Content uploaded by Alexandre Marois
Author content
All content in this area was uploaded by Alexandre Marois on Sep 21, 2021
Content may be subject to copyright.
This is an Accepted Manuscript of an article published by Human Factors and Ergonomics Society: Marois,
Lafond, Gagnon, Vachon, & Cloutier, Predicting Stress among Pedestrian Traffic Workers Using Physiolog-
ical and Situational Measures, Proceedings of the Human Factors and Ergonomics Society 2018 Annual
Meeting (62) pp. 1262-1266. Copyright © 2018 Human Factors and Ergonomics Society. DOI:
10.1177/1541931218621290.
Predicting Stress among Pedestrian Traffic Workers
Using Physiological and Situational Measures
Alexandre Marois1, Daniel Lafond2, Jean-François Gagnon2, François Vachon1, & Marie-Soleil Cloutier3
1Université Laval, Quebec City, Canada
2Thales Research and Technology Canada, Quebec City, Canada
3Institut National de la Recherche Scientifique, Montreal, Canada
Traffic workers are vulnerable to accidents and must make critical decisions to avoid conflicts between road
users. This can lead to high stress levels, which may hinder their capacity to mitigate the occurrence of haz-
ards. Measuring stress on the field could represent an efficient solution to help pinpoint risky situations and
identify factors that increase risk. The goal of this study was to verify whether stress among traffic workers
could be predicted using physiological measures and characteristics of the work situation. Nineteen police
officers in Quebec City and Montreal, Canada, performed traffic duties while their physiological activity was
assessed by a wearable physiological harness. Every 15 minutes, change in subjective stress was also meas-
ured. Results showed that decision-tree models outperformed multifactorial logistic regressions for predict-
ing subjective stress based on both situational factors and physiological measures. This demonstrated the
potential of using such measures to monitor stress among traffic workers.
INTRODUCTION
Pedestrian road workers are highly vulnerable to work-
place accidents. This is especially true for workers, such as flag-
gers or traffic-control police officers, whose job may require
managing traffic on construction sites. Although some preven-
tion tools have been proposed to increase their security (Fan,
Choe, & Leite, 2014), accidents still occur. According to Yu,
Bill, Chitturi, and Noyce (2013), almost half of all the pedes-
trian victims of road accidents in Wisconsin between 2000 and
2010 were flaggers working on construction sites. Moreover,
work zones controlled by flaggers are known to be associated
with more rear-end crashes in comparison with those controlled
by arrow boards (Qi, Srinivasan, Teng, & Baker, 2013). In ad-
dition to the insecurity felt from working surrounded by cars
and heavy machinery, traffic workers must often make critical
decisions to avoid conflict between road users and other con-
struction workers. These decisions must be taken in a complex
dynamic environment that contains myriad information to pay
attention to. Engaging in multiple tasks in which vigilance is
necessary—as it is often the case for traffic management—can
lead to attentional fatigue. Given that attention is a key resource
in executive functioning, this could ultimately impede workers’
job productivity and efficiency (e.g., Kaplan & Berman, 2010).
Such difficult work context can increase stress. Stress, de-
fined as “a particular response of an organism to an identified
demand stimulus”, can indeed arise when one experiences a de-
mand that “exceeds his or her real or perceived abilities to suc-
cessfully cope with the demand, resulting in disturbance to his
or her physiological and psychological equilibrium” (Kolbell,
1995, p. 31). This stress experienced by traffic workers can be
amplified when dealing with non-compliant road users or work-
ing in challenging environments (e.g., with poor visibility or
when a construction site is too close to the road; Debnath,
Blackman, & Haworth, 2015).
Stress can exert a negative impact on worker’s ability to
perform efficient and safe traffic duties. Indeed, the negative
impacts of stress and job demands have been widely demon-
strated (Colligan & Higgins, 2006). For instance, stress in the
workplace, in addition to high workload demands, has been
shown to decrease productivity (MacDonald, 2003). Several
studies have shown that, in other high-risk domains such as air
traffic control (Hodgetts, Tremblay, Vallières, & Vachon,
2015), piloting (Dehais et al., 2014) or command and control
(Grier, 2015), stress and high workload demands may hinder
performance and, in turn, safety of the people. However, no
study has yet focused on the stress of traffic workers while per-
forming their job. Since they may experience stress and high
work demands, solutions must be proposed to prevent any dan-
gerous situation arising from carrying out traffic duties. This
could mitigate chances of hazard for them, as well for road us-
ers and other construction workers. This study aimed at making
continuous assessments of traffic workers’ stress by using phys-
iological and observational factors.
Stress Measurement and Prediction
The online monitoring of workers’ stress levels could rep-
resent a potential solution for risk identification and prevention.
Numerous ways can be used to evaluate stress.
Self-reported measures. Stress can first be assessed
through self-report. Many studies asked subjects to report their
subjective levels of stress using a variety of scales (Damaske,
Zawadzki, & Smyth, 2016). The validity of these measures was
confirmed empirically (Morgan, Umberson, & Hertzog, 2014).
Considering that traffic workers may need to interrupt their
tasks to answer questions about their actual state, self-reported
measures may not represent the optimal tool to be used to assess
stress in real time. Interruptions can indeed hinder performance,
especially in dynamic situations, which continue to evolve dur-
ing the interruption (Labonté, Tremblay, & Vachon, 2016). Alt-
hough self-report can represent a valid way to access workers’
perception of their own stress, one should favor instead meth-
ods that do not disrupt the worker’s task.
Physiological measures. According to Ganster and Rosen
(2013), work stress can be efficiently assessed using physiolog-
ical measures. Stress has been shown to be highly correlated
with several types of physiological responses. For instance,
heart rate (HR) and HR variability (HRV) are known to respec-
tively increase (van Hedger, Necka, Barakzai, & Norman,
2017) and decrease (Pereira, Almeida, Cunha, & Aguiar, 2017)
as stress intensifies. Similarly, measures of respiratory rate
(RR) can also be used to such purpose given their positive cor-
relation with stress (Boiten, Frijda, & Wientjes, 1994). Conse-
quently, stress can be successfully predicted using bio-behav-
ioral data. For example, Sun et al. (2012) showed that stress
could be predicted by using cardiac metrics, galvanic skin re-
sponse and accelerometer data. Likewise, Salai, Vassányi, and
Kósa (2016) achieved significant stress prediction using three
time-domain features of the HR signal.
The key advantages of using physiological measures is
their potential to provide objective and non-intrusive indices of
stress variation continuously, and the relative simplicity to per-
form such measures in applied contexts by using wearable sen-
sors (Brouwer, Zander, van Erp, Korteling, & Bronkhorst,
2015). Physiological data are sometimes considered more “ob-
jective” than other measures used to assess stress given that
they are not easily manipulated by experimenters’ expectations
and may not be biased by participants’ own opinion about their
actual state (Fried, Rowland, & Ferris, 1984). Still, some prob-
lems may be encountered when using physiological measures
to predict work stress. First, some measurement issues ensuing
from permanent, transitory or procedural factors, such as health
or fatigue, can affect the reliability of the physiological re-
sponses used to predict stress level (see Fried et al., 1984). Sec-
ond, such responses cannot be considered as a direct measure
of stress as they rather reflect the activation of the sympathetic
and parasympathetic systems that can be affected by a plethora
of other factors (Semmer, Grebner, & Elfering, 2004). In fact,
despite recent achievements (Parent, Gagnon, Falk, & Trem-
blay, 2016), researchers still have issues at discriminating the
impact of stress on the physiological signal from other compo-
nents such as fatigue and workload.
Observational measures. Stress can also be assessed indi-
rectly, from an external point of view, by using observational
measures of work context and environment. For instance, Holt
(1993) suggested several conditions such as work overload, role
ambiguity or monotony that may induce stress on workers. Ac-
cording to Semmer et al. (2004), assessment by trained job an-
alysts can also be performed to measure workers’ stress levels
by analyzing their work context. However, as stated by the au-
thors, such measures may also be biased considering that a
given context judged stressful by one worker may not always
be considered as such by another worker. Given that each of the
methods presented to measure work stress (self-reported, phys-
iological and observational measures) have their own limits, a
combination of each type should be privileged over the usage
of a single method (cf. Semmer et al., 2004).
Study Objectives
Since stress is known to affect work performance and in-
crease chances of hazards for traffic work, a more systematic
examination of traffic workers’ stress is needed to ensure better
security for both them and road users. Thus, we sought to cali-
brate stress assessment models by determining which physio-
logical responses (given by sensors) or situational factors could
be used to predict self-reported stress in real time for police of-
ficers performing traffic duties. In doing so, we investigated the
applicability of an original method to detect variations in the
level of stress experienced by traffic workers through logistic
regression and decision-tree algorithm techniques.
METHOD
Participants
Eight Quebec City Police officers and 11 Montreal Police
officers took part in the study (6 women). Volunteers were
monitored while performing traffic duties for 1 to 8 sessions (M
= 3.3, SD = 1.8) that lasted between 60 and 285 minutes (M =
150, SD = 45). A total of 54 work shifts was recorded, compris-
ing 614 periods of 15 minutes.
Measures and Pre-processing
Self-reported stress was first assessed at the beginning of
the work shift using a 10-point Likert scale (where 1 = low, 10 =
high). While performing traffic duties, participants were also
asked to indicate every 15 min whether their level of stress had
decreased, increased or remained stable.
Several characteristics of the environment and of the police
officers’ work context were noted by an observer. For each time
period, the current state of the participant (whether on-task,
controlling traffic or off-task, either on a break or watching traf-
fic on the side of the road or in a car), the presence of a poten-
tially-stressful event (e.g., a traffic conflict), and the on-site
traffic density in terms of car, pedestrian and cyclist (high: > 20
for each traffic category) were recorded. The number of lanes
in the intersection, the number of lanes in construction, and the
presence of road construction sites were also noted for every
work shift. Each of these characteristics was used as a situa-
tional predictor.
Physiological measures were garnered and pre-processed
by the built-in system provided by a Zephyr BioModule BH3
(Zephyr Technology) chest strap measuring movement, cardiac
and respiratory activity. Five physiological metrics were used:
1) HR was calculated by using the total number of heartbeats on
the total recording time of the smoothed normal to normal (NN)
signal from which artifacts had been filtered. 2) HRV was cal-
culated by performing a rolling 300 heartbeat-standard devia-
tion of the NN signal (or the so-called SDNN; see Camm et al.,
1996). 3) RR was calculated by using the total number of torso
expansion-contraction cycles on the total recording time. 4) Ac-
tivity, measured in vector magnitude units, in g, represented a
mean acceleration index of the body movements in the three
possible dimensions (x, y and z). 5) Core temperature was esti-
mated by using mean HR of the 60 preceding seconds (see Seo
et al., 2016).
Each measure was resampled at a 1-Hz frequency, fol-
lowed by the removal of invalid data—for which the system in-
dicated a validity index of 80% or less. Remaining data were
averaged over 15-minute periods to match the self-reported
stress 15-minute intervals. These units of analysis were then
aligned with each self-report of change in perceived stress.
Statistical Analyses
In order to verify whether self-reported stress could be pre-
dicted by the situational and physiological measures, multifac-
torial logistic regressions were performed. R implementations
of generalized boosted models (GBM) were also used to train
decision-tree ensembles (Ridgeway, 2017). The training proce-
dure was carried out using the Caret package (Khun, 2017).
This classification algorithm was used to test an alternate pre-
dictive modeling approach, selected for its sensitivity to nonlin-
ear relationships and low-occurrence values (Gagnon, Gagnon,
Lafond, Parent, & Tremblay, 2016).
RESULTS
Subjective Stress
Subjective stress measured at the beginning of the work pe-
riod was of 2.06 points out of 10 (SD = 2.01). Among the 15-
minute periods, stress decreased 8.14% of the time, increased
10.91% of the time and remained constant 78.83% of the time.
Variations in perceived stress over time, starting from the initial
self-report, were calculated. This “anchored-stress” measure
varied between -2 and 17 (M = 2.31, SD = 2.69).
To reduce noise in the data, stress levels were then divided
in two distinct states (low vs. high). The mean anchored-stress
level was used as a split value. This allowed obtaining 331 ob-
servations in which anchored-stress level was considered as low
and 188 observations where stress was considered as high. Such
measure was strongly correlated with anchored-stress level
(rs = .85, p < .001). This classification was used for all subse-
quent analyses involving stress level (namely dichotomous an-
chored-stress; DAS).
Prediction of Stress Levels
Situational measures. Participants were on-task for
73.29% of the recorded periods. Observers identified signifi-
cant stressful events for 7.65% of the 15-minute periods. Car,
pedestrian, and cyclist traffic density was considered high for
62.05%, 32.25%, and 0.65% of the 15-minute periods, respec-
tively. Intersections where police officers worked had on aver-
age 3.64 lanes (SD = 0.51) and 1.92 lanes under construction
(SD = 1.14). On the total amount of traffic duty periods, 83.88%
of them were performed on a construction site.
Binomial logistic regressions were performed on each fac-
tor—except for cyclist traffic whose high occurrences were too
scarce—to verify whether they were significantly related to
DAS. Analyses showed that stress level was associated with be-
ing on-task (Z = -3.81, p < .001), working on a construction site
(Z = 4.40, p < .001), being on a site with one (Z = 3.83, p <
.001), two (Z = 4.90, p < .001), three (Z = 3.42, p < .001), and
four lanes in construction (Z = 3.86, p < .001), high car traffic
(Z = 5.32, p < .001), and high pedestrian traffic (Z = 4.24, p <
.001). A multifactorial logistic regression model permitted to
obtain a predictive accuracy of 66.04% (κ = .24) using a 10-fold
cross-validation procedure. Table 1 shows each level of the sit-
uational factors that were included by the model and whether
they significantly contributed to predict DAS level.
Table 1
Situational characteristics included as contributors to the multifactorial lo-
gistic regression used to classify DAS level
Factora
df residuals
Deviance residuals
Number of lanes in construction
1
2
3
4
501
500
499
498
0.07*
20.32***
4.73*
16.83***
Work state (on-/off-task)
497
10.67**
Car traffic density (low/high)
496
23.82***
Pedestrian traffic density (low/high)
495
08.31**
Stressful event (no/yes)
494
0.74*
Number of lanes
3
4
493
492
0.11*
6.81**
Road construction (no/yes)
492
0.00*
aN – 1 factor levels are reported due to the nature of the analysis.
*p < .05 **p < .01 ***p < .001
Using the same factors than in the first model, superior
cross-validation accuracy (10-fold resampling) of 79.13% (κ =
.53) was observed with a decision-tree model (see the detailed
cross-validation results in Table 2, which includes true nega-
tives, false positives, true positives, and false negatives).
Table 2
Confusion matrix for the DAS level prediction using situational factors per-
formed on 10 folds
Observed classification
Model prediction
Low stress
High stress
Low stress
57.50%
15.30%
High stress
5.60%
21.70%
Physiological measures. The mean HR was of 89.68
beats/min (SD = 13.16). The mean HRV was of 43.30 SDs/ms
(SD = 17.74). On average, RR was 23.23 breaths/min (SD =
5.94). The mean Activity measure was 0.08 g (SD = 0.04) and
the average estimated Core temperature was 37.42 °C (SD =
0.37). Correlation analyses showed that DAS was significantly
associated with HRV (rs = -.13, p = .007), RR (rs = .19, p <
.001), and Activity (rs = .23, p < .001). The correlation between
DAS and HR (rs = .03) and Core temperature (rs = .07) did not
reach significance (ps > .05). A multifactorial logistic regres-
sion was used to predict stress from the psychophysiological
signal data. This model achieved an accuracy of 63.49% (κ =
.11) on 10-fold cross-validation tests with HR (p = .041), RR (p
< .001) and HRV (p = .009) as significant predictors of stress.
Activity (p = .078) and Core temperature (p = .690) did not
reach significance. Once again, higher cross-validation accu-
racy (77.47%, κ = .49) was obtained with a decision-tree clas-
sifier using all physiological features (see Table 3).
Table 3
Confusion matrix for the DAS level prediction using physiological features
performed on 10 folds
Observed classification
Model prediction
Low stress
High stress
Low stress
56.70%
16.20%
High stress
6.30%
20.80%
DISCUSSION
This study aimed at gathering situational and physiological
data to develop and test a stress model-calibration method that
could help measure in real time significant changes in stress
level, and ultimately be used to better detect risk factors and
improve safety conditions of traffic workers. Results showed
that stress measurements were sensitive to potential situational
risk factors such as the police officer’s work state, car and pe-
destrian traffic density, the number of traffic lanes and in con-
struction. To our knowledge, this is the first time that a group
of situational characteristics is used to predict stress among traf-
fic workers. Moreover, the physiological stress model—that
comprised HR, HRV, RR, Activity, and Core temperature—
could also be used to significantly assess perceived stress.
Again, this is the first demonstration of real-time physiological
stress assessment of workers performing traffic duties. Notably,
the demonstration of a positive relationship between stress, HR,
RR and Activity, as well as the negative relationship between
HRV and stress, is in line with previous studies that also meas-
ured these stress proxies (Boiten et al., 1994; Pereira et al.,
2017; Sun et al., 2012; van Hedger et al., 2017).
Interestingly, using multifactorial logistic regressions for
both situational and physiological models provided rather good
predictive accuracy, but did not explain well variations in stress
given the poor observed-predicted agreement indices (κ). In
fact, nonlinear decision-tree classifications achieved better pre-
dictive accuracy and larger κ values, which is in line with pre-
vious work on physiological computing (Gagnon et al., 2016;
Teller, 2004). This is also consistent with Sun et al. (2012) who
achieved significant stress prediction by using a decision-tree
classifier comprising measures of cardiac and electrodermal ac-
tivity as well as accelerometer data. Present findings thus show
that the decision-tree algorithm was able to better predict stress
variations by its ability to capture nonlinear relationships
among predictor variables.
Given that no situational predictive model has been pro-
posed hitherto in the literature, we consider our observational
data model to be of good accuracy as it reached larger or similar
values to that of complex physiological models (Kim, Seo, Cho,
& Cho, 2008; Parent et al., 2016; Salai et al., 2016). This model
could however be improved by obtaining other data of the en-
vironment or by measuring the current features in a more sen-
sitive way. For instance, a context-aware model considering
variables such as time of the day, weather and level of surround-
ing noise could provide further indices that may be positively
associated with stress. A more sensitive analysis of the situation
and work environment could also detect more stressful inci-
dents. This may be useful as only 7.65% of time periods were
considered having a potentially-stressful incident and given that
they did not contribute to increase stress. Accordingly, further
analysis will be performed in the next step of this project using
a video-based intelligent traffic analyzer tool to detect conflicts
or quasi-conflicts between road users, workers and other ele-
ments of the environment.
One could argue that the predictive accuracy of the physi-
ological model is poorer in comparison with that of other simi-
lar studies performed in different contexts. Although superior
to Kim et al.’s (2008) accuracy (66.10%) and similar to that of
Salai et al. (2016) and Parent et al. (2016; respectively 74.60%
and 77.14%), our physiological model’s prediction value of
77.47% is still inferior to that of many other models (Munla,
Khalil, Shahin, & Mourad, 2015; Sun et al., 2012). This can
originate from the measures that were selected to be analyzed
and integrated to the model. For instance, several stress-predic-
tion models with higher accuracy values comprise a large vari-
ety of HRV features of time, frequency, and time-frequency do-
mains (Munla et al., 2015; Sun et al., 2012). In our case, only
the HRV-related SDNN of the time domain was used. Combin-
ing this measure with other cardiac metrics (e.g., percentage of
total intervals successively differing by more than 50 ms in the
NN signal [pNN50], very low, low and high frequency [VLF,
LF, HF] bands, or LF/HF ratio) could have led to greater accu-
racy values.
Implications and Future Directions
Notwithstanding the possibility of enhancing the predic-
tion of experienced stress by adding further features, both of our
models have one common strength: their simplicity to be im-
plemented. Using situational data to predict stress can first be
achieved by observing the context in which traffic is managed.
This could help decision-makers to anticipate that a given work
context may be more stressful for traffic controllers. Conse-
quently, this could lead them to make proper decisions to avoid
hazardous situations such as giving more breaks or scheduling
more workers on a specific zone to alleviate their workload.
The physiological model we propose can also be easily im-
plemented as the measures on which it relied upon were almost
completely calculated by the built-in system of the sensor har-
ness. Indeed, these physiological markers were directly gar-
nered and computed by the Zephyr BioModule BH3, and only
needed to be filtered for invalid data and then averaged in time-
bins. These actions can easily be automatized and performed by
a portable device connected to the harness, or with the advent
of fast wireless networks (e.g., 4G, 5G), on a cloud-based infra-
structure. Such system could hence be implemented in real time
in a high-risk context such as traffic work. For instance, while
the biometric module and portable model-based processing de-
vice would garner data and infer stress levels, feedback could
be send to either the pedestrian road worker or a supervisor to
warn about high-stress circumstances that could be particularly
risk-prone. Such implementation of the model could facilitate
the work of human factors researchers and workplace health
and safety practitioners as stress could be assessed without in-
terrupting traffic work, as it is currently required by self-report
methods. Similar systems have been proven efficient at predict-
ing mental workload by providing relevant information on the
state of operators, the situation and their environment in other
high-risk contexts such as emergency management (cf. Gagnon,
Lafond, Rivest, Couderc, & Tremblay, 2014).
Considering that our models could significantly predict
stress levels, future studies should aim at assessing their poten-
tial limits, such as the impact of social desirability on self-re-
ported stress and on the model’s ability to predict it, as well as
whether they can be applied to other individuals. Eventually,
these models should be calibrated so that they could be used on
any type of pedestrian worker and be implemented to mitigate
potential hazardous situations that could be caused by stressful
work context. Once both models are well adjusted and vali-
dated, they could be exploited as decision-making tools to iden-
tify situations that may need to be prioritised for prevention.
ACKNOWLEDGMENTS
We would like to thank Sylvanie Godillon, Laura Bar-
rachina, Yvette Ishimo, César Tartati and Alexandre Lepage.
This work was funded by a grant from IRSST and by the
MITACS Accelerate program.
REFERENCES
Boiten, F. A., Frijda, N. H., & Wientjes, C. J. E. (1994). Emotions and respira-
tory patterns: review and critical analysis. International Journal of Psy-
chophysiology, 17, 103–128. doi:10.1016/0167-8760(94)90027-2
Brouwer, A.-M., Zander, T. O., van Erp, J. B. F., Korteling, J. E., & Bronkhorst,
A. W. (2015). Using neurophysiological signals that reflect cognitive or
affective state: six recommendations to avoid common pitfalls. Frontiers
in Neuroscience, 9, 136. doi:10.3389/fnins.2015.00136
Camm, A. J., Malik, M., Bigger, J. T., Breithardt, G., Cerutti, S., Cohen, R. J.,
… Singer, D. H. (1996). Heart rate variability — standards of measure-
ment, physiological interpretation, and clinical use. Circulation, 93,
1043–1065. doi:10.1161/01.CIR.93.5.1043
Colligan, T. W., & Higgins, E. M. (2006). Workplace stress. Journal of Work-
place Behavioral Health, 21, 89–97, doi:10.1300/J490v21n02_07
Damaske, S., Zawadzki, M. J., & Smyth, J. M. (2016). Stress at work: Differ-
ential experiences of high versus low SES workers. Social Science &
Medicine, 156, 125–133. doi:10.1016/j.socscimed.2016.03.010
Debnath, A. K., Blackman, R., & Haworth, N. (2014). Towards making in-
formed decisions on selecting promising work zone safety treatments.
Road and Transport Research, 23(4), 41–53.
Dehais, F., Causse, M., Vachon, F., Régis, N., Menant, E., & Tremblay, S.
(2014). Failure to detect critical auditory alerts in the cockpit: Evidence
for inattentional deafness. Human Factors, 56, 631–644.
doi:10.1177/0018720813510735
Fan, W., Choe, S., & Leite, F. (2014). Prevention of backover fatalities in high-
way work zones: A synthesis of current practices and recommendations.
International Journal of Transportation Science, 3, 311–337.
doi:10.1260/2046-0430.3.4.311
Fried, Y., Rowland, K. M., & Ferris, G. R. (1984). The physiological measure-
ment of work stress: A critique. Personnel Psychology, 37, 583–615.
doi:10.1111/j.1744-6570.1984.tb00528.x
Gagnon, J.-F., Gagnon, O., Lafond, D., Parent, M., & Tremblay, S. (2016, July).
A systematic assessment of operational metrics for modeling operator
functional state. 3rd International Conference on Physiological Compu-
ting Systems, Lisbon, Portugal.
Gagnon, J.-F., Lafond, D., Rivest, M., Couderc, F., & Tremblay, S. (2014).
Sensor-Hub: A real-time data integration and processing nexus for adap-
tive C2 systems. In Proceedings of the sixth international conference on
adaptive and self-adaptive systems and applications (pp. 63–67).
Ganster, D. C., & Rosen, C. C. (2013). Work stress and employee health: A
multidisciplinary review. Journal of Management, 39, 1085–1122.
doi:10.1177/0149206313475815
Grier, R. A. (2015). Situation awareness in command and control. In R. R. Hoff-
man, P. A. Hancock, M. W. Scerbo, R. Parasuraman, & J. L. Szalma
(Eds.), The Cambridge handbook of applied perception research (pp.
891–911). New York, NY: Cambridge University Press.
Hodgetts, H. M., Tremblay, S., Vallières, B. R., & Vachon, F. (2015). Decision
support and vulnerability to interruption in a dynamic multitasking envi-
ronment. International Journal of Human-Computer Studies, 79, 106–
117. doi:10.1016/j.ijhcs.2015.01.009
Holt, R. R. (1993). Occupational stress. In L. Goldberger & S. Breznitz (Eds.),
Handbook of stress: Theoretical and clinical aspects (pp. 342–367). New
York, NY: Free Press.
Kaplan, S., & Berman, M. G. (2010). Directed attention as a common resource
for executive functioning and self-regulation. Perspectives on Psycho-
logical Science, 5, 43–57. doi:10.1177/1745691609356784
Kim, D., Seo, Y., Cho, J., & Cho, J.-H. (2008). Detection of subjects with higher
self-reporting stress scores using heart rate variability patterns during the
day. Proceedings of the 30th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society, 2008, 682–685.
doi:10.1109/IEMBS.2008.4649244.
Kolbell, R. M. (1995). When relaxation is not enough. In L. R. Murphy, J. J.
Hurrell, Jr., S. L. Sauter, & G. P. Keita (Eds.), Job stress interventions
(pp. 31–43). doi:10.1037/10183-003
Kuhn, M. (2017). Classification and regression training. Documentation on the
R Package ‘caret’, version 6.0-78. https://cran.r-project.org/web/pack-
ages/caret/caret.pdf, accessed January 2018.
Labonté, K., Tremblay, S., & Vachon, F. (2016). Effects of a warning on inter-
ruption recovery in dynamic settings. Proceedings of the Human Factors
and Ergonomics Society Annual Meeting, 60, 1304–1308.
doi:10.1177/1541931213601302
MacDonald, W. (2003). The impact of job demands and workload on stress and
fatigue. Australian Psychologist, 38, 102–117.
doi:10.1080/00050060310001707107
Morgan, E. S., Umberson, K., & Hertzog, C. (2014). Construct validation of
self-reported stress scales. Psychological Assessment, 26, 90–99.
doi:10.1037/a0034714
Munla, N., Khalil, M., Shahin, A., & Mourad, A. (2015). Driver stress level
detection using HRV analysis. Proceedings of the International Confer-
ence on Advances in Biomedical Engineering, 2015, 61–64.
doi:10.1109/ICABME.2015.7323251
Parent, M., Gagnon, J.-F., Falk, T. H., & Tremblay, S. (2016, May). Modeling
the operator functional state for emergency response management. Pro-
ceedings of the 13th International Conference on Information Systems for
Crisis Response and Management (ISCRAM), Rio de Janeiro, Brazil.
Pereira, T., Almeida, P. R., Cunha, J. P. S., & Aguiar, A. (2017). Heart rate
variability metrics for fine-grained stress level assessment. Computer
Methods and Programs in Biomedicine, 148, 71–80.
doi:10.1016/j.cmpb.2017.06.018
Qi, Y., Srinivasan, R., Teng, H., & Baker, R. (2013). Analysis of the frequency
and severity of rear-end crashes in work zones. Traffic Injury Prevention,
14, 61–72. doi:10.1080/15389588.2012.675109
Ridgeway, G. (2017). Generalized boosted regression models. Documentation
on the R Package ‘gbm’, version 2.1.3. https://cran.r-pro-
ject.org/web/packages/gbm/gbm.pdf, accessed January 2018.
Salai, M., Vassányi, I., & Kósa, I. (2016). Stress detection using low cost heart
rate sensors. Journal of Healthcare Engineering, 2016.
doi:10.1155/2016/5136705
Semmer, N. K., Grebner, S., & Elfering, A. (2004). Beyond self-report: Using
observational, physiological, and situation-based measures in research on
occupational stress. In P. L. Perrewé, & D. C. Ganster (Eds.), Emotional
and physiological processes and positive intervention strategies (Volume
3, pp. 205–263). Amsterdam, Netherlands: Elsevier.
Seo, Y., DiLeo, T., Powell, J. B., Kim, J. H., Roberge, R. J., & Coca, A. (2016).
Comparison of estimated core body temperature measured with the Bio-
Harness and rectal temperature under several heat stress conditions. Jour-
nal of Occupational and Environmental Hygiene, 13, 612–620.
doi:10.1080/15459624.2016.1161199
Sun, F. T., Kuo, C., Cheng, H. T., Buthpitiya, S., Collins, P., & Griss, M.
(2012). Activity-aware mental stress detection using physiological sen-
sors. In M. Gris, & G. Yang (Eds.), Mobile computing, applications, and
services: Lecture Notes of the Institute for Computer Sciences, Social In-
formatics and Telecommunications Engineering (Volume 76, pp. 282–
301). Berlin, Germany: Springer. doi:10.1007/978-3-642-29336-8_16
Teller, A. (2004). A platform for wearable physiological computing. Interact-
ing with Computers, 16, 917–937. doi:10.1016/j.intcom.2004.08.004
van Hedger, K., Necka, E. A., Barakzai, A. K., & Norman, G. J. (2017). The
influence of social stress on time perception and psychophysiological re-
activity. Psychophysiology, 54, 706–712. doi:10.1111/psyp.12836
Yu, L., Bill, A., Chitturi, M., & Noyce, D. (2013). On-duty struck-by crashes.
Transportation Research Record, 2386, 112–120. doi:10.3141/2386-13
