Comparing Machine Learning Approaches for Fall Risk Assessment

Joana Silva

, João Madureira

, Cláudia Tonelo

, Daniela Baltazar

, Catarina Silva

, Anabela Martins

Carlos Alcobia

and Inês Sousa

Fraunhofer Portugal AICOS, Porto, Portugal

Sensing Future Technologies, Coimbra, Portugal

ESTeSC Coimbra Health School, Physiotherapy Department, Coimbra, Portugal

Keywords: Fall Risk Assessment, Inertial Sensors, Pressure Platform, Timed-up and Go Test, Sit-to-Stand, 4-Stage

Test, Machine Learning, Classification, Regression.

Abstract: Traditional fall risk assessment tests are based on timing certain physical tasks, such as the timed up and go

test, counting the number of repetitions in a certain time-frame, as the 30-second sit-to-stand or observation

such as the 4-stage balance test. A systematic comparison of multifactorial assessment tools and their

instrumentation for fall risk classification based on machine learning approaches were studied for a

population of 296 community-dwelling older persons aged above 50 years old. Using features from inertial

sensors and a pressure platform by opposition to using solely the tests scores and personal metrics increased

the F-Score of Naïve Bayes classifier from 72.85% to 92.61%. Functional abilities revealed higher

association with fall level than personal conditions such as gender, age and health conditions.

1 INTRODUCTION

Fall risk assessment methods have been studied

aiming to estimate the risk of falling in order to

identify those at higher risk and timely apply the

appropriate actions to prevent falls. This kind of

assessment can take the form of questionnaires,

simple screenings or more comprehensive

multidimensional fall risk assessments.

Falls are described as a complex phenomenon

caused by the interaction of multiple risk factors. To

assess the risk of falling, it is necessary to identify

the factors that increase an older person’s risk of

falling. Intensive research has been conducted in

order to identify specific risk factors (Ambrose,

Paul, and Hausdorff 2013; Rubenstein 2006; Oliver

et al. 2004), which can increase the likelihood of a

fall occurrence. The idea behind these studies is to

develop preventive strategies based on the identified

risk factors.

According to Rubenstein et al. (Rubenstein and

Josephson 2002) the most common underlying

causes and risk factors for falls include muscle

weakness, gait and balance problems, visual

impairment, cognitive impairment, depression,

functional decline, and particular medications,

especially in the presence of environmental hazards.

Similarly to Physiological Profile Assessment

(PPA) (Lord, Menz, and Tiedemann 2003), we

intend to evaluate the functional ability of subjects

and potential impairments that may be related with

an increased risk of falling, irrespective of the

existence of an underlying disease causing them.

Traditional functional assessment tests are based

on timing certain physical tasks, such as the timed

up and go (TUG) test (Beauchet et al. 2011),

counting the number of repetitions in a certain time-

frame, as the 30-second sit-to-stand (STS) (Jones,

Rikli, and Beam 1999) or observation, such as

balance tests (Agrawa et al. 2011), in particular the

4-stage balance test (4-Stage) (Rossiter-Fornoff et al.

1995; Thomas et al. 2014), or the Tinetti

Performance Oriented Mobility Assessment

(POMA) (Tinetti 1986). A systematic review of

multifactorial and functional mobility assessment

tools for fall risk (Scott et al. 2007) compares

several studies for community settings.

In this paper, the three fall risk assessment

functional tests, TUG, STS and 4-Stage, that have

been also used in the follow-up of the participants of

the Otago Falls Prevention Program, were

instrumented with wearable inertial sensors and a

Silva J., Madureira J., Tonelo C., Baltazar D., Silva C., Martins A., Alcobia C. and Sousa I.

Comparing Machine Learning Approaches for Fall Risk Assessment.

DOI: 10.5220/0006227802230230

In Proceedings of the 10th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2017), pages 223-230

ISBN: 978-989-758-212-7

223

pressure platform for the extraction of several

metrics to perform a comparison with the functional

tests’ scores for the differentiation of fall risk

groups. Machine learning approaches were studied

using a fall level as the classification output.

2 METHODS

2.1 Subjects

A total of 296 subjects voluntarily participated in the

study. Informed consents were obtained from all

participants who responded to personal information,

health, previous falls inquiries and completed the

three instrumented assessment tests: TUG, STS and

4-stage. The data collection took place in different

environments, mostly at community (76.0%), at day-

care centres (15.9%), and at nursing homes (8.1%).

Demographic and anthropometric information

was annotated for all the subjects along with health

related information from two questionnaires: health

conditions and medication intake. Fall related

information was inquired using a history of falls

questionnaire.

The mean age of the sample was 70.2 years (93

persons with age below 65 years), the majority of

the subjects were women (68.2%), 25.0% lived

alone, 51.0% only have primary education and

11.5% use an assistive device. Diabetes was the

most prevalent health condition (15.5%), followed

by osteoarthritis (14.2%) and osteoporosis (10.8%).

Urinary incontinence was reported by 22.3%

(answering the question: do you leak urine when you

cough, laugh, sneeze or lift an object?); fear of

falling was reported by 47.0% (answering the

question: are you afraid of falling?); 57.4% of the

persons referred to intake 4 or more different

medicines per day (mean was 4.52 medicines).

During the previous year 30.7% of the persons

have fallen (18.9% outdoors) and 8.1% underwent to

the emergency service (hospital). The wrist/hand

fracture was the most common injury (2.4%) among

these persons.

2.2 Screening Protocol

This section describes the fall risk assessment tests

applied in this study:

Timed Up and Go Test (TUG) fast pace: the

person is asked to start seated on a chair and when

test starts, the person should stand up, walk straight

for 3 meters, as fast as the person can, turn around,

walk back to the chair and sit down (Beauchet et al.

2011). Test score corresponds to the time needed to

perform TUG test (TUG duration). A threshold of

10s has been found to be associated with falls

occurrence in a 12 months follow up period for

community-dwelling older adults (Rose, Debra J,

Jones, Jessie C, and Lucchese, Nicole 2002).

30 Seconds Sit-to-stand Test (STS): the person is

instructed to sit on a chair and repeatedly stand up

and sit down as many times as possible over 30

seconds (Jones, Rikli, and Beam 1999). The person

must be seated in the middle of the chair, feet should

approximately width apart and placed on the floor,

and arms crossed by the wrists placed against the

chest. Final score of this test is the number of times

the person completes a cycle of sit-to-stand and

stand-to-sit (number of STS cycles). While normative

levels are dependent on age and sex (Rikli and Jones

2010), a score of less than 15 transitions in the 30

seconds test duration has been used to identify

“fallers” in a group of elderlies (Cho et al. 2012).

4 Stage Balance Test “modified”: the person is

instructed to progressively maintain four foot

positions for 10 seconds each, without moving

his/her feet or needing support. The positions are:

side by side stance, semi-tandem stance, tandem

stance and unipedal stance (Rossiter-Fornoff et al.

1995; Thomas et al. 2014). For each position the

subjects were instructed to stand quietly without

shoes on the pressure platform, with their arms along

the body. In this study, except for the one leg stand

position, all positions must be performed with eyes

open and then closed. The final score of this test is

the number of positions a person can hold for 10

seconds without losing balance (number of 4-stage

exercises). The inability to complete the tandem

stance position has been associated with higher risk

of falling (Murphy et al. 2003).

The tests were applied by trained health

professionals. Prior to the execution of tests, the test

procedure was explained to each person and it was

demonstrated how the test should be performed.

Auditory cues were also used to instruct the person

during the execution of the tests. Only persons who

performed the three functional tests (TUG, STS and

4-stage) were included in this study.

2.3 Instrumentation

The participants were instrumented with one

wearable inertial sensor during the execution of

TUG and 30-seconds sit-to-stand tests. The 4-stage

balance test was performed on a pressure platform,

as can be seen in Figure 1.

The wearable sensor was developed and

BIOSIGNALS 2017 - 10th International Conference on Bio-inspired Systems and Signal Processing

224

assembled at Fraunhofer AICOS and was placed at

the lower back. Inertial data was collected using the

built-in 3-axial accelerometer and 3-axis gyroscope,

both sampled at 50 Hz. Raw data from the inertial

sensors were acquired for all the tests in m/s

The pressure distribution data was measured with

PhysioSensing platform (Sensing Future

Technologies, Lda) running at frequency of 50Hz. It

contains 1600 pressure sensors of size 10mm by

10mm with maximum value of 100N/sensor.

Voltage data is converted with an 8-bit A/D

converter and is transmitted via USB (Universal

Serial Bus). In this way it is possible to receive raw

data of each pressure sensor as well as the raw

center of pressure coordinates (CoP), in cm. In order

to obtain more precision in CoP displacements, an

algorithm was employed to obtain CoP positions in

mm, using the matrix of pressure sensors (Hsi 2016).

Figure 1: Example of a test set-up, with the pressure

platform in the floor and an illustration of the inertial

sensor placement of at the lower back, since it is covered

by the clothes.

2.4 Inertial Sensors Data Analysis

The accelerometer and gyroscope signals were

synchronized and used to segment the TUG test into

its several components (stand up, walk forward, turn

around, walk back to the chair and sit down) as

previously described in (Silva and Sousa 2016) and

to identify the stand and sit phases of the STS test.

Identification of the STS transition points was made

analysing the y-axis of the gyroscope signal. After

filtering the signal with a moving average filter of 20

samples window size, zero crossings were identified

(Guimaraes, Ribeiro, and Rosado 2013). In order to

remove outliers, a minimum of 20 samples were

used as difference between consecutive transition

points. Since the score is given by the total number

of complete cycles, it was considered one cycle

between two transitions points, one sit-to-stand and

one stand-to-sit. The number of cycles is therefore

half the number of transitions points identified, as

illustrated in Figure 2.

For each one of the TUG segments and for the

whole STS test, statistical and frequency domain

features were extracted from the magnitude of the

accelerometer signal. The list of features has been

reported in (Silva and Sousa 2016) and corresponds

to: mean, median, maximum, minimum, signal

height, standard deviation, median deviation, root

mean square, inter quartile range, number of times

the magnitude signal crosses the mean value,

energy, entropy, skewness, kurtosis, average of

minima, average of maxima, average signal height,

fundamental harmonic of Fast Fourier Transform

(FFT) spectrum and fundamental amplitude.

Additional metrics for each test were calculated

from the inertial data: for the TUG test, the duration

of the stand segment (duration of the first segment)

and the number of steps (calculated with a step

counter algorithm reported by (Aguiar et al. 2014))

taken during the test; for the STS test, the number of

STS cycles and the STS power (Zhang et al. 2014).

Figure 2: Axis x (red), y (green), z (blue) and magnitude

signals (black) of the accelerometer and gyroscope signals

for STS test with identification of transition points with

blue vertical lines. The interval between two consecutive

lines is considered as one STS cycle. Figures are from a

low risk person.

2.5 Pressure Platform Data Analysis

For each posture of the 4-stage balance test

executed, the pressure values on each sensor of the

Comparing Machine Learning Approaches for Fall Risk Assessment

225

pressure platform were recorded. The centre of

pressure (CoP) coordinates were then obtained and

several parameters, which are typically used in

postural sway and fall risk assessment (Bigelow and

Berme 2011; Guimaraes, Ribeiro, and Rosado 2013;

Raymakers, Samson, and Verhaar 2005) were

calculated.

For all the medio-lateral (ML) and antero-

posterior (AP) CoP position coordinates obtained

during each posture execution, the mean (mean AP

CoP positions, ML mean CoP positions), standard

deviation (std AP CoP positions, std ML CoP

positions), root mean square (rms AP CoP positions,

rms ML CoP positions), maximum (max AP CoP

positions, max ML CoP positions) and minimum

(min AP CoP positions, min ML CoP positions)

were calculated.

The displacement of CoP in each direction per

time unit gave rise to the mean velocity of CoP

displacement (vm CoP position AP, vm CoP position

ML) metrics.

Another metric extracted was the area of a

confidence ellipse containing 95% of the CoP

coordinates projected in a 2D plan (Ellipse area).

Figure 3 shows a comparison of CoP displacements

in ML and AP directions for two persons with

different fall risk levels during the semi-tandem

stance with eyes closed. For a low fall risk person

Figure 3: CoP displacements in ML and AP directions and

95% confidence ellipse area (red line) during semi-tandem

stance with eyes closed of 4-stage test. Top figure is from

a low risk person and botton figure is from a high risk

person, showing more outliers in ML and AP directions.

(top figure) the displacement is concentrated around

the centre, however for a high fall risk person, more

outliers in ML and specially in AP direction are

identified, reflecting unbalance situations.

Sway can be defined, in this scope, as the

amplitude or absolute distance of CoP oscillations.

The sum of all the distances accumulated during the

execution of each posture is computed resulting in

the CoP path length (total Sway distance). The

standard deviation of sway distances (std Sway) and

the maximum and minimum amplitude of CoP

oscillations (maxSway and minSway) were also

included as pressure platform metrics.

2.6 Machine Learning Methods

Classification and regression methods were tested to

differentiate between high and low fall risk groups

using metrics extracted from inertial sensors and

pressure platform. Rapid Miner Toolkit was used for

the train and test processes. Ten-fold cross

validation with random split was used for all the

processes. In order to define a metric to divide the

groups, a fall level was determined based on the

history of falls questionnaire and usage of walking

aid, as presented in Figure 4, since these two factors

have evidence to be more related with risk of falling.

The fall level is merely an indication if the person

shows more or less probability of falling, since the

falls occurrence in a 12 months follow up period was

not possible to measure. The dimension of the

population is 296 subjects. The low risk group

represents 83% of the dataset and is composed by

245 subjects (35% within 50-65y.o. and 65% above

65y.o.). The high risk group represents 17% of the

dataset and contains the remaining 51 subjects (16%

within 50-65y.o. and 84% above 65y.o.). This

distribution is in agreement with the falls incidence

in the elderly population, which is less than 30%

(Bergen, Stevens, and Burns 2016).

Two approaches were compared: first only

personal metrics and tests scores were used to

construct the feature vector, and then this vector was

replaced with features extracted from inertial sensors

and pressure platform. The objective was to study

the added value of the sensors features to

differentiate between fall risk groups.

BIOSIGNALS 2017 - 10th International Conference on Bio-inspired Systems and Signal Processing

226

Figure 4: Fall level definition based on history of falls and

usage of walking aid.

The performance of several classification and

regression methods was compared based on

accuracy, precision, recall and F-Score. It was

considered low risk as the positive class and high

risk as the negative class. TP states for true positive,

FP for false positive, TN for true negative and FN

for false negative. The performance metrics are

calculated as follows:

Precision (P) = TP / (TP+FP)

(1)

Recall (R) = TP / (TP+FN)

(2)

Accuracy = (TP+TN) / (TP+TN+FP+FN)

(3)

F-Score = (2P x R) / (P+R)

(4)

3 RESULTS & DISCUSSION

3.1 Statistical Analysis

A statistical analysis has been conducted for the

variables: gender, age, body mass index (BMI),

number of medicines, number of health conditions,

fear of falling, TUG score, STS score and 4-stage

score. Cut-off values that have been used in previous

studies referred in the introduction section of this

paper to distinguish high and low fall risk levels

were applied to each of these variables. The Fisher’s

exact test was applied with the null hypothesis that

there are no non-random associations between the

two categorical variables: fall level and each of the

variables considered. The Fisher’s exact test p-value

and odds ratio (OR) are reported in Table 1 and were

calculated with Matlab function fishertest.

Table 1: Odds Ratio and Fisher's exact test p-value for

personal metrics and tests scores with the fall level.

Variable Odds Ratio p-value

Feminine Gender 1.04 1.00

Age > 65 2.86 0.01

BMI < 13.7 or BMI > 29.7 1.58 0.18

More than 4 Medicines 1.96 0.05

More than 2 Health

Conditions

1.56 0.38

Has Fear of Fall 3.35 0.00

TUG Duration > 10 s 6.51 0.00

STS Cycles < 15 11.25 0.00

Not completed 10s Tandem

Stance (eyes open)

3.59 0.00

Presence of fear of falling, TUG duration above

10 seconds, number of STS cycles below 15 and not

completed the tandem stance with eyes open were

the metrics with higher odds ratio with the fall level

and p-value below 0.05. Thus, the hypothesis of

random association between fall level and the

variables in shaded lines of Table 1 can be rejected.

Age above 65 years old and take more than 4

medicines per day also showed a p-value below 0.05

but the OR was lower than for the previously

mentioned variables. For the remaining variables,

the conclusion is that female individuals, or

individuals that have BMI lower than 13.7 or higher

than 29.7 or that have more than two health

conditions do not have greater odds of having a high

fall level than individuals that are male, have a

normal BMI and have less than two health

conditions. In general, tests scores showed higher

association with fall level than personal metrics,

reflecting that functional abilities have higher impact

on fall level than personal conditions of a person.

3.2 Machine Learning Approaches

Classification and regression methods were studied

Comparing Machine Learning Approaches for Fall Risk Assessment

227

for the differentiation between low and high fall risk

groups using the fall level as label. All algorithms

applied were retrieved from the Rapid Miner

predictive models.

3.2.1 Functional Tests Scores

As a first analysis, personal metrics (age, gender,

BMI, fear of fall, number of health conditions and

number of medicines) and test scores (TUG

duration, number of STS cycles and number of 4-

stage exercises) were used to define the feature

vector and fall level as label. The results are

summarized in Table 2.

Table 2: Classification and regression results with

personal metrics and functional tests scores. Accuracy,

precision, recall and F-Score are in percentage (%).

Algorithm Accuracy Precision Recall

F-

Score

k-NN, k=4 81.41 69.33 63.00 66.01

Naïve

Bayes

84.82 74.58 71.19 72.85

Random

Forest

83.13 59.37 53.05 56.03

Decision

Tree

81.44 68.28 60.33 64.06

Neural Net 82.45 69.22 64.84 66.96

SVM 82.45 49.08 51.21 50.12

Linear

Regression

83.11 69.01 56.05 61.86

Logistic

Regression

82.13 67.48 64.88 66.15

Naïve Bayes classifier obtained the higher accuracy,

84.82%. Precision was 74.58% and recall was

71.19%. Random Forest and Linear Regression also

obtained acceptable results. In general, all

algorithms showed higher precision than recall.

3.2.2 Sensors Features

In order to compare the previous results based on

tests scores with the features extracted from inertial

sensors and pressure platform, a feature vector

containing 224 sensors features was used. For each

TUG segment (stand, walk, turn and walk back) 19

statistical and frequency domain features were

extracted, yielding 76 features plus 2 metrics, time

to stand and the number of steps. For STS test, the

same 19 features were extracted plus 2 metrics, the

number of STS cycles and the STS power. For the 4-

stage test, 17 CoP metrics were extracted for each

one of the 7 exercises (when available), yielding 119

features. Additionally, 6 personal metrics were

added: age, gender, BMI, fear of fall, number of

health conditions and number of medicines. Fall

level was used as label. Since the number of features

was considerable high, forward feature selection was

applied prior to cross validation. Results are

presented in Table 3.

Table 3: Classification and regression results for personal

metrics and features extracted from sensors. Number of

features selected by forward feature selection follows the

name of the algorithm. Accuracy, precision, recall and F-

Score are in percentage (%).

Algor. Accuracy Precision Recall F-Score

k-NN, k=4

[5 F.]

85.78 87.79 95.88 91.66

Naïve

Bayes

[4 F.]

87.16 88.18 97.50 92.61

Neural Net

[5 F.]

87.20 88.05 97.94 92.73

SVM

[3 F.]

84.82 84.95 99.23 91.54

Random

Forest

[3 F.]

87.48 87.92 98.43 92.88

Decision

Tree [5 F.]

88.17 89.47 97.10 93.13

Linear

Reg. [3 F.]

85.89 85.66 99.55 92.08

Logistic

Reg. [4 F.]

86.54 86.74 98.78 92.37

Decision tree classifier obtained the higher

accuracy, 88.17%. Precision was 89.47% and recall

was 97.10%. Comparing the results of Naïve Bayes

with the previous analysis, the features obtained

from sensors yield higher accuracy than only tests

scores. Moreover, features from TUG and 4-stage

tests were frequently selected with forward feature

selection method. For all algorithms tested, features

from sensors provide higher precision and recall

values. F-Score obtained with features from sensors

were the same across all algorithms tested and

considerable higher than F-Score obtained only with

tests scores and personal metrics (91-93% against

50-72%).

4 DISCUSSION

Previous studies from (Scott et al. 2007) have

compared the accuracy of several functional tests

and fall risk tools to differentiate groups with

BIOSIGNALS 2017 - 10th International Conference on Bio-inspired Systems and Signal Processing

228

different levels of fall risk. Despite the differences in

protocol and population analysed (only for

community settings and validated in a prospective

study), similar accuracy and sensitivity were

reported. Murphy et al. (Murphy et al. 2003)

concluded that ‘floor transfer’ and ‘50 ft walk’ tests

combined can discriminate fallers from non-fallers

with an overall accuracy of 96% (82% sensitivity

and 100% specificity).

A similar study from Liu et al. (Liu et al. 2011)

has used metrics from instrumented TUG, alternate

step test and 5 times STS to classify between fallers

and non fallers and the best models have achieved

70% accuracy (68% sensitivity and 73% specificity).

5 CONCLUSIONS

The objective of this study was to compare the

performance of functional tests scores and features

obtained from inertial sensors and pressure

platforms to discriminate between low and high risk

of fall. A fall level was defined based on history of

falls and usage of walking aid and was used as label

in classification and regression algorithms. Only

subjects who performed the three functional tests

(TUG, STS and 4-stage) were included in this study.

The association between functional tests scores

and fear of falling with fall level are not random

(Fisher’s exact test p-value < 0.05), concluding that

individuals with functional disabilities and fear of

falling have greater odds of having a higher fall level

than individuals without physical disabilities and

without fear of falling. Moreover, when comparing

personal metrics with fall level, it was concluded for

some personal metrics that random association with

fall level cannot be excluded.

The differentiation power of personal metrics

and tests scores was considerable different when

tested with classification and regression methods.

Accuracies above 80% were obtained for all

algorithms. Naïve Bayes outperforms with an

accuracy of 84.82% (74.58% of precision and

71.19% of recall).

However, features from inertial sensors and

pressure platform obtained better results for the

same algorithms than only tests scores. Naïve Bayes

classifier obtained an accuracy of 87.16% (88.18%

of precision and 97.50% of recall).

These results support the conclusion that

instrumentation of fall risk assessment tests with

inertial sensors and pressure platform could better

discriminate the individuals at a higher risk of

falling.

ACKNOWLEDGEMENTS

Authors would like to thank all participants and

centres, clinics and other entities hosting the

screenings. Financial support from project

FallSensing: Technological solution for fall risk

screening and falls prevention (POCI-01-0247-

FEDER-003464), co-funded by Portugal 2020,

framed under the COMPETE 2020 (Operational

Programme Competitiveness and

Internationalization) and European Regional

Development Fund (ERDF) from European Union

(EU).

REFERENCES

Agrawa, Yuri, John P. Carey, Howard J. Hoffman, Daniel

A. Sklare, and Michael C. Schubert. 2011. ‘The

Modified Romberg Balance Test: Normative Data in

U.S. Adults’. Otology & Neurotology 32 (8): 1309–11.

doi:10.1097/MAO.0b013e31822e5bee.

Aguiar, B., J. Silva, T. Rocha, S. Carneiro, and I Sousa.

2014. ‘Monitoring Physical Activity and Energy

Expenditure with Smartphones’. In 2014 IEEE-EMBS

International Conference on Biomedical and Health

Informatics (BHI), 664–67. doi:10.1109/BHI.2014.

6864451.

Ambrose, Anne Felicia, Geet Paul, and Jeffrey M

Hausdorff. 2013. ‘Risk Factors for Falls among Older

Adults: A Review of the Literature’. Maturitas 75 (1):

51–61. doi:10.1016/j.maturitas.2013.02.009.

Beauchet, Olivier, B. Fantino, G. Allali, S. W. Muir, M.

Montero-Odasso, and C. Annweiler. 2011. ‘Timed up

and Go Test and Risk of Falls in Older Adults: A

Systematic Review’. The Journal of Nutrition, Health

& Aging 15 (10): 933–38. doi:10.1007/s12603-011-

0062-0.

Bergen, Gwen, Mark R. Stevens, and Elizabeth R. Burns.

2016. ‘Falls and Fall Injuries Among Adults Aged ≥65

Years — United States, 2014’. MMWR. Morbidity and

Mortality Weekly Report 65 (37): 993–98.

doi:10.15585/mmwr.mm6537a2.

Bigelow, Kimberly Edginton, and Necip Berme. 2011.

‘Development of a Protocol for Improving the Clinical

Utility of Posturography as a Fall-Risk Screening

Tool’. The Journals of Gerontology. Series A,

Biological Sciences and Medical Sciences 66 (2): 228–

33. doi:10.1093/gerona/glq202.

Cho, Kang Hee, Soo Kyung Bok, Young-Jae Kim, and

Seon Lyul Hwang. 2012. ‘Effect of Lower Limb

Strength on Falls and Balance of the Elderly’. Annals

of Rehabilitation Medicine 36 (3): 386–93.

doi:10.5535/arm.2012.36.3.386.

Guimaraes, V., D. Ribeiro, and L. Rosado. 2013. ‘A

Smartphone-Based Fall Risk Assessment Tool:

Measuring One Leg Standing, Sit to Stand and Falls

Efficacy Scale’. In 2013 IEEE 15th International

Comparing Machine Learning Approaches for Fall Risk Assessment

229

Conference on E-Health Networking, Applications

Services (Healthcom), 529–33. doi:10.1109/

HealthCom.2013.6720733.

Hsi, Wei-Li. 2016. ‘Analysis of Medial Deviation of

Center of Pressure after Initial Heel Contact in

Forefoot Varus’. Journal of the Formosan Medical

Association = Taiwan Yi Zhi 115 (3): 203–9.

doi:10.1016/j.jfma.2015.03.004.

Jones, C. Jessie, Roberta E. Rikli, and William C. Beam.

1999. ‘A 30-S Chair-Stand Test as a Measure of

Lower Body Strength in Community-Residing Older

Adults’. Research Quarterly for Exercise and Sport 70

(2): 113–19. doi:10.1080/02701367.1999.10608028.

Liu, Ying, S.J. Redmond, Ning Wang, F. Blumenkron,

M.R. Narayanan, and N.H. Lovell. 2011. ‘Spectral

Analysis of Accelerometry Signals From a Directed-

Routine for Falls-Risk Estimation’. IEEE Transactions

on Biomedical Engineering 58 (8): 2308–15.

doi:10.1109/TBME.2011.2151193.

Lord, Stephen R., Hylton B. Menz, and Anne Tiedemann.

2003. ‘A Physiological Profile Approach to Falls Risk

Assessment and Prevention’. Physical Therapy 83 (3):

237–52.

Murphy, Mary A, Sharon L Olson, Elizabeth J Protas, and

Averell R Overby. 2003. ‘Screening for Falls in

Community-Dwelling Elderly’. Journal of Aging and

Physical Activity 11 (1): 66–81.

Oliver, David, Fergus Daly, Finbarr C. Martin, and

Marion E. T. McMurdo. 2004. ‘Risk Factors and Risk

Assessment Tools for Falls in Hospital in-Patients: A

Systematic Review’. Age and Ageing 33 (2): 122–30.

doi:10.1093/ageing/afh017.

Raymakers, J. A., M. M. Samson, and H. J. J. Verhaar.

2005. ‘The Assessment of Body Sway and the Choice

of the Stability Parameter(s)’. Gait & Posture 21 (1):

48–58. doi:10.1016/j.gaitpost.2003.11.006.

Rikli, RE, and JC Jones. 2010. ‘Functional Fitness

Normative Scores for Community-Residing Older

Adults, Ages 60–94’. Human Kinetics Journals. April

21. http://journals.humankinetics.com/japa-back-

issues/japavolume7issue2april/functional-fitness-

normative-scores-for-community-residing-older-

adults-ages-60-94.

Rose, Debra J, Jones, Jessie C, and Lucchese, Nicole.

2002. ‘Predicting the Probability of Falls in

Community-Residing Older Adults Using the 8-Foot

Up-and-Go: A New Measure of Functional Mobility’.

Journal of Ageing and Physical Activity 10 (4): 466–

75.

Rossiter-Fornoff, J. E., S. L. Wolf, L. I. Wolfson, and D.

M. Buchner. 1995. ‘A Cross-Sectional Validation

Study of the FICSIT Common Data Base Static

Balance Measures. Frailty and Injuries: Cooperative

Studies of Intervention Techniques’. The Journals of

Gerontology. Series A, Biological Sciences and

Medical Sciences 50 (6): M291-297.

Rubenstein, Laurence Z. 2006. ‘Falls in Older People:

Epidemiology, Risk Factors and Strategies for

Prevention’. Age and Ageing 35 (suppl 2): ii37-ii41.

doi:10.1093/ageing/afl084.

Rubenstein, Laurence Z., and Karen R. Josephson. 2002.

‘The Epidemiology of Falls and Syncope’. Clinics in

Geriatric Medicine 18 (2): 141–58.

Scott, Vicky, Kristine Votova, Andria Scanlan, and

Jacqueline Close. 2007. ‘Multifactorial and Functional

Mobility Assessment Tools for Fall Risk among Older

Adults in Community, Home-Support, Long-Term and

Acute Care Settings’. Age and Ageing 36 (2): 130–39.

doi:10.1093/ageing/afl165.

Silva, Joana, and Ines Sousa. 2016. ‘Instrumented Timed

Up and Go: Fall Risk Assessment Based on Inertial

Wearable Sensors’. In 2016 IEEE International

Symposium on Medical Measurements and

Applications (MeMeA) (MeMeA 2016). Benevento,

Italy.

Thomas, Julia C., Charles Odonkor, Laura Griffith, Nicole

Holt, Sanja Percac-Lima, Suzanne Leveille, Pensheng

Ni, Nancy K. Latham, Alan M. Jette, and Jonathan F.

Bean. 2014. ‘Reconceptualizing Balance: Attributes

Associated with Balance Performance’. Experimental

Gerontology 57 (September): 218–23.

doi:10.1016/j.exger.2014.06.012.

Tinetti, Mary E. 1986. ‘Performance-Oriented Assessment

of Mobility Problems in Elderly Patients’. Journal of

the American Geriatrics Society 34 (2): 119–26.

Zhang, Wei, G. Ruben H. Regterschot, Hana Schaabova,

Heribert Baldus, and Wiebren Zijlstra. 2014. ‘Test-

Retest Reliability of a Pendant-Worn Sensor Device in

Measuring Chair Rise Performance in Older Persons’.

Sensors (Basel, Switzerland) 14 (5): 8705–17.

doi:10.3390/s140508705.



BIOSIGNALS 2017 - 10th International Conference on Bio-inspired Systems and Signal Processing

230