Newcastle University ePrints - eprint.ncl.ac.uk
Del Din S, Godfrey A, Rochester L. Validation of an accelerometer to quantify
a comprehensive battery of gait characteristics in healthy older adults and
Parkinson's disease: toward clinical and at home use. IEEE Journal of
Biomedical and Health Informatics 2015, (99).
Copyright:
© 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is
also permitted, but republication/redistribution requires IEEE permission.
DOI link to article:
http://dx.doi.org/10.1109/JBHI.2015.2419317
Date deposited:
02/06/2015
2168-2194 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2015.2419317, IEEE Journal of Biomedical and Health Informatics
1
Abstract—Measurement of gait is becoming important as a
tool to identify disease and disease progression, yet to date its
application is limited largely to specialist centres. Wearable
devices enables gait to be measured in naturalistic
environments however questions remain regarding validity.
Previous research suggests that when compared with a
laboratory reference, measurement accuracy is acceptable for
mean but not variability or asymmetry gait characteristics.
Some fundamental reasons for this have been presented (e.g.
synchronisation, different sampling frequencies) but to date this
has not been systematically examined. The aims of this study
were to: (i) quantify a comprehensive range of gait
characteristics measured using a single tri-axial accelerometer-
based monitor, (ii) examine outcomes and monitor performance
in measuring gait in older adults and those with Parkinson’s
disease (PD) and (iii) carry out a detailed comparison with those
derived from an instrumented walkway to account for any
discrepancies. Fourteen gait characteristics were quantified in
30 people with incident PD and 30 healthy age-matched
controls. Of the 14 gait characteristics compared, agreement
between instruments was excellent for 4 (ICCs 0.913 – 0.983);
moderate for 4 (ICCs 0.508 – 0.766); and poor for 6
characteristics (ICCs -0.637 – 0.370). Further analysis revealed
that differences reflect an increased sensitivity of accelerometry
to detect motion, rather than measurement error. This is most
likely because accelerometry measures gait as a continuous
activity rather than discrete footfall events, per instrumented
tools. The increased sensitivity shown for these characteristics
will be of particular interest to researchers keen to interpret
‘real world’ gait data. In conclusion, use of a body worn
monitor is recommended for the measurement of gait but is
likely to yield more sensitive data for asymmetry and variability
features.
Index Terms — Accelerometer, algorithm, body worn monitor,
Instrumented gait, instrumented walkway.
I. INTRODUCTION
Gait is emerging as a powerful tool in neurodegenerative
disease to identify surrogate markers of incipient disease
manifestation or disease progression [1-5].
However, its widespread adoption for clinical and
research purposes has been limited to date. This is largely
The research was supported by funding from the National Institute for
Health Research (NIHR), Newcastle Biomedical Research Unit based at
Newcastle Hospitals NHS Foundation Trust and Newcastle University and
GlaxoSmithKline (GSK). The views expressed are those of the authors and
not necessarily those of the NHS, the NIHR or the Department of Health.
S. Del Din, A. Godfrey and L. Rochester are with the Institute of
Neuroscience, Newcastle University, Clinical Ageing Research Unit,
Campus for Ageing and Vitality, Newcastle upon Tyne, NE4 5PL, UK
(Corresponding author: L. Rochester phone: +441912081291, fax:
+441912081251, e-mail: lynn.rochester@ncl.ac.uk).
because the majority of studies have been carried out using
specialised gait analysis equipment (most commonly
instrumented walkways such as pressure-sensor activated, e.g.
GaitRite) [6-9] which limits work to specialised centres and a
sparse number of gait cycles [10]. In order to develop the use
of quantitative gait analysis for clinical screening and
research, low-cost tools are required that facilitate
measurement in the clinic and home. This has driven an
interest in the use of accelerometer-based body worn
monitors (BWM) for measuring gait.
BWM can provide a continuous sampling of whole body
movement in controlled or habitual environments [11]. A
BWM worn on the lower back implementing single/numerous
algorithm(s) can provide a simple method to quantify gait; the
adoption of the inverted pendulum model to evaluate step
length and the use of appropriate filtering procedures to
identify initial/final contact events within the gait cycle [12-
14]. However, it is essential to validate the combination of
BMW and implemented algorithms to accurately capture gait
outcomes before widespread adoption. Evidence suggests
suitable validity and reliability of estimated mean values of
gait outcomes to a trusted laboratory reference (GaitRite) [12,
13, 15-19]. Yet moderate to poor agreement has been reported
for step-to-step fluctuations (variability) and bilateral co-
ordination (asymmetry) [12, 14, 15, 17, 20-22]. This leaves
the role of BWM to comprehensively quantify gait in ageing
and pathology unclear.
A comprehensive examination of systems is lacking within
the literature, which is critical to further understand and
explain the poor agreement for asymmetry and variability gait
characteristics. Previous studies have tried to provide a
rationale for the poor agreement such as: difference in
sampling rates; misalignment due to device
orientation/placement; and poor synchronization [12, 13, 15,
23]. However, questions still remain. Furthermore,
weaknesses of previous studies [12-15, 21] include: (i)
limited and inconsistent reporting of gait characteristics [12-
15]; (ii) introduction of newly derived variables [23-25]
which are difficult to interpret making mainstream use
problematic; (iii) restricted testing on a single (small) cohort
(ranging on average from 10 to 23 participants [12, 13, 15,
20, 23]), often with a lack of consideration for the effects of
pathology on BWM performance; and (iv) lack of BWM
signal examination compared to video recordings. Evaluation
of a comprehensive set of gait characteristics is therefore
needed while undertaking a systematic examination of all
data acquired during a validation-based study of a BWM.
The purpose of this study was therefore to take a systematic
approach to address gaps in the literature. Firstly we aimed to
characterise a broad range of gait characteristics using a low
cost BWM in a large cohort of participants. Whilst multiple
Validation of an accelerometer to quantify a comprehensive battery
of gait characteristics in healthy older adults and Parkinson’s
disease: toward clinical and at home use.
Silvia Del Din-IEEE Member, Alan Godfrey-IEEE Member and Lynn Rochester
2168-2194 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2015.2419317, IEEE Journal of Biomedical and Health Informatics
2
characteristics can describe gait, studies have helped to
provide a simplified framework for selection of important gait
characteristics [5, 26, 27]. In this study characteristics were
selected based on our previous work [27, 28] which identified
a comprehensive range of 16 core gait characteristics (using
GaitRite) recognised as indicative of healthy ageing and
pathology (Parkinson’s disease, PD). Here, we used a BWM
and adopted a novel combination of two algorithms to
quantify gait grouped by the properties they measure such as
summary mean values, variability and asymmetry and
including phases of the gait cycle. Secondly, we wanted to
compare results in PD and older adults to see if the
performance of the BWM remained stable in pathology. The
adoption of two cohorts is a key feature of this work,
providing contrasting features of gait, where asymmetric and
variability characteristics are known to differ. Finally, we
compared the findings from the BWM to a common
laboratory reference for comparability with previous work
and where possible carried out a detailed evaluation where
differences were found between systems in order to
characterise the source of error. In adopting this process we
wished to determine the ability of a low cost BWM to
accurately measure a comprehensive set of core gait
characteristics in normal and pathological conditions for
confidence in more widespread adoption. This forms part of
our ongoing work to quantify gait simply and effectively
within laboratory-based instrumented testing sessions, with a
view to quantifying gait in real world environments.
II. METHODS
A. Participants
Thirty PD patients within 4 months of diagnosis and 30
healthy aged matched control subjects (HC) were recruited
from the Incidence of Cognitive Impairment in Cohorts with
Longitudinal Evaluation—GAIT (ICICLE-GAIT) study. This
is a collaborative study with ICICLE-PD, an incident cohort
study (Incidence of Cognitive Impairment in Cohorts with
Longitudinal Evaluation—Parkinson’s disease) conducted
between June 2009 and December 2011 [29]. This study was
conducted according to the declaration of Helsinki and had
ethical approval from the Newcastle and North Tyneside
research ethics committee. All participants signed an
informed consent form prior to testing.
B. Demographic and Clinical Measures
Age, gender and body mass index (BMI) were recorded for
each participant. Cognition was assessed with the Montreal
Cognitive Assessment (MoCA) [30]. Balance confidence was
measured using the self-rated Activities Balance Self
Confidence Scale [31]. The severity of PD motor symptoms
in the PD participants was measured using the Hoehn and
Yahr scale [32], which ranges from 0 (no symptoms) to 5
(wheelchair bound or bedridden if unaided), and section III of
the modified Movement Disorder Society version of the
Unified Parkinson’s Disease Rating Scale (MDS-UPDRS
[33]), which ranges from 0 (no motor symptoms) to 132
(severe motor symptoms). The Postural Instability and Gait
Disorder (PIGD) and Tremor phenotype subscales were also
calculated from the MDS-UPDRS [34]. Levodopa equivalent
daily doses were calculated according to established methods
[35].
C. Equipment
Each participant was asked to wear a low cost (<£100) tri-
axial accelerometer-based device (Axivity AX3, dimensions:
23.0 × 32.5 × 7.6mm, weight: 9g) located on the fifth lumbar
vertebrae (L5), Figure 1. It is a generic movement monitor
that is non-specific for gait or ambulatory assessment and
was held in place by double sided tape (Wig Tape, Natural
Image, UK) and Hypafix (BSN Medical Limited, Hull, UK).
The device measures vertical (a
v
), anteroposterior (a
a
) and
mediolateral (a
m
) accelerations and was programmed to
capture data at 50Hz and 100Hz (16-bit resolution) and at a
range of ±8g. The change in sampling frequency was due to
upgrading of the device during the longitudinal ICICLE-PD
Gait study where updated versions had increased memory
and sampling capabilities. However, for consistency of
analysis all data were down-sampled, where necessary, to 50
Hz.
Gait assessment was conducted concurrently as part of the
ICICLE-GAIT study using a 7.0m long × 0.6m wide
instrumented walkway (Platinum model GaitRite, software
version 4.5, CIR systems, NJ, USA) which was synchronised
with a video camera (Logitech, Webcam Pro 9000, CA,
USA) recording at 25 Hz. The instrumented walkway had a
spatial accuracy of 1.27cm and temporal accuracy of 1
sample (240Hz, ~4.17ms).
The quartz stabilised real time clock of the accelerometer
(accuracy: 20 parts per million) was synchronised with the
computer used for the walkway recordings and for each
walking trial the start and stop time were recorded by the
assessor. Start and stop times were subsequently input to a
bespoke MATLAB
®
program that automatically segmented
and analysed the accelerometer data. Digital synchronization
to identify exact steps (left/right steps) between systems was
not used due to the short distance traversed and lack of
gyroscope within the BWM. However, the variability and
asymmetry equations adopted in this study accounted for all
steps to be used interchangeably (Section D. Laboratory
reference: Instrumented walkway).
C. Protocol and Data Collection
Participants were asked to walk at their preferred speed,
performing four intermittent straight line walking trials over
10.0m. The 7.0m instrumented walkway was placed in the
centre of the 10.0m (Figure 1) to ensure gait was captured at a
steady speed. In addition synchronised video of frontal plane
motion was recorded during each walk. PD participants who
were on medication were tested approximately 1 hour after
medication intake.
Figure 1. The tri-axial accelerometer-based device and site of attachment
on the lower back (L5) and laboratory set up for testing intermittent walks.
2168-2194 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2015.2419317, IEEE Journal of Biomedical and Health Informatics
3
D. Data Analysis
BWM (accelerometer) data processing
The BWM data were downloaded to a computer,
segmented into the 4 different straight line passes using time
stamps and analysed by the MATLAB
®
program.
Accelerometer signals were transformed to a horizontal-
vertical coordinate system [36], and filtered with a 4th order
Butterworth filter at 20 Hz [12, 14] using the MATLAB
functions: detrend, butter, and filtfilt. For the purposes of this
study the program utilised the novel application of a
combination of gait algorithms that have been previously
developed for a single sensor attached to L5 on a cohort of
healthy older adults:
Gait algorithm #1
The first algorithm estimated the initial contact (IC) and
final contact (FC) events within the gait cycle (Figure 2), and
is described in detail elsewhere [12]. In brief the algorithm
consisted of the following:
The IC and FC events are estimated from a continuous
wavelet transform (CWT, using the cwt MATLAB
®
function) of a
v
which was first integrated (cumtrapz)
and then differentiated using a Gaussian CWT: the IC
events were detected as the local minima of the CWT
(findpeaks), Figure 3a. A further differentiation
resulted in the local maxima being defined as the FC
events (Figure 3a, Figure 4).
Previously, the algorithm has been used to estimate step
and stride times only [12]. However, in order to fully
replicate the core characteristics of gait we needed to derive
stance and swing time. This was achieved through the
sequence of IC and FC events in relation to the double
support phase of the gait cycle, Figure 2. From the sequence
(i) of IC and FC events both left and right (opposite) events
were identified where stride and subsequently stance and
swing time were estimated, Equations (1, 2 & 3). For direct
comparison to the steps quantified by the instrumented
walkway, the initial and final 4 steps (walk to/from the
walkway) of the accelerometer data (as determined by the
IC/FC algorithm) were excluded from analysis.
Figure 2. Identification of stride, stance and swing times from the double
support phase of the IC and FC algorithm
FC
IC
Stride time IC
IC
Swing time Stride time Stance time
Gait algorithm #1: optimisation
Upon initial inspection of the signal traces, spurious IC
events (i.e. non-IC events, Figure 4a) were detected in 37%
and 58% of the HC and PD groups, respectively. As a result,
the algorithm to detect IC and FC events was refined to
include a previous methodology for improved step detection:
the updated algorithm only included IC peaks within a
predetermined timed interval similar to Najafi et al. [17]. The
optimisation procedure required IC events to be identified
during a predefined interval (0.25-2.25s) from a previous IC
event. Figure 4b shows an example of the updated algorithm
with the correct estimation of IC and FC events.
Gait algorithm #2
The second gait algorithm estimated step length using the
inverted pendulum model described by Zijlstra et al. [14]
(Equation 4, Figure 3b) where h represents the change in
height (vertical position) of the centre of mass (CoM)
derived using double integration (cumtrapz) of a
v
,, and l the
pendulum length (sensor height from ground).
Gait algorithm #2: optimisation
We evaluated step length using 2 methods: (i) l estimated
using leg length × correction factor [14] and (ii) l estimated
from the height of the BWM located at L5, i.e. the ratio of
the participants’ height (l = height × 0.53) [37]. Preliminary
analysis revealed better agreement between systems for the
second method compared with using the correction factor
and as a result was adopted in this study.
Gait algorithm #3
To estimate a value for step velocity we utilised algorithm
#1 and #2 and the simple ratio between distance (length) and
time, Equation 5.
Figure 3. Flowchart of MATLAB
®
analysis:
algorithm #1 (a), algorithm #2 (b) and algorithm #3 (step velocity).
2168-2194 (c) 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE
permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2015.2419317, IEEE Journal of Biomedical and Health Informatics
4
We replicated 14/16 [27] clinically relevant gait
characteristics (step width and step width variability could not
be measured due to the adoption of a single tri-axial
accelerometer). The mean, variability and asymmetry values
of the gait characteristics were then calculated for direct
comparison to gait characteristics determined by the
instrumented walkway (see following section for details).
Right and left IC’s were previously identified by the sign of
the filtered vertical angular velocity at the instant of IC with
the use of a gyroscope [12]. In this study right and left IC’s in
the accelerometer signal were automatically selected from the
MATLAB
®
program consistently assigning right steps to the
first detected step and alternating with left step assignment.
Laboratory reference: Instrumented walkway
Data for individual steps for each walk were extracted
from the instrumented walkway database using Microsoft
Access 2007 (Microsoft Corp., Redmond, WA, USA). Mean
gait values were calculated for step time, stance time
(duration the stance foot was in contact with the ground for a
given stride), swing time (duration a foot was not in contact
with the ground for a given stride), step length and step
velocity. To calculate step variability, the standard deviation
(SD) from all steps (left and right combined) was calculated,
Equation 6. SD was selected as a measure of variability due
to its robustness and dual use with both the BWM and
walkway. Other measures of variability (harmonic ratio,
coefficient of variation) have been suggested but have shown
low to moderate agreement between systems [23, 24, 38, 39]
and are not all quantifiable by an instrumented walkway.
Figure 4. An example of ICs (squares) and FCs (circles) detection during
a single pass on the walkway. The black solid line represents a
v
, the dashed
line the differentiated with Gaussian CWT of a
v
(a
vd
), and the dotted line the
differentiated with Gaussian CWT of a
vd
(a
vdd
). Panel (a) shows the IC/FC
algorithm with spuriously detected IC events (circled squares). Panel (b)
shows the correct detection of the ICs and FCs with the optimised technique.
We considered each walking pass separately and as a
result, right and left steps were interchangeable. This method
has no impact on evaluation of variability values which were
described as the SD of all steps within walking trials,
Equation 6. Asymmetry was determined as the absolute
difference between left and right steps (alternating) for each
walking pass, averaged across all passes, Equation 7. As
asymmetry represents the absolute mean difference between
right and left steps, it does not depend on the detection of the
“true” right and left steps.
E. Statistical Analysis
Statistical analysis was carried out using SPSS v19 (IBM).
Descriptive statistics (means and standard deviations (SD))
were calculated for all gait characteristics, in PD and HC
pooled across the 4 passes. Normality of data was tested with
a Shapiro-Wilk test. Bland-Altman plots were used to
visually check for non-linear or heteroscedastic distributions
of error between the two systems (instrumented walkway v.
BWM) as a function of the participants' mean gait
performance.
Absolute agreement between the two systems was
formally tested using Intraclass Correlation Coefficients
(ICC
2,1
) and limits of agreement (LoA) expressed both as
absolute values and as a percentage of the mean. Relative
agreement between the two systems was also established
using Pearson's correlations (r). Independent t-tests were
used to examine the difference between groups for
demographical data and systems’ outcomes (p value <0.05
was considered as significant).
III. RESULTS
Demographic data
Participant demographic, clinical and cognitive
descriptors are shown in Table I. Compared to HC, PD
participants were aged matched; included proportionally less
women (CL: 50%, PD: 33%); presented with lower
confidence in their balance (ABCs), and poorer cognition
(MoCA). No differences were found between HC and PD
participants for both height and BMI. Participants with PD
were in the early stages of the disease with mild motor
symptoms.
BWM and Instrumented walkway: agreement
Table II shows the agreement between the two systems
for ICC, r values and LoA (%). There was excellent
agreement between the systems for all the mean gait
variables (step velocity, step time, stance time and step
length) except for swing time. In contrast asymmetry of steps
(bilateral co-ordination) and variability (step to step
fluctuations) showed poor agreement.
Table III shows the results obtained from the derived gait
parameters for both the BWM and instrumented walkway.
Results show that the BWM had systematic longer/greater
gait characteristics and this is significant for 10 of the 14 and
9 of the 14 variables for HC and PD, respectively.
