WRIST-WORN FALL DETECTION DEVICE

Development and Preliminary Evaluation

Mattia Bertschi and Leopoldo Rossini

Swiss Center for Electronics and Microtechnology, Jaquet-Droz 1, Neuchâtel, Switzerland

Keywords: Elderly, Monitoring, Portable device, Wearable device, Automatic fall detection, Accelerometers.

Abstract: Falls are the most important cause of accidents for elderly people and often result in serious physical and

psychological consequences. The rapid growth of the elderly population increases the magnitude of the

problem as well as the generated costs. In order to take care of old people living by themselves or in care

centres and to reduce the consequences of a fall, various technological solutions have been studied, however

none led to a commercial product fulfilling user requirements. In this work we present an automatic fall

detector in the form of a wrist watch which could lead to better life conditions for the elderly. Our device

implements functionalities such as wireless communication, automatic fall detection, manual alarm

triggering, data storage, and simple user interface. Even though the wrist is probably the most difficult

measurement location on the body to discern a fall event, the proposed detection algorithm shows

encouraging results (90% sensitivity, 97% specificity) with the signals of our database.

1 INTRODUCTION

Falls are the most widespread domestic accidents

among the elderly. Their consequences often give

rise to impairments to the health and lifestyle of the

victims (Pérolle et al., 2004). In many cases,

physical after-effects and other injuries are direct

consequences of these accidents and result in

significant medical costs.

Furthermore, it frequently happens that elderly

people who have previously experienced a fall fear a

new fall and sink gradually into inactivity and social

isolation. The reduction in their mobility leads

progressively to an increase in the risk of a fall

(Doughty et al., 2000). Hence, given the growing

part of the elderly people in our modern societies,

the socio-economic impact that this self-imposed

isolation may have should not be neglected.

The most widespread solution for limiting the

apprehension of a fall is provided by social alarms

consisting of portable devices. These are generally

equipped with an alarm triggering button and

endowed with telecommunication means suitable for

alerting the care centre. Nevertheless, because of the

fall, the person may not be able to actuate the button

and to trigger the alarm (unconsciousness, state of

shock, broken arm, etc.).

To alleviate this drawback, autonomous fall

detectors have been developed and are capable of

triggering an alarm automatically without any

intervention of the victim and transferring this

information to a remote site (Doughty et al., 2000).

These autonomous detectors operate essentially on

two principles. The detector is either sensitive to the

person's appearance or impact on its environment

and is based on video (CCD or IR camera) or

vibratory type sensors (acoustic or piezoelectric

layers on the ground) placed in the usual

surroundings of the subject. The benefit of these

devices is that they do not have to be worn. Instead,

they are fixed and integrated in a given spot and

cannot be moved easily when the person changes

location. Moreover, in the case of a video sensor the

person will have the impression of being supervised

and feel inconvenienced. The major drawback of

acoustic based sensors is that they are surface

dependent, while those based on vibration are fragile

and expensive. The detector can also be worn by the

person and thus detect a fall directly as soon as it

occurs, triggering an immediate alarm. In this case,

the information provided by inclinometers,

gyroscopes or accelerometers is exploited. These

devices are generally compact, inexpensive, fairly

non-obtrusive, easy to use, and can be worn at

various body locations.

368

Bertschi M. and Rossini L. (2009).

WRIST-WORN FALL DETECTION DEVICE - Development and Preliminary Evaluation.

In Proceedings of the International Conference on Biomedical Electronics and Devices, pages 368-371

DOI: 10.5220/0001540903680371

 SciTePress

Devices worn close to the centre of gravity

(Depeursinge et al., 2001) are the most reliable ones,

but also the least convenient to wear on a daily basis,

in particular while performing common daily

activities. A device having the form of a wristwatch

would be well tolerated in all situations, despite the

challenge to detect a fall due to changes of position

and accelerations that the wrist experiences during

everyday actions (Degen et al., 2003). Furthermore,

the inclination measurement of the forearm cannot

give reliable information about the person's position.

Generally, fall detectors placed on the wrist give rise

to a large number of false alerts, and this would

generate significant and unnecessary costs.

Our goal is to develop a small, comfortable, and

user friendly device, as well as an automatic fall

detection algorithm that will help elderly people to

handle this problem.

2 METHOD AND MATERIALS

The fall detector that we have developed is a device

capable of automatically detecting various body falls

and sending an alarm to a remote terminal. The user

can manually generate the alarm signal in case of

necessity or, inversely, he can cancel an

automatically generated alarm in case of a false fall

detection.

2.1 Fall Detection System

The detection system is integrated in the case of a

wrist watch. A picture of our prototype is depicted in

Figure 1.

Figure 1: Wrist fall detection system.

The core of the wrist-located device consists of a

microprocessor (MSP430) and two MEMS sensors

arranged perpendicularly to allow the measurement

of acceleration along three axes with a range equal

to ±18g (ADXL321). The user interface is simple

and comprises a small LCD screen (Nokia 3310), a

vibrator which advises the user that a fall has been

detected and an alarm will soon be sent, and a push

button on the front panel to manually trigger the

alarm signal. Data can be transmitted from fixed

and/or mobile devices over short distances utilizing

a short-range communication technology (Bluetooth

protocol).

For experimental purposes, the acceleration signals

can also be stored in a flash memory card. The serial

RS232 port as well as three additional buttons are

also available for debugging and test.

The device is powered by a 3.7 volts rechargeable

Lithium-Cobalt-Polymer battery. The battery life of

the device varies from about 15 days to one month,

depending on the sampling frequency and the details

of the implemented data handling and storage

functionalities.

2.2 Fall Detection Algorithm

The three-axes acceleration signals are recorded and

stored in the flash memory, with a sampling

frequency of 910 Hz and 12-bit resolution. Although

the algorithm has not been implemented yet in real-

time, it has been tested offline using Matlab.

Initially, the device is slowly rotated a few times

in order to project the gravity vector on the three

axes in various configurations. The resulting

acceleration signals roughly define the surface of a

sphere in a three dimensional space and are used to

calibrate the sensors. We define the time-dependent

acceleration vector

(

)

zyx

aaa

whose entries

are defined as

iii

caa

−

, for zyxi ,,= , where

represents the acceleration in the i-direction while

is the i-coordinate of the centre of the sphere

corresponding to a zero acceleration. Therefore, the

equation of the sphere is defined using Cartesian

coordinates as:

=− rA

(1)

where the radius r corresponds to the gravity

acceleration and defines the sensor gain while the

centre (c

) defines the accelerometer offset.

Notice that one can use the Gauss-Newton method

to estimate (c

,r) by minimizing the left-hand

part of equation (1) and use those parameters to

calibrate the acceleration signals (see Figure 2).

The three-axes acceleration experienced at the

wrist of the person wearing the device can not be

directly linked to the specific posture of the body.

However, it has been observed that the distribution

WRIST-WORN FALL DETECTION DEVICE - Development and Preliminary Evaluation

369

of the acceleration norm

A experienced at the

wrist over a time window ΔT provides a particularly

reliable signature, allowing the identification of a

given event happening to the person wearing the

device.

-6

-3

(g)

-6

-3

(g)

0 0.2 0.4 0.6 0.8 1.0 1.2 1.4

-6

-3

(g)

Time (s)

Figure 2: Typical recording of a fall event. Each signal

corresponds to the acceleration in the x, y, or z direction.

We define S

as the signature at instant t that

represents the distribution of the acceleration norm

A over a time window ΔT. The signature S

can

be extracted via a simple amplitude threshold and

compared to a reference signature S

ref

. To obtain S

ref

a subset of all the fall signatures recorded in the

database is used to obtain statistics in the form of

means and variances. Specifically, a particular value

of each signature in the subset is identified so as to

carry out an alignment, for example on the

maximum. This particular value is used to align the

signatures, so as to obtain a group of aligned fall

signatures. On the basis of the various aligned

signatures, the reference signature S

ref

is obtained by

averaging the values of the aligned signatures.

In order to construct a similarity measure d

between S

and S

ref

, the squared value of the

acceleration due to gravity is subtracted to remove

its influence. For this application, d is calculated

according to the Mahalanobis distance definition.

The variances of the aligned signatures are placed on

the diagonal of a matrix θ that is used together with

ref

to estimate if the instantaneous signature S

arises from a fall:

)(

),(

ref

−

−−

(2)

The influence of the least reliable samples is reduced

due to the use of the variances. One can see that the

value of d is in the interval [0,1] and therefore can

be considered to be a fall probability. A threshold

was used to discriminate between the two classes:

“fall” and “no fall”.

2.3 Experimental Setup

A preliminary validation has been performed with

three adult healthy adult volunteers to assess the

reliability of our system for fall detection tasks. The

study consists in recording the three-axial

accelerometer signals (wrist-located) during

controlled fall events and daily life activities. Each

case was repeated three times and stored in a

separate file. In our database we have documented

180 situations by recording acceleration signals and

video sequences. The latter are particularly useful to

understand the dynamics of the wrist movement.

To evaluate the sensitivity of the fall detection a

first subset consisting of six kinds of falls is created.

The kind of falls selected for this subset are: front,

back, left, right, falling backwards while sitting

down, and falling forward while standing up. Each

fall was intentional and a mattress was used to

protect volunteers from injuries.

A second subset was used to estimate the

specificity of the fall detector. It includes 14

common situations representing challenges for the

detection algorithm (slow/fast walking, walking

upstairs/downstairs, hit on a table with the fist,

sitting down, lying down, applauding…).

3 RESULTS

To generate the reference signature of a fall event

we have taken 8 recordings representing fall events

randomly selected from the 54 falls of the database

and we have computed the mean and the standard

deviation for each event. Figure 3 shows a typical

reference signature that we obtain with the signals of

our database. The mean and standard deviation

values of the reference signature are then used to

compute a normalized similarity measurement for

each recording of the database and a binary classifier

is used to separate falls from other events.

BIODEVICES 2009 - International Conference on Biomedical Electronics and Devices

370

0 0.5 1 1.5

Time

(

)

Acceleration norm (g)

0 0.5 1 1.5

Time

(

)

Acceleration norm (g)

Figure 3: Reference signature in terms of mean and

standard deviation.

A binary classifier system is generally assessed

in terms of sensitivity and specificity. A useful

graphical tool to represent the sensitivity versus (1-

specificity) for a binary classifier system as its

discrimination threshold is varied is the receiving

operating characteristic (ROC). Figure 4 shows the

ROC curve obtained with the signals of the database.

0 2 4 6 8 10

100

1 - Specificity (%)

Sensitivity (%)

0 2 4 6 8 10

100

1 - Specificity (%)

Sensitivity (%)

Figure 4: The receiver operating characteristic shows the

relationship between the sensitivity and (1-specificity).

In order to become independent from the

recordings used to generate the reference signal, we

repeated the same procedure ten times: random

selection of 8 fall events, estimation of the mean and

the standard deviation, classification of the 172

recordings, and estimation of the ROC curve. The

circles in the plot of Figure 4 are thus a mean value

of the ten simulated results.

One can see that there is a tradeoff between

sensitivity and specificity: the larger the sensitivity,

the smaller the specificity. We can decide to favour

one characteristic over the other, but a good balance

is normally to privilege the upper-left corner of the

ROC curve. In the present case, this particular point

allows us to detect about 90% of the falls while

keeping the false alarm rate below 3%.

4 CONCLUSIONS

We have developed a small, light, and comfortable

fall detector device which is worn at the wrist like an

ordinary watch, removing the social stigma of

wearing a medical device. The real advantage of

fixing the fall detector at the wrist is the possibility

of wearing the device at night, when falls can also

occur. The device is thus easy to wear continuously

without any specific constraints. The major

drawback is the signal processing challenge of

estimating a fall from wrist acceleration data, due to

the strong accelerations experienced by the forearm

during daily-life activities.

The proposed algorithm to detect falls is based

on accelerometric signals and has the advantage to

be simple and robust showing promising results with

our present database. Results demonstrate high

sensitivity (90%) as well as high specificity (97%)

for the detection of intentional falls.

REFERENCES

Pérolle, G., Sánchez, D., Abarrategui, M. I., Eizmendi, G.,

Buiza, C., Etxeberria, I., Yanguas, J. J., 2004. Fall

Detection: Project of an Improved Solution. In

TELECARE 2004, 1st International Workshop on

Tele-Care and Collaborative Virtual Communities in

Elderly Care.

Doughty, K., Lewis, R., McIntosh, A., 2000. The Design

of a Practical and Reliable Fall Detector for

Community and Institutional Telecare. In Journal of

Telemedicine and Telecare.

Doughty, K., 2000. Fall Prevention and Management

Strategies Based on Intelligent Detection, Monitoring

and Assessment. In New Technologies in Medicine for

the Elderly.

Depeursinge, Y., Krauss, J., El-Khoury, M., 2001. Device

for monitoring the activity of a person and/or detecting

a fall, inparticulaer with view to providing help in the

event on an incident hazardous to life or limb. In

Patent US 6,201,476 B1.

Degen, T., Jaeckel, H., Rufer, M., Wyss, S., 2003.

SPEEDY: A Fall Detector in a Wrist Watch. In

IEEE2003, 7

International Symposium on Wearable

Computers.

WRIST-WORN FALL DETECTION DEVICE - Development and Preliminary Evaluation

371