Fall Risk Assessment Through Automatic Combination of Clinical Fall Risk Factors and Body-Worn Sensor Data

Abstract

Falls are the leading global cause of accidental death and disability in older adults, and are the most common cause of injury and hospitalization. Accurate, early identification of patients at risk of falling, could lead to timely intervention and a reduction in the incidence of fall-related injury and associated costs.
We report a statistical method for fall risk assessment using standard clinical fall risk factors (N=748). We also report a means of improving this method by automatically combining it, with a fall risk assessment algorithm based on inertial sensor data and the timed up and go (TUG) test. Furthermore, we provide validation data on the sensor based fall risk assessment method using a statistically independent data set.
Results obtained using cross-validation on a sample of 292 community dwelling older adults, suggest that a combined clinical and sensor-based approach yields a classification accuracy of 76.0%, compared to either 73.6% for sensor-based assessment alone, or 68.8% for clinical risk factors alone. Increasing the cohort size by adding an additional 130 subjects from a separate recruitment wave (N=422), and applying the same model building and validation method, resulted in a decrease in classification performance (68.5% for combined classifier, 66.8% for sensor data alone, and 58.5% for clinical data alone). This suggests heterogeneity between cohorts may be a major challenge when attempting to develop fall risk assessment algorithms which generalize well. Independent validation of the sensor-based fall risk assessment algorithm on an independent cohort of 22 community dwelling older adults yielded a classification accuracy of 72.7%.
Results suggest that the present method compares well to previously reported sensor based fall risk assessment methods in assessing falls risk. Implementation of objective fall risk assessment methods on a large scale has the potential to improve quality of care and lead to a reduction in associated hospital costs, due to fewer admissions and reduced injuries due to falling.

Full text

 Abstract— Falls are the leading global cause of accidental
death and disability in older adults, and are the most common
cause of injury and hospitalization. Accurate, early identification
of patients at risk of falling, could lead to timely intervention and
a reduction in the incidence of fall-related injury and associated
costs.
We report a statistical method for fall risk assessment using
standard clinical fall risk factors (N=748). We also report a
means of improving this method by automatically combining it,
with a fall risk assessment algorithm based on inertial sensor
data and the timed up and go (TUG) test. Furthermore, we
provide validation data on the sensor based fall risk assessment
method using a statistically independent data set.
Results obtained using cross-validation on a sample of 292
community dwelling older adults, suggest that a combined
clinical and sensor-based approach yields a classification
accuracy of 76.0%, compared to either 73.6% for sensor-based
assessment alone, or 68.8% for clinical risk factors alone.
Increasing the cohort size by adding an additional 130 subjects
from a separate recruitment wave (N=422), and applying the
same model building and validation method, resulted in a
decrease in classification performance (68.5% for combined
classifier, 66.8% for sensor data alone, and 58.5% for clinical
data alone). This suggests heterogeneity between cohorts may be
a major challenge when attempting to develop fall risk
assessment algorithms which generalize well. Independent
validation of the sensor-based fall risk assessment algorithm on
an independent cohort of 22 community dwelling older adults
yielded a classification accuracy of 72.7%.
Results suggest that the present method compares well to
previously reported sensor based fall risk assessment methods in
assessing falls risk. Implementation of objective fall risk
assessment methods on a large scale has the potential to improve
quality of care and lead to a reduction in associated hospital
costs, due to fewer admissions and reduced injuries due to
falling.
BRG is a director of Kinesis Health Technologies Ltd, a company with a
license to commercialise this technology.
We would like to acknowledge the help and support of the staff of the
Irishtown & Ringsend Primary Care Centre Physiotherapy Department, and
the TRIL Clinic, St James’ Hospital, Dublin, and the participants involved in
this study.
B.R Greene is with Kinesis Health Technologies, Dublin, Ireland (e-mail:
barry.greene@kinesis.ie).
S. J Redmond is with the Graduate School of Biomedical Engineering
University of New South Wales, Sydney, Australia. (e-mail:
s.redmond@unsw.edu.au).
B. Caulfield, Insight centre and School of Physiotherapy and
Performance Science, University College Dublin, Ireland (e-mail:
b.caulfield@ucd.ie).
I. INTRODUCTION
An ageing global population coupled with increasing
healthcare costs has led to prevention of falls in the elderly to
become an increasingly pressing healthcare issue. Roughly
one in three adults over 65 fall each year [1] and it has been
estimated that the annual medical costs of fall-related injuries
for people aged 65 years and older, in the USA alone, is
US$30 billion [2, 3]. The incidence of falls and their
associated costs will increase significantly [3, 4] with the
continued ageing demographics. The causes of falls are
complex with a wide variety of social [5], physical and
environmental [6] risk factors. However, gait and mobility
impairments are considered to be leading risk factors for falls
[7, 8], but crucially have been shown to be modifiable
through exercise intervention [9].
In current clinical practice patient’s risk of falling is usually
assessed by a variety of specialists: physical therapists,
geriatric specialists, and community nurses or occupational
therapists. There are a number of widely used clinical
assessment tools for assessing risk of falls [10-12]. However
these current methods are somewhat subjective in nature and
often require specialist clinical expertise. An objective,
evidence-based method that facilitates screening for fall risk,
suitable for use by non-specialists, would free up specialists’
time and allow for identification of people who need onward
referral to specialist services. This would allow such
specialist services to concentrate their efforts on detailed
examination to identify the specific factors leading to the
level of risk identified, and subsequent implementation of
appropriate remedial interventions to ameliorate the risk. This
should lead to improved quality of care and a reduction in
associated hospital costs, due to fewer admissions and
reduced injuries due to falling.
In recent years, there has been a proliferation of research
reporting a variety of inertial sensor-based methods for
assessing fall risk in older adults [13-18]. These are usually
based on the use of quantitative data obtained during
prescribed motor tasks, such as quiet standing, walking or a
TUG (timed-up-and-go) test. The majority of studies have
validated their methods using cross-sectional (retrospective)
fall data, i.e., each patient’s self-reported history of falling
prior to assessment [19, 20]. A smaller number of trials have
reported a prospective validation of their methods, i.e., each
subject was followed-up and their falls tracked, for a period
of time after the assessment is performed; these fall data were
then used to validate the methods’ performance in predicting
falls [15, 21]. Previous research [22] has reported that fall
Fall risk assessment through automatic combination of clinical fall
risk factors and body-worn sensor data
Barry R. Greene, Senior IEEE Member, Stephen J. Redmond Senior IEEE Member, Brian Caulfield,
IEEE Member

history data were more effective in classifying healthy
controls than two year prospective data (fall counts during a
follow-up period).
Recently Shany et al. [23] and Howcroft et al. [19] have
pointed out weaknesses in the methodologies of many of
these reported papers which aim to assess fall risk and/or
predict falls. Specifically, they noted insufficient validation
and inappropriate use of statistical modelling methods in
some or all of the papers published to date. In this paper we
attempt to address some of these concerns by means of a re-
examination of previous research findings from our group, as
well as validation of risk assessment models with independent
data. In doing so, we will address three issues that were raised
by Shany et al. [23], as well as highlighting a potentially
dominant data analysis issue that should be considered in
future studies. Firstly, we examine the impact of applying best
practice statistical modelling methods, specifically, a nested
cross-validation approach. We also examine the impact that
heterogeneity in terms of patient data and in outcome
measures, has on the efficacy of sensor-based fall risk
assessment algorithms.
This paper introduces two new methods for fall risk
assessment; the first is based on standard clinical risk factors
obtained from questionnaires; the second is a means of
improving this questionnaire-based method by automatically
combining it with a sensor-based method for fall risk
assessment. In addition, we quantify the validity of a
previously-published sensor-based fall risk assessment
algorithm from our group using a fully independent dataset.
II. DATA
A. Training and validation data set
Seven hundred and forty eight (519 female, age 73.6±7.4
years) subjects were assessed as part of the TRIL project, a
large ageing research project in Dublin, Ireland. All subjects
provided informed consent and had their mobility assessed
using the TUG test instrumented with inertial sensors in the
TRIL clinic, St James’s Hospital (SJH), Dublin, Ireland.
Subjects were assessed between 2007 and 2012; data from
two separate phases of the trial (waves) are included here.
The Wave 1 data set contained 616 subjects, while Wave 2
contained 132 subjects. Inclusion criteria were subjects 60
years and older, with no history of stroke, able to walk
without assistance. Ethical approval was received from the
local research ethics committee in each instance as well as
informed consent from each subject.
History of falls in the past 5 years for each subject was
obtained by means of a questionnaire. A fall was defined as
“an event which resulted in a person coming to rest on the
lower level regardless of whether an injury was sustained,
and not as a result of a major intrinsic event or
overwhelming hazard” [24]. Fall outcome data were verified
using available hospital records as well as information
provided by relatives.
Each subject completed at least one TUG test,
instrumented with inertial sensors. Due to issues with
hardware failure and data storage during sensor data
acquisition, the sensor data for some subjects were
incomplete. Subjects with incomplete sensor data were
excluded from the analysis. Inertial sensor data were available
for 422 subjects (308 female; 292 from Wave 1, 130 from
Wave 2).
B. Clinical assessment
Clinical fall risk factors were captured for each subject by
means of a clinical assessment [21]. Eyesight was assessed
using the Pelli-Robson contrast sensitivity scale and the
binocular logMAR scale. Each subject was checked for
orthostatic hypotension (defined as orthostatic systolic blood
pressure drop > 20mmHg), using a Finometer (Finapres
Medical Systems, Amsterdam, Netherlands). Each subject’s
prescription medications were reviewed to determine if they
had a polypharmacy issue, where polypharmacy is defined as
the use of four or more prescription medications.
C. Independent data set
An independent data set of community-dwelling older adults
was obtained to validate the expected performance of the
sensor-based fall risk assessment on unseen data. Twenty-five
older adults (22 female), were assessed at the Irishtown &
Ringsend Primary Care Centre (Dublin 4, Ireland). Subjects
were aged 64-87 years (mean 75.4 years) and recruited
through a community fall prevention service. TUG tests were
conducted on 25 subjects using the inertial sensor method
described above. Sensor data from one recording was not
usable due to improper sensor placement while falls history
data from two subjects were not available, leaving a total of
22 valid tests, one test per participant.
Each subject completed a questionnaire on their history of
falls in the past 12 months. 11 of 22 patients reported a
history of falls. These data were used to validate the
performance of the fall risk assessment tool. Prospective data
were not available, and previous falls history is considered a
good surrogate for future falls [25, 26]. The use of an
independent data set means that a valid confidence interval
can be calculated. A binomial proportion confidence interval
was used to estimate the confidence interval for classification
accuracy [27].
Figure 1: Medio-lateral shank angular velocity signal obtained from an 85
year old female with a history of falls while performing a TUG test. Initial
contact (IC) and terminal contact (TC), as well as mid-swing points, are
shown.

III. METHODS
A. TUG test protocol
The TUG test is a standard test of mobility, used all over the
world to screen for gait and balance issues in the elderly [28-
30]. The test requires the subject to get up from a chair, walk
three metres, turn through 180º at a designated spot, return to
the seat and sit back down. The time taken to perform the test
is recorded by a clinical staff member using a stopwatch. The
performance of older adults prone to falling can be very
different from those who do not fall [31-33]; longer TUG
times are thought be associated with falling. However,
research has shown that the TUG test alone is only
moderately accurate in assessing fall risk [31, 33]. Subjects
were asked to complete the TUG test, ‘as fast as safely
possible’. A standard chair (46 cm high seat, 65 cm) with
armrests was used. The timer was started the moment the
clinician said ‘go’, and stopped the moment the subject’s
back touched the back rest of the chair. Each subject was
given time to become familiar with the test and the test was
demonstrated to them beforehand.
B. Sensor data acquisition
For each TUG test, subjects were fitted with two wireless
inertial sensors which were attached by a clinician or research
assistant, using elasticated bandages, to the mid-point of the
left and right anterior shank (shin)[13]. Sensors were oriented
to capture movement about the anatomical medio-lateral axis.
Inertial sensor data (see Figure 1) were synchronously
acquired in real-time via Bluetooth using dedicated software
(QTUG, Kinesis Health Technologies; Dublin, Ireland). Each
sensor was tri-axial and contained an accelerometer and
gyroscope. Sensor data were sampled at 102.4 Hz with a full
scale range of 500 °/s and a sensitivity of 2 mV/°/s. Sensors
were calibrated using a standard method [34]. The raw tri-
axial gyroscope signals were low-pass filtered (zero-phase 2nd
order Butterworth filter, 20 Hz corner frequency). Data
acquisition software automatically connected to the relevant
sensors, allowing the clinician to start and stop the recording
manually as with a stopwatch. After each test, data were
saved to text format for analysis.
C. Clinical risk factor based fall risk assessment
The AGS/BGS offer guidelines aimed to capture the main
clinical risk factors linked to falls in older adults [35]. These
guidelines concentrate on self-reported data on standard fall
risk factors, captured as part of a clinical assessment. We
aimed to create an easily reproducible classifier model based
on these standard self-reported risk factors.
A logistic regression model was created using a number of the
self-reported factors discussed in the AGS/BGS guidelines,
the factors selected were those that were also present in the
TRIL data set. All available data were used. The features
included and used to classify subjects according to falls
history were as follows: gender, height, weight and age on
date of assessment, polypharmacy, vision problems and
orthostatic hypotension.
Two candidate logistic regression models, used to produce a
fall risk estimate (FREclinical), are trained using either data
from Wave 1 (2007) alone or Wave 1 and Wave 2 (2012)
combined. An estimate of the classifier performance of each
model on unseen data was obtained using ten repetitions of
ten-fold cross-validation [36].
D. Inertial sensor-based fall risk assessment
Previously, we reported an inertial sensor-based method to
quantify, standing, walking and turning during the TUG test.
We used these quantitative TUG parameters along with a
supervised classifier model to obtain a statistical fall risk
estimate (FREsensor) for older adults. The method was
validated using both prospective (two year follow-up) [21]
and cross-sectional (falls history) falls data [13]. Previously,
testing this method on two sets of healthy control data found
that classifier models trained using falls history data was far
more accurate in correctly distinguishing healthy control
subjects (defined as healthy older adults, aged greater than
50, with no history of falls) from fallers, and so less prone to
false positives than models trained using prospective falls
data [22]. In this study, only falls history data were used to
validate performance.
The sensor-based fall risk assessment method used a
regularized discriminant (RD) classifier model [37], with
regularization parameter values set to λ=0.1 and r=0.1 prior
to analysis. Ten repetitions of ten-fold cross validation [36]
was used to estimate the generalized classifier performance.
Using only the training data for each iteration of the cross-
validation routine, a potential feature set was evaluated using
a second inner cross-validation loop. Once a feature set is
identified using the training data, it is tested using the
withheld data for this iteration of the outer cross-validation
loop [38], a process known as ‘nested’ cross-validation.
Training and testing sets were randomly selected for each
repetition.
Classifier models were trained using the subset of subjects
with sensor data, from either data from Wave 1 (2007) alone
or Wave 1 and Wave 2 (2012) combined.
E. Combined clinical and inertial sensor fall risk assessment
A combined fall risk estimate (FREcombined) is obtained by
applying classifier combination theory [37] and averaging the
posterior probabilities produced for a given subject by the
sensor-based FRE (FREsensor) and the clinical FRE (FREclinical)
to produces a combined FRE (FREcombined):
 
2/
clinicalsensorcombined FREFREFRE 
To examine the performance of FREcombined relative to
FREsensor and FREclinical, the classification performance is
compared (see Table 1 below) on two common data sets
(Wave 1 as well as the combination of Wave 1 and Wave 2).
The data sets used were the subset that had both sensor and
clinical data. The classifier performance for each model was
estimated using Leave One Out (LOO) cross-validation,
where N-1 samples were used to train the classifier model and
the remaining sample used to test the performance, with this
process repeated for each sample. The FREsensor feature and
model selection was conducted using the nested cross-
validation as discussed above, where model selection is

conducted using 10 repetitions of 10-fold cross-validation
using only the training data, within the LOO procedure.
F. Independent validation of inertial sensor fall risk
assessment
Cross-sectional data (self-reported fall history) from the two
independent data sets were used to validate the previously
reported sensor-based fall risk assessment method (FREsensor)
[22] to obtain a statistically independent evaluation of the
algorithm performance on a cohort of independent subjects.
G. Classifier performance metrics
Classifier performance for each fall risk assessment algorithm
was assessed using standard classifier performance measures.
Accuracy (Acc) is defined as the percentage of subjects
correctly classified by the algorithm as being a ‘faller’ or
‘non-faller’ (a faller is defined as having one injurious fall or,
more than one previous fall); sensitivity (Sens) is defined as
the percentage of the fallers identified correctly; similarly,
specificity (Spec) is defined as the percentage of the non-
fallers correctly identified. Positive predictive value (PPV) is
defined as the proportion of subjects the algorithm classified
as fallers, who are correctly classified; negative predictive
value (NPV) is the proportion of subjects the algorithm
classified as non-fallers, who are correctly classified. The
values reported for each performance metric were averaged
across the folds of the cross-validation routine.
IV. RESULTS
A. Clinical risk factor based fall risk assessment
Cross-validated results for the FREclinical logistic regression
classifier model are detailed in Table 1. Wave 1 data (N=616)
yielded a classification accuracy of 70.37%, while combining
Wave 1 and Wave 2 data (N=748) yielded a slightly lower
accuracy of 67.13%.
B. Combination of clinical risk factor classifier with sensor
data classifier
Cross-validated results for the FREcombined, FREsensor, and
FREclinical classifier models are detailed in Table 2. Results
are shown for the subsets of Wave 1 data (N=292) and Wave
1 and 2 data combined (N=422), with both clinical and sensor
data. The combined model was 76.03% accurate for Wave 1
and 68.48% for the combined Wave 1 and Wave 2 data set.
The inertial sensor-based method achieved 73.63% accuracy
for Wave 1 data and was 66.82% accurate for the
combination of Wave 1 and Wave 2. The clinical risk factor-
based model was 68.84% accurate for Wave 1 data and
58.53% accurate for Wave 1 and Wave 2 data combined.
C. Validation of sensor fall risk assessment using
independent data set
Results for a statistically independent validation of the
sensor-based fall risk assessment algorithm, using fall history
data from a community-dwelling population are detailed in
Table 3, below. Sixteen of the 22 tests were correctly
identified as being a ‘faller’ or a ‘non-faller’ yielding a
72.70% classification accuracy. A binomial proportion
confidence interval calculation yielded a confidence interval
for classification accuracy of 54.12-91.34%.
V. DISCUSSION
This paper aimed to introduce a number of new results to the
field of sensor-based assessment of fall risk. Specifically, we
report a method for obtaining a statistical estimate of fall risk
based on a questionnaire on standard fall risk factors. We also
report a classifier fusion approach that combines the classifier
derived from clinical risk factors, with a classifier derived
from inertial sensor data obtained during the TUG test. In
TABLE II: PERFORMANCE OF COMBINED FRE, CLINICAL RISK FACTOR BASED
FRE AND QTUG FRE CLASSIFIER MODELS. MODELS WERE VALIDATED USING
NESTED CROSS-VALIDATION.
Wave 1
Wave1/2
FREcombined
FREsensor
FREclinical
FREcombined
FREsensor
FREclinical
N
292
292
292
422
422
422
Acc (%)
76.03
73.63
68.84
68.48
66.82
58.53
Sens (%)
66.99
72.82
46.81
68.36
74.01
35.93
Spec (%)
80.95
74.07
86.26
68.57
61.63
78.90
PPV (%)
65.71
60.48
63.77
61.11
58.22
54.55
NPV (%)
81.82
83.33
75.85
75.00
76.65
63.61
TABLE I: CROSS-VALIDATED PERFORMANCE ESTIMATES FOR CLINICAL
RISK FACTOR BASED FALL RISK ASSESSMENT CLASSIFIER MODEL
(FRECLINICAL) USING WAVE 1 DATA OR COMBINED WAVE 1 AND 2
DATA SETS. FEATURES INCLUDED IN BOTH LOGISTIC REGRESSION
MODELS ARE GENDER, HEIGHT, WEIGHT, AGE, POLYPHARMACY,
IMPAIRED VISUAL FIELDS AND ORTHOSTATIC HYPERTENSION.
Wave 1
Wave 1/2
N
616
748
Acc (%)
70.37
67.13
Sens (%)
78.59
76.66
Spec (%)
38.89
40.40
PPV (%)
69.09
65.51
NPV (%)
51.11
53.96
TABLE III: PERFORMANCE ESTIMATES FOR VALIDATION OF FRESENSOR
USING AN INDEPENDENT DATA SET OF COMMUNITY DWELLING OLDER
ADULTS. 95% CONFIDENCE INTERVALS ARE PROVIDED. 12 MONTH
FALLS HISTORY DATA WERE USED AS THE OUTCOME MEASURE.
FREsensor
N
22
Acc (%) [95% CI]
72.70 [54.12, 91.34]
Sens (%) [95% CI]
90.91 [78.90, 100.0]
Spec (%) [95% CI]
54.50 [33.69, 75.31]
PPV (%) [95% CI]
66.67 [46.97, 86.37]
NPV (%) [95% CI]
85.71 [71.09, 100.0]

addition, we report a validation of a sensor-based falls risk
assessment algorithm on a completely independent data set.
For a single wave (Wave 1) of 293 subjects, a classifier
trained using only questionnaire based clinical risk factors
was 68.84% accurate in classifying subjects as fallers or non-
fallers; these results are consistent with results (0.64 area
under the receiver operating curve (AUC)) reported for the
FRAT-up questionnaire-based classifier model, by Cattelani
et al. [39].
For the subset of Wave 1 data that had sensor data, the
sensor-based method [22] for assessing fall risk was 73.63%
accurate in classifying subjects according to history of falls.
Results from a study [13] on the same (Wave 1) data set
showed that TUG time taken alone was 60.1% accurate in
classifying subjects according to falls history, while the Berg
balance scale [10] was 61.4% accurate.
Combining the sensor-based fall risk assessment classifier and
clinical fall risk factor classifier using classifier fusion led to
an improved classification accuracy of 76.03%, compared to
either clinical or sensor data taken alone. Combining subjects
from two waves (N=422) and applying the same methods
yielded reduced classification accuracies for all three
classifiers; 58.53% (FREclinical), 66.82% (FREsensor) and
68.48% (FREcombined).
Validation of the sensor based fall risk assessment algorithm
on an independent data set (N=22) of community-dwelling
older adults yielded classification accuracy of 72.7% [54.12,
91.34].
This study sought to address some deficiencies in the
literature as highlighted by recent commentary [23]. We
employed rigorous statistical validation (nested cross-
validation) on a large data set to obtain performance
estimates. We also provide a validation of the sensor based
algorithm on a statistically independent data set; this is
significant as a criticism of much of the sensor-based fall risk
assessment research to date, is that there has been no
independent validation of methods on completely
independent data sets. The results of this validation using
independent data show high sensitivity but relatively low
specificity, suggesting that the sensor-based method
recognize some subjects as being at risk of falling although
they had no prior history of falling. Prospective validation of
the method on a large data set, using best practice for
prospective studies on falls (i.e., one year follow-up for falls
with daily falls diaries, collected on a weekly basis) would be
required to definitively assess the method’s performance in
predicting falls.
Our results suggest that improved automatic assessment of
fall risk can be achieved through the automatic combination
of clinical fall risk factors and a sensor-based physical
assessment. It has been noted in the literature that different
classifier models, trained on the same outcomes with different
data sets, offer potentially complementary information about
the patterns to be classified, which can be harnessed to
improve overall performance [37, 40]. Theoretical results
demonstrate that combining classifiers from different modes
with generalized knowledge of the patterns to be classified
often yields improved, more robust, classification
performance.
Classification results from Wave 1 and 2 combined were
consistently lower than those from Wave 1 alone. This
suggests that heterogeneity between waves is a major issue
affecting classifier performance. Sources of heterogeneity
include: the fact that Wave 2 data were captured several years
after Wave 1; and the assessments performed by different
personnel (with potential differences in the instructions given
on how to complete the test); different sensors (with
variations in calibration) were used; Wave 1 data used a 5
year interval for falls history, while Wave 2 used a 12 month
interval; while subjects in both waves consisted of community
dwelling older adults. Another significant source of
heterogeneity is the fact that Wave 1 subjects were largely
recruited through the hospital system, while Wave 2 subjects
were almost entirely recruited through advertisement in the
community.
Results for the sensor-based fall risk assessment method
reported here agree broadly in terms of accuracy with
previously reported results on the same data [13, 22],
suggesting that the effect of bias/overlearning arising from
model selection deviating from best practice (as discussed by
Shany et al. [23]), on previous results that did not employ
nested cross-validation, was relatively small. We would
suggest instead that heterogeneity is a bigger issue to consider
in the application of sensor-based fall risk assessment
algorithms to broader subject populations.
The cross-sectional falls data is a significant limitation of this
study. Subjects were classified according to their history of
falling, which is an established risk factor for future falls [8,
25] and so aims to produce a noisy estimate of the subjects’
propensity to fall. However, this retrospective design is a
clear limitation, given that self-reported historical falls data
can be unreliable [41] in terms of subject recall bias, and gait
changes due to previous falls. The current best practice for
the use of outcome data in falls injury prevention trials, as
recommended by the consensus guidelines [42], is the use of
prospective fall diaries. Prospective fall data can suffer from
issues with recall; however, if data are collected according to
best practice, the inaccuracy due to recall error can be
minimized. An additional limitation of this study is the small
size of the independent data set. Future studies may allow for
a larger scale independent validation. A significant number of
the subjects in this study were self-referred, which might
indicate differences when compared to a cohort of hospital in-
patients or patients in long term care. The results reported
here generalized well across the study cohort, however if a
comparison were made against broader patient populations,
differences may be apparently due to the fact the study cohort
were considered to quite physically robust.
In summary, we have introduced a statistical method for
assessing fall risk using standard clinical risk factors, as well
as an algorithm that automatically combined classifiers of
clinical risk factors and inertial sensors parameters, to
produce improved assessment of fall risk in older adults.
While it is difficult to compare performance estimates across
studies, the present results suggest our method may compare

well with previously reported methods for sensor based
assessment of fall risk and compares favourably with the
theoretical maximal accuracy for fall risk assessment
algorithms of 81% suggest by Palumbo et al. [43]. The
algorithm was validated using nested cross-validation to
reduce the possibility of biased performance estimates due to
model overfitting. In addition, we report validation results,
using fall history, of a sensor based fall risk assessment
algorithm on an independent data set. We also investigated
the effect of heterogeneity on classification results for sensor-
based fall risk assessment algorithms using data from two
waves of a large ageing research study. Heterogeneity appears
to account for a large discrepancy in accuracy. Future studies
employing a statistically meaningful representative sample
would be necessary to overcome the effect of heterogeneity, a
topic which has not been adequately explored in this field to
date.
Implementation of objective fall risk assessment methods on a
large scale has the potential to prevent falls through early
intervention, reduce health care costs, and improve the quality
of care offered to patients at risk of falling. Although
identification of those at risk is but a first step in reducing the
incidence of falls, it is a critical element of an effective and
efficient end-to-end falls prevention service. Early
identification of patients at risk of falling would allow more
timely intervention, such as targeted physiotherapy,
medication changes, or occupational therapy. It is important
to recognize that such improvements in care and reductions in
cost will not be realized, unless the identification of falls risk
is coupled with an effective service that further explores the
specific factors leading to risk at an individual level, leading
to implementation of appropriate remedial strategies.
VI. REFERENCES
[1] T. Masud and R. O. Morris, "Epidemiology of falls," Age Ageing,
vol. 30, pp. 3-7, 2001.
[2] J. A. Stevens, P. S. Corso, E. A. Finkelstein, and T. R. Miller,
"The costs of fatal and non-fatal falls among older adults," Injury
Prevention, vol. 12, pp. 290-295, 2006.
[3] WHO. (2007) WHO global report on falls prevention in older
age. . WHO Department of Ageing and Life Course.
[4] L. Gillespie, W. Gillespie, R. Cumming, S. Lamb, and B. Rowe,
"American Geriatrics Society; British Geriatrics Society;
American Academy of Orthopaedic Surgeons Panel on Falls
Prevention. Guideline for the prevention of falls in older persons
Interventions for preventing falls in the elderly," J Am Geriatr
Soc, vol. 49, pp. 664 - 672, 2001.
[5] F. Bloch, M. Thibaud, B. Dugué, C. Brèque, A. S. Rigaud, and
G. Kemoun, "Episodes of falling among elderly people: a
systematic review and meta-analysis of social and demographic
pre-disposing characteristics," Clinics, vol. 65, pp. 895-903,
2010.
[6] A. Bueno-Cavanillas, F. Padilla-Ruiz, J. Jimenez-Moleon, C.
Peinado-Alonso, and R. Galvez-Vargas, "Risk factors in falls
among the elderly according to extrinsic and intrinsic
precipitating causes," Eur J Epidemiol, vol. 16, pp. 849 - 859,
2000.
[7] J. Hausdorff, H. Edelberg, M. Cudkowicz, M. Fiatarone Singh,
and J. Wei, "The relationship between gait changes and falls," J
Am Geriatr Soc, vol. 45, p. 1406, 1997.
[8] A. M. Tromp, S. M. F. Pluijm, J. H. Smit, D. J. H. Deeg, L. M.
Bouter, and P. Lips, "Fall-risk screening test: A prospective study
on predictors for falls in community-dwelling elderly," Journal of
Clinical Epidemiology, vol. 54, pp. 837-844, 8// 2001.
[9] Gillespie LD, Robertson MC, Gillespie WJ, Sherrington C, Gates
S, Clemson LM, et al., "Interventions for preventing falls in older
people living in the community," Cochrane Database of
Systematic Reviews, 2012.
[10] K. Berg, "Measuring balance in the elderly: preliminary
development of an instrument," Physiotherapy Canada, vol. 41,
pp. 304-311, 1989.
[11] S. R. Lord, H. B. Menz, and A. Tiedemann, "A Physiological
Profile Approach to Falls Risk Assessment and Prevention,"
Physical Therapy, vol. 83, pp. 237-252, March 2003 2003.
[12] M. E. Tinetti, T. Franklin Williams, and R. Mayewski, "Fall risk
index for elderly patients based on number of chronic
disabilities," The American Journal of Medicine, vol. 80, pp.
429-434, 1986.
[13] B. R. Greene, A. O’Donovan, R. Romero-Ortuno, L. Cogan, C.
Ni Scanaill, and R. A. Kenny, "Quantitative falls risk assessment
using the timed up and go test," IEEE Trans. Biomed. Eng., vol.
57, pp. 2918-26, 2010.
[14] M. R. Narayanan, S. J. Redmond, M. E. Scalzi, S. R. Lord, B. G.
Celler, and N. H. Lovell, "Longitudinal Falls-Risk Estimation
Using Triaxial Accelerometry," IEEE Trans. Biomed. Eng., vol.
57, pp. 534-541, 2010.
[15] M. Marschollek, A. Rehwald, K. H. Wolf, M. Gietzelt, G.
Nemitz, H. Meyer Zu Schwabedissen, et al., "Sensor-based fall
risk assessment--an expert 'to go'," Methods Inf Med, vol. 50, pp.
420-6, 2011.
[16] D. Giansanti, "Investigation of fall-risk using a wearable device
with accelerometers and rate gyroscopes," Phys. Meas, vol. 27,
pp. 1081-90, 2006.
[17] B. Najafi, K. Aminian, F. Loew, Y. Blanc, and P. A. Robert,
"Measurement of stand-sit and sit-stand transitions using a
miniature gyroscope and its application in fall risk evaluation in
the elderly," IEEE Trans. Biomed. Eng., vol. 49, pp. 843-851,
2002.
[18] A. Weiss, T. Herman, M. Plotnik, M. Brozgol, N. Giladi, and J.
M. Hausdorff, "An instrumented timed up and go: the added
value of an accelerometer for identifying fall risk in idiopathic
fallers," Physiological Measurement, vol. 32, p. 2003, 2011.
[19] J. Howcroft, J. Kofman, and E. D. Lemaire, "Review of fall risk
assessment in geriatric populations using inertial sensors," 2013.
[20] T. Shany, S. J. Redmond, M. R. Narayanan, and N. H. Lovell,
"Sensors-Based Wearable Systems for Monitoring of Human
Movement and Falls," Sensors Journal, IEEE, vol. 12, pp. 658-
670, 2012.
[21] B. R. Greene, E. P. Doheny, C. Walsh, C. Cunningham, L.
Crosby, and R. A. Kenny, "Evaluation of falls risk in community-
dwelling older adults using body-worn sensors " Gerontology,
vol. 58 pp. 472-80, 2012.
[22] B. R. Greene, D. McGrath, and B. Caulfield, "A comparison of
cross-sectional and prospective algorithms for falls risk
assessment," IEEE Engineering in Medicine & Biology
Conference, pp. 4527-4530, 26-30 Aug. 2014.
[23] T. Shany, K. Wang, Y. Liu, N. H. Lovell, and S. J. Redmond,
"Are we stumbling in our quest to find the best predictor? Over-
optimism in sensor-based models for predicting falls in older
adults," Healthcare Technology Letters, vol. 2, 2015.
[24] M. E. Tinetti, M. Speechley, and S. F. Ginter, "Risk factors for
falls among elderly persons living in the community," The New
England journal of medicine, vol. 319, pp. 1701-7, Dec 29 1988.
[25] P. A. Stalenhoef, J. P. M. Diederiks, J. A. Knottnerus, A. D. M.
Kester, and H. F. J. M. Crebolder, "A risk model for the
prediction of recurrent falls in community-dwelling elderly: A
prospective cohort study," Journal of clinical epidemiology, vol.
55, pp. 1088-1094, 2002.
[26] A. Tiedemann, C. Sherrington, T. Orr, J. Hallen, D. Lewis, A.
Kelly, et al., "Identifying older people at high risk of future falls:
development and validation of a screening tool for use in
emergency departments," Emergency Medicine Journal, vol. 30,
pp. 918-922, November 1, 2013 2013.

[27] B. Kirkwood and J. Sterne, Essential Medical Statistics, 2nd
Edition ed.: Wiley-Blackwell, 2003.
[28] S. Mathias, U. Nayak, and B. Isaacs, "Balance in elderly patients:
the "get-up and go" test. ," Arch. Phys. Med. Rehabil., vol. 67,
pp. 387-9, 1986.
[29] D. Podsiadlo and S. Richardson, "The timed "Up & Go": a test of
basic functional mobility for frail elderly persons," J Am Geriatr
Soc, vol. 39, pp. 142-148, 1991.
[30] A. Shumway-Cook, S. Brauer, and M. Woollacott, "Predicting
the probability for falls in community-dwelling older adults using
the Timed Up & Go Test," Phys Ther, vol. 80, pp. 896 - 903,
2000.
[31] G. Thrane, R. Joakimsen, and E. Thornquist, "The association
between timed up and go test and history of falls: The Tromso
study," BMC Geriatrics, vol. 7, p. 1, 2007.
[32] S. L. Whitney, G. F. Marchetti, A. Schade, and D. M. Wrisley,
"The sensitivity and specificity of the Timed "Up & Go" and the
dynamic gait index for self-reported falls in persons with
vestibular disorders," Journal of Vestibular Research, vol. 14,
pp. 397-409, 2004.
[33] Barry E, Galvin R, Keogh C, Horgan F, and F. T., "Is the Timed
Up and Go test a useful predictor of risk of falls in community
dwelling older adults: a systematic review and meta-analysis.,"
BMC Geriatr., vol. 1, p. 14, 2014.
[34] F. Ferraris, U. Grimaldi, and M. Parvis, "Procedure for effortless
in-field calibration of three-axis rate gyros and accelerometers,"
Sens. Mater, vol. 7, pp. 311-330, 1995.
[35] A. G. S. Panel on Prevention of Falls in Older Persons, British
Geriatrics Society, "Summary of the Updated American
Geriatrics Society/British Geriatrics Society Clinical Practice
Guideline for Prevention of Falls in Older Persons," Journal of
the American Geriatrics Society, vol. 59, pp. 148-157, 2011.
[36] T. Hastie, R. Tibshirani, and J. H. Friedman, The Elements of
Statistical Learning, 2nd ed.: Springer, 2009.
[37] L. I. Kuncheva, Combining Pattern Classifiers; Methods and
Algorithms.: Wiley, 2004.
[38] R. Kohavi and G. H. John, "Wrappers for feature subset
selection," Artificial Intelligence, vol. 97, pp. 273-324, 1997.
[39] L. Cattelani, P. Palumbo, L. Palmerini, S. Bandinelli, C. Becker,
F. Chesani, et al., "FRAT-up, a Web-based Fall-Risk Assessment
Tool for Elderly People Living in the Community," Journal of
Medical Internet Research, vol. 17, p. 41, 2015.
[40] Kittler J., Hatef M., Duin R.P.W., and M. J., "On combining
classifiers. IEEE," Trans Patt Anal Mach Intell, vol. 20, pp. 226–
39, 1998.
[41] S. Cummings, M. Nevitt, and S. Kidd, "Forgetting falls. The
limited accuracy of recall of falls in the elderly," J Am Geriatr
Soc, vol. 36, pp. 613 - 616, 1988.
[42] S. E. Lamb, E. C. Jørstad-Stein, K. Hauer, C. Becker, E. on
behalf of the Prevention of Falls Network, and G. Outcomes
Consensus, "Development of a Common Outcome Data Set for
Fall Injury Prevention Trials: The Prevention of Falls Network
Europe Consensus," Journal of the American Geriatrics Society,
vol. 53, pp. 1618-1622, 2005.
[43] P. Palumbo, L. Palmerini, and L. Chiari, "A Probabilistic Model
to Investigate the Properties of Prognostic Tools for Falls,"
Methods of Information in Medicine, vol. 54, pp. 189-197, 2015.