfortable,

any momentary aberration in HRV are due to either

work stress or thermal discomfort. We tested this as-

sumption by adding a dataset_id as a control predic-

tion feature to the datasets in order to indicate the

parent dataset of each HRV value before merging the

datasets. This would allow us to probe how much the

model is biased by analyzing how much the dataset_id

contributes to the classiﬁcation model. This evaluation

was done by comparing dataset_id’s mean decrease in

impurity (MDI) (Equ.1) to that of other features of the

model.

∑

k=1



1 −p



(1)

where K is the total number of HRV indices used for

classiﬁcation and

the proportion of a single HRV

feature k. After merging the datasets, we obtained two

new datasets:

PHRENIC plus SWELL —dataset resulting from

combining the PHRENIC and the SWELL

datasets. In this case, for the PHRENIC dataset, we

assumed that subjects in group A were under time

pressure and those in group B were not stressed.

SWELL plus PHRENIC plus COMFORT —

dataset resulting from combining the SWELL, the

PHRENIC and the COMFORT dataset

2.2.5 Dataset Re-sampling

The dataset resulting in merging the PHRENIC and

the SWELL dataset and that resulting from merging

the PHRENIC, the SWELL and COMFORT dataset

are inherently unbalanced because they have differ-

ent duration —thus, have different number of samples.

We addressed this issue by oversampling the minority

classes using synthetic minority over-sampling tech-

nique (SMOTE) (Chawla et al., 2002). The new over-

sampled dataset may, however, contain noisy samples.

Thus, it was, in turn, under-sampled using edited near-

est neighbor rule (ENN). The under-sampling consists

in discarding any samples whose class label differs

from at least two of its three nearest samples (Wilson,

1972).

2.3 Machine Learning Models

Each HRV dataset was divided into a training and

a test set. Thereafter, a machine learning model was

trained on a 10-folds cross validation of the training

set, i.e. each of the 10 folds is used to train a random

forest classiﬁer on the remaining 9 folds. We tested

various machine learning classiﬁers (and their predic-

tion performances were similar) but settled to using

a Decision Jungle (DJ) (Shotton et al., 2013) because

DJ models tend to generalize better and require less

memory; thus, can run on limited computing devices

such as a user’s smart-phone. We evaluated the perfor-

mance of the model by computing its accuracy, preci-

sion, recall and the support of each class.

3 RESULT AND DISCUSSION

We observed that, in general, there is a visible change

in HRV between stressed and relaxed subjects. For ex-

ample, on the PHRENIC dataset, the subjects in group

A (i.e., the one that knew about the task difﬁculty

ahead of time) took less time to ﬁnish their tasks. We

believe these subjects were more in a hurry compared

to those in group B, who did not know the amount of

work they were supposed to do. However, the subjects

in group A made more mistakes compared to those in

group B because, perhaps, they did not take time to

carefully check their answers before submitting their

work. Unsurprisingly, the HRV power spectra distri-

bution of the two groups is also strikingly different

(Fig. 3). In HRV studies, LF and LF/HF ratio are in-

dicators of the sympathetic component of the nervous

system while HF reﬂects the cardiac parasympathetic

nerve activity (Malik et al., 1996). Overall, subjects in

group A exhibited a higher LF; thus, a higher sympa-

thetic activity compared to those in group B (Fig. 2).

Our ﬁndings are congruent with existing research that

showed that LF increases (Kim et al., 2018) and VLF

decreases (Usui and Nishida, 2017) in stressful situa-

tions. Thus, one can conclude that subjects in group A

are more stressed than those in group B.

After applying the trained machine learning mod-

els on the test datasets, we observed that all models per-

HEALTHINF 2019 - 12th International Conference on Health Informatics

260

Figure 2: Comparison of LF and VHF component of the

HRV of subjects in groups A and B.

Figure 3: Comparison of spectra HRV between subjects in

groups A and B. The solid line indicates the mean values

while the shaded area represent the data variation ranges

(expressed in terms of the

and

quartile). Subjects

in group A exhibit a higher sympathetic activity (high LF)

while those in group B have a higher stress recovery compo-

nent (higher VLF).

formed well in predicting stress and/or thermal com-

fort (Tables 2). In summary:

•

The stress prediction model on the PHRENIC

dataset achieved a perfect classiﬁcation accuracy

in distinguishing stressed subjects (group A) and

non-stress subject (group B).

•

The model that predicts the work stress vs non-

stress conditions (i.e., interruption and time pres-

sure vs no stress) achieved a near perfect classiﬁ-

cation both in terms of precision and recall. The

achieved performance is the highest among other

published works, and outperforms, by a large mar-

gin, even sophisticated models that use multimodal

set of sensor data (e.g., (Koldijk et al., 2018)).

•

After merging the PHRENIC and the SWELL

dataset (with the assumption that the subjects in

group A were under “time pressure” and those in

group B were not stressed), the prediction model

achieved equally a good stress prediction accuracy

both in terms of precision and recall. This accu-

racy slightly improves when tested on a model that

was trained on the dataset resulting from the com-

bination of the PHRENIC and SWELL datasets

after being over-sampled using SMOTE and down-

sampled using ENN. Further, after analyzing the

impact of the dataset_id feature that was added to

inspect any bias that might be introduced by merg-

ing the two datasets, we noticed that the dataset_id

has the lowest MDI compared to any HRV fea-

tures. It is 9 lower than that of the highest HRV

feature and 2 times smaller than that of the HRV

feature with next lowest MDI. Thus, we concluded

that there are no signiﬁcant bias in merging the

SWELL and the PHRENIC datasets.

•

The model trained on the dataset resulting in

merging the PHRENIC, SWELL and COMFORT

datasets also achieved a high accuracy in distin-

guishing thermal comfort (cold, neutral, and hot)

from work stress (no stress, time pressure and

interruption). Nevertheless, there are some mis-

classiﬁcations (especially in classifying the “neu-

tral” condition). In this case, we noticed that the

dataset_id’s MDI value is signiﬁcant. Thus, while

the impact of the dataset_id feature is not the most

decisive stress/comfort prediction attribute, it is

exigent to note that merging the three datasets re-

sulted in some unknown biases, and there is a need

for further studies to elucidate and limit these bi-

ases.

•

The stress/comfort prediction accuracy was notice-

ably improved when using a model trained on the

re-sampled dataset (see section 2.2.5). Notably, it

achieved a near-perfect stress vs. thermal comfort

prediction accuracy both in terms of precision and

recall and the log loss is signiﬁcantly lower com-

pared to the model that was not sampled. However,

this high prediction performance was achieved at

the cost of a more complex model (a decision

jungle with 50 decision directed acyclic graphs

(DAGs)). Further, this model also suffers from

some dataset merging biases; thus, its performance

should be taken with a grain of salt.

4 CONCLUSION

Modern work culture is stressful and the number of

stressed ofﬁce workers keeps increasing. At the same

time, current thermal comfort provision methods are

inefﬁcient both in the quality they provide and the

amount of energy they consume. Recent policies to

reduce energy use in buildings can only increase ther-

mal comfort dissatisfaction. In our previous works,

we proposed to provide thermal comfort based on the

person

s physiological changes due to his surround-

ing environment. We developed a system that pre-

dicts thermal comfort based on the a person’s HRV.

Thermal Comfort and Stress Recognition in Ofﬁce Environment

261

Table 2: Stress thermal comfort prediction model performance evaluation.

Dataset name

PHRENIC

SWELL

PHRENIC

plus SWELL

SWELL plus

PHRENIC

plus COMFORT

SMOTEENN SWELL

plus PHRENIC

plus COMFORT

overall accuracy 100.00 99.8 99.25 97.28 99.25

average accuracy 100.00 99.86 99.5 99.09 99.75

macro-averaged precision 100.00 99.84 99.31 95.99 99.27

macro-averaged recall 100.00 99.76 99.09 95.49 99.60

α see dataset description in section 2.2.1

β see dataset description in section 2.2.2

the dataset was obtained by combining the SWELL and the PHRENIC

dataset as described in section 2.2.4

the dataset is a combination of the SWELL, PHRENIC and the thermal

comfort datasets as described in section 2.2.4

an oversampled version of the dataset described in δ above using SMOTEENN method

and noise cleaning using ENN. See detailed explanations in section 2.2.5

There were, however, concerns about its performance

because work stress can also affect HRV. This paper

investigated the possibility to distinguish HRV due to

thermal distress from that due to work stress. The fol-

lowing ﬁndings emerged from our study:

•

We predicted stress with a near perfect accuracy

(99.25% accuracy and 99.5% recall)

•

Although both thermal comfort and work stress af-

fect HRV, we surmise that, in an ofﬁce environ-

ment, unless a person is both stressed and ther-

mally dis-comfortable, most ephemeral changes

in HRV are due to either work stress or thermal

discomfort. Although we believe there are biases

in our models, it is possible to differentiate stress-

ful working conditions (interruption and time pres-

sure) from non-stressful ones with a high accuracy

(99.25% accuracy and 99.75% average recall).

•

Our result suggests that it could be possible to de-

sign an unobtrusive system that delivers thermal

comfort and predicts work stress based on people

variation in their HRV. The system may predict a

person

s stress and thermal comfort from his pho-

toplethysmogram (PPG) signal recorded using an

unobtrusive wristband device. In the case of ther-

mal comfort, thermal comfort/energy constrained

optimization algorithms could be used to control

the most suitable thermal comfort provision mech-

anism in order to achieve a optimum thermal com-

fort at lowest energy consumption. Further, the

person’s stress level is continuously monitored

and the system could give adequate stress reduc-

tion recommendations. A practical system, how-

ever, needs further improvements. For example,

our approach does not address situations in which

a person is stressed and thermally dis-comfortable

and it fails to address other potential factors that

can affect a person’s HRV (e.g., anger, anxiety,

fear, sadness, sickness, happiness, etc ...). Further,

a generic, one-size-ﬁts-all model may not work in

a real-world ofﬁce environment because of individ-

ual differences that should not be ignored (Koldijk

et al., 2018). Instead, a personalized system could

be considered since people have different physi-

ologies. A generic model could only be used as a

baseline that must be further tweaked to satisfy the

uniqueness of each ofﬁce worker.

REFERENCES

Alberdi, A., Aztiria, A., and Basarab, A. (2016). Towards

an automatic early stress recognition system for ofﬁce

environments based on multimodal measurements: A

review. J. Biomed. Inform., 59:49–75.

ANSI/ASHRAE (2013). Thermal environmental conditions

for human occupancy standard 55-2013. Ashrae.

Brager, G. S., Zhang, H., and Arens, E. (2015). Evolving

opportunities for providing thermal comfort. Build.

Res. Inf., 43(3):274–287.

Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer,

W. P. (2002). Smote: Synthetic minority over-sampling

technique. J. Artif. Intell. Res., 16:321–357.

EU-OSHA (2017). Psychosocial risks and stress at work -

safety and health at work.

Gomes, R. L., Vanderlei, L. C. M., Garner, D. M., Vanderlei,

F. M., and Valenti, V. E. (2017). Higuchi fractal analy-

sis of heart rate variability is sensitive during recovery

from exercise in physically active men. Med. Express,

4(3):1–8.

Harvard Health (2011). Understanding the stress response.

Kim, H.-g., Cheon, E.-j., Bai, D.-s., Lee, Y. H., and Koo, B.-

h. (2018). Stress and heart rate variability: A meta-

analysis and review of the literature. Psychiatry Inves-

tig., 15(3):235–245.

Koldijk, S., Neerincx, M. A., and Kraaij, W. (2018). Detect-

ing work stress in ofﬁces by combining unobtrusive

sensors. IEEE Trans. Affect. Comput., 9(2):227–239.

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

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

HEALTHINF 2019 - 12th International Conference on Health Informatics

262

for stress and user modeling research. Proc. 16th Int.

Conf. Multimodal Interact. - ICMI ’14, pages 291–298.

Liu, W., Lian, Z., and Liu, Y. (2008). Heart rate variability at

different thermal comfort levels. Eur. J. Appl. Physiol.,

103(3):361–366.

Lopez, G., Ide, H., Shuzo, M., Warisawa, S., and Yamada,

I. (2014). Workplace stress estimation from physio-

logical indices in real situation. In Lect. Notes Inst.

Comput. Sci. Soc. Telecommun. Eng. LNICST, volume

100, pages 13–22. Springer International Publishing.

Lopez, G., Takahashi, K., Nkurikiyeyezu, K., and Yokokubo,

A. (2018). Development of a Wearable Thermo-

Conditioning Device Controlled by Human Factors

Based Thermal Comfort Estimation. In 2018 12th

France-Japan and 10th Europe-Asia Congress on

Mechatronics, pages 255–259. IEEE.

Malik, M., Bigger, J. T., Camm, A. J., Kleiger, R. E.,

Malliani, A., Moss, A. J., and Schwartz, P. J. (1996).

Heart rate variability: Standards of measurement, phys-

iological interpretation, and clinical use. Eur. Heart J.,

17(3):354–381.

Mizuno, K. and Watanabe, Y. (2008). Utility of an advanced

trail making test as a neuropsychological tool for an

objective evaluation of work efﬁ ciency during mental

fatigue. Fatigue Sci. Hum. Heal., pages 47–54.

Morrison, S. F. (2011). Central neural pathways for ther-

moregulation. Front. Biosci., 16(1):74.

Nicol, J. F. and Roaf, S. (2017). Rethinking thermal comfort.

Build. Res. Inf., 45(7):1–5.

Nkurikiyeyezu, K. and Lopez, G. (2018). Toward a real-

time and physiologically controlled thermal comfort

provision in ofﬁce buildings. In Intell. Environ., pages

168–177. IOS Press.

Nkurikiyeyezu, K., Suzuki, Y., Maret, P., Lopez, G., and Itao,

K. (2018). Conceptual design of a collective energy-

efﬁ cient physiologically-controlled system for ther-

mal comfort delivery in an o fﬁ ce environment. SICE

J. Control. Meas. Syst. Integr., 11(4):1–9.

Nkurikiyeyezu, K. N., Suzuki, Y., and Lopez, G. F. (2017).

Heart rate variability as a predictive biomarker of ther-

mal comfort. J. Ambient Intell. Humaniz. Comput.,

pages 1–13.

Sassi, R., Cerutti, S., Lombardi, F., Malik, M., Huikuri, H. V.,

Peng, C. K., Schmidt, G., and Yamamoto, Y. (2015).

Advances in heart rate variability signal analysis: Joint

position statement by the e-Cardiology ESC Working

Group and the European Heart Rhythm Association

co-endorsed by the Asia Paciﬁc Heart Rhythm Society.

Europace, 17(9):1341–1353.

Schulz, P., Kirschbaum, C., Prüßner, J., and Hellhammer, D.

(1998). Increased free cortisol secretion after awak-

ening in chronically stressed individuals due to work

overload. Stress Med., 14(2):91–97.

Shotton, J., Sharp, T., Kohli, P., Nowozin, S., Winn, J., and

Criminisi, A. (2013). Decision jungles: Compact and

rich models for classiﬁcation. Adv. Neural Inf. Process.

Syst. 26, pages 234–242.

Thayer, J. F., Nabors-Oberg, R., and Sollers, J. J. (1997).

Thermoregulation and cardiac variability: a time-

frequency analysis. Biomed. Sci. Instrum., 34:252–6.

Usui, H. and Nishida, Y. (2017). The very low-frequency

band of heart rate variability represents the slow recov-

ery component after a mental stress task. PLoS One,

12(8):1–9.

Vollmer, M. (2015). A robust, simple and reliable measure

of heart rate variability using relative rr intervals. In

2015 Comput. Cardiol. Conf., volume 42, pages 609–

612. IEEE.

Wikilivres (2018). Neurosciences/La thermorégulation —

Wikilivres.

Wilson, D. L. (1972). Asymptotic properties of nearest

neighbor rules using edited data. IEEE Trans. Syst.

Man Cybern., 2(3):408–421.

Thermal Comfort and Stress Recognition in Ofﬁce Environment

263