Predictive Behavioural Monitoring and Deviation Detection in Activities

of Daily Living of Older Adults

Jiawei Zheng

and Petros Papapanagiotou

School of Informatics, University of Edinburgh, 10 Crichton Street, Edinburgh, EH8 9AB, U.K.

Keywords:

Activities of Daily Living, Behaviour Monitoring, Deviation Detection, Predictive Modelling, Sensor Data.

Abstract:

Predictive behaviour monitoring of Activities of Daily Living (ADLs) can provide unique, personalised in-

sights about an older person’s physical and cognitive health and lead to unique opportunities to support self-

management, proactive intervention and promote independent living. In this paper, we analyse ADL data from

ambient sensors to model behaviour markers on a daily basis. Using a number of machine learning and statis-

tical methods we model a predicted daily routine for each marker, detect deviations based on a set of relative

thresholds and calculate long-term drifts. We further analyse the causal factors of deviations by investigating

relationships between different activities. We demonstrate our results using data from a sample of 11 partici-

pants from the CASAS dataset. Finally, we develop a dashboard to visualize our computed daily routines and

quantiﬁed deviations in an attempt to offer useful feedback to the monitored person and their caregivers.

1 INTRODUCTION

Loss of independence in Activities of Daily Living

(ADL) is associated with adverse health outcomes,

both physical and mental, and mortality in older

adults (Albanese et al., 2020; Cohen-Mansﬁeld and

Perach, 2012). Adverse health events, including heart

failures, falls, strokes, etc. and the onset of cogni-

tive and physical frailty are not random occurrences,

but a consequence of long-term health deterioration

or unhealthy lifestyle. Gradual cognitive and phys-

ical decline can signiﬁcantly affect the capacity of

people in advanced age to perform ADLs indepen-

dently (Akram et al., 2020). Proactive monitoring and

analysis of short and long term deviations from a reg-

ular routine in ADLs can provide vital insights on an

older person’s health and a continuous evaluation of

their physical and cognitive ability (Sepesy Mau

cec

and Donaj, 2021). These can not only inform the de-

cision making of care givers towards timely, proactive

interventions, but also support person-centred self-

management and motivate healthier living.

Modern and emerging smart home and wearable

sensor technologies allow us to collect continuous

data on daily living in an unobtrusive and affordable

manner. This involves time series data with status in-

https://orcid.org/0000-0002-6515-6423

https://orcid.org/0000-0003-0928-6108

formation and timestamped event data when a status

change occurs (Cook et al., 2013a). Activity Recogni-

tion and Machine Learning (ML) techniques can then

produce ﬁne-grained daily activity data with tempo-

ral and spatial information. Further AI modelling can

allow us to develop rich temporal and spatial proﬁles

of daily routines, including sleep duration, number of

meals, and levels of active movement. We can detect

temporal and spatial deviations on individual days or

in the long term, such as staying in the toilet too long,

too frequent toilet visits during the night, activity de-

lays due to reduced mobility, sleep disruption or de-

cline over time etc. Ultimately, our goal is to link

these detected deviations to health outcomes, towards

health monitoring and timely, proactive interventions.

More importantly, ﬁguring out the causes of the de-

viations and the relationships between daily activities

can provide key insights towards preventative instruc-

tions and effective care provisions, both by care givers

and the people themselves, to avoid potential negative

effects and prevent adverse events.

Until recently, studies that investigate the relation-

ship between ADLs and health outcomes were based

on questionnaires and self-reporting (Kanti Majum-

dar, 2014; Cook and Schmitter-Edgecombe, 2021).

The result of this may be affected by experi-

menter observations or retrospective memory limita-

tions (Palmer, 2018). Instead, sensor-based, passive

and continuous monitoring of ADLs, associated with

Zheng, J. and Papapanagiotou, P.

Predictive Behavioural Monitoring and Deviation Detection in Activities of Daily Living of Older Adults.

DOI: 10.5220/0010974500003123

In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - Volume 5: HEALTHINF, pages 899-910

ISBN: 978-989-758-552-4; ISSN: 2184-4305

899

behavioural analysis of patterns and routines, can pro-

vide concrete, objective insights about the relation-

ship between daily activities and health.

Our work is part of the Advanced Care Research

Centre, a large multi-disciplinary project focused on

the support of people in later life living in their own

homes and in supported care environments. (Univer-

sity of Edinburgh, 2021). The main contributions

of this paper are the following: Firstly, we extract

behaviour markers from activity-labelled time series

sensor data to model daily behaviours with rich tem-

poral information. Secondly, we propose a devia-

tion detection approach for daily behaviours, includ-

ing a deviation score for a given day and a long-term

drift from the normal routine. Thirdly, we investigate

the relationship between different activities towards

a causal explanation of deviations. Finally, we de-

velop an interactive dashboard, which can visualize

personal temporal proﬁles of daily behaviours includ-

ing daily routines, trends of each behaviour marker

and potential deviations. The dashboard provides in-

tuitive behavioural statistics of the monitored older

adults that may be useful both for self-monitoring

and management and as useful information to care

providers. We demonstrate the results of our approach

using ADL data from 11 participants over 2 months

from the CASAS project (Cook et al., 2013a).

2 RELATED WORK

Research in the ﬁeld of ADLs has recently been re-

ceiving increasing attention, especially in the context

of supporting independent living for older people and

providing effective care. Aminikhanghahi et al. pro-

pose a change point detection algorithm for identify-

ing activity transition points, which is used to improve

the performance of activity recognition (Aminikhang-

hahi et al., 2019). In this paper, we focus on predic-

tive daily behaviour monitoring and deviation detec-

tion based on the activity-labelled time series data.

Yahaya et al. provide a comprehensive list of recent

efforts on deviation detection (Yahaya et al., 2019).

Classiﬁcation methods treat deviation detection

as a binary classiﬁcation problem. They require ADL

data labelled as normal or abnormal, where the lat-

ter reﬂects either a pre-speciﬁed pattern of behaviour,

such as the behavioural difﬁculties of people with de-

mentia (Arifoglu and Bouchachia, 2019). Due to the

scarcity of abnormal data in real datasets, it is com-

mon to train and generate synthetic data for the ab-

normal class. More recent approaches consider as

abnormal any deviation from a normal routine that

is learned from historical data (Yahaya et al., 2019;

Pazhoumand-Dar et al., 2020; Yahaya et al., 2021b).

Some classiﬁcation approaches rely on detect-

ing whether the sensor information exceeds a ﬁxed

threshold of values. Collected data is used as train-

ing data representing the normal behaviours and sub-

sequent activity data are used as testing data for the

learned model. Data that have signiﬁcant variations

past certain thresholds are deﬁned as outliers. For ex-

ample, Pierleoni et al. propose a fall detection method

based on the fusion data collected from a triaxial ac-

celerometer, gyroscope, and magnetometer on wear-

able devices (Pierleoni et al., 2015). If the body orien-

tation falls below a pre-deﬁned threshold for a certain

period of time, the system will issue an alarm.

However, approaches that rely on wearable sen-

sors are not always applicable in practice. For in-

stance, some people may not feel comfortable con-

stantly wearing a device or may forget to put it on

or charge the battery. Moreover, missing data and

false positives can make these approaches less reli-

able (Pazhoumand-Dar et al., 2020). For example,

lying down on a bed suddenly may be mistaken for

the movement pattern of an accidental fall. Some

approaches have overcome these limitations by us-

ing ambient sensors in a smart home (Cook et al.,

2013a). The data are collected from environmental

sensors when the subjects have interactions with their

environment. For example, Pazhoumand-Dar et al.

use Kinect sensors composed with power consump-

tion data to monitor daily behaviours (Pazhoumand-

Dar et al., 2020). Their training data is aimed to

model the regularity and frequency of important ac-

tivities and does not need to be labelled in advance.

Howedi et al. use ADL data collected from ambient

sensors to detect deviations based on entropy mea-

sures (Howedi et al., 2020). Activities with entropy

values exceeding a certain range are detected as ab-

normalities. Similarly, Yahaya et al. propose an en-

semble of abnormal detection approach by detecting

if the test data differs signiﬁcantly from the training

data based on a threshold for a deﬁned Normality

Score (Yahaya et al., 2019).

Finally, clinical score-based approaches involve

an assessment of older people by clinical experts

through various factors of their daily activities, such

as cognitive health, functional mobility, etc (Dawadi

et al., 2016; Alberdi Aramendi et al., 2018; Cook and

Schmitter-Edgecombe, 2021). The assessments are

usually conducted at regular time intervals, for ex-

ample every 6 months, and a total score representing

the health status of the participant is calculated. Af-

ter data collection, a computation model is trained to

map the clinical score to the data collected by ambi-

ent sensors and predict future scores based on that.

Smart CommuniCare 2022 - Special Session on Smart Living Environments to Support Aging-in-Place in Vulnerable Older Adults

900

Figure 1: An overview of the proposed approach.

The assessments are carried out by self-reporting, us-

ing an appropriate questionnaire. However, the result

from this type of assessment may be affected by the

respondent’s subjective view and state of mind when

ﬁlling in the questionnaire (Yahaya et al., 2021a). As

a consequence, the score may not accurately reﬂect

the actual health condition of the respondent.

In our work, we also adopt a threshold-based clas-

siﬁcation approach using unlabelled data.

3 PROPOSED APPROACH

Our work is aimed towards proactive health monitor-

ing and predictive deterioration and deviation detec-

tion. Our proposed approach consists of four stages,

as shown in Figure 1: (1) data preprocessing, (2)

extraction of behaviour markers, (3) predictive mod-

elling, and (4) deviation detection, including investi-

gating causal factors of deviations. We describe each

stage in more detail next.

3.1 Data Description

In this paper, we analyse available ADL data collected

from ambient sensors in smart homes, published by

Cook et al. (Cook et al., 2013a). The data sets contain

continuous data from unobtrusive sensors including

motion sensors, door sensors, light switches and light

sensors, deployed in single-family residences.

Each sensor event is labelled with a corresponding

activity, including sleeping, cooking, eating, napping,

going to the toilet, and working, in a total of 33 differ-

ent activities. A sample of data is shown in Table 1a.

Our paper is based on data from 11 individuals

over 2 months. The raw sensor data is recorded in a

time-series format with the following ﬁelds:

• Timestamp: The date and time of the event.

• Sensor: The name of the sensor, as found in the

sensor ﬂoor plan.

• Room: The room-level sensor location.

• Location: The ﬁne-grained location of the sensor,

such as bed, chair, etc.

• Message: The value generated by the sensor, such

as on, off, etc.

• Sensor Type: The type of sensor gener-

ating the event (e.g. Control4-Motion,

Control4-LightSensor, etc.), such that

provides context to the generated message.

• Activity: A manual label of the corresponding ac-

tivity of this event, such as sleeping, eating, etc.

Due to inherent uncertainty in the environment

and human behaviours, the dataset is noisy. We

speciﬁcally identiﬁed 4 categories of potential noise

below:

• Accuracy of Labelling: Since the particular

dataset we are examining was manually labelled

post-hoc, there is no measure of the accuracy

of the labelling. In fact, newer versions of the

dataset seem to have updated some of the labels.

More generally, even with an automated activity

recognition algorithm, such as the one by Cook et

al. (Cook et al., 2013b), there is still some level of

uncertainty that may affect the results.

• Lack of end Event: The observed activity events

indicate when an activity is detected. However,

there is no clear indication of the exact end time

of the activity or the transition time to the next

one. This causes some inaccuracy in our calcu-

lated activity intervals (see Section 3.2).

• Distinguishing Activities: In some cases, the

same small set of 1-2 sensors may be used to de-

tect multiple types of activities with similar action

Predictive Behavioural Monitoring and Deviation Detection in Activities of Daily Living of Older Adults

901

Table 1: Sample of ADL sensor data.

(a) Raw ADL sensor data labelled with activities.

Timestamp Sensor Room Location Message Sensor type Activity

1 2011-06-15 09:58:27 M007 Kitchen Kitchen ON Control4-Motion Cook Breakfast

2 2011-06-15 09:58:42 M007 Kitchen Kitchen OFF Control4-Motion Cook Breakfast

3 2011-06-15 09:58:45 D005 Kitchen Refrigerator Close Control4-Door Cook Breakfast

4 2011-06-15 09:58:46 M005 DiningRoom DiningRoom ON Control4-Motion Eat Breakfast

5 2011-06-15 09:58:47 M008 Kitchen Kitchen ON Control4-Motion Cook Breakfast

6 2011-06-15 09:59:05 LS001 Kitchen Kitchen 4 Control4-LightSensor Cook Breakfast

7 2011-06-15 09:59:06 LS007 Kitchen Kitchen 6 Control4-LightSensor Cook Breakfast

8 2011-06-15 09:59:10 M005 DiningRoom DiningRoom ON Control4-Motion Eat Breakfast

9 2011-06-15 10:00:29 LS005 DiningRoom DiningRoom 13 Control4-LightSensor Eat Breakfast

10 2011-06-15 10:00:46 LS015 DiningRoom DiningRoom 14 Control4-LightSensor Eat Breakfast

11 2011-06-15 10:00:47 M008 Kitchen Kitchen OFF Control4-Motion Cook breakfast

(b) ADL data resulting after pre-processing.

Activity Start time End time Duration Daycase Interval

Cook Breakfast 2011-06-15 09:58:27 2011-06-15 09:59:06 00:00:39 2011-06-15 00:00:04

Eat Breakfast 2011-06-15 09:59:10 2011-06-15 10:00:46 00:01:36 2011-06-15 00:00:01

patterns or taking place in the same location, such

as washing the dishes and preparing a meal. This

can lead to some misclassiﬁcation.

• Sensor Modalities: Given the selection of sen-

sors, there are only a few modalities available,

mainly movement and lighting. This means that

certain activities, such as the exact time an indi-

vidual falls asleep, cannot be detected accurately.

Additional modalities, such as the ones offered by

wearable devices, can help provide a ﬁner grained

activity detection. However, there will always be

some uncertainty in activity recognition due to the

granularity of the involved sensors.

Whilst we do not address this noise explicitly in

our modelling, it needs to be taken into considera-

tion when making potential decisions or interventions

based on our produced insights.

3.2 Data Preprocessing

The ADL dataset consists of raw sensor data, which

represents all sensor events that occur during a pe-

riod of time, along with their speciﬁc location, times-

tamps, sensor ID, message, etc, as shown in Table 1a.

The ﬁrst processing step involves the removal of

noise, as described in the previous section, to the ex-

tent possible, as well as the removal of unknown ac-

tivities labelled as “other activity”.

We then focus on particular patterns of sequences

of events representing two different, interleaved ac-

tivities during the same period of time, as shown in

Table 1a. A person is cooking breakfast when a se-

quence of eating breakfast events occurs. Subsequent

events show that the person continues to cook break-

fast. This means that the person is cooking while eat-

ing and the two activities are interleaved during this

time.

In row 4 of Table 1a, the eat breakfast event is a

single orphan event, which prevents us from measur-

ing the (assumed to be short) duration of the corre-

sponding activity. We treat such events as false oc-

currences and remove them from the dataset.

In rows 8-10 of Table 1a, we observe a continuous

sequence of eat breakfast events. We consider such

cases as an interleaved activity and we treat them as

though the initial activity ended in row 7 and a new

cook breakfast activity started in row 11.

Based on the above, we detect consecutive events

of the same activity and choose the timestamp of the

ﬁrst event as the start of the activity interval and the

timestamp of the last event as the end of the activity

interval. This results in a processed dataset (shown

in Table 1b) that includes the following temporal fea-

tures of each activity in order:

• Activity: the name of the ADL.

• Start Time: the start time of the activity.

• End time: the end time of the activity.

• Duration: the duration of the activity in seconds.

• Daycase: the date of the activity.

• Interval: the interval in seconds until the next ac-

tivity starts.

3.3 Modelling Behaviour Markers

Our aim is to process activities performed on a daily

basis in order to detect daily routines and deviations.

We model behaviour markers for each day.

Smart CommuniCare 2022 - Special Session on Smart Living Environments to Support Aging-in-Place in Vulnerable Older Adults

902

This daily segmentation requires us to split events

occurring at midnight into two events, each be-

longing to the previous and new day respectively.

For example, a Watch TV event from 22:10:00 to

01:12:23 is split into two events, one from 22:10:00 to

23:59:59 for the previous day and one from 00:00:00

to 01:12:23 belonging to the new day. The only ex-

ception is the sleep activity, which we count from

noon to noon of the next day, as we consider overnight

sleep to be a signiﬁcant indicator of healthy living.

The modelled behaviour markers are extracted ac-

cording to the features related to each daily activ-

ity such that provide key insights of the individual’s

lifestyle and potential links to health indicators.

The total number of behaviour markers we ob-

served is 26, with some examples shown in Table 2.

We subsequently focus our efforts to detect deviations

on these particular behaviour markers.

Table 2: Examples of daily behaviour markers.

Behaviour markers Description

nap duration total nap duration

sleep time time of sleep

wake up time

morning wake up

time

last personal hygiene

time of the last

personal hygiene

last nap endtime

end time of the last

nap

if eat breakfast

whether eat breakfast

occurred or not

bed toilet times

times of toilet visits

during sleep

3.4 Detecting Deviations

The main goal of our approach is to detect and mea-

sure short and long term deviations of ADLs from a

routine, such that may indicate health deterioration. A

key input here is the ground truth compared to which

a certain behaviour is considered a deviation. Each

individual may have a signiﬁcantly different routine

and lifestyle, particularly at a later age. This vari-

ance can be exacerbated by the consideration of the

large variety of different health conditions that may

apply to each person. For instance, the connections

between frailty and sleep disturbances have been well

studied (Piovezan et al., 2013). We therefore adopted

a personalised approach and observe daily living de-

viations from an individual routine, irrespective of its

relation to a healthy standard. This means we are de-

tecting and measuring when and individual is not be-

having according to their regular routine, even if that

routine is unhealthy.

More speciﬁcally, given a set of daily behaviour

markers for each individual, we create a computa-

tional model of their expected measures each day.

Taking nap duration as an example, we develop a

model that can predict the total duration of a person’s

naps during the day given their historical data. We

also account for seasonality by including the day of

the week (e.g. some people may nap more during the

weekend) and the month (e.g. people tend to sleep

longer during the winter, particularly if affected by

a seasonal affective disorder) as features (Anderson

et al., 1994). This model is then used as the ground

truth of expected values, against which any new ob-

served behaviour is compared.

The technical aspects of this approach are de-

scribed in more detail next.

3.4.1 Building the Predictive Model

Our aim is to build a predictive model of the nor-

mal daily behaviour markers of an individual. Such

a model needs to be trained with historical data of

routine daily living. We therefore begin by remov-

ing outliers in the data, consisting of values that de-

viate from the mean by 2 standard deviations, which

is a common cut-off for outliers in practice in a small

dataset (Ilyas and Chu, 2019).

We also consider special cases of daily behaviours

that we may want to ﬁlter out. For example, we use

speciﬁc patterns for detecting if the person is away

from home overnight and sleeps elsewhere. If the last

event of that day is leave home and sleep duration is

zero, we consider that day an outlier and remove it.

We use the remaining data to train a regression

model to predict the future value of behaviour mark-

ers. We only consider seasonality features in the

model, i.e. the full date, the day of the week, the

month, and the order of events.

Our next goal is to maximize the accuracy of the

model. For this, we compare the predictive accuracy

of different regression models in a model selection

process. In this process, we use one of the partici-

pants as the validation set for choosing best perform-

ing model and other participant as the testing set. The

best performing model on the validation set is eval-

uated on the testing set. Finally, the most accurate

model is chosen as the predictive model.

For the data of each behaviour marker of a indi-

vidual, we use K-fold time series split technique in

Scikit-learn (Pedregosa et al., 2011) to split 3-fold

training and testing sets. It returns 3 split of data.

In the kth split, it returns the ﬁrst k folds as train-

ing set and the (k + 1)th fold as testing set. We train

and evaluate different ML models and statistical time

Predictive Behavioural Monitoring and Deviation Detection in Activities of Daily Living of Older Adults

903

series models based on the training and testing sets.

Note that we build separate model for each behaviour

marker of a individual. The evaluation result is ob-

tained by computing the average performance of 3

pairs of training/testing sets.

All experiments are evaluated using the mean ab-

solute error (MAE) and the root mean squared error

(RMSE) measures. MAE, shown in (1), and RMSE,

shown in (2) both measure the average magnitude of

the errors in a set of predictions, with 0 corresponding

to perfect accuracy, while RMSE magniﬁes the im-

pact of large errors. In these equations, for each pre-

dicted value i, ˆy

represents a size-n vector of the pre-

dicted values, y

is the vector of actual values, and n

is the number of test instances. All performance eval-

uations are conducted using 3-fold time series cross

validation (Bergmeir and Ben

ıtez, 2012).

MAE =

∑

i=1

| ˆy

− y

| (1)

RMSE =

∑

i=1

( ˆy

− y

)

(2)

3.4.2 Deviation Detection

After we build the predictive model, we can get the

predicted value of behaviour markers by ﬁtting them

into the training set. To detect potential deviations

based on the predicted value, we calculate the dis-

tance between the predicted routine value and the ac-

tual value as shown in (3).

z =

predict

− y

true

MAE

(3)

MAE is the mean absolute error of the predictive

model (see also Section 3.4.1). For example, if the

MAE of a model of sleep duration is 1800 seconds,

then the model predicts on average 30 minutes more

or less sleep than the actual value. A lower MAE in-

dicates a better performing algorithm with 0 corre-

sponding to perfect accuracy. When comparing mod-

els trained with datasets from 2 distinct individuals

with the same algorithm, a lower MAE score is an in-

dication of less variability in the data and, therefore,

a more predictable and steady daily routine.

The z value in (3) represents the distance of the

actual value from the predicted value as a fraction of

MAE. If the calculated z value of a predicted be-

haviour marker exceeds a chosen threshold, a devi-

ation is detected. The chosen threshold corresponds

to a time window proportional to MAE within which

we consider the behaviour as normal or routine. A

distinct threshold can be selected for each behaviour

marker in the daily routine, to account for the ﬂexibil-

ity we want to allow for each activity. For instance, if

a participant decides to read a book at a considerably

different time than usual, we might not want to con-

sider this as a signiﬁcant deviation. However, the time

they go to sleep is much more important as a health in-

dicator, and therefore we may want to ﬂag smaller de-

viations. We could therefore choose a higher z thresh-

old for the former behaviour marker and a lower value

for the latter.

For example, given the sleep duration model with

MAE of 0.5 hours discussed above, assume a pre-

dicted sleep duration y

predict

of 8 hours for a partic-

ular day. Setting the threshold of z to 1 means that

an actual sleep y

rue of less than 7.5 or more than 8.5

hours will be considered a deviation.

Given that the MAE reﬂects the variability in an

individual’s routine, choosing the same z value across

all individuals allows us to account for that variability.

In the example above, an MAE of 1 hour would lead

us to only consider sleep of less than 7 hours or more

than 9 hours to be a deviation.

The deviations of an individual across all be-

haviour markers in a particular day can be quanti-

ﬁed in terms of a deviation cost. In this, deviations

in each behaviour marker may have a different cost,

for instance in terms of its potential impact to the per-

son’s health. For example, a deviation from the ex-

pected sleep duration is likely considered more im-

pactful to health compared to a deviation in the time

one chooses to read a book.

For this purpose, we set a customized weight for

each behaviour marker to adjust the impact of speciﬁc

behaviours. For the detected deviations above, we

then calculate both the absolute total deviation cost

and the weighted deviation cost. The total deviation

cost is the sum of the cost of each deviation behaviour

marker. The weighted deviation cost is shown in (4),

where i represents each behaviour marker, z

is the

z value from (3) for each behaviour marker and w

the selected custom weight of the marker. The combi-

nation of total deviation cost and weighted deviation

cost provides a quantiﬁcation of deviations for a sin-

gle person on an individual day. This has the potential

to be used as a behaviour performance score for care

givers when monitoring the daily living routine of an

older adult at a glance.

C =

∑

∗ w

(4)

3.4.3 Long-term Deterioration

In addition to short-term deviations over a single day,

we also analyse the long-term trend of behaviour

Smart CommuniCare 2022 - Special Session on Smart Living Environments to Support Aging-in-Place in Vulnerable Older Adults

904

markers. This type of analysis can help us detect

long-term changes such as reduced mobility (some

activities taking longer), reduced or disrupted sleep

patterns, etc. that may be linked to health outcomes,

such as deterioration, physical and cognitive frailty,

and reduced independence.

We calculate long-term trends of behavioural

markers based on the individual’s normal daily rou-

tine. In this context, we consider days with detected

deviations through the previous analysis as abnormal

and ﬁlter them out of the dataset.

3.5 Investigating Relationship between

Activities

The last step in our current work is aimed towards in-

sights on why the detected deviations occur, focusing

on individual behaviour markers. Taking sleep dura-

tion as an example, we can use the approach described

so far to classify the sleep duration of each day as ab-

normal (deviating) or normal based on the individ-

ual’s historical data. Taking this labelling as ground

truth, we train a new classiﬁer to use the other be-

haviour markers as features for normal/abnormal clas-

siﬁcation. We then calculate the feature importance

of each marker using the feature selection technique

in Scikit-learn (Pedregosa et al., 2011). This tech-

nique can assign a score to each feature of the clas-

siﬁer based on their impact in predicting the target

label. The higher the score, the more important the

corresponding feature is. Features with high impor-

tance are then considered to have a higher correlation

with sleep duration.

4 EXPERIMENTAL RESULTS

In this section, we present the results we obtained in

each stage of our approach using the CASAS dataset.

We develop and evaluate our predictive model of daily

routines by comparing the performance of different

learning models. Additionally, we show the results of

detected deviations and relationships between sleep

duration and other daily behaviours. Finally, we de-

velop an interactive dashboard for visualizing per-

sonal temporal proﬁles of daily living.

4.1 Predictive Model

In our effort to develop a predictive model for individ-

ual ADL routines, we performed an array of experi-

ments to select the best performing algorithm.

First, we evaluated 19 well known ML regression

models, including Random Forest Regressor, Lin-

ear Regressor, Logistic Regressor, Extreme Gradient

Boosting, etc. As mentioned earlier, the features used

for regression only contain time-related characteris-

tics, i.e. month, day, weekday and chronological or-

der. Then the top 3 best performing among them are

integrated as an ensemble using Stacking and Blend-

ing techniques (Maclin and Opitz, 1999) to further

improve the performance. We also evaluate one statis-

tical time series approach named Prophet (Taylor and

Letham, 2017), which is speciﬁcally tailored to deal

with seasonal effects in time-series data, and compare

its performance to the ML models.

We randomly chose hh102 as an example to

present our results. The other participants show simi-

larly interesting results. The performance of different

models of sleep duration in terms of the two metrics

(MAE and RMSE) is shown in Table 3. The units of

these two metrics are both seconds.

Table 3: MAE and RMSE of different predictive models of

sleep duration for individual hh102.

Model MAE RMSE

Prophet 3683 4532

RandomForestRegressor 4200 5065

KNeighborsRegressor 4306 5394

ExtraTreesRegressor 4552 5436

StackingRegressor (top 3) 5943 6551

BlendingRegressor (top 3) 4214 5127

As the result shows, Prophet is the best model for

forecasting sleep duration. In fact, Prophet outper-

formed the ML algorithms in all behaviour markers

across our dataset, so we selected that algorithm for

all our predictive models. This came with added ben-

eﬁts of Prophet, such as the calculation of long-term

trends (see Section 4.3).

The model performance on the behaviour mark-

ers of the individual hh102 is shown in Table 4. As

an example, the mean absolute error in the predic-

tion of the time when breakfast was cooked is 3745

seconds, so approximately one hour, which is a rea-

sonable level of variability for that activity. Similarly,

all of the obtained results show a reasonable level of

accuracy, given the high variability in people’s daily

lives.

4.2 Detected Deviations

We present the results of our deviation detection ap-

proach using hh102 as an example. In this, we set the

threshold of the z value of each behaviour marker to

1. We set the weight cost of sleep duration, bathe du-

ration and leave home duration, which we considered

more important in our particular use case, to 2, and

the weights of the rest of the behaviour markers to 1.

Predictive Behavioural Monitoring and Deviation Detection in Activities of Daily Living of Older Adults

905

Table 4: Performance measures of the models on the be-

haviour markers of hh102.

Behaviour markers MAE RMSE

cook breakfast time 3745.318 3938.721

eat breakfast time 4185.226 4611.391

cook lunch time 4232.662 5127.29

eat lunch time 6120.533 6651.227

cook dinner time 1263.975 1580.311

eat dinner time 1687.354 2115.673

sleep time 2726.95 3160.01

sleep duration 3875.959 4543.112

take medicine time 4615.162 4933.677

morning medicine time 2250.793 2825.039

bathe duration 187.1277 274.4289

leave home duration 3415.511 4944.483

Figure 2: Detected deviations of sleep duration for hh102.

Figure 2 shows the calculated z values for sleep

duration on different dates. Negative values mean that

the actual sleep duration was longer than the predicted

value, while positive values mean the expected sleep

duration was longer than the actual value. We detect

deviations on 15 days, which we label abnormal.

We also visualize the actual values and predicted

values of behaviour markers related to time of the day

and their MAE range in a timetable in Figure 3. Red

marks mean the actual value and the yellow marks

mean the predicted value with a line showing the cor-

responding fault tolerance range based on the thresh-

old of the z value. We can readily ﬁgure out which be-

haviour markers are deviations and the magnitude of

each one. For example, we see that participant hh102

took their morning medicine earlier than the predicted

time on this particular day.

We summarize the computed personal temporal

proﬁle of the individual person, including their daily

routines, long-term trends and detected deviations in

an interactive dashboard. Figure 3 shows one of the

views of our dashboard that visualizes the absolute

deviation cost, the weighted deviation cost, the de-

tected deviations, and the predicted ranges (in yellow)

and actual values (in red) of behavioural markers re-

lated to the time of the day for a particular, selected

date. Other views of our dashboard show a timeline

of the detected day routine and the long-term trends of

the behavioural markers, but due to space limitations

we do not present these here.

4.3 Long-term Deviations

Based on the 2 month data of the individual hh102,

we analyse the sleep duration trend to detect whether

it is deteriorating. The general trend of sleep duration

for hh102 is shown in Figure 4, where the horizontal

axis and the vertical axis represent the observed date

period and the sleep duration in seconds respectively.

The black dots represent the actual sleep duration of

that day, while the blue shaded area shows the predic-

tion range (lower bound and upper bound) provided

by Prophet. The predicted trend shows the sleep du-

ration follows a relatively steady pattern over time.

We further explore the long-term trend by ﬁltering

out deviations that may be skewing the trend of the

normal routine. Firstly, we classify the sleep duration

as abnormal or normal based on the detected devia-

tions. Secondly, we ﬁlter out the dates with abnormal

sleep duration and investigate the normal pattern over

an extended period of time. More speciﬁcally, we cal-

culated the overall linear trend and weekly seasonal-

ity (the relative effect of each day of the week to the

predicted value) of our predictive model, the results

are shown in Figure 5. The normal linear trend (Fig-

ure 5a) shows the sleep duration of hh102 is increas-

ing during this period. The weekly seasonality shows

the person gets the least sleep time on Wednesday.

We have also included the trend calculated using

all of the data, without ﬁltering deviations (Figure 5a).

The 2 trends are visibly different, which demonstrates

how deviating behaviour can signiﬁcantly affect the

observed trend. In the context of analysing long term

sleep routines with the aim of exploring implications

to health, abnormal days become outliers that skew

the results. Instead, we choose to ﬁlter those out and

focus on the “normal” or routine behaviour and how

it evolves through time.

4.4 Relationships between Activities

We take sleep duration as an example to investigate

the relationship with other activities. Through tempo-

ral deviation detection, the sleep duration of each day

can be labelled as abnormal and normal. We combine

all the labelled sleep duration data of 11 individuals

to train a classiﬁcation model. The goal is to classify

sleep duration as normal or abnormal using the other

behaviour markers as features.

Smart CommuniCare 2022 - Special Session on Smart Living Environments to Support Aging-in-Place in Vulnerable Older Adults

906

Figure 3: The dashboard for visualizing the temporal proﬁle of each individual.

Figure 4: The predicted trend of sleep duration for hh102

using data from 15 June to 26 July 2011.

To this end, we ﬁrst train separate predictive mod-

els of sleep duration for each person in the dataset.

The performance of each model is shown in Table 5.

The results show that the error of each predictive

model is less than 1 hour, which we consider a rea-

sonable error range for a sleep duration prediction.

Next, we experiment with different z values be-

tween 1 and 2 to choose the best classiﬁcation method

and investigate how different z values impact the ac-

curacy. We train separate classiﬁcation models based

on the results of deviations calculated by different z

values. A 10-fold cross validation is used to calculate

the performance. We choose accuracy, Auc, recall,

precision and F1 score of the cross-validated results as

(a) The overall trend of sleep duration for hh102.

(b) The weekly seasonality of sleep duration for hh102.

Figure 5: Long term pattern of sleep duration for hh102.

our metrics to evaluate the classiﬁcation model. The

Auc metric can make a reasonable evaluation of a bi-

nary classiﬁcation problem on imbalanced datasets,

i.e. datasets where one class (normal) has a much

Predictive Behavioural Monitoring and Deviation Detection in Activities of Daily Living of Older Adults

907

Table 5: Individual performance of sleep duration models.

Participant MAE RMSE

hh101 3959.30 4863.67

hh102 2802.01 3152.73

hh103 1177.24 1443.22

hh104 2974.59 3491.83

hh105 2013.01 2181.18

hh106 1218.23 1442.67

hh107 1939.70 1956.81

hh108 2671.29 2962.98

hh109 1537.11 2286.03

hh110 559.72 804.56

hh111 1533.05 1784.17

larger population than the other (abnormal).

The results show that the Random Forest Classi-

ﬁer performs best, so we use it as the main model.

The comparison of performance for different z values

is shown in Table 6. It shows that the higher z value

we choose, the better classiﬁcation performance we

get. The higher the performance of the model means

the features are more predictable and the results on

feature importance are more reliable. We get the high-

est Auc when the z value is set to 1.8, which means the

classiﬁcation performs best on this imbalanced case.

Therefore, we investigate the relationships between

sleep duration and other activities for a z value of 1.8.

The calculated (impurity-based) feature impor-

tance of the Random Forest Classiﬁcation model is

shown in Figure 6. These particular results show that

the total duration and the last time of personal hy-

giene activities have a strong inﬂuence on sleep du-

ration. Also, the morning wake up time has a high

correlation with sleep duration, meaning that the time

one wakes up in the morning can signiﬁcantly affect

their sleep duration compared to their usual routine at

night. Surprisingly, nap duration during the day has a

weaker correlation with the nightly sleep duration.

5 CONCLUSIONS

In this paper, a predictive behaviour monitoring ap-

proach is proposed for ADLs coupled with a method-

ology to detect short-term deviations and long-term

trends. Moreover, we investigate the causal factors

of deviations by exploring the relationship between

different daily activities in terms of predictive power.

Experiments are conducted on a sample of 11 individ-

uals’ ADLs data that is publicly available.

Due to the dynamic nature of human behaviour

and the variability between different people, we set

an adjustable deviation threshold for each behaviour

marker such that the model is tailored to individual

Figure 6: Feature importance of the Random Forest Classi-

ﬁer.

human activities. We can also modify the individual

weights of each behaviour marker in order to obtain

an aggregate score that reﬂects the importance of each

deviation, for instance in terms of health outcomes.

Finally, we develop a dashboard which visualizes the

various information of the human behaviour, such as

daily routine, potential deviations, trend of sleep du-

ration, etc. The dashboard provides both the person

involved and their caregivers with key behavioural in-

sights.

The work presented in this paper is indicative of

the useful insights that sensor data on ADLs can pro-

vide towards health monitoring of older adults. We

believe that such insights have the potential to en-

able an unprecedented capacity for self monitoring

and management as well as inform proactive interven-

tions far in advance of any adverse health events.

However, further challenges still exist in dealing

with the inherent noise in the data and the variability

of people’s routines. The models used are naturally

sensitive to noise. We therefore believe that future

improvements in ADL data collection, especially in

the 4 types of noise we identiﬁed (see Section 3.1),

will bring forth signiﬁcantly better and more accurate

insights. Moreover, results need to be further contex-

tualised to the needs, lifestyles, and medical condi-

tions of the individual participants. Further compar-

ison experiments with literature could get more con-

vincing insights for providing proactive interventions

Smart CommuniCare 2022 - Special Session on Smart Living Environments to Support Aging-in-Place in Vulnerable Older Adults

908

Table 6: The performance of classiﬁcation model with different z value.

z value Normal cases Abnormal cases Accuracy Auc Recall Precision F1

1 268 226 0.5478 0.5892 0.5914 0.5815 0.5858

1.3 330 164 0.6812 0.6145 0.8707 0.7067 0.7799

1.5 359 135 0.7304 0.6705 0.9145 0.7565 0.8279

1.8 390 104 0.8058 0.7001 0.9814 0.8094 0.8871

2.0 405 89 0.8348 0.6703 0.9857 0.8395 0.9067

2.3 422 72 0.8580 0.6709 0.9966 0.8581 0.9222

and making decisions. Data collection and analysis

beyond the limited dataset we have explored so far is

likely to improve the quality of our algorithms and

lead to new types of insights particularly for long-

term predictions.

REFERENCES

Akram, U., Gardani, M., Riemann, D., Akram, A., Allen,

S. F., Lazuras, L., and Johann, A. F. (2020). Dys-

functional sleep-related cognition and anxiety mediate

the relationship between multidimensional perfection-

ism and insomnia symptoms. Cognitive Processing,

21(1):141–148.

Albanese, A. M., Bartz-Overman, C., Parikh, MD, T.,

and Thielke, S. M. (2020). Associations Between

Activities of Daily Living Independence and Mental

Health Status Among Medicare Managed Care Pa-

tients. Journal of the American Geriatrics Society,

68(6):1301–1306.

Alberdi Aramendi, A., Weakley, A., Aztiria Goenaga, A.,

Schmitter-Edgecombe, M., and Cook, D. J. (2018).

Automatic assessment of functional health decline in

older adults based on smart home data. Journal of

Biomedical Informatics, 81:119–130.

Aminikhanghahi, S., Wang, T., and Cook, D. J. (2019).

Real-Time Change Point Detection with Application

to Smart Home Time Series Data. IEEE Transactions

on Knowledge and Data Engineering, 31(5):1010–

1023.

Anderson, J. L., Rosen, L. N., Mendelson, W. B., Jacobsen,

F. M., Skwerer, R. G., Joseph-Vanderpool, J. R., Dun-

can, C. C., Wehr, T. A., and Rosenthal, N. E. (1994).

Sleep in fall/winter seasonal affective disorder: Ef-

fects of light and changing seasons. Journal of Psy-

chosomatic Research, 38(4):323–337.

Arifoglu, D. and Bouchachia, A. (2019). Detection of ab-

normal behaviour for dementia sufferers using Con-

volutional Neural Networks. Artiﬁcial Intelligence in

Medicine, 94:88–95.

Bergmeir, C. and Ben

ıtez, J. M. (2012). On the use of cross-

validation for time series predictor evaluation. Infor-

mation Sciences, 191:192–213.

Cohen-Mansﬁeld, J. and Perach, R. (2012). Sleep Duration,

Nap Habits, and Mortality in Older Persons. Sleep,

35(7):1003–1009.

Cook, D. J., Crandall, A. S., Thomas, B. L., and Krish-

nan, N. C. (2013a). CASAS: A Smart Home in a Box.

Computer, 46(7):62–69.

Cook, D. J., Krishnan, N. C., and Rashidi, P. (2013b). Ac-

tivity Discovery and Activity Recognition: A New

Partnership. IEEE Transactions on Cybernetics,

43(3):820–828.

Cook, D. J. and Schmitter-Edgecombe, M. (2021). Fusing

Ambient and Mobile Sensor Features Into a Behav-

iorome for Predicting Clinical Health Scores. IEEE

Access, 9:65033–65043.

Dawadi, P. N., Cook, D. J., and Schmitter-Edgecombe,

M. (2016). Automated Cognitive Health As-

sessment From Smart Home-Based Behavior Data.

IEEE Journal of Biomedical and Health Informatics,

20(4):1188–1194.

Howedi, A., Lotﬁ, A., and Pourabdollah, A. (2020). An

Entropy-Based Approach for Anomaly Detection in

Activities of Daily Living in the Presence of a Visi-

tor. Entropy, 22(8):845.

Ilyas, I. F. and Chu, X. (2019). Data Cleaning. Association

for Computing Machinery, New York, NY, USA.

Kanti Majumdar, K. (2014). Relationship of Activity of

Daily Living with Quality of Life. British Medical

Bulletin, BBB[2][4][2014]:757–764.

Maclin, R. and Opitz, D. (1999). Popular Ensemble Meth-

ods: An Empirical Study. Journal of Artiﬁcial Intelli-

gence Research, 11:169–198.

Palmer, M. G. (2018). Experimenter presence in human be-

havior analytic laboratory studies: Confound it? Be-

havior Analysis: Research and Practice, 19(4):303.

Pazhoumand-Dar, H., Armstrong, L. J., and Tripathy, A. K.

(2020). Detecting deviations from activities of daily

living routines using kinect depth maps and power

consumption data. Journal of Ambient Intelligence

and Humanized Computing, 11(4):1727–1747.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P.,

Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,

Cournapeau, D., Brucher, M., Perrot, M., and Duch-

esnay,

E. (2011). Scikit-learn: Machine Learning

in Python. Journal of Machine Learning Research,

12(85):2825–2830.

Pierleoni, P., Belli, A., Palma, L., Pellegrini, M., Pernini,

L., and Valenti, S. (2015). A High Reliability Wear-

able Device for Elderly Fall Detection. IEEE Sensors

Journal, 15(8):4544–4553.

Piovezan, R. D., Poyares, D., and Tuﬁk, S. (2013). Frailty

and sleep disturbances in the elderly: possible con-

nections and clinical implications. Sleep Science,

6(4):175–179.

Predictive Behavioural Monitoring and Deviation Detection in Activities of Daily Living of Older Adults

909

Sepesy Mau

cec, M. and Donaj, G. (2021). Discovering

Daily Activity Patterns from Sensor Data Sequences

and Activity Sequences. Sensors, 21(20):6920.

Taylor, S. J. and Letham, B. (2017). Forecasting at scale.

Technical Report e3190v2, PeerJ Inc.

University of Edinburgh (2021). Advanced Care Research

Centre. https://edin.care. Accessed on Dec 12, 2021.

Yahaya, S. W., Lotﬁ, A., and Mahmud, M. (2019). A

Consensus Novelty Detection Ensemble Approach for

Anomaly Detection in Activities of Daily Living. Ap-

plied Soft Computing, 83:105613.

Yahaya, S. W., Lotﬁ, A., and Mahmud, M. (2021a). Detect-

ing Anomaly and Its Sources in Activities of Daily

Living. SN Computer Science, 2(1):14.

Yahaya, S. W., Lotﬁ, A., and Mahmud, M. (2021b). To-

wards a data-driven adaptive anomaly detection sys-

tem for human activity. Pattern Recognition Letters,

145:200–207.

Smart CommuniCare 2022 - Special Session on Smart Living Environments to Support Aging-in-Place in Vulnerable Older Adults

910