Fall Prediction Amongst the Elderly Using Data from an Ambient

Assisted Living System

Philip Branch

1 a

, Divya Balasubramanyam Sridharam

, Andre Ferretto

and Tim Carroll

School of Science, Computing and Engineering Technologies, Swinburne University of Technology, Melbourne, Australia

HalleyAssist Pty Ltd, Melbourne, Australia

Keywords:

Assistive Technologies, Home Monitoring and Smart Homes, eHealth Applications, Pattern Recognition and

Machine Learning, Data Mining and Data Analysis.

Abstract:

Falls amongst the elderly are life threatening. Being able to predict falls means steps could be taken to reduce

fall likelihood or severity. In this paper we report on our work using data generated by HalleyAssist, an

advanced Ambient Assisted Living System, to predict falls amongst the elderly. HalleyAssist unobtrusively

monitors older people using sensors to provide services to help them with their day-to-day activities. We

conducted a three-month trial of the HalleyAssist system with six households of older people primarily to

gauge acceptance and utility of the system. During the trial we also asked participants to keep a ’falls diary’

in which they recorded the date, time and location of any falls. After the initial trial we continued monitoring

one of the participants (with her consent) who was susceptible to falls, for an additional seven months. Over

the ten months of the trial she fell 32 times on 28 days. None of the other participants fell during the trial. We

analysed data from the sensors and correlated it with whether she fell later in the day. Using techniques from

machine learning we were able to identify features that enabled a fall to be predicted with 64.9 % accuracy.

1 INTRODUCTION

A consequence of people living longer, smaller fami-

lies, the difﬁculties of ﬁnding enough aged care sup-

port staff, and families that are geographically dis-

persed means that older people who might otherwise

be able to live comfortably in their own home with

only a small amount of assistance are more likely to

be institutionalised if there is no support.

Consequently, there has been interest in how the

Internet of Things (IoT) might be used to provide

support to older people living alone. The Internet of

Things comprises sensors and actuators connected to

the Internet with sensors generating data that is pro-

cessed by cloud or fog computing and then generating

actions for actuators to carry out.

A particularly signiﬁcant development of the In-

ternet of Things are ambient Assisted Living Systems

(AAL). These are unobtrusive systems that that oper-

ate in the background of a home, helping with day to

day actions that may be difﬁcult for an older or dis-

abled person. In an AAL motion sensors can turn on

lights, heat sensors can turn on air heating or cooling,

https://orcid.org/000-0003-2188-1452

medication reminders can be scheduled, security can

be automated and a care giver can be informed if there

is an unusual event requiring attention such as a fall

or an unusual change in behaviour such as the older

person becoming bed-ridden.

We have over the past ﬁve years developed and

commercialised the HalleyAssist System, an ad-

vanced Ambient Assisted Living system developed

by SP Tech Solutions Pty Ltd with contributions from

Swinburne University of Technology. One of the fea-

tures of HalleyAssist that distinguishes it from other

Ambient Assisted Living systems is that it incorpo-

rates Artiﬁcial Intelligence features to identify both

acute and chronic conditions that require assistance

or other form of response from a carer. HalleyAssist

Artiﬁcial Intelligence is based on using sensors de-

ployed as part of the system in order to detect unusual

events (anomalies) when they happen and to act upon

them. Usually the action is to report the anomaly to a

care giver via the HalleyAssist App.

HalleyAssist detects anomalies in two broad ways.

It includes both statistical learning where it detects

what sensor activations are normal for this person,

and rule based anomaly detection where speciﬁc sen-

sor activations are interpreted as having a particular

218

Branch, P., Sridharam, D., Ferretto, A. and Carroll, T.

Fall Prediction Amongst the Elderly Using Data from an Ambient Assisted Living System.

DOI: 10.5220/0011603400003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 5: HEALTHINF, pages 218-223

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

meaning, such as a fall.

For example, the system detect falls based on the

location and activation of motion sensors. Falls are

particularly dangerous for elderly people and rapid

detection of them is important in minimising the dam-

age caused by them. The system is also able to detect

signiﬁcant changes in behaviour. For example, if a

person usually leaves his or her bed fewer than two

times a night but starts leaving it much more than that,

that may be indicative of an acute condition, for ex-

ample a urinary tract infection requiring them to visit

the bathroom multiple times. Of course there may be

any number of other explanations but unusual changes

in behaviour may well be signiﬁcant and require some

form of intervention by a carer.

One of the areas of research we have been explor-

ing is predicting falls based on sensor activations. Are

there patterns of sensor activation, such as might be

generated by broken or restless sleep, that are predic-

tive of falls the following day? Being able to deter-

mine that a fall is more likely than normal could be

particularly useful in preventing them or in making

them less likely. For example if the system suggested

that a fall was quite likely on a particular day, the per-

son might modify their schedule so that they are less

likely to have a fall, or make the consequences of a

fall less damaging.

However, research in this area is very difﬁcult be-

cause there are many obstacles to obtaining any useful

data. Trials of AALs during which falls are likely are

difﬁcult to arrange, have ethical implications and re-

quire a reasonable length of time and number of par-

ticipants. Correlating any falls with sensor activations

requires the participant to recall accurately when the

fall occured. Also, although the consequences of falls

can be devastating, they are, fortunately, quite rare

amongst people living alone.

Consequently, when we ran an ethically approved

major trial of HalleyAssist we took the opportunity

to ask participants to record the date, time and loca-

tion of any falls. The primary purpose of the trial was

to assess the acceptance of the system. We certainly

were not hoping for falls. However, if falls occurred

we wanted to be in a position to analyse the data as-

sociated with them. To that end we asked participants

of the trial to keep a ’falls diary’ where they recorded

when and where any falls happened.

The trial ran for ten months starting 28th February

2020 till 31st December 2020.

During this time, one of the participants, an el-

derly woman who had a history of falls, fell 31 times

during the trial. We have used her records of when

falls occured and records from HalleyAssist to see if

it is possible to see if there is any predictive power

in sensor activations. In this initial trial we explored

the association between a poor night’s sleep and the

likelihood of a fall the following day. Poor sleep has

been associated with falls (Noh et al., 2017). We were

interested in whether sensors indicating a poor night’s

sleep could be used to predict an increased likelihood

of a fall the following day.

Based on the sensor activations we can success-

fully predict days when a fall will occur 61% of the

time and days when a fall will not occur 65% of the

time.

We think these results are interesting but acknowl-

edge that they are hardly conclusive. There are many

caveats associated with these results which we discuss

later in the paper. Nevertheless, they are encouraging

and we plan to continue with further collection and

analysis of data.

The remainder of this paper is structured as fol-

lows. Section 2 discusses related research in this area.

Section 3 describes the trials and their outcomes. Sec-

tion 4 provides an analysis of the sensor and fall data.

Finally, Section 5 is our conclusion where we sum-

marise our work and discuss future research.

2 RELATED RESEARCH

Ambient Assisted Living Systems (AALs) have at-

tracted a great deal of research attention over the past

ten years (Forkan et al., 2019a; Demiris and Hensel,

2008; Forkan et al., 2019b). These systems are an

application of the Internet of Things where data from

sensors is processed by a central hub or a cloud pro-

cess which then issues commands to actuators to con-

trol the living environment in an unobtrusive manner.

HalleyAssist is a AAL system that helps older peo-

ple remain living in their own home or in a minimal

supported accommodation by providing unobtrusive

assistance and monitoring (Forkan et al., 2019b).

As well as supporting day to day activities such

as providing lighting, heating and cooling AALs can

also provide reminders of appointments, notiﬁcation

of the weather as well as reminders to take medi-

cation, eat sufﬁcient food and drink sufﬁcient water

(Demiris and Hensel, 2008).

However, they can also be used to identify events

and trends that a carer should be notiﬁed of. So for

example, long term trends such as isolation, wander-

ing or sleeplessness might be detected and the carer

informed. Such changes might be due to chronic or

acute illnesses such as dementia or infection or they

might report on some particular event requiring im-

mediate response, the most common of which is a fall

(Forkan et al., 2019b).

Fall Prediction Amongst the Elderly Using Data from an Ambient Assisted Living System

219

A fall by an older person can have devastating

consequences (Berg and Cassells, 1992). Reduced re-

action times, weaker muscles and more fragile bones

mean a fall can result in serious injury. It is also

possible that the older person may be unable to get

back up after a fall leading to further serious conse-

quences. Consequently, rooms need to be designed so

as to make falls less likely but if they do happen, make

the consequences less damaging than they might be if

they occurred in another environment (Bianco et al.,

2015).

Ideally it would be best to prevent a fall from hap-

pening in the ﬁrst place (Tinetti, 2003). If it were

known that on a certain day a person was more likely

to fall than on other days then it might be possible to

take preemptive steps to prevent a fall or minimize its

consequences.

AALs have potential in fall detection and possibly

in fall prevention. Camera based systems have been

trialed and provide useful information for researchers

about the nature of a fall, but camera based AALs are

strongly resisted by older people (De Miguel et al.,

2017). Wearables have also been trialed but also face

resistance from older people. There is also the difﬁ-

culty of people, particularly those with dementia for-

getting to wear the device. Sensor based systems

where sensors such as motion based sensors, bed sen-

sors, temperature sensors are located throughout the

living space are much better accepted and have been

trialed with some success (Liu et al., 2014).

There has been very little research into the poten-

tial of AALs to predict falls. Yet there may well be be-

haviourial factors that can be detected by sensors that

are associated with falls. In particular, a poor night’s

sleep has been associated with falls the following day

(Noh et al., 2017). If an AAL can detect that a person

has slept poorly then it may be that the system could

issue a warning to them to take extra care or to inform

their carer.

In section 4 we demonstrate that for the data we

collected during our trials a fall was associated with

a poor night’s sleep as detected by a bed sensor at-

tached to the system. This in turn, can be used to

predict when a person has a heightened risk of falling

the following day.

3 HalleyAssist TRIAL

HalleyAssist Pty Ltd was formed to develop the Hal-

leyAssist product. It is an AAL intended to enable

older people to live independently with with assis-

tance from sensors and actuators to control warm-

ing and cooling, provide medication, food and liquid

reminders, manage security, and monitor the elderly

person for events that might require the attention of a

carer.

HalleyAssist comprises ﬁve modules (Forkan

et al., 2019b):

• Ambient Assisted Living Subsystem

• Learning Module

• Anomaly Detection Module

• Caregiver Module

• Reporting Module

The Ambient Assisted Living Subsystem controls

heating, cooling, lighting, security and other activities

to help the user in their day-to-day lviing. The Learn-

ing Module and the Anomaly Detection Module are

AI based subsystems that learn what is normal for the

person and detects whether or not data observed over

the past reporting interval (which is conﬁgurable) is

anomalous. The Reporting Module and the Caregiver

Module enable a carer to monitor the user in an un-

obtrusive manner. The Caregiver Module is a smart

phone App that displays data generated by the Re-

porting Module.

The system is designed to be ﬂexible in most of its

functions. In particular new sensors and actuators can

be easily added to the system, anomaly detection can

be ﬁne tuned and adapted to detect new situations and

monitoring and reporting periods and sensor thresh-

olds can be adjusted.

A diagram showing the key components and their

interactions is shown in Figure 1.

New sensors can be readily added using a variety

of communications protocols. The system supports

ZigBee, WiFi and cellular communications. All the

sensors are connected to the central hub via WiFi or

ZigBee. The system is able to support large numbers

of sensors but in the installations used in the trials of it

each household has had less than ten sensors attached.

These are usually motion detection, bed activity and

door closing and opening sensors.

A major trial of the system was carried out from

March to December 2020 primarily to determine ac-

ceptance and to ﬁeld test the system for reliability

and effectiveness. The purpose of this paper is not

to discuss those trials, but to discuss the prediction

of falls based on data generated by the system. Nev-

ertheless a brief summary of the project is appropri-

ate. The project received ethical approval from Swin-

burne University of Technology. The system was in-

stalled in six households. These households com-

prised single people living alone with some limited

support from family member carers who did not live

with the person. The participants and carers were sur-

veyed before and after the trial. Feedback from the

HEALTHINF 2023 - 16th International Conference on Health Informatics

220

Figure 1: HalleyAssist Architecture (Forkan et al., 2019b).

participants was excellent. The system was well ac-

cepted and proved very reliable and useful for carers

and the elderly participants.

4 ANALYSIS

4.1 Data Attributes Used to Predict

Falls

As noted in (Noh et al., 2017) falls amongst the el-

derly are often associated with poor sleep. We made

use of sensor activations that indicated that the user

had probably slept poorly. We made use of a sensor on

the bed that detected excessive movement when com-

pared with usual sleep patterns, number of absences

from the bed and total duration of absences from the

bed. The rationale behind choosing these features as a

proxy for a poor night’s sleep was that they were read-

ily collected and seems to capture a prolonged and

interrupted night’s sleep. Data captured between the

hours of midnight and 6 a.m. were used in the training

and testing of the classiﬁer.

We developed two metrics that captured the mag-

nituded of these features and so acted as proxies for a

poor night’s sleep. The ﬁrst was the number of sleep

cycles based on tossing and turning during the night.

The second was the number of absences from the bed.

Both of these were obtained from a bed sensor in-

stalled beneath the mattress.

The sleep cycles metric was based on the number

of episodes of ’tossing and turning’ and their duration

during the night. A period of restful sleep would typ-

ically be broken with a period of tossing and turning.

The duration of the tossing and turning episodes was

averaged over the number of the episodes to give a

metric for sleep disturbance. The bed sensor was able

to detect the absence of the person from the bed.

We used the WEKA system with the above fea-

tures to train a Naive Baye’s classiﬁer using the data

from the falls diary which said whether or not a fall

occurred in the following day (Hall et al., 2009).

4.2 Results

Amongst elderly people living alone falls are, fortu-

nately, rare. Only one participant of the trial experi-

enced falls. However, she fell frequently. Of a total of

308 days of the trial, she fell at least once, occasion-

ally more, on 28 days.

Using the Naive Baye’s classiﬁer trained on the

sleep data and using k-fold testing we were able to

predict whether or not a fall would occur with an Ac-

curacy of 64.9%. We were able to predict days when

a fall occured 60.7% of the time and days when a fall

will not occur 64.9% of the time. We obtained an

overall Accuracy of 64.9%. The overall F-Score was

0.73. The Accuracy is shown in Table 1 while the con-

fusion matrix is in Table 2 and the accuracy by class

is in Table 3.

Table 1: Overall Accuracy.

Correctly Classiﬁed Instances 200 64.9%

Incorrectly Classiﬁed Instances 108 35.1%

Total Number of Instances 308

Table 2: Confusion Matrix for Fall Data.

a b classiﬁed as

183 97 a = No fall

11 17 b = Fall

The confusion matrix tells us that the system was

able to predict a fall 17 of the 28 days when a fall did

not occur and 183 of the 210 days when one did occur.

The accuracy by class table gives us the True Pos-

itive (TP) and False Positive (FP) probabilities. Other

frequently used metrics are also included which also

use False Negative (FN) and True Negative (TN).

Fall Prediction Amongst the Elderly Using Data from an Ambient Assisted Living System

221

Table 3: Accuracy by Class.

TP FP Precision Recall F-Measure

Days without a fall 0.65 0.39 0.65 0.65 0.77

Days with a fall 0.61 0.35 0.15 0.61 0.24

Weighted Average 0.65 0.39 0.87 0.65 0.73

These are Accuracy, Precision, Recall and F-Measure.

These are deﬁned below:

Accuracy =

T P + T N

T P + T N + FP + FN

. (1)

Precision =

T P

T P + FP

. (2)

Recall =

T P

T P + FN

. (3)

F − measure = 2.

precision.recall

precision + recall

. (4)

Accuracy is the simplest measure. It tells us how

many times the system made the correct prediction.

However for datasets such as this where there are

many more days without a fall than with a fall it can

be misleading. Consequently, other metrics need to be

considered. Precision is a measure of consistency and

Recall measures the number of True Positives com-

pared with the number of Predicted Positives. Recall

is sometimes referred to as the True Positive Rate.

The F-measure is the harmonic mean of Accuracy and

Precision.

For our data all metrics apart from one Precision

measure are 0.6 or above. However, the Precision for

Days with a fall is 0.15, caused by the large number of

False Positives which is a consequence of there being

many more days without a fall than days with a fall.

4.3 Discussion

Using only one general concept, that of a poor night’s

sleep, as a predictor of falls we can correctly deter-

mine whether or not a fall is likely to occur approx-

imately 65 % of the time. Of course there are many

limitations to this study. Only one person fell, the

number of days on which she fell was only 28 and

the only behaviour used was an indication of a poor

night’s sleep. Nevertheless, given how scarce be-

havioural data before a fall is, the results are very en-

couraging. It may be that other behavioural data may

also be indicative of a fall. It is known that multiple

factors affect the probability of a fall. For example,

low blood sugar and dehydration have also been noted

as precursors to a fall. Perhaps sensor data that can act

as proxy for these biological markers may also help

improve the accuracy of the fall prediction. For exam-

ple, a sensor on the kitchen tap that is activated less

than usual may indicate potential dehydration. Per-

haps the refrigerator not being opened may be an in-

dicator of low blood sugar. We are not claiming that

this is the case but that if other factors are associated

with falls, sensors installed at strategic locations in

the home may be able to be used as proxies for these

markers and used to predict whether a fall is likely the

next day.

5 CONCLUSION

We have taken a simple to state behaviour (a poor

night’s sleep), used sensor data that acts as a proxy

to identify that behaviour and demonstrated that using

that it is possible to predict an increased likelihood of

a fall the following day. While the results are inter-

esting and potentially signiﬁcant, it is important not

to overstate the signiﬁcance of the result. Only one

person fell and she fell on only 28 days. Nevertheless

the results are promising. As far as we can determine

there has been no comparable collection of such data

before this work.

Perhaps more sophisticated behaviours such as de-

hydration, hunger and stress which are implicated in

falls can also be detected using sensors that act as

proxies for these biological markers. Perhaps tap,

cupboard and refrigerator door opening sensors can

be used as proxies for dehydration and low blood

sugar which have also been implicated in falls.

Our interest in this research is in developing tech-

niques for predicting an increased likelihood of a

fall. However, another perspective is that it gives re-

searchers into falls amongst the elderly data to explore

what the factors are that contribute to a fall. It may be

useful for researchers into falls to know what are and

are not factors that contribute to a fall.

This is very much a report of work in progress.

We believe there is evidence that systems such as Hal-

leyAssist can not only help people live more indepen-

dent lives than might otherwise be possible but also

enable prediction of health threats of which falling is

just one example. Are there sensor activations that

can be used as proxies that suggest increased likeli-

hood of illnesses? We intend continuing research into

HEALTHINF 2023 - 16th International Conference on Health Informatics

222

this area in the hope of making life for the elderly

more independent and safer.

ACKNOWLEDGMENT

Funding and support from the Australian Govern-

ment, National Ageing Research Institutue, Baptcare

and HalleyAssist is gratefully acknowledged in the

preparation of this paper.

REFERENCES

Berg, R. L. and Cassells, J. S. (1992). The second ﬁfty

years: Promoting health and preventing disability.

Bianco, M. L., Pedell, S., Renda, G., and Kapoor, A. (2015).

A person-centered approach for fall prevention: Em-

bodying the goals of older adults in personas. Pro-

ceedings of IASDR.

De Miguel, K., Brunete, A., Hernando, M., and Gambao, E.

(2017). Home camera-based fall detection system for

the elderly. Sensors, 17(12):2864.

Demiris, G. and Hensel, B. K. (2008). Technologies

for an aging society: a systematic review of “smart

home” applications. Yearbook of medical informatics,

17(01):33–40.

Forkan, A. R. M., Branch, P., Jayaraman, P. P., and Ferretto,

A. (2019a). Halleyassist: A personalised internet of

things technology to assist the elderly in daily living.

Forkan, A. R. M., Branch, P., Jayaraman, P. P., and Ferretto,

A. (2019b). An internet-of-things solution to assist in-

dependent living and social connectedness in elderly.

ACM Transactions on Social Computing, 2(4):1–24.

Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann,

P., and Witten, I. H. (2009). The weka data min-

ing software: An update. SIGKDD Explor. Newsl.,

11(1):10–18.

Liu, L., Popescu, M., Skubic, M., and Rantz, M. (2014). An

automatic fall detection framework using data fusion

of doppler radar and motion sensor network. In 2014

36th Annual International Conference of the IEEE

Engineering in Medicine and Biology Society, pages

5940–5943. IEEE.

Noh, J.-W., Kim, K.-B., Lee, J. H., Lee, Y., Lee, B.-H., and

Kwon, Y. D. (2017). Association between sleep du-

ration and injury from falling among older adults: a

cross-sectional analysis of korean community health

survey data. Yonsei medical journal, 58(6):1222–

1228.

Tinetti, M. E. (2003). Preventing falls in elderly persons.

New England journal of medicine, 348(1):42–49.

Fall Prediction Amongst the Elderly Using Data from an Ambient Assisted Living System

223