Journals of Gerontology: MEDICAL SCIENCES, 2015, 608–615
doi:10.1093/gerona/glu225
Reaearch Article
Advance Access publication January 7, 2015
Research Article
Ambulatory Fall-Risk Assessment: Amount
and Quality of Daily-Life Gait Predict Falls in
Older Adults
Kimberley S. van Schooten,1 Mirjam Pijnappels,1 Sietse M. Rispens,1
Petra J. M. Elders,2 Paul Lips,3 and Jaap H. van Dieën1
1
MOVE Research Institute Amsterdam, Faculty of Human Movement Sciences, VU University Amsterdam, the
Netherlands. 2EMGO+ Institute, Department of General Practice and Elderly Care and 3MOVE Research Institute
Amsterdam, Department of Internal Medicine, Endocrine section, VU University Medical Center, Amsterdam, the
Netherlands.
Address correspondence to Prof. Dr. Jaap van Dieën, MOVE Research Institute Amsterdam, Faculty of Human Movement Sciences, VU University Amsterdam, Van der Boechorststraat 9, 1081 BT Amsterdam, The Netherlands. Email: [email protected]
Abstract
Background. Ambulatory measurements of trunk accelerations can provide valuable information
on the amount and quality of daily-life activities and contribute to the identification of individuals at
risk of falls. We compared associations between retrospective and prospective falls with potential
risk factors as measured by daily-life accelerometry. In addition, we investigated predictive value
of these parameters for 6-month prospective falls.
Methods. One week of trunk accelerometry (DynaPort MoveMonitor) was obtained in 169 older
adults (mean age 75).The amount of daily activity and quality of gait were determined and validated
questionnaires on fall-risk factors, grip strength, and trail making test were obtained. Six-month
fall incidence was obtained retrospectively by recall and prospectively by fall diaries and monthly
telephone contact.
Results. Among all participants, 35.5% had a history of ≥1 falls and 34.9% experienced ≥1 falls
during 6-month follow-up. Logistic regressions showed that questionnaires, grip strength, and
trail making test, as well as the amount and quality of gait, were significantly associated with
falls. Significant associations differed between retrospective and prospective analyses although
odds ratios indicated similar patterns. Predictive ability based on questionnaires, grip strength,
and trail making test (area under the curve .68) improved substantially by accelerometry-derived
parameters of the amount of gait (number of strides), gait quality (complexity, intensity, and
smoothness), and their interactions (area under the curve .82).
Conclusions. Daily-life accelerometry contributes substantially to the identification of individuals
at risk of falls, and can predict falls in 6 months with good accuracy.
Key Words: Accidental falls—Accelerometry—Activities of daily living—Activity monitoring—Elderly.
Decision Editor: Stephen Kritchevsky, PhD
Falls are among the leading causes of disability in older adults.
Each year, one-third of the adults aged 65 and older experience
at least one fall and 10–15% of these falls result in serious injuries or even death (1,2). Multifactorial interventions, consisting
© The Author 2015. Published by Oxford University Press on behalf of The Gerontological Society of America. All rights reserved. For
permissions, please e-mail: [email protected].
608
Journals of Gerontology: MEDICAL SCIENCES, 2015, Vol. 70, No. 5
of medication management, physical training, and elimination of
hazards in the home environment, can prevent up to 40% of falls
(3,4). However, to efficiently allocate individuals to these fall-prevention interventions, early identification of those at risk of falling
is necessary.
Fall-risk assessment in the clinical setting is commonly based on
questionnaires or functional tests. Questionnaires are often qualitative and can be biased by fall history, while functional tests, such as
the Timed-Up and Go or the Berg Balance Scale, are objective but
appear to lack responsiveness and discriminative ability in relatively
healthy populations (5). A promising method to assess intrinsic fall
risk is the evaluation of gait dynamics. Previous studies have shown
that walking speed, stride frequency, movement intensity, stride-tostride variability, index of harmonicity, harmonic ratios, entropy, and
local dynamic stability differentiate fallers and nonfallers (6–12).
However, studies have often used fall history to identify fallers, while
recall difficulties, sustained injuries, or fear of falling might bias these
retrospective designs (13,14). Gait characteristics associated with fall
history may, therefore, not be related to actual fall risk but could be
indicators of adaptations after a fall. In this study, we compared the
associations of questionnaires and gait characteristics with falls in
both a retrospective as well as a prospective manner in order to identify true fall-risk indicators.
Physical activity is generally encouraged in fall prevention
interventions to increase physical capacity and reduce fall risk
(3,4,15,16). Although an increase in physical activity has indeed
been shown to reduce the risk of falls in community-dwelling older
adults, it actually increased the risk of falls in a more frail population (3,4). This suggests that the amount of physical activity reflects
exposure to environmental and task-related hazards (17). Most fallrisk assessments neglect this extrinsic fall risk, which might explain
why current clinical prediction models provide only poor to fair
predictive ability (18–20). The exposure to environmental hazards,
such as curbs, carpets and pets, can only be investigated in daily life.
Therefore, measurements during daily life may offer more accurate
estimates of fall risk. Trunk accelerations during daily-life activities
can be used to infer types and durations of activities such as lying,
sitting, standing, and locomotion in a valid and reliable manner
(21,22). These trunk accelerations can also be used to obtain gait
characteristics under daily-life circumstances (11,12). To our knowledge, two studies addressed the association between accelerometerbased gait characteristics obtained in daily life and prospective falls
(11,23). The number of fallers in the prospective part of these studies was small, and possibly as a consequence of that, only spectral
parameters related to stride-to-stride variability were reported to be
associated with fall risk, while prospective lab-based studies also
identified walking speed, stride frequency, harmonic ratio, and gait
intensity as fall-risk indicators (9,10). Hence, we investigated the
predictive value of parameters describing the amount and quality of
gait to discriminate between fallers and nonfallers in a larger cohort
and determined to what extent fall prediction models based on questionnaires, grip strength, and trail making test could be improved
by incorporating gait characteristics as obtained from accelerometry
during daily life.
Methods
Participants
This study was part of a larger project concerning fall-risk
assessment in older adults (FARAO) performed at the VU
University Amsterdam. Participants were community dwelling or
609
living in a residential home and were recruited in Amsterdam (the
Netherlands) and surroundings via general practitioners, pharmacies, hospitals, and residential care facilities. Participants were
included if they were between 65 and 99 years of age, their Mini
mental state exam score [MMSE (24)] exceeded 18 out of 30 points
(actual range 21–30), and they were able to walk at least 20 m
(with walking aid if needed). A total of 169 older individuals were
included for the analyses. The medical ethical committee of the VU
medical centre approved the protocol and all participants signed
informed consent.
Fall Incidence and Clinically Used Risk Factors
Fall history was obtained as the self-reported number of falls in the
past 6 months. Prospective fall incidence was obtained by monthly
telephone contact in addition to a daily fall diary. Participants were
identified as retrospective and prospective fallers based on the number of falls (≥1) in the 6 months preceding and following the measurements, respectively.
Descriptive characteristics such as age, weight, height, and
the use of a walking aid were obtained from all participants. In
addition, participants were asked to complete validated questionnaires and tests on fall-risk factors [LASA fall-risk profile (25)],
cognitive function (MMSE), executive functioning [trail making
test parts A & B; TMT-A & TMT-B (26)], fear of falling [16-item
fall efficacy scale; FES-I (27)], and depression [30-item geriatric
depression scale; GDS (28)]. The LASA fall-risk profile comprised
questions concerning dizziness, independence in daily life, having
pets, alcohol consumption, education duration and required the
measurement of hand grip strength, which was quantified using a
handgrip dynamometer (TKK 5401, Takei Scientific Instruments,
Tokyo, Japan).
Ambulatory Assessment
Participants wore a tri-axial accelerometer (DynaPort MoveMonitor,
McRoberts, The Hague, the Netherlands) with a sampling frequency
of 100 Hz and a range from −6 g to +6 g, for 8 consecutive days.
This accelerometer was placed dorsally on the trunk at the level of
L5 using an elastic belt. The participants were instructed to wear the
accelerometer at all times, except during aquatic activities such as
showering. On the first and last day of the measurement period, the
accelerometers were delivered to and collected from the participant’s
home. Hence, the first and last 6 hours of the measurements were
omitted from analysis to discard any possible artefacts caused by
transportation.
Amount of physical (in-)activity.—Based on accelerometry, periods
of nonwearing, locomotion, sitting, lying, and standing were identified using the manufacturer’s algorithm. This algorithm is described
in more detail in the supplementary material of Dijkstra and colleagues (21). We calculated the total duration of locomotion, sitting,
standing, and lying per day. In addition, we calculated the amount of
strides, the average number of short periods of locomotion, henceforth called walking bouts, the median and maximum duration of
walking bouts, and the number of transitions to stance per day.
These estimates were averaged over days on which the accelerometer was worn for more than 75% of the time, to obtain reliable, and
valid indicators of the amount of daily activity (22). This criterion
was complied by 91% of the measurement days, and accelerometrydata of five participants had to be excluded due to noncompliance
with this criterion.
610
Gait characteristics.—Gait characteristics were estimated using
MATLAB R2011a (MathWorks, Natick, MA) for all locomotion
bouts that exceeded 10 seconds. Data of four additional participants was excluded (excluding in total nine participants for gait
characteristics) since they did not exhibit sufficient (>50) locomotion bouts exceeding 10 seconds (12). The locomotion bouts were
cut into 10-second windows and median characteristics over all
10-second windows were used to improve stationarity and avoid
effects of differences in data series length (29–31). Prior to estimation of gait characteristics, the raw accelerations were realigned
with the anatomical axes based on the sensor’s orientation with
respect to gravity (32) and optimization of the left–right symmetry
(33).
The gait characteristics estimated in this study were selected
based on previously established associations with fall risk (6,12)
and algorithms to estimate these parameters are described in more
detail by Rispens and colleagues (12) and the references below. We
estimated walking speed (34), stride time, stride length (34); gait
intensity expressed as the standard deviation and range; gait variability expressed as the variability in walking speed, stride time, and
stride length, the autocorrelation at the dominant period (35), the
magnitude (power) and width of the dominant peak in the frequency
domain (10,11) and the percentage of power below 0.7 Hz (12);
gait symmetry expressed as the harmonic ratio (9,36); gait smoothness expressed as the index of harmonicity (37); and gait complexity
expressed as the mean logarithmic rate of divergence per stride based
on 10-samples delayed embedding in seven dimensions using Wolf’s
method (38) and sample entropy with embedding dimension 5 and
tolerance .3 (39). All characteristics were determined for each of the
three directions of acceleration, that is, anteroposterior (AP), mediolateral (ML), and vertical (VT).
Statistical Analysis
Statistical analysis was performed using SPSS Statistics 21 (IBM,
Armonk, NY). To facilitate interpretations, all continuous parameters were transformed to z-scores prior to analyses. We used univariate logistic regressions to identify parameters that were associated
with either or both retrospective or prospective falls and had a
p < .05. Due to the exploratory aim of this study on a relatively
large number of parameters, no correction for repeated testing was
employed. Subsequently, three stepwise, forward, logistic regression analyses predicting prospective falls were performed including
all participants with complete data; first to establish the predictive
value of commonly used questionnaires, trail making test, and grip
strength, secondly to investigate the predictive value of accelerometry-derived characteristics, and thirdly to investigate their combined predictive value. All parameters were considered candidates
Journals of Gerontology: MEDICAL SCIENCES, 2015, Vol. 70, No. 5
and parameters were entered in the prediction model until p ≥
.10. Highly correlated parameters (Spearman’s ρ ≥ .7) were not
entered together but separate prediction models were preformed.
Interactions between parameters in the model that were deemed
plausible, that is between the amount of physical activity and quality of gait, were tested by adding the product of these parameters
to the model. Prediction models with the highest area under the
receiver operating curve, c-statistic, are reported. A Hosmer–
Lemeshow test was used to test for miscalibration between the
models’ predicted probability and actual fall incidences. Sensitivity,
specificity, and positive and negative predictive values are reported
for the cutoff value optimizing sensitivity plus specificity − 1, that
is, Youden’s index. All three prediction models were compared using
Hanley and McNeil test (40) for areas under the receiver operating
curve derived from the same sample.
Results
Of the 169 participants, 35.5% had a history of falling and 34.9%
experienced falls during follow-up. Twenty-eight out of 169 participants experienced multiple falls during follow-up. The distribution
of fall history and prospective falls over participants are shown in
Figure 1 and descriptive characteristics of the participants are summarized in Table 1.
Having a history of falls was significantly associated with the
inability to use public transportation, a lower grip strength, a higher
fear of falling, a higher depression score, having a walking aid, a
lower number of strides per day, a lower total duration of daily
locomotion, and a higher power in the dominant frequency in ML
direction (for significant associations see Table 2, for a list of all
associations consult Supplementary Material).
Although the overall pattern of the odds ratios was similar,
associations were more often statistically significant in prospective analysis. The odds of experiencing a fall during follow-up was
significantly higher for individuals with a history of falls, a higher
LASA fall-risk score, higher fear of falling, a higher depression score,
a longer duration on the TMT-A, a lower walking speed, a lower
stride frequency, shorter stride length, a lower standard deviation in
VT and AP directions, a lower range of the signal in the VT direction, a lower harmonic ratio in VT and AP directions, a higher index
of harmonicity in AP direction, and a higher logarithmic rate of
divergence, that is, lower local stability, in VT, ML, and AP directions (Table 2; Supplementary Material).
The best prediction model for prospective falls based on questionnaires, trail making test, and grip strength comprised, in order
of variance accounted for, fall history and depression score (Table 3,
model 1). The area under the receiver operating curve was .68 (95%
4 7 .9 %
N o h is t o r y o f fa lls , n o p r o s p e c tiv e fa lls
1 6 .6 %
N o h is to r y o f fa lls , p r o s p e c t iv e fa lls
1 7 .2 %
H is to r y o f fa lls , n o p r o s p e c t iv e fa lls
1 8 .3 %
H is to r y o f fa lls , p r o s p e c tiv e fa lls
Figure 1. Percentages of participants reporting 6-months retrospective and/or prospective falls during 6 months follow-up.
Journals of Gerontology: MEDICAL SCIENCES, 2015, Vol. 70, No. 5
CI = .60–.77) and the predicted probability ranged from .209 to
.736. There was no significant miscalibration between the models’
predicted probability and actual fall incidences (p = .87). The optimal cutoff of predicted probability was .279 and corresponding
sensitivity and specificity were 57.7% and 72.4% and the positive
and negative predictive values were 48.7% and 78.9%, respectively.
The prediction model with solely accelerometry-derived parameters
comprised local dynamic stability in AP direction, range of the signal
in VT direction, total number of strides, duration of lying, and the
interaction between range of the signal in VT direction and the total
Table 1. Descriptive Characteristics (means and SD) of the 169 Participants
Descriptives
Age (y)
Gender (% female)
Weight (kg)
Height (cm)
Mini mental state exam, MMSE (score)
Home situation (% independent)
6-months history of falls (number)
6-months prospective falls (number)
75.4 (6.8)
52.1
74.1 (14.2)
170.8 (8.9)
27.7 (2.2)
88.1
0.6 (1.1)
0.6 (1.0)
Note: Participants were mainly community-dwelling and requited from Amsterdam and surroundings.
611
number of strides (Table 3, model 2). The area under the receiver
operating curve was .71 (95% CI = .63–.79), which was comparable to the model with fall history and depression score (p = .55).
The predictive probability ranged from .014 to .885 and there was
no significant miscalibration between the model and the actual fall
incidences (p = .68). The optimal cutoff of predicted probability was
.352 and corresponding sensitivity and specificity were 67.9% and
66.3% and the positive and negative predictive values were 52.1%
and 79.3%, respectively. When all parameters were combined, the
best prediction model comprised fall history, local dynamic stability
in AP direction, range of the signal in VT direction, total number
of strides, depression score, index of harmonicity in ML direction,
sample entropy in VT direction, the interaction between range of
the signal in VT direction and the total number of strides, and the
interaction between index of harmonicity in ML direction and the
total number of strides (Table 3, model 3). The resulting predicted
probability ranged from .017 to .979, and the area under the receiver
operating curve was .82 (95% CI = .75–.89). There was no significant miscalibration between the model and actual fall incidences
(p = .53). The optimal cutoff of predicted probability was .396 and
corresponding sensitivity and specificity were 70.0% and 80.9% and
the positive and negative predictive values were 66.0% and 82.6%,
respectively. The receiver operating curves for all three prediction
models are shown in Figure 2. The prediction model with both questionnaires and accelerometry was significantly better than the separate models (both p < .001).
Table 2. Prevalence or Mean (SD) Values of Parameters Significantly Associated With Retrospective or Prospective Falls and Odds Ratio
(95% confidence interval) of Univariate Associations
Retrospective Falls OR (95% CI)
Questionnaires
6-months history of falls
Inability to use public transportation
Grip strength (kg)
Use of a walking aid
Fall efficacy scale, FES-I (score)
LASA fall-risk profile (score)
Trail making test, TMT-A duration ≥45 s
Geriatric depression score, GDS (score)
Daily activities
Number of strides (N)
Locomotion duration (h)
Gait characteristics
Walking speed (m/s)
Stride frequency (hz)
Stride length (m)
Standard deviation VT (m/s2)
Standard deviation AP (m/s2)
Range VT (m/s2)
Power at dominant frequency ML (psd)
Harmonic ratio VT
Harmonic ratio AP
Index of harmonicity AP
Logarithmic divergence rate VT (/stride)
Logarithmic divergence rate ML (/stride)
Logarithmic divergence rate AP (/stride)
Prospective Falls OR (95% CI)
Prevalence or Mean (SD)
Fallers N = 60, nonfallers N = 109 Fallers N = 59, nonfallers N = 110
35.5%
7.1%
58.4 (20.5)
17.2%
4.8 (4.4)
5.5 (4.7)
50.0%
4.6 (4.2)
N/A
4.12 (1.15–14.68)*
0.61 (0.42–0.88)**
2.26 (1.00–5.17)*
1.08 (1.02–1.14)**
N/A
1.90 (0.99–3.64)
1.83 (1.29–2.61)**
3.09 (1.57–6.09)**
2.80 (0.83–9.48)
0.97 (0.69–1.36)
1.40 (0.61–3.22)
1.06 (1.00–1.12)*
1.56 (1.11–2.19)**
2.02 (1.05–3.89)*
1.55 (1.11–2.18)**
3060 (1527)
1.20 (0.53)
0.69 (0.48–1.00)*
0.71 (0.50–1.00)*
0.92 (0.66–1.29)
0.75 (0.70–1.35)
0.87 (0.22)
0.87 (0.08)
1.11 (0.20)
1.84 (0.87)
1.31 (0.39)
11.80 (5.60)
0.37 (0.17)
2.25 (0.62)
1.84 (0.41)
0.71 (0.08)
1.71 (0.42)
2.02 (0.31)
1.91 (0.31)
0.74 (0.50–1.09)
0.85 (0.60–1.19)
0.78 (0.56–1.10)
0.66 (0.36–1.20)
0.78 (0.52–1.18)
0.69 (0.39–1.22)
1.46 (1.05–2.05)*
0.90 (0.64–1.29)
0.75 (0.53–1.06)
1.16 (0.82–1.62)
1.26 (0.90–1.77)
0.86 (0.62–1.20)
1.17 (0.84–1.63)
0.64 (0.42–0.95)*
0.67 (0.46–0.97)*
0.70 (0.50–0.99)*
0.39 (0.19–0.77)**
0.61 (0.38–0.96)*
0.42 (0.22–0.82)**
1.06 (0.76–1.47)
0.67 (0.46–0.95)*
0.68 (0.48–0.97)*
1.45 (1.02–2.07)*
1.48 (1.05–2.09)*
1.41 (1.00–2.00)*
1.65 (1.15–2.36)**
Notes: AP = anteroposterior direction; ML = mediolateral direction; N/A = not applicable; VT = vertical direction. Continuous parameters are z-transformed.
*p < 0.05, **p < 0.01.
Journals of Gerontology: MEDICAL SCIENCES, 2015, Vol. 70, No. 5
612
Table 3. Loading of Predictors, Beta (standard error), in a Fall Prediction Model With Only Questionnaires (model 1); Only Accelerometry
(model 2); or Their Combination (model 3)
Predictor
Constant
6-months history of falls
Logarithmic divergence rate AP
Range VT
Number of strides
Total duration of lying
GDS
Index of harmonicity ML
Sample entropy VT
Range VT × Number of strides
Index of harmonicity ML × N umber of strides
Model 1
Model 2
Model 3
B (SE)
B (SE)
B (SE)
−0.98 (0.23)
0.97 (0.36)
N/A
N/A
N/A
N/A
0.32 (0.18)
N/A
N/A
N/A
N/A
−0.90 (0.20)
N/A
0.58 (0.22)
−1.18 (0.40)
0.34 (0.23)
−0.35 (0.20)
N/A
−1.21 (0.30)
1.32 (0.47)
0.90 (0.30)
−1.80 (0.56)
0.65 (0.29)
0.34 (0.17)
0.64 (0.26)
−0.64 (0.28)
0.38 (0.23)
0.69 (0.22)
0.74 (0.30)
Notes: AP = anteroposterior direction; ML = mediolateral direction; N/A = not applicable; VT = vertical direction. Continuous parameters are z-transformed. Positive betas indicate a positive relation between the predictor and falls.
ROC for logistic prediction model
100
90
80
70
Sensitivity
60
50
40
30
20
10
0
Only questionnaires
Only accelerometry
Questionnaires and accelerometry
0
20
40
60
80
100
100−Specificity
Figure 2. Receiver operating curves (ROC) for the three fall prediction models. The model with questionnaires and accelerometry [area under the curve (AUC) of
.82] outperformed both other models (AUCs of 0.68 and 0.71; p < .001).
Discussion
We determined associations of fall history and falls during follow-up
in older adults with questionnaires, trail making test, grip strength,
physical activity, and gait characteristics as determined from dailylife trunk accelerometry. In addition, we investigated the added value
of these accelerometry-derived measures for prospective fall prediction over commonly used questionnaires, trail making test, and grip
strength.
Most of the risk indicators obtained with questionnaires, trail
making test, or grip strength were significantly associated with falls.
Having a history of falls, a high LASA fall-risk score, fear of falling, poor executive function (as assessed by TMT-A), and depressive
symptoms were associated with prospective falls, as was expected
based on literature (25,26,41). The associations between the ability
to use public transportation, the use of a walking aid, and grip
strength were only significant when investigated retrospectively. An
explanation could be that these parameters are affected by the occurrence of a fall, possibly by a subsequent decline in the amount of
daily activity as is suggested by the observed association with locomotion duration. Such declines in activity after a fall may in turn
have negative health effects (14,42,43).
Of the accelerometer-derived measures, only the amount of locomotion and gait variability as expressed by the power at the dominant frequency of the ML accelerations were significantly associated
with fall history. A higher power at the dominant frequency in ML
direction for persons with a history of falls was also observed by
Weiss and colleagues (11), and—although not significantly—Rispens
and colleagues (12) and Weiss and colleagues (23), and might
Journals of Gerontology: MEDICAL SCIENCES, 2015, Vol. 70, No. 5
indicate a more rigid, stereotypical gait pattern. The power at the
dominant frequency in ML direction was not significantly associated
with prospective falls, which suggests that this adaptation decreased,
or at least did not increase, prospective fall risk. Unfortunately,
Weiss and colleagues (11,23) did not report on the association of
this parameter with prospective falls, rendering comparison impossible. Previously, we (12) and others (11,23) identified associations
between fall history and index of harmonicity, local dynamic stability, power at the dominant frequency, width of the dominant frequency and harmonic ratios derived from daily-life accelerometry,
while in the current study these did not reach significance. This
might partially be explained by a slightly younger and less frail population in the current study, differences in statistical techniques, and
recall periods but might also suggest that different characteristics
may prove to be valuable in subgroups of older adults.
Several gait characteristics were significantly associated with
prospective falls. Participants who walked slower at a lower stride
frequency and with shorter strides were more likely to fall, in
accordance with literature (6,7). In addition, a lower standard deviation and range of the signal, two measures of intensity of locomotion, were also associated with prospective falls. As these measures
are moderately to highly correlated with walking speed (Pearson’s
r = .60–.85), they might reflect the influence of walking speed and
relate to the adoption of a more conservative walking pattern to
maintain balance (42), or to a common underlying factor, such as
a less vigorous movement pattern, possibly reflecting one’s general
health status (44). In contrast with Weiss and collagues (11,23),
the width of the dominant frequency, a measure of gait variability,
was not associated with prospective falls. A possible explanation
may be differences in gait detection algorithms and durations of
the analyzed bouts or the group under study. A lower gait symmetry, that is, lower harmonic ratio, was also associated with future
falls, which is in line with previous findings in a laboratory study by
Doi and colleagues (9). The association with prospective falls was
not significant for harmonic ratio in the ML direction, although the
odds ratio was comparable to those of the VT and AP directions.
Possibly, adaptations to increase balance occurred mainly in ML
direction or changes in progression direction, which will affect ML
trunk accelerations, effaced effects in ML direction. Unexpectedly,
a higher index of harmonicity in AP direction, indicating smoother
gait, was significantly associated with prospective falls, which might
reflect a more fluent, less adaptive, and forward progression. This
parameter in ML direction interacted with the amount of locomotion, which will be discussed later. Finally, in accordance with
previous retrospective findings (7,8,12) a higher logarithmic rate
of divergence, that is, lower local stability, was associated with prospective falls. This might indicate less and slower responses to small
perturbations caused by, for example, noise in the neuromuscular
system or uneven flooring (45).
The best prediction model for future falls comprised the amount
of gait, gait characteristics describing complexity, intensity, and
smoothness of gait, and the interaction between these gait characteristics. It outperformed the models with only questionnaires or only
accelerometry and its area under the curve (AUC) of .82 is very high
in comparison with other fall risk assessment methods, which generally achieve AUCs ranging from .55 to .74 (18–20), suggesting that
daily-life trunk accelerometry contributed substantially to the ability
to predict falls. We also investigated the predictive ability of gait
characteristics obtained from the Euclidian or summed power of the
three acceleration directions. The univariate associations were comparable to those of the three separate acceleration directions and the
613
orientation invariant measures were not favored over the separate
directions for the prediction model.
The amount of physical activity was not univariately associated with prospective falls, but was associated with prospective
falls after correcting for gait quality. Schwenk and colleagues (46)
showed that a shorter duration of gait bouts was predictive for falls
in older adults with dementia and suggested that shorter bout may
reflect disorientation, wandering, or dual task difficulties. However,
the exclusion of older adults with dementia in the current renders
direct comparison impossible. The number of strides did interact
with gait smoothness and intensity. These interactions illustrate that
the amount of physical activity is a protective factor for persons with
higher gait intensity or a smoother gait pattern in ML direction,
while it is a risk factor for persons with lower gait intensity and has
no effect for persons with a less smooth gait pattern. Overall, these
results suggest that for persons with an energetic, smooth walking
pattern, walking poses no threat, while for persons with a less vigorous and less smooth walking pattern, more walking, that is, more
exposure to environmental and task-related hazards, increases fall
risk. Future research should focus on this interaction, as this might
provide insights allowing more appropriate, individualized fall prevention interventions.
Our study has some limitations that have to be taken into
account. There was a large number of parameters available for the
prediction models. Although we tried to control for overfitting by a
forward selection procedure and despite the high agreement between
the three prediction models, overfitting may have occurred. Hence
replication of the results of this study is required. Furthermore, we
employed a fall history and follow up period of 6 months, while
1 year is generally employed. We did also obtain and analyzed 1-year
recall of fall incidences and although associations were slightly
weaker, our conclusions for the retrospective analysis would not
have been different. It remains to be shown whether accelerometry-derived measures are equally predictive for falls during 1 year
follow-up. Moreover, participants in the current study were mainly
community-dwelling older adults, which might limit generalizability
of the prediction model to frail or diseased populations. However,
stronger associations with retrospective falls in an older and more
frail group (12) suggest that adaptations to maintain balance are less
successful in this population and hence stronger associations with
prospective falls could be expected. Nevertheless, external validation of predictive ability of fall prediction models comprising quality
and amount of daily-life gait in other populations is necessary. In
addition, we analysed all locomotion episodes exceeding 10 seconds,
which may have comprised other cyclic activities than walking, such
as stair negotiation, as these are difficult to differentiate based on
accelerometry (21). Moreover, gait in daily life is often not at a constant velocity or straight-lined, while stationarity is an assumption
underlying the calculation of many of the employed gait measures.
To overcome this, we analyzed locomotion episodes in 10-second
windows and used the median over all windows. Extreme values
instead of medians, although more prone to artefacts, might be more
indicative of high fall risk. Gait detection algorithms can be further
improved and the use of more extreme values of characteristics for
the prediction of falls should be explored. In addition, predictive
ability of a fall-risk model including daily-life accelerometry might
be further improved by extracting information during other types of
activities, such as transfers, and using more advanced methods such
as support vector machines or learning vector quantization, which
do not assume linear relations between parameters and allow for
complex interactions.
614
In conclusion, parameters significantly associated with falls differed between retrospective and prospective analyses, although odds
ratios indicated a similar pattern of associations. Prospective falls
can be predicted with good accuracy based on questionnaires and
daily-life accelerometry, to which parameters describing quality and
quantity of daily-life gait contribute substantially.
Supplementary Material
Supplementary material can be found at: http://biomedgerontology.
oxfordjournals.org/
Funding
Kim van Schooten, Sietse Rispens, and Mirjam Pijnappels were
financially supported by a TOP-NIG grant (#91209021) from the
Dutch Organization for Scientific Research (NWO).
References
1. Milat AJ, Watson WL, Monger C, Barr M, Giffin M, Reid M. Prevalence,
circumstances and consequences of falls among community-dwelling older
people: results of the 2009 NSW Falls Prevention Baseline Survey. N S W
Public Health Bull. 2011;22:43–48.
2. Salvà A, Bolíbar I, Pera G, Arias C. Incidence and consequences of
falls among elderly people living in the community. Med Clin (Barc).
2004;122:172–176.
3. Gillespie LD, Robertson MC, Gillespie WJ, et al. Interventions for preventing falls in older people living in the community. Cochrane Database Syst
Rev. 2009;2:CD007146.
4. Cameron ID, Murray GR, Gillespie LD, et al. Interventions for preventing falls in older people in nursing care facilities and hospitals. Cochrane
Database Syst Rev. 2010;1:CD005465.
5. Hamacher D, Singh NB, Van Dieën JH, Heller MO, Taylor WR. Kinematic
measures for assessing gait stability in elderly individuals: a systematic
review. J R Soc Interface. 2011;8:1682–1698.
6. Howcroft J, Kofman J, Lemaire ED. Review of fall risk assessment in geriatric populations using inertial sensors. J Neuroeng Rehabil. 2013;10:91.
7. Lockhart TE, Liu J. Differentiating fall-prone and healthy adults using
local dynamic stability. Ergonomics. 2008;51:1860–1872.
8. Toebes MJ, Hoozemans MJ, Furrer R, Dekker J, van Dieën JH. Local
dynamic stability and variability of gait are associated with fall history in
elderly subjects. Gait Posture. 2012;36:527–531.
9. Doi T, Hirata S, Ono R, Tsutsumimoto K, Misu S, Ando H. The harmonic
ratio of trunk acceleration predicts falling among older people: results of
a 1-year prospective study. J Neuroeng Rehabil. 2013;10:7.
10.Marschollek M, Rehwald A, Wolf KH, et al. Sensor-based fall risk assessment—an expert ‘to go’. Methods Inf Med. 2011;50:420–426.
11. Weiss A, Brozgol M, Dorfman M, et al. Does the evaluation of gait quality
during daily life provide insight into fall risk? A novel approach using 3-day
accelerometer recordings. Neurorehabil Neural Repair. 2013;27:742–752.
12.Rispens S, van Schooten KS, Pijnappels M, Daffertshofer A, Beek PJ, Van
Dieën JH. Identification of fall risk predictors in daily life measurements:
gait characteristics’ reliability and association with self-reported fall history. Neurorehabil Neural Repair. April 23, 2014. Epub ahead of print.
13.Cummings SR, Nevitt MC, Kidd S. Forgetting falls. The limited accuracy
of recall of falls in the elderly. J Am Geriatr Soc. 1988;36:613–616.
14.Maki BE. Gait changes in older adults: predictors of falls or indicators of
fear. J Am Geriatr Soc. 1997;45:313–320.
15.Peeters GM, van Schoor NM, Pluijm SM, Deeg DJ, Lips P. Is there a
U-shaped association between physical activity and falling in older persons? Osteoporos Int. 2010;21:1189–1195.
16.Graafmans WC, Lips P, Wijlhuizen GJ, Pluijm SM, Bouter LM. Daily
physical activity and the use of a walking aid in relation to falls in elderly
people in a residential care setting. Z Gerontol Geriatr. 2003;36:23–28.
Journals of Gerontology: MEDICAL SCIENCES, 2015, Vol. 70, No. 5
17.Skelton DA. Effects of physical activity on postural stability. Age Ageing.
2001;30(suppl 4):33–39.
18.Gates S, Smith LA, Fisher JD, Lamb SE. Systematic review of accuracy of
screening instruments for predicting fall risk among independently living
older adults. J Rehabil Res Dev. 2008;45:1105–1116.
19.Scott V, Votova K, Scanlan A, Close J. Multifactorial and functional
mobility assessment tools for fall risk among older adults in community, home-support, long-term and acute care settings. Age Ageing.
2007;36:130–139.
20.Lee J, Geller AI, Strasser DC. Analytical review: focus on fall screening
assessments. PM R. 2013;5:609–621.
21.Dijkstra B, Kamsma Y, Zijlstra W. Detection of gait and postures using a
miniaturised triaxial accelerometer-based system: accuracy in communitydwelling older adults. Age Ageing. 2010;39:259–262.
22. van Schooten KS, Rispens SM, Petra J, Lips P, van Dieën JH, Pijnappels M.
Assessing physical activity in older adults: required days of trunk accelerometer measurements for reliable estimation. J Aging Phys Act. December
4, 2013. Epub ahead of print.
23.Weiss A, Herman T, Giladi N, Hausdorff JM. Objective assessment of
fall risk in Parkinson’s disease using a body-fixed sensor worn for 3 days.
PLoS One. 2014;9:e96675.
24.Folstein MF, Folstein SE, McHugh PR. “Mini-mental state”: a practical
method for grading the cognitive state of patients for the clinician. J Psychiatr Res. 1975;12:189–198.
25.Tromp AM, Pluijm SM, Smit JH, Deeg DJ, Bouter LM, Lips P. Fall-risk
screening test: a prospective study on predictors for falls in communitydwelling elderly. J Clin Epidemiol. 2001;54:837–844.
26.Holtzer R, Friedman R, Lipton RB, Katz M, Xue X, Verghese J. The relationship between specific cognitive functions and falls in aging. Neuropsychology. 2007;21:540–548.
27. Yardley L, Beyer N, Hauer K, Kempen G, Piot-Ziegler C, Todd C. Development and initial validation of the Falls Efficacy Scale-International (FES-I).
Age Ageing. 2005;34:614–619.
28.Yesavage JA, Brink TL, Rose TL, et al. Development and validation of a
geriatric depression screening scale: a preliminary report. J Psychiatr Res.
1983;17:37–49.
29.van Schooten KS, Rispens SM, Elders PJ, van Dieën JH, Pijnappels M.
Toward ambulatory balance assessment: estimating variability and stability from short bouts of gait. Gait Posture. 2014;39:695–699.
30.England SA, Granata KP. The influence of gait speed on local dynamic
stability of walking. Gait Posture. 2007;25:172–178.
31.Bruijn SM, van Dieën JH, Meijer OG, Beek PJ. Statistical precision and
sensitivity of measures of dynamic gait stability. J Neurosci Methods.
2009;178:327–333.
32.Moe-Nilssen R. A new method for evaluating motor control in gait under
real-life environmental conditions. Part 1: the instrument. Clin Biomech
(Bristol, Avon). 1998;13:320–327.
33.Rispens SM, Pijnappels M, van Schooten KS, Beek PJ, Daffertshofer A,
van Dieën JH. Consistency of gait characteristics as determined from
acceleration data collected at different trunk locations. Gait Posture.
2014;40:187–192.
34.Zijlstra W, Hof AL. Assessment of spatio-temporal gait parameters from
trunk accelerations during human walking. Gait Posture. 2003;18:1–10.
35.Moe-Nilssen R, Helbostad JL. Estimation of gait cycle characteristics by
trunk accelerometry. J Biomech. 2004;37:121–126.
36. Yack HJ, Berger RC. Dynamic stability in the elderly: identifying a possible
measure. J Gerontol. 1993;48:M225–M230.
37.Lamoth CJ, Beek PJ, Meijer OG. Pelvis-thorax coordination in the transverse plane during gait. Gait Posture. 2002;16:101–114.
38.Wolf A, Swift JB, Swinney HL, Vastano JA. Determining Lyapunov exponents from a time-series. Physica D. 1985;16:285–317.
39.Richman JS, Moorman JR. Physiological time-series analysis using
approximate entropy and sample entropy. Am J Physiol Heart Circ Physiol. 2000;278:H2039–H2049.
40.Hanley JA, McNeil BJ. A method of comparing the areas under receiver
operating characteristic curves derived from the same cases. Radiology.
1983;148:839–843.
Journals of Gerontology: MEDICAL SCIENCES, 2015, Vol. 70, No. 5
41.Pluijm SM, Smit JH, Tromp EA, et al. A risk profile for identifying community-dwelling elderly with a high risk of recurrent falling: results of a
3-year prospective study. Osteoporos Int. 2006;17:417–425.
42.Menz HB, Lord SR, Fitzpatrick RC. Age-related differences in walking
stability. Age Ageing. 2003;32:137–142.
43.Delbaere K, Crombez G, Vanderstraeten G, Willems T, Cambier D. Fearrelated avoidance of activities, falls and physical frailty. A prospective
community-based cohort study. Age Ageing. 2004;33:368–373.
615
44. Studenski S, Perera S, Patel K, et al. Gait speed and survival in older adults.
JAMA. 2011;305:50–58.
45.Dingwell JB, Cusumano JP. Nonlinear time series analysis of normal and
pathological human walking. Chaos. 2000;10:848–863.
46.Schwenk M, Hauer K, Zieschang T, Englert S, Mohler J, Najafi B. Sensorderived physical activity parameters can predict future falls in people with
dementia. Gerontology. 2014;60:483–492.