Wearable Monitoring for the Detection of Nocturnal Agitation in

Dementia

Ana Cristina Marc´en

, Jes´us Carro

1,2,3

and Violeta Monasterio

Universidad San Jorge, Campus Universitario, Autov A23 km 299, 50830, Villanueva de G´allego, Zaragoza, Spain

Aragon Institute for Engineering Research (I3A), University of Zaragoza, Zaragoza, Spain

CIBER in Bioengineering, Biomaterials and Nanomedicine (CIBER-BBN), Madrid, Spain

Keywords:

Wearable Computers, Pervasive Health, Support Vector Machines, Nocturnal Agitation, Accelerometry.

Abstract:

Nocturnal agitation is one of the symptoms exhibited by dementia patients. Diagnosing and monitoring the

evolution of agitation is difﬁcult because patient monitoring requires a doctor, nurse or caregiver observing

patients for extended periods of time. In this work, we propose to use an automatic monitoring system based

on wearable technology that complements the caregiver’s work. The proposed system uses a wrist wearable

device to record agitation data, which are subsequently classiﬁed through machine learning techniques as

quantiﬁable indexes of nocturnal agitation. Preliminary tests performed with volunteers showed that the clas-

siﬁcation of recorded movements between nocturnal agitation or quiet periods was successful in 78.86% of the

cases. This proof of concept demonstrates the feasibility of using wearable technology to monitor nocturnal

agitation.

1 INTRODUCTION

According to recent studies, 35.6 million people lived

with dementia worldwide in 2010, and estimations

predict that these numbers will almost be doubled

every 20 years, 65.7 million living with dementia in

2030 and 115.4 million living with dementia in 2050

(Prince et al., 2013). Approximately one-quarter of

adults with Alzheimer’s disease (AD) and with other

dementias suffer from sleep disturbances (Rose et al.,

2010).

Patients with dementia usually present sleep dis-

turbances such as insomnia, sleep disruption, or

movements that can escalate to become agitation. The

treatment of these disturbances is complex because

they involve multiple factors, such as neurodegen-

erative changes in the brain, the patient’s environ-

ment, medical or psychiatric morbidity, and medi-

cations used to treat chronic illnesses and dementia-

related behavioral symptoms (Deschenes and Mc-

Curry, 2009).

In particular, changes in nocturnal agitation be-

haviors may provide information about the evolution

of dementia. However, it is difﬁcult to get objective

and quantiﬁable information about the nocturnal be-

havior of dementia patients because they are not usu-

ally aware of their behavior and their caregivers can-

not monitor them 24 hours a day (Cooke and Ancoli-

Israel, 2006).

Nocturnal agitation is generally assessed using ob-

servational scales, such as the Cohen-Mansﬁeld Ag-

itation Inventory (CMAI) (Cohen-mansﬁeld et al.,

1989), which is particularly difﬁcult to apply in out-

of-hospital settings. This work presents the ﬁrst

steps towards a pervasive health tool for automati-

cally monitoring nocturnal agitation. In order to de-

tect the nocturnal activity, we propose a system based

on wearable wristband computers similar to watches

because most patients feel comfortable using them.

Then, the collected data is analyzed using machine

learning techniques. In particular, Support Vector

Machines (SVMs), which are widely used for min-

ing physiological data in medical applications (Ba-

naee et al., 2013), are used to identify nocturnal ab-

normal behaviors by classifying movements as nor-

mal or agitated. The aim of this work is to provide an

objective and quantiﬁable characterization of noctur-

nal movements to help medical staff in their diagno-

sis.

The remainder of this article is organized as fol-

lows. Section 2 summarizes related work. Section

3 describes the development of the proposed system:

selection of the wearable device, creation of the refer-

ence dataset, and classiﬁcation algorithm. Finally, in

Marcén, A., Carro, J. and Monasterio, V.

Wearable Monitoring for the Detection of Nocturnal Agitation in Dementia.

DOI: 10.5220/0005938500630069

In Proceedings of the 6th International Joint Conference on Pervasive and Embedded Computing and Communication Systems (PECCS 2016), pages 63-69

ISBN: 978-989-758-195-3

Section 4 and Section 5, we present and discuss the

results respectively.

2 RELATED WORK

Our approach is based on two main aspects: the

use of wearable technology to characterize movement

through collected accelerometer data, and the use of

machine learning techniques to detect nocturnal agi-

tation as a complementary diagnostic tool.

Several works focus on studying the feasibility of

replacing traditional monitoring and diagnostic sys-

tems by less intrusive and more reliable ones in the

case of dementia patients. In (Ancoli-Israel et al.,

1997) and (Van Someren, 1997), actigraphy is a

medium to collect objective sleep data. However, nei-

ther one focuses on nocturnal agitation which is a spe-

ciﬁc symptom of dementia sleep activity (Sink et al.,

2005).

Other works focus on detecting nocturnal agita-

tion of patients with dementia, but they tend to present

complex multi-sensor systems. In (Sakr et al., 2010),

the agitation is detected using a heart rate sensor, a

galvanic skin response sensor and a skin temperature

sensor. In (Fook et al., 2007), the authors propose

a video camera for detecting agitated behavior. In

(Biswas et al., 2006), the system used to detect agi-

tation is composed by acoustic sensors, pressure sen-

sors and ultrasound sensors.

In contrast to previous studies, in this work we

propose a simple system using only accelerometer to

quantify nocturnal agitation in dementia. Accelerom-

eter measures have been found to correlate well with

the CMAI scale (Nagels et al., 2006), so our hypoth-

esis is that accelerometry can be a simple and low-

cost solution for detecting agitation in patients with

dementia.

3 METHODS

To assess the viability of using wearable accelerome-

ter device for nocturnal agitation monitoring, we de-

signed a three-phase proof of concept. During the ﬁrst

phase we studied the wearable accelerometer devices

available in market and their features taking into ac-

count the needs of our research. The second phase

consisted on the creation of the reference dataset,

which involved the deﬁnition of the experimental pro-

tocols, the recruitment of participants, and the record-

ing of data. Finally, the third phase consisted on the

selection and application of a classiﬁcation algorithm

in order to extract useful information about nocturnal

agitation.

3.1 Technological Survey

Nowadays, there is a great diversity of wearable de-

vices in the market, and they will continue to grow

(Intelligence, 2015). Their functionality opens a wide

range of possibilities; phone calls, activity monitor-

ing and geolocation are some of the most common

features of these devices. Unfortunately, most com-

mercial devices do not allow access to raw data. Pop-

ular devices such as Jawbone, Misﬁt, Fitbit or Garmin

trackers use proprietary algorithms to translate the

raw data into generic activity statistics, and their ac-

curacy against gold standard methods remains under-

studied (Kolla et al., 2016). Even if they were accu-

rate enough, generic activity measures such as num-

ber of steps or hours of sleep may not necessarily be

the best surrogate for nocturnal agitation.

Thus, in order to ﬁnd the most appropriate wear-

able device for our research we considered only those

devices that allowed access to raw data. In addition

to that, we considered additional features. Since the

agitation is characterized for a restlessness state and

movements with different intensity degree, a 3-axis

accelerometer was considered to be the most appro-

priate sensor to measure the periods and the intensity

of the movements.

Another feature required in the device is a sam-

pling rate greater or equal than 20 Hz, since the fastest

body movements, such as spasms and tremor, are in

the order of 10 Hz (Bendersky et al., 2014). Other de-

sirable characteristics are ofﬂine data storage capabil-

ities, exportable data in standard format, and low cost.

The memory has to be able to store the activity of a

complete night, so the device logging period needs to

be at least 8 hours at 20 Hz. The possibility of export-

ing data in a standard format is an advantage because

it can be used easily as input data for statistical tools,

signal processing tools or simulation tools. Finally,

in 2010, the total estimated worldwide costs of de-

mentia were US$604 billion (Wimo et al., 2013), so

cost-effective monitoring systems would help reduce

that cost.

Taking into account all of those requirements,

we selected Original GENEActiv by Activeinsights

whose sampling frequency range is 10-100Hz, its

maximum logging period is 22 days @ 20Hz, its

exporting formats are BIN (Binary ﬁle) and CSV

(Comma-separated values), and its cost includes GE-

NEActiv software.

PEC 2016 - International Conference on Pervasive and Embedded Computing

3.2 Creation of the Reference Dataset

In order to develop and validate the classiﬁcation al-

gorithm, a reference dataset was created by record-

ing controlled movements imitating nocturnal agita-

tion and quiet periods in a realistic way.

3.2.1 Protocols

In order to understand the real behavior of dementia

patients, we had the cooperation of a medical advi-

sor, the responsible of the Geriatrics Department of

the Hospital San Juan de Dios, Zaragoza, Spain. She

described for us the usual movements of dementia pa-

tients when the night is quiet, agitated or with insom-

nia.

Taking into account these descriptions, we deﬁned

the movements to be performed by our test subjects

(see Tables 1 and 2). The result was reviewed and

approved by the medical advisor, who conﬁrmed that

the movements closely resembled the nocturnal agita-

tion movements found in dementia patients. Two pro-

tocols were deﬁned for subjects to follow during the

tests. The ﬁrst protocol contained movements of the

extremities, and the second protocol contained move-

ments with the whole body.

Table 1: Protocol P1. Movements of the extremities and

their duration.

Duration Movement

15 s. Remain laying down

10 s. Move right arm from bottom to top

15 s. Remain laying down

10 s. Move left arm from bottom to top

15 s. Remain laying down

10 s. Try to sit up

15 s. Remain laying down

10 s. Move right leg from bottom to top

15 s. Remain laying down

10 s. Move left leg from bottom to top

15 s. Remain laying down

15 s. Stand up and walk

5 s. Move the wearable device on your

wrist

Table 2: Protocol P2. Movements of the whole body and

their duration.

Duration Movement

15 s. Remain laying down

10 s. Sit up on the bed

15 s. Remain laying down

10 s. Laying put in bed

15 s. Remain laying down

10 s. Try to sit up

15 s. Remain laying down

10 s. Stand up

15 s. Remain laying down

10 s. Walk

3.2.2 Recruitment of Volunteers and Data

Recording

Eleven healthy volunteers participated in the experi-

ments, 4 females (mean age 35.5 years, SD 5.17) and

7 males (mean age 30.6 years, SD 9.29). All volun-

teers signed informed consent.

Tests were conducted in a sleep lab that is spe-

cially equipped for medical and nursing students (Fig-

ure 1). Therefore, volunteers posed as patients in the

hospital beds available in this room. Most patients

with dementia are elderly, whose movements are lim-

ited not only by the illness, but also by age. For this

reason, participants wore 1.5 kg weights on their an-

kles and wrists to limit their movements, in order to

simulate a real scenario as far as possible. Figure 1

shows a volunteer on the hospital bed waiting to start

the tests.

Figure 1: Volunteer with weights in his extremities on a

hospital bed.

Wearable Monitoring for the Detection of Nocturnal Agitation in Dementia

Figure 2: Segment of accelerometry traces recorded during

protocol P1. Vertical lines indicate the division between

movements and quiet periods.

Participants performed the movements in proto-

cols P1 and P2. Accelerometry data was recorded

by the wearable accelerometer device. Both proto-

cols could be correctly executed, recorded and stored

in 8 out of the 11 cases. Figure 2 shows a segment of

recorded data.

3.2.3 Signal Processing

To classify the movements of the recorded dataset

into agitated periods and normal/quiet periods, we

used Support Vector Machines (SVMs) following the

guidelines provided in (Hsu et al., 2008). The ﬁrst

step was to deﬁne the features to be fed to the SVM

classiﬁer. These features have to be representative

of the quiet and agitation periods based on available

measurements (Guyon et al., 2006). Classical signal

processing techniques were applied to compute the

features as follows.

Acceleration on x, y and z axes was recorded at

a sampling frequency of 100 Hz. Recordings were

processed in segments from 5 to 15 s., depending on

the duration of the movement as speciﬁed in the pro-

tocol. In order to eliminate the effect of gravity, the

mean acceleration on each axis was subtracted from

its corresponding trace. Then, traces were low-pass

ﬁltered with a cut-off frequency of 10 Hz. The result-

ing traces, denoted as a

, a

, where used to com-

pute the classiﬁcation features as follows. For axis

i = x, y, z the following features were computed:

• Peak acceleration

= max(a

) (1)

• Energy

∑

n=1

(n)|

(2)

where n indicates the sample within the trace.

• Peak amplitude of the frequency spectrum

= max(FFT(a

)) (3)

where FFT stands for Fast Fourier Transform.

• Peak frequency, F

, computed as the frequency

corresponding to S

• Relative peak amplitude, R

, computed as the dif-

ference between S

and the mean amplitude of the

spectrum between 9 and 10 Hz.

Furthermore, the modulus of the acceleration vec-

tor was computed as m=

+ a

, and the same

features as in S

, F

and R

were computed from m,

thus obtaining features S

, F

and R

As a result, we obtained a set of 18 features that

contained information about the amplitude and tem-

poral variations of the wrist acceleration.

3.2.4 Feature Selection

In order to perform a selection of the most relevant

features for agitation detection, the 18 extracted fea-

tures were divided in 3 sets:

• Temporal Axis Features (TAF): A

and E

where

i = x, y, z.

• Frequency Axis Features (FAF): S

, F

and R

where i = x, y, z.

• Frequency Modulus Features (FMF): S

, F

and

The possible combinations (see Table 3) of these

sets were evaluated using cross-validation in order to

ﬁnd which ones contain relevant, irrelevant or redun-

dant information for agitation detection.

Table 3: Sets of features resulting of the extracted features

combination. The combination depends on the processing

signal: temporal or frequency axis, and, single axis signal

or modulus signal.

Feature Set Combinations

TAF

FAF

FMF

TAF & FAF

TAF & FMF

FAF & FMF

TAF & FAF &FMF

PEC 2016 - International Conference on Pervasive and Embedded Computing

3.2.5 Classiﬁcation of Movements

Following the guidelines in (Burges, 1998; Hsu et al.,

2008), movements were classiﬁed following the next

steps: division of the data in training and testing

subsets, transformation of movement data into SVM

package format, scaling data, selection of a kernel

function, search of the optimum parameters using

cross-validation, and training and testing the data.

The reference dataset was divided in two subsets,

one for training and one for testing. Knowing that

the imbalanced datasets where the number of nega-

tive instances far outnumbers the positive instances

declines the performance of SVM signiﬁcantly (Wu

and Chang, 2003), the division was made at 50%, that

is, the recordings of four volunteers (50%) were used

for training and the rest were used for testing. On

each subset, the number of agitated movements (pos-

itive instances) was similar to the number of normal

movements/quiet periods (negative instances) (Table

4).

Table 4: Number of agitated and normal/quiet movements

that compose the dataset of protocol P1 and protocol P2.

Protocol P1 Protocol P2

Training

Agitation 48 37

Normal 50 24

Testing

Agitation 48 40

Normal 50 32

For each instance, a instance-label pair is deﬁned

as (x

, y

) i = 1, ..., n where x

∈ R

is the instance with

feature values for a movement, and y

∈ {+1, −1} is a

label that determines the class of x

, negative or posi-

tive instance. Given the instance-label pairs for a sub-

set, the Support Vector Machine (SVM) (Cortes and

Vapnik, 1995) can be expressed as the following opti-

mization problem:

min

w,b,ξ

w+ C

∑

i=1

, (4)

subject to y

φ(x

) + b) ≥ 1− ξ

, (5)

≥ 0, (6)

where w and b deﬁne the separating hyper plane,

are ’slack’ variables which allow for misclassiﬁed

vectors and φ is a function that maps the training vec-

tors x

into a higher dimensional space.

Scaling avoids that greater numeric ranges take

precedence over smaller ones. At the same time,

numerical difﬁculties during the calculation are de-

creased (Sarle et al., 1997). For these reasons, in-

stances were scaled to [-1,1]. On the other hand, a

kernel function K(x

, x

) = φ(x

)·φ(x

) can be used to

avoid the explicit deﬁnition of a mapping function φ.

The Radial Basis Function (RBF) is the kernel used

in this work, deﬁned as:

K(x

, x

) = exp(−γ||x

− x

), γ > 0. (7)

A RBF kernel improves classiﬁcation results over

a linear kernel in most cases (Chang and Lin, 2011),

but it’s necessary to select suitable values for γ and

C parameters. In order to ﬁnd the optimum values, a

grid-search was carried out by creating a grid space of

(C, γ) pairs with log

C ∈ −5, −4, ..., 15 and log

γ ∈

−15, −14, ..., 3. For each pair (C, γ) in the space, we

performed a ﬁve-fold cross validation (CV) on each

training subset B

. The pair (C, γ) for each set B

chosen taking into account the maximum mean CV

accuracy.

Each selected pair (C

, γ

) together with the train-

ing subset B

, where i = 1, .., 7, are used to create a

SVM classiﬁer. Then, testing subsets B

were classi-

ﬁed using the corresponding classiﬁer i.

4 RESULTS

The optimum RBF parameters found during the grid

search are presented in Table 5. Taking into account

the results of Table 5, the feature set B

is the most

relevant when all movements are evaluated together.

The classiﬁcation results were evaluated by com-

paring the accuracy (Acc), which is the percentage

of instances correctly classiﬁed, the sensitivity (Se),

which is the percentage of positive instances correctly

classiﬁed, and the speciﬁcity (Sp), which is the per-

centage of negative instances correctly classiﬁed. Ta-

ble 6 presents the classiﬁcation results in testing sets.

The maximum accuracy for movements in proto-

col P1 is 69.90% for feature set B

. In contrast, the

feature sets B

, B

and B

returned the best accuracy,

94.44% for movements in protocol P2. And the union

of both protocols returns 78.86% of accuracy in fea-

ture sets B

and B

5 DISCUSSION AND

CONCLUSIONS

In this study, we analyzed nocturnal agitation using

a wearable device and machine learning techniques.

Results showed that nocturnal agitation movements

Wearable Monitoring for the Detection of Nocturnal Agitation in Dementia

Table 5: Optimization parameters for RBF-SVM obtained with the whole training set (protocols P1 and P2). The accuracy

(%) resulting from cross-validation is reported as mean ± standard deviation.

Feature Set Accuracy (%) log

C log

78.67 ± 7.98 4 -2

79.23 ± 7.88 12 3

74.84 ± 4.43 2 -5

74.84 ± 6.26 12 3

78.45 ± 8.11 4 1

77.42 ± 10.07 14 4

76.09 ± 5.73 15 5

Table 6: RBF-SVM classiﬁcation results for the testing set. For each B

, C and γ were set as in Table 5 for the SVM. Se:

Sensitivity, Sp: Speciﬁcity, Acc: Accuracy.

Feature Set

P1 P2 P1&P2

Se (%) Sp (%) Acc (%) Se (%) Sp (%) Acc (%) Se (%) Sp (%) Acc (%)

45.83 67.27 57.28 100.0 87.50 94.44 70.45 77.01 73.71

52.08 85.45 69.90 90.00 78.13 84.72 62.50 95.00 78.86

68.58 34.55 48.54 100.0 28.13 68.06 82.95 40.23 61.71

35.42 90.91 65.05 97.50 59.38 80.56 54.55 93.10 73.71

27.08 81.82 56.31 82.50 90.63 86.11 59.09 87.36 73.14

50.00 81.82 66.99 95.00 93.75 94.44 69.32 88.51 78.86

58.33 77.55 66.99 97.56 90.63 94.44 56.82 90.80 73.71

can be successfully detected with an accuracy of

78.86%. Classiﬁcation results were better for proto-

col P2 than for protocol P1 or for the whole dataset.

Protocol P1 involves some movements that are difﬁ-

cult to detect using the wearable device on a wrist,

such as move the legs, or movements that can be eas-

ily confused with agitation although they are normal,

such as move the clock. On the other hand, proto-

col P2 contains more energetic movements, which are

easier to classify. The applicability of these results to

real settings will depend on the realism of the sim-

ulated movements. The protocols described here in-

cluded clearly deﬁned movements, with obvious dif-

ferences between agitation and normal/quiet periods.

An evolution of the system should be trained with

more realistic protocols, which could include move-

ments such as roll over the bed or cover with bed

sheets. These movements are more complex than re-

main laying down and they introduce a greater con-

fusing factor, but at the same time they may better

reﬂect actual nocturnal behavior.

Although preliminary results are promising, sev-

eral limitations of the work need to be acknowledged.

First, the use of different kernel functions could be ex-

plored, in order to ﬁnd a classiﬁer with better results,

or to conﬁrm that the radial basis function is the best

for this particular classiﬁcation problem. Also, by

comparing the results of table 6, the relevance of fre-

quency features over temporal features is clear, since

the best results for the whole dataset correspond to

B2 and B6 feature sets. These feature sets only con-

tain frequency features, and the main difference be-

tween both is the modulus features. B6 contains all

the features of B2 more the frequency features of the

modulus signal. While the B2 feature set increases

the accuracy of protocol P1 up to 69.90%, the accu-

racy of protocol P2 is increased by the B6 feature set

up to 94.44%. Thus, modulus features include rel-

evant information to classify movements in protocol

P2, but some of the modulus features may include

redundant information for the classiﬁcation of move-

ments in protocol P1. The inclusion of more sophis-

PEC 2016 - International Conference on Pervasive and Embedded Computing

ticated feature selection methods may result in a net

improvement of the solution.

The presented proof of concept was designed in

coordination with medical staff in order to study the

feasibility to improving dementia diagnostics by us-

ing wearable devices. Since the preliminary results

demonstrated that it is possible to detect agitation us-

ing a wearable accelerometer, the next step towards

clinical translation of our research will be to perform

a pilot study with hospitalized patients, after includ-

ing the above mentioned classiﬁcation improvements.

ACKNOWLEDGEMENTS

The authors would like to thank the cooperation of Dr.

Mercedes Gim´enez, Responsible for the Geriatrics

Department of Hospital San Juan de Dios (Zaragoza,

Spain). This work is supported by project AEI-

010500-2015-200(MINETUR, Spain) and by Grupos

BSICoS (T96) and SVIT (T92) from DGA (Arag´on)

and European Social Fund (EU). Partially supported

by the Aragonian Government and the European So-

cial Fund ”Building Europe from Aragon”. This work

has been supported by research fellowship from the

Universidad San Jorge.

REFERENCES

Ancoli-Israel, S., Clopton, P., Klauber, M. R., Fell, R., and

Mason, W. (1997). Use of wrist activity for moni-

toring sleep/wake in demented nursing-home patients.

Sleep, 20(1):24–27.

Banaee, H., Ahmed, M. U., and Loutﬁ, A. (2013). Data

mining for wearable sensors in health monitoring sys-

tems: a review of recent trends and challenges. Sen-

sors, 13(12):17472–17500.

Bendersky, D., Ajler, P., and Yampolsky, C. (2014). The

use of neuromodulation for the treatment of tremor.

Surgical Neurology International, 5(6):232.

Biswas, J., Jayachandran, M., Thang, P. V., Fook, V. F. S.,

Choo, T. S., Qiang, Q., Takahashi, S., Jianzhong,

E. H., Feng, C. J., and Kiat, P. (2006). Agitation moni-

toring of persons with dementia based on acoustic sen-

sors, pressure sensors and ultrasound sensors: a feasi-

bility study. In International Conference on Ageing,

Disability, and Independence, pages 3–15.

Burges, C. J. C. (1998). A tutorial on support vector

machines for pattern recognition. Data Mining and

Knowledge Discovery, 2(2):121–167.

Chang, C.-C. and Lin, C.-J. (2011). Libsvm: A library for

support vector machines. ACM Transactions on Intel-

ligent Systems and Technology (TIST), 2(3):27.

Cohen-mansﬁeld, J., Marx, M. S., and Rosenthal, A. S.

(1989). A description of agitation in a nursing home.

Journal of Gerontology, 44(3):M77–M84.

Cooke, J. R. and Ancoli-Israel, S. (2006). Sleep and its

disorders in older adults. Psychiatric Clinics of North

America, 29(4):1077–1093.

Cortes, C. and Vapnik, V. (1995). Support-vector networks.

Machine Learning, 20(3):273–297.

Deschenes, C. L. and McCurry, S. M. (2009). Current treat-

ments for sleep disturbances in individuals with de-

mentia. Current Psychiatry Reports, 11(1):20–26.

Fook, V. F. S., Thang, P. V., Htwe, T. M., Qiang, Q., Wai,

A. A. P., Jayachandran, M., Biswas, J., and Yap, P.

(2007). Automated recognition of complex agitation

behavior of dementia patients using video camera. In

2007 9th International Conference on e-Health Net-

working, Application and Services, pages 68–73.

Guyon, I., Gunn, S., Nikravesh, M., and Zadeh, L. A.

(2006). Feature Extraction, Foundations and Appli-

cations. Springer, Berlin, 1st edition.

Hsu, C.-W., Chang, C.-C., and Lin, C.-J. (2008). A practi-

cal guide to support vector classiﬁcation. BJU inter-

national, 101(1):1396–400.

Intelligence, B. I. (2015). The wearables report: Growth

trends, consumer attitudes, and why smartwatches

will dominate. Website. http://goo.gl/ZF3ZiN.

Kolla, B. P., Mansukhani, S., and Mansukhani, M. P. (2016).

Consumer sleep tracking devices: a review of mecha-

nisms, validity and utility. Expert Review of Medical

Devices, 12:497–506.

Nagels, G., Engelborghs, S., Vloeberghs, E., Van Dam, D.,

Pickut, B. A., and De Deyn, P. P. (2006). Actigraphic

measurement of agitated behaviour in dementia. Inter-

national Journal of Geriatric Psychiatry, 21(4):388–

393.

Prince, M., Bryce, R., Albanese, E., Wimo, A., Ribeiro,

W., and Ferri, C. P. (2013). The global prevalence

of dementia: A systematic review and metaanalysis.

Alzheimer’s & Dementia, 9(1):63–75.e2.

Rose, K. M., Fagin, C. M., and Lorenz, R. (2010). Sleep

disturbances in dementia: What they are and what to

do. Journal of gerontological nursing, 36(5):9–14.

Sakr, G., Elhajj, I., and Huijer, H.-S. (2010). Support vec-

tor machines to deﬁne and detect agitation transition.

IEEE Transactions on Affective Computing, 1(2):98–

108.

Sarle, W. S. et al. (1997). Neural network faq. Periodic post-

ing to the Usenet newsgroup comp. ai. neural-nets.

Sink, K. M., Holden, K. F., and Yaffe, K. (2005). Pharma-

cological treatment of neuropsychiatric symptoms of

dementia: a review of the evidence. JAMA : The Jour-

nal of the American Medical Association, 293(5):596–

608.

Van Someren, E. (1997). Actigraphic monitoring of move-

ment and rest-activity rhythms in aging, alzheimer’s

disease, and parkinson’s disease. IEEE Transactions

on Rehabilitation Engineering, 5(4):394–398.

Wimo, A., Jnsson, L., Bond, J., Prince, M., and Winblad, B.

(2013). The worldwide economic impact of dementia

2010. Alzheimer’s & Dementia, 9(1):1–11.e3.

Wu, G. and Chang, E. Y. (2003). Class-boundary alignment

for imbalanced dataset learning. In ICML 2003 work-

shop on learning from imbalanced data sets II, pages

49–56.

Wearable Monitoring for the Detection of Nocturnal Agitation in Dementia