sensors

Article
Utilizing a Wristband Sensor to Measure the Stress
Level for People with Dementia
Basel Kikhia 1, *, Thanos G. Stavropoulos 2 , Stelios Andreadis 2 , Niklas Karvonen 3 ,
Ioannis Kompatsiaris 2 , Stefan Sävenstedt 1 , Marten Pijl 4 and Catharina Melander 1
1 Department of Health Sciences, Luleå University of Technology, 97187 Luleå, Sweden;
stefan.savenstedt@ltu.se (S.S.); catharina.melander@ltu.se (C.M.)
2 Information Technologies Institute, Centre for Research & Technology Hellas, 57001 Thessaloniki, Greece;
athstavr@iti.gr (T.G.S.); andreadisst@iti.gr (S.A.); ikom@iti.gr (I.K.)
3 Department of Computer Science, Electrical and Space Engineering, Luleå University of Technology,
97187 Luleå, Sweden; niklas.karvonen@ltu.se
4 Personal Health Solutions, Philips Research, 5656 AE Eindhoven, The Netherlands; marten.pijl@philips.com
* Correspondence: basel.kikhia@ltu.se; Tel.: +46-73-698-9928

Academic Editors: Octavian Adrian Postolache, Alex Casson and Subhas Chandra Mukhopadhyay
Received: 3 August 2016; Accepted: 18 November 2016; Published: 24 November 2016

Abstract: Stress is a common problem that affects most people with dementia and their caregivers.
Stress symptoms for people with dementia are often measured by answering a checklist of questions
by the clinical staff who work closely with the person with the dementia. This process requires a lot
of effort with continuous observation of the person with dementia over the long term. This article
investigates the effectiveness of using a straightforward method, based on a single wristband sensor
to classify events of “Stressed” and “Not stressed” for people with dementia. The presented system
calculates the stress level as an integer value from zero to five, providing clinical information of
behavioral patterns to the clinical staff. Thirty staff members participated in this experiment, together
with six residents suffering from dementia, from two nursing homes. The residents were equipped
with the wristband sensor during the day, and the staff were writing observation notes during the
experiment to serve as ground truth. Experimental evaluation showed relationships between staff
observations and sensor analysis, while stress level thresholds adjusted to each individual can serve
different scenarios.

Keywords: clinical assessment; nursing homes; sensors; dementia; stress monitoring

1. Introduction
Stress is a common problem that is experienced by people with dementia and their caregivers [1].
There are different symptoms that people with dementia show when they are stressed such as
apathy, aggression, sleep disturbances, wandering, and depression [2]. In general, Behavioral and
Psychological Symptoms of Dementia, labeled as (BPSD), affect more than 90% of people who suffer
from dementia [3]. BPSD are divided into symptoms of agitation, delusions, euphoria, hallucinations,
apathy, depression, aberrant motor behavior, irritability, eating problems and sleep problems [4].
BPSD affect not only people with dementia, but also caregivers as they are closely connected to the
person with dementia on a daily basis. Statistics show that 74% of caregivers were concerned about
maintaining their own health since becoming a caregiver. In addition, 40% of caregivers reported that
the emotional stress of their role is high or very high [5].
Understanding BPSD and the changes of behavior of the person with dementia helps in assessing
the development of the disease. Clinical staff will also be able to monitor the treatment of behavioral
symptoms, and in planning their clinical interventions to reduce the BPSD and improve the status

Sensors 2016, 16, 1989; doi:10.3390/s16121989 www.mdpi.com/journal/sensors

Sensors 2016, 16, 1989 2 of 17

of the person with dementia. One key factor when assessing BPSD is to identify moments when the
person with dementia is stressed, and then search for the possible causes of these stress episodes to
reduce it. Stress and anxiety can be a result of social isolation, unfamiliar surroundings, environmental
factors such as light and noise, mental confusion, physical inactivity, etc. [6,7].
The common method in assessing BPSD in nursing homes is to rely on the staff observations
in defining different symptoms that the person with dementia has. This is done with the help of
the Neuropsychiatric Inventory-Nursing Home version (NPI-NH) instrument [8]. The instrument
provides a list of questions about the person with dementia that covers the BPSD domains. Answering
the questions by the caregiver will result in a score for each domain and an overall score that gives
hints about the status of the person with dementia. From this assessment, the clinical staff can locate
possible problems and plan their clinical interventions to reduce the BPSD. The NPI-NH instrument,
however, requires a lot of effort with continuous observation of the resident over long periods of time
in order for the caregiver to get an understanding of the behavior [9]. The whole process of using the
NPI-NH instrument is thus both complicated and time-consuming, which makes it impractical to use
for the clinical staff.
To help caregivers and clinical staff understand the behavior of the person with dementia, and to
simplify the assessment of BPSD, this work presents a sensor-based system that measures the stress
level of people with dementia. The system detects stress patterns, which will provide objective clinical
information that helps caregivers in planning clinical interventions to reduce stress and aggression for
people with dementia. Reducing stress and aggression for people with dementia will also result in
reducing the stress for the caregivers, and for other people in their environment. These other people
include residents in the nursing home, or family members living with the person with dementia
at home. The system thus could contribute to improving the status of the person with dementia,
and in the nursing homes it could significantly improve the quality of care and reduce the care needs,
and further reduce costs.
The system relies on a single wristband that is worn by the person with dementia during the day.
The use of a single sensor is important to simplify the system, so it does not create any burden for the
caregivers, or for the person with dementia [10]. Having multiple sensors will increase the complexity
of the monitoring system and make it more cumbersome for the users [11]. While many advanced
machine learning methods have been used to measure stress using multiple information sources [12],
we aim for a more straightforward method based on such limited input. The data from the wristband
is transferred and processed in an online service that calculates the stress level that the person with
dementia had. This data is presented to the caregivers and the clinical staff in a timeline, including
a single-day view and a multi-day view. The system also marks patterns of high stress over a longer
period, so the clinical staff can observe those patterns and identify regular stress episodes that the
user experiences.
The usability and the effectiveness of the above-described sensor-based system were explored by
the authors in a previous study [4]. In this article, the authors extend their previous work by validating
the proposed system. Episodes of stress that are detected by the system are compared to observation
notes written by the caregivers to evaluate the accuracy of the system. The work investigates if it is
feasible to rely on a wristband band sensor for detecting stress episodes that people with dementia
experience. The remainder of this article is organized as follows: Section 2 presents a review of the
related work. In Section 3, the design of the study within this work is discussed. The design section
includes a presentation of the wristband and the data collection protocol. Data processing using the
proposed system is presented in Section 4. The evaluation results from the study as well as clinical user
interfaces are discussed in Section 5. Finally, conclusions and future works are presented in Section 6.
Sensors 2016, 16, 1989 3 of 17

2. Related Work
Many works have focused on other aspects of wearable sensors with success. Gjoreski et al. [13]
examined fall detection from accelerometer measurements on wearable sensors. Such insights can be
useful to provide real-time alerts for caregivers to intervene. On the other hand, our work focuses
on treatment over a longer period and quality of life by reducing and intervening in stress by using
wearables. Therefore, this section considers the state-of-the-art in stress detection rather than activity
recognition and fall detection.
Many studies suggested the use of physiological signals to detect stress and measure
emotions of people, such as Skin Conductance [14], Blood Volume Pulse (BVP) [15] and
Electroencephalography (EGG) [16]. Kirschbaum et al. [17] measured the cortisol in saliva to investigate
the hypothalamus-pituitary-adrenal (HPA) axis activity. The HPA is usually activated when the person
is experiencing a stressful situation, and it is easily measurable in saliva, urine, and blood. Blood
pressure can be also used to give information of the stress level that people have, and it can be measured
by a vital signs monitor. Lupien [18] explained in her book that two numbers appear when measuring
the blood pressure. The first one represents the systolic pressure, and the second one represents the
diastolic pressure. These numbers could be a source for detecting a stressful moment for the person.
However, measuring the cortisol in saliva and blood pressure cannot be done unobtrusively with the
technology available today. This makes it unsuitable for continuous observations of a person over
longer periods of time in everyday life.
Villarejo et al. [19] introduced a stress sensor based on Galvanic Skin Response (GSR), also known
as skin conductance, and controlled by ZigBee. The sensor has two electrodes that are placed
on the fingers, and it sends the data via ZigBee to a coordinator that forwards it to a computer.
Sixteen adults participated in the experiment and the results showed a success rate of detecting stress
by 76.56%. Bakker et al. [20] proposed the use of a watch-style stress measurement device that captures
stress-related physiological signs. The aim of the framework is to manage stress at work by making the
workers aware of the past, current or expected stress. Their watch had two electrodes that require skin
contact in order to produce a reliable signal that is used to give two labels: “stressed” or “not stressed”.
The results indicated that additional data to GSR is needed to make a detection of the stress level,
as there are varieties of patterns in the GSR data.
Perala et al. [21] presented a study to measure the soldiers’ stress in military using GSR.
The authors in [21] compared the stress results from the GSR method with the traditional survey
method. The aim was to verify if GSR is actually measuring stress. GSR data was collected by
using a small, lightweight, unobtrusive body monitor armband. The armband was worn by the
soldiers on the back of the upper arm during the day. The results indicate that GSR is a promising
continuous measure of stress that does not entirely rely on biased self-reporting. Another experiment
was conducted by Hernandez et al. [22] to measure the stress of call center employees after each call
they receive using a wrist skin conductance sensor. Nine employees participated in the experiment,
who wore the sensor and made self-report ratings at the end of each call. The reports were used
to verify if the skin conductance results were accurate, and the results showed 78.03% accuracy in
detecting stressful moments.
The aforementioned works, however, focused on healthy adults as a target group, and not on
people with medical problems, e.g., dementia. In addition, not many studies had a continuous
observation of the user, as they investigated methods to give an indication of the stress level when
performing a specific test for the physiological signals. Our work performs an experiment with people
with dementia in the natural settings of nursing homes. The aim is to have a continuous observation
of the user and test the GSR response on this target group in uncontrolled environment.
A very recent study by Sarker et al. [23] uses sophisticated time-series mining to predict significant
stress episodes for just-in-time interventions. For this, it utilizes a flexible band worn around the
chest, providing rich data (accelerometer, electrocardiograph, and respiration). It also examines
several cofounding factors, such as the varying recovery time and physical activity. The need for
Sensors 2016, 16, 1989 4 of 17

personalization for both stress measurement and intervention is showcased, as the two most stressed
participants are on average twice as stressed as the two least stressed ones. On the other hand, our work
aims for long-term monitoring and interventions in nursing homes, instead of just-in-time ones. In this
framework, it utilizes a single comfortable wrist-worn sensor for simplicity, which provides less input,
and thus employs less sophisticated processing.
Another study utilizes various sensors and information on smartphones to assess stress in school
students [12]. Information input includes activity (smartphone accelerometer), Wi-Fi usage, light levels,
ambient audio such as conversations or noise, battery charge and call logs. Questionnaires were filled
out by the students to self-report severe, mild, or no stress. Learning algorithms were employed
using the above information as features along with school schedule information (such as midterm
exams). Building a model for each student using the rest of the students yielded poor results (up to
43% accuracy). Clustering students of similar traits demonstrated the same performance. Both of
these results led to the need for a personalized model, possibly due to high subjectivity of perceived
stress, i.e., different people appraising the severity of the same negative effects differently, according to
diverse personalities, coping resources and support [24]. Indeed, higher accuracy (60%) was achieved
when using the initial data for each student as a calibration (test) phase. To some extent, our method is
in-line with this approach, calibrating the model with the initial data for each user. Notably, many
dissimilarities do exist when targeting the elderly instead of students, such as logs from the use of
technology (battery, calls, and Wi-Fi), as well as the school schedule information (e.g., midterms).
Focusing on measuring stress in the dementia domain, Rodney [25] did a study to explore the
relationship between the stress of the caregivers and the stress/aggression that people with dementia
show. The author measured the stress of people with dementia using a checklist of questions that
was introduced by Mackay et al. [26]. The checklist was given to the caregivers in a self-report
format, and they were instructed to answer based on their experience with the person with dementia.
The checklist contains 19 adjectives such as “tense” and “peaceful”, and the answers yield a maximum
score of 57. This score was used to give an estimation of the level of stress/aggression of the person
with dementia. The results of the study indicated that the increased stress of the caregivers is strongly
related to the stress/aggression level of people with dementia.
Another study was conducted by Vedhara et al. [27] to investigate the effect of the stress on
the immune system for people with dementia and their caregivers. The stress level of people with
dementia was measured using the Global Measure of Perceived Stress scale [28]. This scale consists of
14 items that measure if a moment of the person’s life is considered stressful. The scale also has queries
about an experienced stress during the day of the person. These studies used the traditional method of
a list of questions that the caregivers should answer to have an understanding of the mental state of
the person with dementia. The work presented in this article relies on a sensor to measure the stress,
so it eliminates the efforts and the time that is consumed by the staff in answering the questions.
There are very few studies that focused on the use of sensors to measure the stress for people
with dementia. Algase et al. [29] used the step watch to measure the wandering behavior of people
with dementia. The wandering activity was used as an indication of stress for the person. There were,
however, no studies that illustrated the use of sensor technology for people with dementia in nursing
homes to measure the stress level. We are not familiar with a similar study to the one presented in
this article, as we are focusing on measuring stress for a special target group in a natural living setting.
Table 1 compares our work to the related work presented before.
Sensors 2016, 16, 1989 5 of 17

Table 1. Comparison of the related work.

Authors Sensors Analysis Application Domain

Wearable electrodermal activity Calculate the stress level from Personal health system at
(Setz et al., 2010)
(EDA) device. the EDA data the workplace
Sensors placed on the hand: galvanic
skin response (GSR), blood volume
Calculate the stress level from
(Zhai et al., 2006) pulse (BVP), skin temperature (ST), Emotion recognition system
the collected data
and eye gaze tracking instrument:
pupil diameter (PD)
Flexcom Infiniti biofeedback device:
Calculate the stress
skin conductance (SC), Emotional stress
(Hosseini et al., 2010) level by analyzing
photoplethysmograph (PPG), recognition system
multi-modal bio-signals
respiratory rate (RR) and EEG.
Calculate the stress level from
(Kirschbaum et al., 1989) N/A N/A
the cortisol level in saliva
Calculate the stress level from
(Lupien, 2013) Simple vital signs monitor N/A
the blood pressure
Two electrodes placed on the fingers: Calculate the stress level based
(Villarejo et al., 2012) Stress monitoring system
Galvanic Skin Response (GSR) on GSR data
Watch-style stress measurement Calculate the stress level based
(Bakker et al., 2011) Stress management system
device: Galvanic Skin Response (GSR) on GSR data
Armband worn on the back of the
Calculate the stress level based
(Perala et al., 2007) upper arm: Galvanic Skin Stress monitoring system
on GSR data
Response (GSR)
Wrist skin conductance sensor: Calculate the stress level based
(Hernandez et al., 2011) Stress monitoring system
Galvanic Skin Response (GSR) on GSR data
Chest-worn band (Accelerometer, Predict significant stress Just-in-time interventions at
(Sarker et al., 2016)
respiration, electrocardiogram (ECG)) episodes from time-series data work or daily life
Smartphone: Accelerometer, Audio
Assess student behavior from School student
(Gjoreski et al., 2015) (ambient noise), Wi-Fi, Call logs,
sensors and questionnaires stress assessment
Battery level, Light
Calculate the stress/aggression
Stress monitoring system,
(Vic Rodney., 2000) N/A level based a checklist
Dementia Care
of questions
Calculate the stress/aggression
Stress monitoring system,
(Mackay et al., 1978) N/A level based on a checklist
Dementia Care
of questions
Calculate the stress using Global
(Vedhara et al., 1999) N/A Measure of Perceived Dementia Care
Stress scale
Calculate the stress using Global
(Cohen et al., 1983) N/A Measure of Perceived N/A
Stress scale
Assess the wandering behavior, Stress monitoring system,
(Algase et al., 2003) Step watch
and use it as a sign for stress Dementia Care
Wristband sensor: Galvanic Skin Stress monitoring system,
Calculate the stress level based
Our Approach Response (GSR), accelerometers Context management, User
on GSR and ACC data
data (ACC) profiling, Dementia Care

3. Design
The study was conducted in two nursing homes in the northern part of Sweden. The staff
members in the nursing homes are familiar with the NPI-NH instrument [8], and they usually use
a computerized system for manually registering BPSD domains and visualizing them long-term.
This manual registration is part of a national program in Sweden for improving the quality of care
of people with BPSD [30]. Residents of the nursing homes live in their own apartments, and the
study was carried out in natural settings. Trained research staff, both technical experts and experts
in dementia care, set up the test including training and supervising the staff in the nursing homes.
The pilot involved thirty staff members and six participants with dementia. The participants are
diagnosed with Alzheimer’s disease, and they were recruited to take part in the data collection based
on their BPSD problems. Problems include, for example, sleeping problems, anxiety, difficulty with
orientation, disturbance by other residents, agitation, and wandering.
Sensors 2016, 16, 1989 6 of 17

3.1. Deployment
The platform, called DemaWare@NH, was explained in details by the authors in [4].
The DemaWare@NH system highlights the most common behavioral and psychological problems of
people with dementia that are related to stress [1]. It consists of a wristband sensor and an online service
that processes the sensor data and visualizes the assessment results. The wristband sensor is a wearable
skin sensor that is worn at the wrist by the participant, and activated by the staff. The wristband collects
a number
Sensors 2016, of
16, different
1989 parameters, typically GSR (skin conductance), accelerometers (for capturing 6 of 16
motion), skin temperature, environment temperature and environment light. Figure 1 shows the
environment
wristband light.
sensor Figure
that 1 shows the
was deployed wristband
during sensor
the pilot, thatthe
namely was deployed
Philips sensorduring the pilot,Research,
DTI-2 (Philips namely
the Philips sensor DTI-2 (Philips
Eindhoven, The Netherlands) [31]. Research, Eindhoven, The Netherlands) [31].

Figure 1. Philips DTI-2 wristband sensor.

Figure 1. Philips DTI-2 wristband sensor.

3.2. Data Collection

3.2. Data Collection
3.2.1. Sensors
3.2.1. Sensors
The DTI-2
The DTI-2sensor
sensorwaswas activated
activated andand puttheon
put on the participant’s
participant’s wrist bywrist by when
the staff the staff when the
the participant
participant was awake, and was put on a charger when the participant was sleeping.
was awake, and was put on a charger when the participant was sleeping. The aim was to observe The aim wasthe
to
observe
stress theof
level stress level of the when
the participant participant when
the user wasthe user during
awake was awake during
the day thenight.
or the day orEach
the night. Each
participant
had the sensor for a period of two months. The researchers collected the data from the sensor the
participant had the sensor for a period of two months. The researchers collected the data from on
sensor on a weekly basis and processed it using
a weekly basis and processed it using the platform. the platform.

3.2.2. Observation
3.2.2. Observation Notes
Notes
The staff
The staffwere
wereprovided
provided with
with observation
observation notesnotes to indicate
to indicate the of
the status status of thewith
the person person with
dementia
dementia based on their own observations. The notes had a timeline of 24
based on their own observations. The notes had a timeline of 24 h and four categories, includingh and four categories,
includingAggression,
Sleeping, Sleeping, Aggression,
Stress, and Stress,
Normal. andTheNormal.
staff wereTheasked
staff to
were
mark asked
momentsto mark moments
of the of the
day based on
day based on the categories using specific colors for each one. For instance, an hour of
the categories using specific colors for each one. For instance, an hour of stress for the participant will stress for the
participant
be marked onwillthebetimeline
markedwith
on the
red,timeline
while thewith red, while
periods thewill
of sleep periods of sleep
be marked willblue.
with be marked with
An example
blue.
of the An examplenotes
observation of the
forobservation
a participantnotes for in
is shown a Figure
participant
2. As is
theshown in Figure
data collection was2. conducted
As the datain
collection was conducted in natural settings, there were moments that the staff
natural settings, there were moments that the staff forgot to mark the observation notes. The notes, forgot to mark the
observation
however, notes. The
provided notes,
a wealth of however,
information provided a wealth oftoinformation
for the researchers compare with forthe
thesystem’s
researchers to
results
compare
after with the
processing system’s
the sensor’sresults
data. after processing the sensor’s data.
Figure 3 summarizes the overall design
Figure 3 summarizes the overall design ofof the
the study.
study.
 day based on the categories using specific colors for each one. For instance, an hour of stress for the
participant will be marked on the timeline with red, while the periods of sleep will be marked with
blue. An example of the observation notes for a participant is shown in Figure 2. As the data
collection was conducted in natural settings, there were moments that the staff forgot to mark the
observation notes. The notes, however, provided a wealth of information for the researchers to
compare
Sensors 2016, 16,with
1989 the system’s results after processing the sensor’s data. 7 of 17
Figure 3 summarizes the overall design of the study.

Figure 2. An observation note filled in by the staff for a participant over a week.
Figure 2. An observation note filled in by the staff for a participant over a week.
Sensors 2016, 16, 1989 7 of 16

Figure
Figure 3.
3. The
The overall
overall design
design of
of the
the study.
study.

4. Data Processing
4. Data Processing
The data of the DTI-2 was processed by the researchers for each participant on a weekly basis.
The data of the DTI-2 was processed by the researchers for each participant on a weekly basis.
Data was analyzed by the platform based on the GSR measurements. GSR is one of the known
Data was analyzed by the platform based on the GSR measurements. GSR is one of the known
indicators for stress and other emotional states, such as agitation or aggression [32]. Stress is often
indicators for stress and other emotional states, such as agitation or aggression [32]. Stress is often
accompanied by a physiological response, which results in an increase of GSR, which can be measured
accompanied by a physiological response, which results in an increase of GSR, which can be measured
between two electrodes placed on the skin, in nano-Siemens (nS). The signal entails two components:
between two electrodes placed on the skin, in nano-Siemens (nS). The signal entails two components:
the rapidly-changing skin conductance response, which corresponds to short-term external stimuli,
the rapidly-changing skin conductance response, which corresponds to short-term external stimuli,
and skin conductance changing more slowly (after several seconds or more). The latter is the signal of
and skin conductance changing more slowly (after several seconds or more). The latter is the signal of
interest here, as it reflects longer-term emotional changes regardless of external stimuli.
interest here, as it reflects longer-term emotional changes regardless of external stimuli.
The GSR signal is often noisy, as it contains many artifacts due to movement, short-term spikes
The GSR signal is often noisy, as it contains many artifacts due to movement, short-term spikes
and noise due to various causes. Another drawback is the high variability of the measurements
and noise due to various causes. Another drawback is the high variability of the measurements
between individuals, as apart from arousal and stress, GSR is also influenced by other physiological
between individuals, as apart from arousal and stress, GSR is also influenced by other physiological
processes and external factors such as ambient temperature. Thus, a baseline needs to be established
processes and external factors such as ambient temperature. Thus, a baseline needs to be established
per individual after a short period of observation. In this approach, we create a baseline to
per individual after a short period of observation. In this approach, we create a baseline to statistically
statistically classify each measurement as high or low stress. Initially, the method applies filtering to
clear out artifacts and noise from the signal. The filtering steps are given below:
 No-contact zero values: the signal often contains instances of values equal to zero or very close
to zero. These values correspond to moments that the sensor lost skin contact, either
intentionally or unintentionally, due to movement. Therefore, if 90% of values within a 5 s
window do not exceed the lower bound threshold (experimentally found to be 0.001 nS), then
Sensors 2016, 16, 1989 8 of 17

classify each measurement as high or low stress. Initially, the method applies filtering to clear out
artifacts and noise from the signal. The filtering steps are given below:

• No-contact zero values: the signal often contains instances of values equal to zero or very close to
zero. These values correspond to moments that the sensor lost skin contact, either intentionally or
unintentionally, due to movement. Therefore, if 90% of values within a 5 s window do not exceed
the lower bound threshold (experimentally found to be 0.001 nS), then they are removed.
• Removing Spikes: generally, GSR levels were experimentally found to be changing no more than
20% when increasing and 10% when decreasing, each second. A moving one-second median
filter is used for an initial interpolation to even out the signal. Outliers exceeding those values are
considered noise.

These filtering methods mostly remove noisy intervals from the signal, such as the initial period
of sensor calibration and final periods of moving and taking the sensor off.

• Obtain the long-term skin conductance component: in order to obtain the slowly changing
(after several seconds) component of skin conductance, a low-pass filter is applied, in the form of
a sliding window filter, with a window size of one minute, accepting values when at least 60% of
the window is not noise.
• Filling the noise gaps: linear interpolation is used to reconstruct the signal, filling the gaps due to
noise. However, the values generated in this step are not considered when setting a baseline or
categorizing stress.

After the signal has been cleared, the stress measurement method entails baseline construction
and actual classification:

• Baseline Construction: a baseline is needed for each individual in order to achieve a personalized
stress interpretation. The baseline is kept in the form of a 600-bin histogram storing the frequency
of the appearance of each skin conductance level. This frequency is measured over the entire
signal. If a previous histogram entry exists, the new histogram is merged into it.
• Stress level measurement: The method follows state-of-the-art guidelines to determine the lowest,
i.e., resting stress level [32]. Specifically, it assumes that a five-minute window that the person is
at rest can be found, for which the maximum skin conductance level within it is the smallest of
all other maximum values of five-minute windows in the signal (excluding noise). This lowest
maximum is called the zero level, i.e., ‘10’, and serves as the lowest level of stress encountered
in the recordings. Combined with the skin conductance level histogram, l0 is stored as part of
the baseline.

The next step is to find a reasonable cut-off point for high stress, i.e., the ‘l5’ level, as the algorithm
considers five distinct stress levels. The l5 level of skin conductance is selected using the baseline
histogram as the first level (greater than the mean), for which no values in the baseline have been
observed. To avoid setting a value of l5 that is too low, less than 40% of observed levels should be
greater than the chosen value.
The remaining levels are distributed equally, according to the formulae below:

−l0
δ = l54.5
l1 = l0 + 12 δ
l2 = l1 + δ (1)
l3 = l2 + δ
l4 = l3 + δ

Finally, a given value is classified within each stress level boundary. Notably, a single value is
not selected to serve as a threshold. As a result, the algorithm does not actually highlight events
Sensors 2016, 16, 1989 9 of 17

of stress, but rather stops at defining a stress level from one to five for each time segment. Indeed,
as thresholds can yield different outcomes, this allows for further tailoring the method at later stages,
e.g., by setting a threshold at the GUI. The evaluation section shows an investigation of how threshold
values affect performance.

5. Evaluation and Results

The goal of this experimental evaluation is to show how well the sensor analysis methods, as
described in Section 4, manage to align to clinical observation notes. First of all, we start with a
description of the data set collected and utilized in this study. Then, the process of evaluation is
described. To begin with, clinical notes and sensor data are incompatible with one another, both as
modalities and temporal entries. Therefore, we establish a common information schema and mapping
methods to evaluate stress predictions across corresponding clinical notes.
In detail, the evaluation framework is established as follows:

• A common information schema is devised as a common base between both notes

and measurements.
• Clinical notes are translated/mapped upon this schema.
• Stress measurements from sensor data analysis are mapped and translated to the information
schema using a threshold.
• Corresponding measurements are evaluated across corresponding clinical notes using standard
metrics, for various threshold settings and individuals.

5.1. Data Set

The evaluation process utilizes a collected dataset from six individuals living in the nursing
home, along with ground truth obtained from clinical notes. In detail, the dataset contains skin
conductivity measurements per minute, while clinical notes mention the segments of stress incidents
for the involved individuals. While the former type of data is quite commonly found online with the
emergence of wearables and the Internet of Things (IoT), the latter is quite rare and hard to obtain,
as it involves a clinical, ethically-approved and well-defined process. Using this dataset as the only
benchmarking resource, at least to the best of our knowledge, prohibits comparisons with existing
techniques at this stage.
Table 2 provides some insights to the data set collected and used in this evaluation. In detail, it
shows positive, negative, and total instances for each user, treated as a separate set. Each instance
corresponds to an hour-long observation marked as either “Stressed” (Positive) or “Not Stressed”
(Negative) by clinicians, reaching up to more than 2400 instances. However, the percentage of positive
instances to the total is in most cases quite low. This is shown clearly on Figure 4, which displays the
percentage of stressed versus not stressed observation instances, total and per person.

Table 2. Observation note instances per user, showing the positive and negative count, and positive
instances percentage to total.

User Positive (Stressed) Negative (Not Stressed) User Total Positive/User Total
User 1 42 464 506 8.30%
User 2 8 379 387 2.07%
User 3 29 193 222 13.06%
User 4 41 223 264 15.53%
User 5 101 254 355 28.45%
User 6 19 689 708 2.68%
All Users 240 2202 2442 9.83%
 User 2 8 379 387 2.07%
User 3 29 193 222 13.06%
User 4 41 223 264 15.53%
User 5 101 254 355 28.45%
User 6 19 689 708 2.68%
Sensors 2016, 16, 1989 10 of 17
All Users 240 2202 2442 9.83%

Figure 4.4.Class
Classdistribution
distribution in dataset
in the the dataset as a percentage
as a percentage of instances
of instances of stressin reported
of stress reported in
observation
notes out of the
observation total
notes outhours
of thelogged.
total hours logged.

5.2. Information Schema

While the skin conductivity processing method provides stress measurements, the evaluation
process provides a series of transformation steps for them and the clinical notes alike in order to
synchronize them
synchronize themandandyield
yieldperformance
performance metrics.
metrics. Initially,
Initially, clinical
clinical notesnotes in a piece
in a piece of paper,
of paper, as shown as
shown in Figure 2, have to be transformed into a machine-interpretable format
in Figure 2, have to be transformed into a machine-interpretable format and also strictly defined and also strictly
defined
to serve to
asserve
groundas ground truth. Therefore,
truth. Therefore, we defined
we defined a structured
a structured schema schema for clinical
for clinical observations
observations to be
to be mapped
mapped to ground
to ground truth. Clinical
truth. Clinical notes
notes do not do not necessarily
necessarily accountaccount for themonitoring
for the exact exact monitoring
period,
period,
but, on but, on the contrary,
the contrary, overlapoverlap it. Hence,
it. Hence, the schema
the schema shouldshould specify
specify whether
whether an incident
an incident refers
refers to
‘Stressed’, ‘Not Stressed’ or no observation (lacking notes). The same schema should similarly be able
to accommodate the observations derived from sensor measurements. The resulting schema is defined
as shown in Table 3:

Table 3. Information schema for representing ground truth from clinical notes and analyzed sensor input.

Field Person ID Start Time End Time Type Provider

Datetime as a UNIX Datetime as a UNIX ‘Not Stressed’ ‘Clinical Notes’
Type String
Timestamp Timestamp or ‘Stressed’ or ‘Sensor’

In this schema, each event is characterized by a starting and an ending timestamp, and a unique
(anonymous) person ID. The event type is either a ‘Stressed’ or ‘Not Stressed’ label while the
information provider field is either the ‘Clinical Notes’ or the ‘Sensor’ analysis component, covering
effectively all requirements for evaluation.

5.3. Mapping Clinical Notes to Information Schema

In order to transform handwritten notes to the information schema for ground truth, clinical staff
used simple spreadsheets (Excel), as a human-friendly authoring tool. The notes are divided into
hourly segments. The researchers used Excel to insert one row for each instance of consecutive hours
with the same observation (type), either stressed or not stressed. Following the given example, Sleeping
and Normal are considered no stress observation, while Aggression and Stress are considered simply
stress. The resulting Comma Separated Values (CSV) files were generated from the spreadsheets and
imported into the evaluation tool’s database to serve as ground truth.
Sensors 2016, 16, 1989 11 of 17

5.4. Mapping Stress Measurements to Information Schema

In general, stress measurements extracted from sensor measurements processed by skin
conductivity analysis also have to align to the common information schema in terms of structure.
For this reason, the outcomes from the analysis have to either signify stressed or not stressed instead
of providing stress level as a numerical value. Since the values are integers between one and five,
they are actually investigated in this method as a parameter whose value could yield the best results.
This investigation is presented further in this section.
Apart from aligning to the schema, the information between sensors and ground truth should
also align temporally in order to perform direct evaluation comparisons. For this purpose, analysis
outcomes are aggregated and rounded to provide a single value per hour. Each hour-long segment,
furthermore, begins at minute zero and ends accordingly as provided in the clinical notes. To aggregate
the values, the method averages values within each segment.

5.5. Metrics
After aligning actual stress incidents from ground truth and predicted outcomes, they can be
directly compared to measure the following attributes: TP (true positive), FP (false positive), TN (true
negative), and FN (false negative). In turn, these attributes can be combined to provide a comprehensive
set of metrics. The method considers the following metrics, as established in the field of information
retrieval and binary classification:
TP
Precision = , (2)
TP + FP
TP
Recall = , (3)
TP + FN
TP + TN
Accuracy = , (4)
TP + FP + FN + TN

5.6. Results
The sensor recordings from the six participants added up to 142 h across 37 (calendar) days,
with corresponding clinical notes of an equal number of (full) days. The goal of this study is to evaluate
the effectiveness of classifying events i.e., time segments as either ‘stressed’ or ‘not stressed’. To do
so, we treated time interval (sampling rate) and stress value threshold as parameters to be optimized.
Experimentation with the time interval parameter, from one minute, thirty minutes, an hour, two hours
etc., resulted in a clear, monotonous optimization of all metrics when sampling events per hour. Due to
this result being straightforward, we do not thoroughly present performance for other sampling rates,
but rather consider the one hour period as a given for the rest of this study.
The second parameter to optimize is the numeric threshold above which we consider stress level
measurements to signify a stress event. As presented previously, the current stress algorithm outputs
stress level as an integer value from zero to five. Indeed, this is suitable for cases where thresholds
have to be adjusted to certain conditions, e.g., environmental or physiological. In order to find the
optimal stress threshold for this particular study, we performed the evaluation for all metrics and all
participants for each distinct integer value from zero to five. For each value, an interval was marked
with the ‘stressed’ label if the average value within the interval was larger or equal to that. Note that
for a threshold equal to zero, all intervals are considered as stress.
Table 4 summarizes the results for all metrics and for each distinct stress level threshold. Generally,
the outcome helps support the choice of using a different threshold according to the situation.
As expected, raising the threshold causes fewer events to be classified as stress, due to harder criteria.
This, in turn, increases precision, i.e., the percentage of true positive events within all positive events.
Still, it peaks at level four, dropping low at level five. This goes to show that perceived (observed)
stress corresponds with middle-to-high physical responses and not high ones. Meanwhile, recall drops
Sensors 2016, 16, 1989 12 of 17

monotonously, as many actual stress events are filtered out by the increased threshold, i.e., fewer true
positive events are found out of the total actual stress events. Accuracy, much like precision, is also
increased as the threshold increases. This is due to more accurately identifying true negatives along
with true positives. On the other hand, the F-measure clearly does not favor extremes, neither the
lowest nor the highest value, offering some balance. Combined with precision peaking before level
five, a threshold of three to four offers a well-balanced solution.

Table 4. Performance metrics per stress level threshold.

Metric/Stress Level Threshold 0 1 2 3 4 5

Precision 9.9% 12.1% 13.9% 19.1% 20.9% 4.2%
Recall 100.0% 81.7% 65.9% 48.2% 20.4% 0.9%
Accuracy 9.9% 43.3% 59.8% 75.9% 85.0% 89.4%
F-measure 17.6% 20.7% 22.6% 26.8% 18.9% 1.4%

Figure 5 shows the continuous change of metrics alongside the threshold. As F-measure also
confirms, an optimal threshold for increased precision, accuracy and decent recall is level 3. However,
this does not mean that this threshold should be used in all cases. Presenting all the metrics in this
study gives a holistic picture of the quality each threshold gives. This allows clinicians to pre-select
a threshold according to scenario. For example, if interested in seeing all stress events while tolerating
false positives, a low threshold can be chosen, such as three. On the contrary, if interested in receiving
severe events only, or else getting a high percentage of actual stress events amongst the selected ones,
aSensors
high2016, 16,such
value 1989 as four or five would suit the requirements better. 12 of 16

Figure 5. Performance metrics across different stress level thresholds.

thresholds.

These outcomes
These outcomes havehave led
led to
to aa further
further personalized
personalized method,
method, where
where aa stress
stress level
level threshold
threshold isis
previously set. In this case, the threshold to optimize the F-measure metric was chosen
previously set. In this case, the threshold to optimize the F-measure metric was chosen for each user, for each user,
yielding the
yielding theoutcomes
outcomesshownshowninin Figure
Figure 6. Indeed,
6. Indeed, F-measure
F-measure score
score has been
has been increased
increased for users.
for most most
users. The best performing user-set is User 5, which also holds the highest
The best performing user-set is User 5, which also holds the highest ratio of positive to negative ratio of positive to
negative instances in his set. The lowest F-measure is met in Users 2 and 6, which,
instances in his set. The lowest F-measure is met in Users 2 and 6, which, at the same time, hold at the same time,
ahold a quite
quite low positive
low positive to negative
to negative instanceinstance ratio, hinting
ratio, hinting at howat thehow
twothe two might
factors factorsbemight be Indeed,
linked. linked.
Indeed, the method utilizes the initial measurements of a user to classify stress events and not the
entire set. This means that adequate stress events should have been noted in this initial measurement
period.
 These outcomes have led to a further personalized method, where a stress level threshold is
previously set. In this case, the threshold to optimize the F-measure metric was chosen for each user,
yielding the outcomes shown in Figure 6. Indeed, F-measure score has been increased for most
users. The best performing user-set is User 5, which also holds the highest ratio of positive to
negative instances in his set. The lowest F-measure is met in Users 2 and 6, which, at the same time,
Sensors 2016, 16, 1989 13 of 17
hold a quite low positive to negative instance ratio, hinting at how the two factors might be linked.
Indeed, the method utilizes the initial measurements of a user to classify stress events and not the
entire
the set. This
method means
utilizes that
the adequate
initial stress events
measurements of ashould
user tohave been
classify noted
stress in thisand
events initial
not measurement
the entire set.
period.
This means that adequate stress events should have been noted in this initial measurement period.

Figure 6.
Figure 6. Performance metrics
Performance of the
metrics of personalized threshold
the personalized methodmethod
threshold to optimize F-Measure,
to optimize per each
F-Measure,
user.
per each user.

5.7. Clinical
5.7. Clinical View
View
Based on
Based on the
the above
above results,
results, two
two methods
methods for
for the
the clinicians
clinicians to
to monitor
monitor individuals
individuals are
are apparent:
apparent:
the general method, where no threshold for stress level is set or known, and the personalized
the general method, where no threshold for stress level is set or known, and the personalized method,
where the clinician picks an optimized stress level threshold for the scenario or individual. This was
done to provide clinicians with more flexibility. Most often, the stress level threshold varies across
individuals, as shown during experimentation. Furthermore, it may also vary between time periods
(although this is not investigated in the present study). Finally, clinicians at times follow different
investigation use cases. They may want to have a quick overview of a large period in time or go
in-depth and investigate stress fluctuations in a large resolution, such as per minute or per hour.
In order to support all of these scenarios and flexibility, the web application which exposes stress
measurements to the clinicians, provides many functions.
As shown in Figure 7, it initially supports a slider to pick sampling resolution, ranging from per
minute to per month averages for the metrics shown. Then, the general view scenario is supported by
showing a line chart of stress level averages, for the selected period and scale. As Figure 7 shows, daily
average stress level fluctuations are presented as measured for the clinician to be able to investigate
the patterns themselves.
Alternatively, the web application also serves for personalized investigations of stress fluctuations
as follows. The Dashboard, shown on Figure 8, allows a clinician to run a stress analysis, namely
set a stress threshold for a given individual (anonymized in the figure) and a given time period.
This simply marks “High Stress” events in the log of measurements where they exceed the given,
personalized threshold.
 daily
stress average stresstolevel
measurements fluctuations
the clinicians, are presented
provides as measured for the clinician to be able to
many functions.
investigate the patterns themselves.
As shown in Figure 7, it initially supports a slider to pick sampling resolution, ranging from per
minute to per month averages for the metrics shown. Then, the general view scenario is supported
by showing a line chart of stress level averages, for the selected period and scale. As Figure 7 shows,
daily 2016,
Sensors average stress level fluctuations are presented as measured for the clinician to be able
16, 1989 14 of to
17
investigate the patterns themselves.

Figure 7. Web application showing daily average stress level.

Alternatively, the web application also serves for personalized investigations of stress
fluctuations as follows. The Dashboard, shown on Figure 8, allows a clinician to run a stress analysis,
namely set a stress threshold for a given individual (anonymized in the figure) and a given time
period. This simply marks “High Stress” events in the log of measurements where they exceed the
Figure 7. Web application showing daily average stress level.
Figure 7. Web application showing daily average stress level.
given, personalized threshold.

Alternatively, the web application also serves for personalized investigations of stress
fluctuations as follows. The Dashboard, shown on Figure 8, allows a clinician to run a stress analysis,
namely set a stress threshold for a given individual (anonymized in the figure) and a given time
period. This simply marks “High Stress” events in the log of measurements where they exceed the
given, personalized threshold.

Figure
Figure 8. 8. Dashboardfor
Dashboard forthe
theclinicians
clinicians to
to run
run analysis,
analysis, setting
setting aa personalized
personalizedstress
stresslevel threshold
level threshold
(“Threshold”) for each user (anonymized drop-down menu) and time period (“From”–“To”
(“Threshold”) for each user (anonymized drop-down menu) and time period (“From”–“To” fields). fields).
Sensors 2016, 16, 1989 14 of 16

The
Theresult
resultof
ofthe
thestress
stressanalysis
analysis is
is shown
shown on on Figure
Figure 9. A timeline
9. A timeline ofof stress
stressevents
eventsisisdisplayed,
displayed,
essentially hiding all detailed stress level measurements
essentially hiding all detailed stress level measurements from the clinician. This view is particularly
clinician. This view is particularly
useful
usefulwhen
when reviewing
reviewingan individual’s log at log
an individual’s a glance.
at a Stress level
glance. threshold
Stress level and period combinations
threshold and period
may Figure 8. Dashboard
be changed at any for the clinicians to run analysis, setting a personalized stress level threshold
time.
combinations may be changed at any time.
(“Threshold”) for each user (anonymized drop-down menu) and time period (“From”–“To” fields).

Figure9.9. View
Figure Viewof
of High
High Stress
Stress level
level problems,
problems, according
according to
to personalized
personalized thresholds.
thresholds.

6. Conclusions
This paper presented a framework to measure and highlight stress events and also presented it
as a sign of behavioral patterns of people with dementia to clinical nursing home staff. The system
utilizes a wearable sensor that measures GSR. GSR data is processed using a per-user trained
statistical method, extracting stress measurements classified in five levels of severity. The method
Sensors 2016, 16, 1989 15 of 17

6. Conclusions
This paper presented a framework to measure and highlight stress events and also presented it
as a sign of behavioral patterns of people with dementia to clinical nursing home staff. The system
utilizes a wearable sensor that measures GSR. GSR data is processed using a per-user trained statistical
method, extracting stress measurements classified in five levels of severity. The method was evaluated
by six participants wearing the equipment over a period of two months with thirty staff members
taking clinical notes to serve as ground truth. Experimenting with stress level thresholds showed
that different thresholds can serve different scenarios. For instance, a high threshold yields high
accuracy and high precision, a low threshold yields high recall, while the medium threshold achieves
the best balance between them. This stimulates research for more tailored prediction methods based
on machine learning.
Future work will focus on collecting more data from real users to further validate the presented
method in this article. A pilot is currently running with a new wristband, called Empatica E4 [33],
which measures the GSR level, accelerometers, and pulses. The data from the new sensor will be used
to further develop the stress detection algorithm to approach better results. Pulse measurements, for
instance, will be integrated into the algorithm to provide better detection of stress episodes and to
achieve the best fit per scenario and even per user.

Acknowledgments: This work is funded by a grant from Vårdalsstiftelsen at Luleå University of Technology.
The work has been also supported by the EU FP7 project Dem@Care: Dementia Ambient Care—Multi-Sensing
Monitoring for Intelligent Remote Management and Decision Support (contract No. 288199).
Author Contributions: Basel Kikhia initiated the work together with Thanos Stavropoulos in the beginning and
participated in writing all sections. Basel Kikhia had the responsibility of the design section and in running the
experiment in the nursing homes and reporting the results. Thanos Stavropoulos performed experiments and
wrote the integration and experimental parts of the paper. Stelios Andreadis developed the tools and software for
evaluation. Niklas Karvonen has been part of the discussion prior to analysis, helped formatting the user-input
data for processing, and has also contributed in writing the data processing section and the revision of the
work. Ioannis Kompatsiaris conceived and coordinated the software engineering aspects and the experimental
evaluation in this study. Stefan Sävenstedt and Catharina Melander participated in setting up the pilot and
they reviewed the work and gave feedbacks to improve the quality of the paper. Marten Pijl contributed in the
development of raw sensor signal transformation to stress level.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Levenson, R.W.; Sturm, V.E.; Haase, C.M. Emotional and behavioral symptoms in neurodegenerative disease:
A model for studying the neural bases of psychopathology. Annu. Rev. Clin. Psychol. 2014, 28, 581–606.
[CrossRef] [PubMed]
2. Ferretti, L.; McCurry, S.M.; Logsdon, R.; Gibbons, L.; Teri, L. Anxiety and Alzheimer’s disease. J. Geriatr.
Psychiatry Neurol. 2001, 14, 52–58. [CrossRef] [PubMed]
3. Bergh, S.; Engedal, K.; Røen, I.; Selbæk, G. The course of neuropsychiatric symptoms in patients with
dementia in Norwegian nursing homes. Int. Psychogeriatr. 2011, 23, 1231–1239. [CrossRef] [PubMed]
4. Kikhia, B.; Stavropoulos, T.G.; Meditskos, G.; Kompatsiaris, I.; Hallberg, J.; Sävenstedt, S.; Melander, C.
Utilizing ambient and wearable sensors to monitor sleep and stress for people with BPSD in nursing homes.
J. Ambient Intell. Humaniz. Comput. 2015, 1–13. [CrossRef]
5. Alzheimer’s Statistics. Available online: http://www.alzheimers.net/resources/alzheimers-statistics/
(accessed on 19 June 2016).
6. Moyle, W.; Cooke, M.L.; Beattie, L.; Shum, D.H.K.; O’Dwyer, S.T.; Barrett, S.; Sung, B. Foot massage and
physiological stress in people with dementia: A randomized controlled trial. J. Altern. Complement. Med.
2014, 20, 305–311. [CrossRef] [PubMed]
7. Gomoll, B.P.; Kumar, A. Managing anxiety associated with neurodegenerative disorders. F1000Prime Rep.
2015, 7. [CrossRef] [PubMed]
8. Cummings, J.L.; McPherson, S. Neuropsychiatric assessment of Alzheimer’s disease and related dementias.
Aging 2001, 13, 240–246. [CrossRef] [PubMed]
Sensors 2016, 16, 1989 16 of 17

9. Cummings, J.L.; Mega, M.; Gray, K.; Rosenberg-Thompson, S.; Carusi, D.A.; Gornbein, J. The Neuropsychiatric
Inventory: comprehensive assessment of psychopathology in dementia. Neurology 1994, 44, 2308–2314.
[CrossRef] [PubMed]
10. Kikhia, B. Acceptance of Ambient Intelligence (AmI) in Supporting Elderly People and People with Dementia.
Master’s Thesis, Luleå University of Technology, Luleå, Sweden, 2008.
11. Kikhia, B.; Gomez, M.; Jiménez, L.L.; Hallberg, J.; Karvonen, N.; Synnes, K. Analyzing body movements
within the Laban effort framework using a single accelerometer. Sensors 2014, 14, 5725–5741. [CrossRef]
[PubMed]
12. Gjoreski, M.; Gjoreski, H.; Lustrek, M.; Gams, M. Automatic detection of perceived stress in campus students
using smartphones. In Proceedings of the 2015 International Conference on Intelligent Environments (IE),
Prague, Czech Republic, 15–17 July 2015; pp. 132–135.
13. Gjoreski, M.; Gjoreski, H.; Lustrek, M.; Gams, M. How accurately can your wrist device recognize daily
activities and detect falls? Sensors 2016, 16, 800. [CrossRef] [PubMed]
14. Setz, C.; Arnrich, B.; Schumm, J.; La Marca, R.; Tröster, G.; Ehlert, U. Discriminating stress from cognitive
load using a wearable EDA device. IEEE Trans. Inf. Technol. Biomed. 2010, 14, 410–417. [CrossRef] [PubMed]
15. Zhai, J.; Barreto, A. Stress detection in computer users based on digital signal processing of noninvasive
physiological variables. In Proceedings of the 28th Annual International Conference Engineering in Medicine
and Biology Society (EMBS’06), New York, NY, USA, 31 August–3 September 2006; pp. 1355–1358.
16. Hosseini, S.A.; Khalilzadeh, M.A. Emotional stress recognition system using EEG and psychophysiological
signals: Using new labelling process of EEG signals in emotional stress state. In Proceedings of the Conference
on Biomedical Engineering and Computer Science (ICBECS), Wuhan, China, 23–25 April 2010; pp. 1–6.
17. Kirschbaum, C.; Hellhammer, D.H. Salivary cortisol in psychobiological research: An overview. Neuropsychobiology
1989, 22, 150–169. [CrossRef] [PubMed]
18. Lupien, S.J. How to Measure Stress in Humans, 2nd ed.; Centre for Studies on Human Stress, Cambridge
University Press: Cambridge, UK, 2013.
19. Villarejo, M.V.; Zapirain, B.G.; Zorrilla, A.M. A Stress sensor based on Galvanic Skin Response (GSR)
controlled by ZigBee. Sensors 2012, 12, 6075–6101. [CrossRef] [PubMed]
20. Bakker, J.; Pechenizkiy, M.; Sidorova, N. What’s your current stress level? Detection of stress patterns from
GSR sensor data. In Proceedings of the 2011 IEEE 11th International Conference on Data Mining Workshops,
Vancouver, BC, Canada, 11 December 2011; pp. 573–580.
21. Perala, C.H.; Sterling, B.S. Galvanic Skin Response as a Measure of Soldier Stress; Technical Report No.
ARL-TR-4114; Army Research Laboratory: Adelphi, MD, USA, 2007.
22. Hernandez, J.; Morris, R.R.; Picard, R.W. Call center stress recognition with person-specific models.
Affect. Comput. Intell. Interact. 2011, 6974, 125–134.
23. Sarker, H.; Tyburski, M.; Rahman, M.; Hovsepian, K.; Sharmin, M. Finding significant stress episodes in
a discontinuous time series of rapidly varying mobile sensor data. In Proceedings of the 2016 CHI Conference
on Human Factors in Computing Systems, San Jose, CA, USA, 7–12 May 2016; pp. 4489–4501.
24. Gellman, M.D.; Turner, J.R. Encyclopedia of Behavioral Medicine; Springer: New York, NY, USA, 2013.
25. Rodney, V. Nurse stress associated with aggression in people with dementia: Its relationship to hardiness,
cognitive appraisal and coping. J. Adv. Nurs. 2000, 31, 172–180. [CrossRef] [PubMed]
26. Mackay, C.; Cox, T.; Burrows, G.; Lazzerini, T. An inventory for the measurement of self-reported stress and
arousal. Br. J. Soc. Clin. Psychol. 1978, 17, 283–284. [CrossRef] [PubMed]
27. Vedhara, K.; Cox, N.K.M.; Wilcock, G.K.; Perks, P.; Hunt, M.; Anderson, S.; Lightman, S.L.; Shanks, N.M.
Chronic stress in elderly carers of dementia patients and antibody response to influenza vaccination. Lancet
1999, 353, 627–631. [CrossRef]
28. Cohen, S.; Kamarck, T.; Mermelstein, R. A global measure of perceived stress. J. Health Soc. Behav. 1983, 24,
385–396. [CrossRef] [PubMed]
29. Algase, D.L.; Beattie, E.R.A.; Leitsch, S.A.; Beel-Bates, C.A. Biomechanical activity devices to index wandering
behaviour in dementia. Am. J. Alzheimer’s Dis. Other Dement. 2003, 18, 285–292. [CrossRef]
30. Behavioral and Psychological Symptoms of Dementia Register. Available online: http://www.bpsd.se
(accessed on 21 November 2016).
Sensors 2016, 16, 1989 17 of 17

31. Kocielnik, R.; Sidorova, N.; Maggi, F.M.; Ouwerkerk, M.; Westerink, J. Smart technologies for long-term
stress monitoring at work. In Proceedings of the 26th IEEE International Symposium on Computer-Based
Medical Systems, Porto, Portugal, 20–22 June 2013; pp. 53–58.
32. Kuijpers, E.; Nijman, H.; Bongers, I.; Lubberding, M.; Ouwerkerk, M. Can mobile skin conductance
assessments be helpful in signalling imminent inpatient aggression? Acta Neuropsychiatr. 2012, 24, 56–59.
[CrossRef] [PubMed]
33. E4 Wristband. Available online: https://www.empatica.com/e4-wristband (accessed on 21 November 2016).

© 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).