Technological Approach for Behavior Change Detection toward Better

Adaptation of Services for Elderly People

Firas Kaddachi

, Hamdi Aloulou

, Bessam Abdulrazak

1,2

, Joaquim Bellmunt

, Romain Endelin

Mounir Mokhtari

3,4

and Philippe Fraisse

Montpellier Laboratory of Informatics, Robotics and Microelectronics (LIRMM), Montpellier, France

University of Sherbrooke, Sherbrooke, Canada

Image and Pervasive Access Lab (IPAL), Singapore, Singapore

Institut Mines-Telecom (IMT), Paris, France

Keywords:

Behavior, Behavior Change Detection, Elderly People, Sensors.

Abstract:

Aging process is associated with behavior change and continuous decline in physical and cognitive abilities.

Therefore, early detection of behavior change is major enabler for providing adapted services to elderly people.

Today, different psychogeriatric methods target behavior change detection. However, these methods require

presence of caregivers and manual analysis. In this paper, we present our technological approach for early be-

havior change detection. It consists in monitoring and analyzing individual activities using pervasive sensors,

as well as detecting possible changes in early stages of their evolution. We also present a ﬁrst validation of the

approach with real data from nursing home deployment.

1 INTRODUCTION

Early detection of behavior change is keystone for ser-

vice providers to better adapt their services for elderly

people. Existing psychogeriatric methods for behav-

ior change detection are inconvenient, as they are

time-consuming and require manual analysis work

from caregivers.

According to existing deﬁnitions (Cao, 2010), be-

havior change is deﬁned as any continuous modiﬁca-

tion or transformation in way and manner of behav-

ior execution. Behavior change characterizes possi-

ble instabilities, variations, impairments, declines, in-

creases or improvements in behavior performance.

Behavior change has signiﬁcant impact on quality

of life. For example, emergence of orientation prob-

lems (Cockrell and Folstein, 2002), eating difﬁculties

(Vellas et al., 1999) and mood impairments (Parmelee

and Katz, 1990) leads to serious decline in quality of

life. On the other hand, any improvement in manag-

ing personal ﬁnances (Barberger-Gateau et al., 1992),

managing household (Lafont et al., 1999) and mobil-

ity (Mathias et al., 1986) has positive inﬂuence on

quality of life.

Early detection of behavior change is major en-

abler for more efﬁcient intervention, by taking neces-

sary actions in early stages of behavior change. Au-

tonomy of elderly people is consequently improved,

by reducing symptoms and evolution of sensor, motor

and cognitive diseases.

In this paper, we propose a technological approach

for behavior change detection. Changes are detected

at temporal scale; i.e., compared to past habits of one

particular person.

Our employed technologies (e.g., movement and

contact sensors) do not interfere with monitored be-

havior. These technologies are ubiquitous. They dis-

appear in the environment, without generating un-

wanted behavior change, and without affecting indi-

vidual privacy.

Our approach conducts long-term analysis of be-

havior for detection of continuous changes. We do not

study snapshots of behavior, but we analyze overall

behavior over long periods. This enables to differen-

tiate between transient and continuous deviations.

Following, section 2 discusses state of the art of

behavior deﬁnitions and change detection methods.

Sections 3 and 4 present our methodology for be-

havior change detection and our implementation ap-

proach. Section 5 introduces a ﬁrst validation of the

proposed approach through real results from nursing

home deployment. Section 6 concludes this paper.

Kaddachi F., Aloulou H., Abdulrazak B., Bellmunt J., Endelin R., Mokhtari M. and Fraisse P.

Technological Approach for Behavior Change Detection toward Better Adaptation of Services for Elderly People.

DOI: 10.5220/0006145100960105

In Proceedings of the 10th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2017), pages 96-105

ISBN: 978-989-758-213-4

2 RELATED WORK

Researchers study behavior change from two different

perspectives: initiation and maintenance of behavior

change, and detection of behavior change.

2.1 Behavior Deﬁnitions

Cao deﬁnes behavior as actions or reactions made by

individuals (Cao, 2010). Behavior is a response to

internal or external stimuli or inputs.

From sociological point of view, Wilson consid-

ers behavior as interactions between individuals (Wil-

son, 2000). It can be inﬂuenced by family structure,

work or school environment relationships, health con-

ditions and psychiatric issues.

Economists recognize behavior as processes con-

sumers go through or reactions they have toward pur-

chasing or consuming products or services (Perner,

2008; Szwacka-Mokrzycka, 2015). It is inﬂuenced by

internal factors, such as attitudes, needs, motives and

preferences. External factors have also signiﬁcant in-

ﬂuence, such as marketing activities, social, econom-

ical and cultural aspects.

In the medical ﬁeld, behavior refers to persons’

beliefs and actions regarding their health (Miller et al.,

2007; Lavikainen et al., 2009). While positive behav-

iors promote healthy life (e.g, maintain moderate al-

cohol intake, not smoke and avoid snacks), negative

behaviors present health risks.

These deﬁnitions consider behavior as a response

to internal or external factors, such as intentions,

desires, social interactions and marketing activities.

However, individuals respond differently to these fac-

tors; e.g., being hungry stimulates individuals to pre-

pare meals with different duration, frequency and dif-

ﬁculty. Therefore, we deﬁne behavior as the way and

manner individuals perform actions, inactions and be-

liefs.

2.2 Behavior Change Initiation and

Maintenance Models

Numerous models have been proposed to predict

the amount of effort individuals require for behavior

change initiation and maintenance (Ormrod, 2013).

In fact, initiating and maintaining behavior changes

are related to individuals’ perception of their own

ability to perform demanding or challenging tasks.

This perception is inﬂuenced by individuals’ prior

success in those tasks or related tasks, their psycho-

logical state and outside sources of persuasion.

In the medical ﬁeld, behavior change refers

to abandoning health-compromising behaviors and

maintaining health-improving behaviors. Rosenstock

suggests that individuals’ belief about health prob-

lems and perceived beneﬁts of actions plays important

role in adopting health-promoting behaviors (Rosen-

stock, 1974).

Schwarzer considers behavior change as two con-

tinuous processes: goal setting and goal pursuit

(Schwarzer, 2008). While goal setting is related to

factors that motivate behavior change, goal pursuit

consists in planning and performing intended change.

Prochaska et al. propose a ﬁve-step model of

behavior change (Prochaska and DiClemente, 2005).

In the ﬁrst step, individuals have not thought about

changing their behaviors. Then, individuals begin

thinking about changing particular behaviors. After-

wards, they prepare their plans for behavior change.

In the fourth step, individuals adopt and perform new

behaviors. Finally, they consistently conserve their

new behaviors.

While these models target behavior change initi-

ation and maintenance, detection of behavior change

enables better fulﬁllment of both objectives. In deed,

methods for behavior change detection allow to make

better decisions of when to initiate new behavior

changes, and which services to select for behavior

change initiation and maintenance.

2.3 Behavior Change Detection

Methods

In the literature, we distinguish psychogeriatric and

technological methods for behavior change detection.

While psychogeriatric methods use formal tests and

questionnaires, technological solutions are developed

to automate detection of anomalies.

2.3.1 Psychogeriatric Methods

Psychologists and geriatricians propose several inter-

nationally validated methods for behavior change de-

tection (Table 1). Using formal scales and question-

naires, trained clinicians and caregivers request that

seniors reply to speciﬁc questions and perform spe-

ciﬁc tasks, such as ”How many falls did you have in

the last six months?” (Tardieu et al., 2016) and ”Could

you please get up and walk three meters away!”

(Mathias et al., 1986).

Following, we present the psychogeriatric tests of

Table 1:

• Short Emergency Geriatric Assessment

(SEGA) allows to evaluate frailty of elderly

people (Tardieu et al., 2016). It considers mul-

tiple behavior changes, such as falls, nutrition

Technological Approach for Behavior Change Detection toward Better Adaptation of Services for Elderly People

Table 1: Examples of Psychogeriatric Tests for Behavior Change Detection.

SEGA MMSE 4Tests GDS IADL AGGIR GetUp MNA BEHA NPI

AndGo VEAD

ADL X X X X X

Mobility X X X

Cognition X X X X X X X X

Social Life X X X X X X

Nutritional Status X X X

Biological Status X X X X

Mood and Emotions X X X X

problems, mobility impairments and memory

troubles.

• Mini Mental State Examination (MMSE) tar-

gets detection of changes in cognitive abilities,

such as orientation problems, attention difﬁcul-

ties and language troubles (Cockrell and Folstein,

2002).

• Benton, Five Word, Clock and Verbal Fluency

Tests (4Tests) target changes in cognitive func-

tions, such as learning problems, memory troubles

and construction difﬁculties (Neuropsy, 2016).

• Geriatric Depression Scale (GDS) investigates

changes in mood and emotions (Parmelee and

Katz, 1990); e.g., feeling sad and that one’s life

is empty is associated with possible depression.

• Instrumental Activities of Daily Living (IADL)

identiﬁes changes in activities of daily living that

are associated with autonomy loss, such as us-

ing telephone, using means of transport, tak-

ing medicines and managing personal ﬁnances

(Barberger-Gateau et al., 1992).

• Autonomie Gerontologique et Groupes Iso-

Ressources (AGGIR) investigates changes in

autonomy of seniors, such as movement trou-

bles, household difﬁculties and orientation im-

pairments (Lafont et al., 1999).

• Get-Up and Go targets motor behavior changes

(Mathias et al., 1986). This test asks elderly peo-

ple to get up, walk and turn around, in order to

analyze task execution and identify possible mo-

bility impairments.

• Mini Nutritional Assessment (MNA) investi-

gates changes in nutritional status, such as eating

difﬁculties, weight loss and protein intake insufﬁ-

ciency (Vellas et al., 1999).

• Behavioral Pathology in Alzheimers Disease

(BEHAVE-AD) and Neuropsychiatric Inven-

tory (NPI) allow to detect possible behavioral im-

pairments for elderly people, such as presence of

hallucinations, aggressiveness and anxiety (Reis-

berg et al., 1997) (Cummings et al., 1994).

Using these tests, clinicians observe task execu-

tion and analyze senior behavior, in order to iden-

tify cognitive impairments, autonomy problems, rapid

mood changes, nutritional and behavioral anomalies.

Certain inconveniences limit the efﬁciency of psy-

chogeriatric tests. In fact, it is inconvenient for se-

niors to recall past events with full details at assess-

ment time. It is also often not convenient for elderly

people to move to assessment place.

Besides, requesting that individuals reply to given

questions and perform determined tasks has potential

negative impact on their future behaviors after assess-

ment. For example, anxiety of seniors can increase

in case they feel their inability to correctly reply to

orientation questions or perform mobility tasks. Fur-

thermore, subjective evaluation of assessment results

cause possible assessment inaccuracies.

2.3.2 Technological Methods

Different technological methods target behavior

change detection. They employ technologies de-

ployed in the environment (e.g., movement sensors,

bed sensors, cameras and microphones) or worn by

seniors (e.g., smart phone, smart watch and neurosen-

sors). These methods conduct advanced analysis of

acquired data, in order to detect changes in monitored

behaviors.

Allin et al. propose technological method for so-

cial behavior change detection (Allin et al., 2003).

This method detects emergence of physically and ver-

bally aggressive interactions. It employs cameras and

microphones for continuous collection of video and

audio recordings. Using hidden markov models, com-

plex analysis of these recordings allows to build typ-

ical movement patterns for anomaly detection. How-

ever, employed cameras and microphones affect pri-

vacy of individuals.

Avvenuti et al. target detection of wandering and

falls from bed during sleep (Avvenuti et al., 2010).

HEALTHINF 2017 - 10th International Conference on Health Informatics

They study correlation between brain activity and

body movement, in order to deﬁne rules and derive

threshold values for anomaly detection. This method

employs modern neurosensors placed on person’s

scalp to record brain activities. Yet, neurosensors

limit individual movements indoors and outdoors.

Another technological method studies mental

state changes that lead to possible depression (Mag-

ill and Blum, 2012). It provides objective feedback

to patients using body and environmental sensors, in

addition to subjective questionnaire-based records for

their health. Content and timing of questionnaires are

personalized for individuals and altered over time as

individual’s mental health changes. However, it has

been reported that technical trial of developed system

reveals acceptability issues from participants regard-

ing questionnaires.

Kaye et al. investigate changes in computer use

(Kaye et al., 2014). Based on statistical analysis of

mouse events, they compare frequency and duration

of computer use between two aging populations with

or without mild cognitive impairments (MCI). They

conclude that MCI patients use computers less than

regular persons.

Hayes et al. target detection of medication ad-

herence changes (Hayes et al., 2009). Using elec-

tronic pillbox, subjects take medicines twice per day

at speciﬁc times. This method monitors number of

days when subjects take both medicines, and veriﬁes

whether volunteers adhere to given times. It compares

two senior groups with low or high performance in

given cognitive tests, and concludes that lower per-

forming group has risk of non-adherence.

Another technological study investigates motor

behavior changes (Hayes et al., 2008). Careful in-

series placement of wireless infrared sensors at home

identiﬁes how quickly and frequently seniors pass

through sensor lines per day. Comparing two aging

populations with or without MCI, MCI patients show

a coefﬁcient of variation in median walking speed as

twice as high compared to regular subjects.

Petersen et al. propose further solution to detect

changes in telephone use (Petersen et al., 2014). Em-

ployed land-line phone monitors record phone events,

such as dialed numbers and ring rate. These record-

ings allow to have a picture on size and contact fre-

quency of friend, family and acquaintance network.

Results show that seniors with high cognitive abilities

receive signiﬁcantly more phone calls.

These last four studies detect behavior changes be-

tween different individuals. However, they do not tar-

get detection of behavior changes that affect one par-

ticular individual.

3 BEHAVIOR CHANGE

DETECTION METHODOLOGY

We target behavior change detection at temporal

scale. Over long periods, we analyze behavior of el-

derly people, in order to identify changes compared

to past habits. Our technologies do not interfere with

monitored behavior and do not affect individual pri-

vacy.

Our behavior analysis identiﬁes indicators of be-

havior change, such as activities of daily living, mo-

bility and social life (Figure 1). These indicators

are associated with changes in physical and cognitive

abilities of elderly people.

Activities of Daily Living

(Prepare Meals)

Nutritional Status (Eating) Mood and Emotions

Health and Biological Status

(Take Medicines)

Social Life (Shopping)

Mobility (Get Up)

Cognition (Learn)

Figure 1: Examples of Behavior Change Indicators.

We also analyze these indicators considering dif-

ferent dimensions, such as quantity, duration, time

and location. These dimensions are metrics that quan-

tify collected data and allow to apply algorithms on

these data for change detection.

Furthermore, we correlate identiﬁed changes with

global context, such as weather conditions, family sta-

tus and house architecture. Considering these factors

provides better understanding of detected changes;

e.g., senior stays at home for seven days due to heavy

snow and not due to eventual social isolation.

3.1 Behavior Change Indicators

We have considered different validated psychogeri-

atric scales (e.g., SEGA, AGGIR, MNA and NPI)

to identify indicators of behavior change. That can

be captured via ambient technologies (Table 2). An-

alyzing these indicators allows to detect signiﬁcant

changes in physical and cognitive abilities. Figure 1

shows following examples of indicators:

Technological Approach for Behavior Change Detection toward Better Adaptation of Services for Elderly People

• Activities of Daily Living are essential complex

tasks of daily living that demand important physi-

cal and cognitive capacities, such as performing

household, preparing meals, dressing, hygiene,

and urinary and fecal elimination.

• Mobility refers to motor behaviors, such as mov-

ing indoors and outdoors, getting up, turning

around and walking.

• Cognition includes essential cognitive tasks such

as learning, language and managing ﬁnancial sit-

uation. These complex tasks are associated with

temporal orientation, spatial orientation, atten-

tion, calculation and construction.

• Social Life refers to social behaviors, such as

communicating with others, using means of trans-

port, shopping and participating in collective free

time activities.

• Nutritional Status is related to serving oneself

and eating.

• Health and Biological Status targets health be-

haviors that indicate vision, audition and vital

sign impairments, such as irregularities in taking

medicines and increased hospitalization number.

• Mood and Emotions correlate with physical and

cognitive functions and are signiﬁcant depression

and stress indicators.

Table 2: Examples of Ambient Technologies for Behavior

Change Indicator Monitoring.

Environment Technologies

Indoor Movement, contact, proximity,

vibration and pressure sensors

Outdoor Smart phone and smart watch

with beacons

3.2 Metrics

We analyze selected behavior change indicators re-

garding different dimensions. These dimensions are

metrics that quantify way and manner of performing

these indicators, and allow to apply algorithms on col-

lected data for change detection. Following, we dis-

cuss four signiﬁcant metrics:

• Quantity refers to number and amount of behav-

ior execution; e.g., number of friend visits de-

creases due to social isolation, number of move-

ments decreases due to walk impairments, number

of sport center visits increases thanks to raised in-

terest in physical exercise and number of hospital-

izations decreases thanks to health status improve-

ment.

• Duration is related to length of behavior execu-

tion; e.g., duration of preparing meals increases

due to cognitive impairments, duration of stair

climbing increases due to walk impairments, time

spent out of home increases thanks to raised in-

terest in social interactions and time spent in free

time activities considerably increases thanks to

raised interest in active aging.

• Time refers to start and end times of behavior

execution; e.g., sleep hours are irregular due to

sleep troubles, eating meal hours are inappropri-

ate due to nutritional problems, going out hours

are changing thanks to raised interest in social

activities and taking medicine hours are adhered

thanks to cognitive status improvement.

• Place describes where behavior is executed; e.g.,

detected falls outdoors become more frequent due

to fear of going outside, visiting senior activ-

ity center becomes less usual due to social isola-

tion and visiting city park becomes more frequent

thanks to raised interest in physical exercise.

3.3 Global Context

Analyzing behavior in correlation with global con-

text enables better understanding of behavior change.

Following, we discuss inﬂuence of general personal

information, general context information and speciﬁc

temporary information on behavior change.

General personal information are general descrip-

tors of persons; e.g., age over 85, health care his-

tory including more than three physical and mental

diseases, and inconvenient family status increase the

probability of behavior changes (Tardieu et al., 2016).

General context information describe the envi-

ronment of behavior execution; e.g., changing one’s

house affects activities of daily living, moving televi-

sion in room not easily accessible by elderly people

reduces watching television frequency, opening smart

city subways adapted for elderly people has positive

inﬂuence on outdoor activities and building senior ac-

tivity centers raises interest in social interactions.

Speciﬁc temporary information refer to short-term

events, such as several consecutive days of heavy

snow that obligate senior to stay at home, recent hos-

pitalization of husband that raises wife’s anxiety, and

recent friend visits that improve emotional state.

HEALTHINF 2017 - 10th International Conference on Health Informatics

100

4 IMPLEMENTATION

APPROACH

We perform a ﬁrst implementation of our behavior

change detection methodology in our ambient as-

sisted living platform UbiSMART (Aloulou, 2013;

Aloulou et al., 2013). This platform uses environ-

mental sensor data for activity recognition, detection

of abnormal activity change and provision of person-

alized services for elderly people (Figure 2).

HOME

UbiSMART

Deployed Sensors at Home

Database

Data

Analysis

Change

Detection

Sensor Data

Services for

Elderly People

Figure 2: Overview on our Ambient Assisted Living Plat-

form.

Our implementation approach considers following

stages:

• Deployment consists in installing our hardware

infrastructure. This includes environmental sen-

sors (e.g., movement and contact sensors), gate-

ways, receivers and internet access points.

• Data Acquisition is essential to build our data-

base. Via internet, data are transmitted to our ded-

icated server for permanent storage.

• Data Pre-processing allows to discard inaccurate

and erroneous data for better analysis quality.

• Data Analysis quantiﬁes data by considering dif-

ferent metrics, such as daily number and duration

of shopping activity. Afterwards, we apply algo-

rithms on these data to detect possible changes at

temporal scale; e.g., these algorithms identify de-

crease in shopping activity periods.

4.1 Algorithms

We select statistical algorithms for our data analysis,

as they differentiate between transient and continuous

deviations; e.g., these statistical algorithms ignore oc-

casional decreases in going out frequency, and con-

sider only continuous decreases as signiﬁcant changes

in going out frequency.

We can distinguish ofﬂine and online algorithms

for change detection in the literature (Basseville et al.,

1993; Liu et al., 2013). Ofﬂine algorithms require

fully available data as input, such as full history of

free time activity number and duration. However, on-

line algorithms iteratively operate on data one by one,

such as number and duration of free time activity day

by day.

Existing online algorithms use probabilistic mod-

els (Takeuchi and Yamanishi, 2006), singular spec-

trum analysis (Moskvina and Zhigljavsky, 2003) and

cumulative sum control charts (Mesnil and Petitgas,

2009). Existing ofﬂine algorithms apply relative

density-ratio estimation (Liu et al., 2013) and cumula-

tive sum control charts with binary segmentation (An-

dersson, 2014; Cho, 2015) or bootstrapping (Taylor,

2000).

In order to detect changes as early as possible, we

select online algorithms. Following, we discuss two

algorithms investigated in our research; i.e., cusum-

based (Page, 1954) and window-based (Bland and

Altman, 1995) algorithms. These algorithms apply

different ﬁlters on detected deviations and identify

changes with different granularity.

4.1.1 CUSUM-based Algorithm

Page proposes Cumulative Sum Control Chart

(CUSUM) algorithm for change detection in time

series (Page, 1954). This algorithm considers two

phases: reference phase and analysis phase.

In reference phase, initial data allow to compute

parameters that will condition change detection:

• M refers to mean of reference data.

• SD is standard deviation of reference data.

• SHIFT is related to shift of interest, that de-

termines smallest deviation we target to detect.

Ledolter et al. set SHIFT to 1x SD (Ledolter and

Kardon, 2013).

• K refers to allowance parameter, that is related to

shift of interest. Mesnil et al. set K to 0,5x SHIFT

(Mesnil and Petitgas, 2009).

• H is decision parameter that determines whether

change occurs or not. In the literature, researchers

deﬁne H with published tables, speciﬁc software

or set it to 5x SD (Mesnil and Petitgas, 2009; Kib-

ria, 2016).

In the analysis phase, mean and standard devia-

tion parameters allow to standardize data by applying

formula 1:

data[i] = (data[i] − M)/SD (1)

For each datum, cumulative sums recursively ac-

cumulate positive and negative deviations, using for-

mula 2 and 3:

Technological Approach for Behavior Change Detection toward Better Adaptation of Services for Elderly People

101

HIGH

[i] = max(0, S

HIGH

[i − 1] + data[i] − K) (2)

LOW

[i] = min(0,S

LOW

[i − 1] + data[i] + K) (3)

In case S

HIGH

is higher then +H or S

LOW

is lower

than -H, positive or negative change occurs.

4.1.2 Window-based Algorithm

Based on Bland-Altman analysis, window-based al-

gorithm applies moving window on input data to dis-

tinguish between transient deviations and continuous

change (Bland and Altman, 1995). Only in case

selected number of deviations are consecutively de-

tected without interruption, change occurs.

Positive or negative deviations are data values that

are higher or lower than M ± SD, where M and SD

correspond respectively to mean and standard devi-

ation of all previously observed data including cur-

rently observed datum.

Window length (N) depends on analyzed behav-

ior; e.g., seven consecutive days of staying at home

correspond to change in going out frequency or three

consecutive months of loosing weight indicate change

in nutritional status. Positive or negative changes are

detected in case N consecutive positive or negative de-

viations occur.

5 VALIDATION

Following, we present a ﬁrst validation of our ap-

proach through real data from nursing home deploy-

ment. Considering mobility as indicator of behavior

change, collected data allow to analyze movements of

patients inside their individual rooms.

5.1 Data Collection

Table 3: Patient Gender, Age and Monitoring Period.

Patient Gender Age Period(months)

A M 90 6

B M 89 5

C M 81 2

D F 84 11

E F 95 2

F F 85 13

G F 87 13

H F 92 9

I F 92 4

Over one year, we deploy movement sensors in

bedrooms and bathrooms of 9 patients in a french

nursing home in Occagnes (Table 3). Average age of

patients is 88 years.

5.2 Data Analysis

We use movement sensor data to analyze physical ac-

tivity periods (PAP) of persons. We simply deﬁne a

PAP as period of consecutive movements, that are de-

tected with time difference less than 3 minutes.

We do not consider days of inactivity, that cor-

respond to hospitalizations or holidays outside indi-

vidual rooms. In our analysis, we quantify collected

movement sensor data using following metrics:

• Number refers to quantity of detected movements

and PAPs.

• Duration is total length of detected PAPs.

• Intensity measures mean number of detected

movements per PAP. This corresponds to number

of detected movements divided by number of de-

tected PAPs.

5.3 Results

Figure 3 shows our analysis results for patient F over

13 months. For each month, we compute average of

daily number of movements, PAPs, their duration and

intensity. We also study inﬂuence of mean ambient

temperature on physical activities.

We observe decrease in movement and PAP num-

ber in case ambient temperature increases. However,

PAP duration grows and PAP intensity is quite stable.

This is also observed for other patients. Higher tem-

perature stimulate them to perform less activities with

longer total duration inside individual rooms.

For early change detection, we apply cusum-based

and window-based algorithms on collected data after

each day. In order to validate their results, we also

apply an ofﬂine algorithm on full months of data.

In the literature, ofﬂine algorithms provide more

robust results than online algorithms, as they retro-

spectively analyze longer periods of data (Basseville

et al., 1993; Liu et al., 2013).

We select ofﬂine algorithm of Change Point Ana-

lyzer (CPA) tool (Taylor, 2000). This algorithm im-

plements an iterative combination of cumulative sum

control charts and bootstrapping to detect changes.

Table 4 shows dates and values of identiﬁed changes

in movement number data of patient F.

Results of cusum-based and window-based algo-

rithms are compared to those obtained with CPA tool

in Figure 4, considering true positive rate (TPR), pre-

cision (P), true negative rate (TNR) and accuracy (A).

Cusum-based and window-based (N=5 and N=4)

algorithms show true positive rate of 28%, 40% and

45% respectively, as they do not detect all changes.

Their precision is 64%, 56% and 33% respectively,

HEALTHINF 2017 - 10th International Conference on Health Informatics

102

Figure 3: Monthly Average of Movement Number, PAP Number, Duration and Intensity for Patient F.

Table 4: Change Dates and Values of Movement Number

for Patient F.

Change Date From To

2014, October 3 239 354

2015, March 15 354 253

2015, April 3 253 180

2015, September 13 180 282

2015, September 18 282 176

which indicates that not all identiﬁed changes are rel-

evant.

However, true negative rate is 99%, 99% and 98%

respectively, as they correctly identify almost all nor-

mal data. Their accuracy is 97%, 98% and 97% re-

spectively, which corresponds to good overall results.

6 CONCLUSION

We propose a technological approach for behavior

change detection at temporal scale. We analyze over-

all behavior to identify changes compared to past

habits over long periods. Our technologies disappear

in the environment, in order to avoid generation of

unwanted changes and protect individual privacy.

We also present a ﬁrst validation of our methodol-

ogy through real data from nursing home deployment.

Over months, employed movement sensors allow to

TPR

TNR

A0%

50%

100%

Cusum

Window (N=5)

Window (N=4)

Figure 4: Comparison of Cusum-based and Window-based

Algorithms to CPA Tool.

monitor physical activities of patients. Collected data

are quantiﬁed considering different metrics, such as

number and duration. Our selected statistical change

detection algorithms provide good overall results.

We are working on improving our behavior

change detection in the context of the European

project City4Age (City4Age, 2016). The City4Age

project target using data generated by technologies

deployed in urban areas, in order to provide new

adaptable services for elderly people. These services

target capturing frailty of elderly people, and provi-

sioning subsequent individualized interventions.

Further technologies are investigated for more di-

versiﬁed analysis of behavior; e.g., bed sensors can be

Technological Approach for Behavior Change Detection toward Better Adaptation of Services for Elderly People

103

used for sleep period and vital sign recognition, kinect

sensors enable more accurate monitoring of walking

activity, and beacon sensors with smart phones allow

more precise understanding of outdoor activities.

New reasoning techniques are studied to corre-

late identiﬁed statistical changes with overall changes

in behavior toward better adaptation of provided ser-

vices; e.g., decrease in weight indicates negative nu-

tritional change and triggers sending of personalized

notiﬁcations to improve nutritional status.

ACKNOWLEDGEMENT

We give our special thanks to Saint Vincent de Paul

nursing home in Occagnes, France. Our deployment

in this nursing home is also supported by VHP in-

ter@ctive project and the Quality Of Life chair.

Our work is part of the European project City4Age

that received funding from the Horizon 2020 research

and innovation program under grant agreement num-

ber 689731.

REFERENCES

Allin, S., Bharucha, A., Zimmerman, J., Wilson, D., Robin-

son, M., Stevens, S., Wactlar, H., and Atkeson, C.

(2003). Toward the automatic assessment of behav-

ioral distrubances of dementia.

Aloulou, H. (2013). Framework for ambient assistive liv-

ing: handling dynamism and uncertainty in real time

semantic services provisioning. PhD thesis, Evry, In-

stitut national des t

ecommunications.

Aloulou, H., Mokhtari, M., Tiberghien, T., Biswas, J., Phua,

C., Lin, J. H. K., and Yap, P. (2013). Deployment of

assistive living technology in a nursing home environ-

ment: methods and lessons learned. BMC medical

informatics and decision making, 13(1):1.

Andersson, J. (2014). Locating multiple change-points us-

ing a combination of methods.

Avvenuti, M., Baker, C., Light, J., Tulpan, D., and Vec-

chio, A. (2010). Non-intrusive patient monitoring of

alzheimers disease subjects using wireless sensor net-

works. In 2009 World Congress on Privacy, Security,

Trust and the Management of e-Business.

Barberger-Gateau, P., Commenges, D., Gagnon, M., Leten-

neur, L., Sauvel, C., and Dartigues, J.-F. (1992). In-

strumental activities of daily living as a screening

tool for cognitive impairment and dementia in elderly

community dwellers. Journal of the American Geri-

atrics Society, 40(11):1129–1134.

Basseville, M., Nikiforov, I. V., et al. (1993). Detection of

abrupt changes: theory and application, volume 104.

Prentice Hall Englewood Cliffs.

Bland, J. M. and Altman, D. G. (1995). Comparing

methods of measurement: why plotting difference

against standard method is misleading. The lancet,

346(8982):1085–1087.

Cao, L. (2010). In-depth behavior understanding and use:

the behavior informatics approach. Information Sci-

ences, 180(17):3067–3085.

Cho, H. (2015). Change-point detection in panel data via

double cusum statistic.

City4Age (2016). Elderly-friendly city services

for active and healthy aging. available as

http://www.city4ageproject.eu/.

Cockrell, J. R. and Folstein, M. F. (2002). Mini-mental state

examination. Principles and practice of geriatric psy-

chiatry, pages 140–141.

Cummings, J. L., Mega, M., Gray, K., Rosenberg-

Thompson, S., Carusi, D. A., and Gornbein, J. (1994).

The neuropsychiatric inventory comprehensive as-

sessment of psychopathology in dementia. Neurology,

44(12):2308–2308.

Hayes, T. L., Abendroth, F., Adami, A., Pavel, M., Zitzel-

berger, T. A., and Kaye, J. A. (2008). Unobtru-

sive assessment of activity patterns associated with

mild cognitive impairment. Alzheimer’s & Dementia,

4(6):395–405.

Hayes, T. L., Larimer, N., Adami, A., and Kaye, J. A.

(2009). Medication adherence in healthy elders: small

cognitive changes make a big difference. Journal of

aging and health.

Kaye, J., Mattek, N., Dodge, H. H., Campbell, I., Hayes,

T., Austin, D., Hatt, W., Wild, K., Jimison, H., and

Pavel, M. (2014). Unobtrusive measurement of daily

computer use to detect mild cognitive impairment.

Alzheimer’s & Dementia, 10(1):10–17.

Kibria, G. (2016). Cumulative sum and exponen-

tially weighted moving average control charts.

preprint, available as http://www2.ﬁu.edu/ kibriag

/Stat5666/Handout/Chapter99.pdf.

Lafont, S., Barberger-Gateau, P., Sourgen, C., and Dar-

tigues, J. (1999). Relation entre performances cog-

nitives globales et d

ependance

evalu

ee par la grille

aggir. Revue d’

epid

emiologie et de sant

e publique,

47(1):7–17.

Lavikainen, H. M., Lintonen, T., and Kosunen, E. (2009).

Sexual behavior and drinking style among teenagers:

a population-based study in ﬁnland. Health promotion

international, 24(2):108–119.

Ledolter, J. and Kardon, R. (2013). Detecting the pro-

gression of eye disease: Cusum charts for assessing

the visual ﬁeld and retinal nerve ﬁber layer thickness.

Translational vision science & technology, 2(6):2–2.

Liu, S., Yamada, M., Collier, N., and Sugiyama, M. (2013).

Change-point detection in time-series data by relative

density-ratio estimation. Neural Networks, 43:72–83.

Magill, E. and Blum, J. M. (2012). Personalised ambi-

ent monitoring: supporting mental health at home.

Advances in home care technologies: Results of the

Match project, pages 67–85.

Mathias, S., Nayak, U., and Isaacs, B. (1986). Balance in

elderly patients: the” get-up and go” test. Archives of

physical medicine and rehabilitation, 67(6):387–389.

Mesnil, B. and Petitgas, P. (2009). Detection of changes in

HEALTHINF 2017 - 10th International Conference on Health Informatics

104

time-series of indicators using cusum control charts.

Aquatic Living Resources, 22(2):187–192.

Miller, J. W., Naimi, T. S., Brewer, R. D., and Jones, S. E.

(2007). Binge drinking and associated health risk

behaviors among high school students. Pediatrics,

119(1):76–85.

Moskvina, V. and Zhigljavsky, A. (2003). An algorithm

based on singular spectrum analysis for change-point

detection. Communications in Statistics-Simulation

and Computation, 32(2):319–352.

Neuropsy, R. (2016). Four

Tests. http://fmc31200.free.fr/MG-

liens/Neurologie/quatre tests.pdf. [Online; accessed

September-2016].

Ormrod, J. E. (2013). Educational Psychology: Pear-

son New International Edition: Developing Learners.

Pearson Higher Ed.

Page, E. (1954). Continuous inspection schemes.

Biometrika, 41(1/2):100–115.

Parmelee, P. A. and Katz, I. R. (1990). Geriatric depression

scale. Journal of the American Geriatrics Society.

Perner, L. (2008). Consumer behaviour and marketing strat-

egy.

Petersen, J., Austin, D., Yeargers, J., and Kaye, J. (2014).

Unobtrusive phone monitoring as a novel measure of

cognitive function. Alzheimer’s & Dementia: The

Journal of the Alzheimer’s Association, 10(4):P366–

P367.

Prochaska, J. O. and DiClemente, C. C. (2005). The trans-

theoretical approach. Handbook of psychotherapy in-

tegration, 2:147–171.

Reisberg, B., Auer, S. R., and Monteiro, I. M. (1997).

Behavioral pathology in alzheimer’s disease (behave-

ad) rating scale. International Psychogeriatrics,

8(S3):301–308.

Rosenstock, I. M. (1974). Historical origins of the

health belief model. Health Education & Behavior,

2(4):328–335.

Schwarzer, R. (2008). Modeling health behavior change:

How to predict and modify the adoption and main-

tenance of health behaviors. Applied Psychology,

57(1):1–29.

Szwacka-Mokrzycka, J. (2015). Trends in consumer be-

haviour changes. overview of concepts. Acta Scien-

tiarum Polonorum. Oeconomia, 14(3).

Takeuchi, J.-i. and Yamanishi, K. (2006). A unifying frame-

work for detecting outliers and change points from

time series. IEEE transactions on Knowledge and

Data Engineering, 18(4):482–492.

Tardieu,

E., Mahmoudi, R., Novella, J.-L., Oubaya, N.,

Blanchard, F., Jolly, D., and Drame, M. (2016). Ex-

ternal validation of the short emergency geriatric as-

sessment (sega) instrument on the safes cohort. Geri-

atrie et psychologie neuropsychiatrie du vieillisse-

ment, 14(1):49–55.

Taylor, W. A. (2000). Change-point analysis: a powerful

new tool for detecting changes. preprint, available

as http://www. variation. com/cpa/tech/changepoint.

html.

Vellas, B., Guigoz, Y., Garry, P. J., Nourhashemi, F., Ben-

nahum, D., Lauque, S., and Albarede, J.-L. (1999).

The mini nutritional assessment (mna) and its use in

grading the nutritional state of elderly patients. Nutri-

tion, 15(2):116–122.

Wilson, E. O. (2000). Sociobiology. Harvard University

Press.

Technological Approach for Behavior Change Detection toward Better Adaptation of Services for Elderly People

105