Objective Classification of Dynamic Balance Using a Single Wearable
Sensor
William Johnston
1,2
, Martin O’Reilly
1,2
, Kara Dolan
1,2
, Niamh Reid
1,2
,
Garrett F Coughlan
3
and Brian Caulfield
1,2
1
Insight Centre for Data Analytics, University College Dublin, Dublin, Ireland
2
School of Public Health, Physiotherapy & Sports Science, University College Dublin, Dublin, Ireland
3
Connacht Rugby, The Sportsground, College Road, Galway, Ireland
Keywords: Dynamic Balance, Inertial Measurement Unit, Y Balance Test, Fatigue, Lumbar.
Abstract: The Y Balance Test (YBT) is one of the most commonly used dynamic balance assessments in clinical and
research settings. This study sought to investigate the ability of a single lumbar inertial measurement unit
(IMU) to discriminate between the three YBT reach directions, and between pre and post-fatigue balance
performance during the YBT. Fifteen subjects (age: 23±4, weight: 67.5±8, height: 175±8, BMI: 22±2) were
fitted with a lumbar IMU. Three YBTs were performed on the dominant leg at 0, 10 and 20 minutes. A
modified Wingate fatiguing intervention was conducted to introduce a balance deficit. This was followed
immediately by three post-fatigue YBTs. Features were extracted from the IMU, and used to train and evaluate
the random-forest classifiers. Reach direction classification achieved an accuracy of 97.80%, sensitivity of
97.86±0.89% and specificity of 98.90±0.56%. “Normal” and “abnormal” balance performance, as influenced
by fatigue, was classified with an accuracy of 61.90%-71.43%, sensitivity of 61.90%-69.04% and specificity
of 61.90%-78.57% depending on which reach direction was chosen. These results demonstrate that a single
lumbar IMU is capable of accurately distinguishing between the different YBT reach directions and can
classify between pre and post-fatigue balance with moderate levels of accuracy.
1 INTRODUCTION
Dynamic balance requires the maintenance of
equilibrium during tasks that involve movement of
the centre of mass outside of the base of support
(Gribble et al., 2012). The Star Excursion Balance
Test (SEBT) is one of the most commonly used
dynamic balance assessment tools (Holden et al.,
2016, Doherty et al., 2016, Gribble et al., 2012, Smith
et al., 2015). It assesses many facets of the
sensorimotor spectrum, including strength, proprio-
ception and dynamic balance, closely mimicking the
functional demands required for optimal sports
performance. The SEBT requires the individual to
maintain their balance, while reaching as far as
possible in eight directions (Gribble et al., 2012).
Large bodies of research have demonstrated
dynamic balance deficits, as measured by the SEBT,
between control and pathological groups with
conditions such as acute ankle injuries (Doherty et al.,
2015), chronic ankle instability (Doherty et al., 2016)
and anterior cruciate ligament injuries (Herrington et
al., 2009). Additionally, researchers have attempted
to establish the role these assessments play in the
detection of risk factors that may predispose
individuals to lower limb injuries (Gribble et al.,
2015, Plisky et al., 2006). Despite this, there are a
number of limitations to the SEBT which should be
considered. These include the non-standard stance
surface, the lack of a definite starting point reference,
the time consuming nature of completing eight reach
directions and the requirements of the assessor to
visually monitor the stance foot, while marking the
maximal reach distance (Gribble et al., 2012, Plisky
et al., 2009). In an attempt to address some of these
limitations, improve the reliability and the uptake of
dynamic balance tests in clinical practice, the
redundancy of five of the eight reach directions was
demonstrated. This resulted in the development of the
commercially available Y Balance Test (YBT)
(functionalmovement.com, Danville, VA) which
incorporates the anterior (ANT), posteromedial (PM)
and posterolateral (PL) reach directions of the SEBT
(Plisky et al., 2009).
Johnston, W., O’Reilly, M., Dolan, K., Reid, N., Coughlan, G. and Caulfield, B.
Objective Classification of Dynamic Balance Using a Single Wearable Sensor.
DOI: 10.5220/0006079400150024
In Proceedings of the 4th International Congress on Sport Sciences Research and Technology Support (icSPORTS 2016), pages 15-24
ISBN: 978-989-758-205-9
Copyright
c
2016 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
15
While the YBT does address some of these
aforementioned limitations, there are a number of
challenges which continue to restrict its use in clinical
practice. Firstly, while research has shown that these
assessments are capable of demonstrating statistically
significant differences in reach distances between
groups (Plisky et al., 2006, Gribble et al., 2015,
Doherty et al., 2016, Doherty et al., 2015, Herrington
et al., 2009), it has been difficult to determine
clinically relevant cut off points. Plisky and
colleagues (2006) and Smith and colleagues (2015)
reported that a right/left asymmetry of greater than
4cm on the ANT reach direction of the SEBT and
YBT respectively is associated with an increased risk
of a lower limb injury. While Gribble and colleagues
(2015) reported that a reduced ANT reach distance, in
combination with high BMI, is associated with
increased risk of lower limb injury. Schaefer and
colleagues (2012) reported that the minimally
detectable change for normalised reach distances
ranged from 4.9-5.4% for the different reach
directions, while Munro et al (2010) showed that the
smallest detectable difference ranged from 6.87-
8.15% of leg length depending on the reach direction.
While these thresholds provide guidance for
clinicians on the reach distances that can be
considered clinically relevant, they are population
specific, and only provide a small amount of
clinically relevant information. Another is the time
consuming nature of the YBT testing protocol, which
requires the individual to complete 4 practice trials
followed by 3 recorded trials in order to obtain a
reliable and repeatable score (Gribble et al., 2012).
An additional strategy which has been employed
to improve the accuracy and objectivity of the SEBT
and YBT is the use of marker based motion analysis
and force platform systems, providing information on
the control of movement and balance strategy
employed during the task (Coughlan et al., 2012,
Fullam et al., 2014, Doherty et al., 2015). However,
these methods have a number of major limitations,
restricting their application in clinical practice.
Firstly, the set-up is time intensive and requires
training, increasing the overall testing time and
limiting the number of clinicians with the experience
required to use the systems with efficacy. The
systems are expensive (> €100,000). They are
commonly not accessible outside of a laboratory
environment. The application of markers may hinder
natural movement during dynamic tasks (Bonnechère
et al., 2014, Ahmadi et al., 2014). The data recorded
from such systems also requires extensive processing
and analysis, which is time consuming.
In recent times, there has been a shift away from
traditional motion capture systems towards
unobtrusive systems that incorporate inertial
measurement units (IMUs) (Ahmadi et al., 2014).
Such systems address some of the aforementioned
limitations of traditional motion capture, as they
allow for inexpensive, accessible quantification of
human movement, in an unconstrained environment
(Giggins et al., 2013). These IMU systems have been
used in the objective quantification of a range of
activities, from static balance tasks (King et al., 2014,
Alberts et al., 2015, Furman et al., 2013), to dynamic
tasks such as the squat (O'Reilly et al., 2015) and
single leg squat (Whelan et al., 2015), walking
(Zijlstra and Hof, 2003, Yang et al., 2013) and
running (Lee et al., 2010). Early work investigating
the use of IMUs in balance assessment has shown that
a static balance assessment, instrumented with an
IMU mounted on the lumbar spine, was not as
effective as the traditional subjectively scored
assessment in identifying balance deficits post-
concussion (Furman et al., 2013). More recently,
King et al (2014) demonstrated improved levels of
sensitivity and specificity from the instrumented
balance error scoring system (BESS). It is likely that
the conflicting results are due to the different
quantified variables selected in the two studies. King
et al (2014) utilised root mean squared acceleration,
whereas Furman et al (2013) used sway path length,
which may not be capable of detecting subtle changes
in balance, when measured using a lumbar mounted
IMU. While these initial studies have demonstrated
the ability of IMUs to detect differences in static
balance between groups, there is a paucity of
evidence surrounding their ability to classify dynamic
balance performance during tasks such as the YBT.
Previous research has established the effect
various forms of muscle fatigue such as high intensity
intermittent exercise (Whyte et al., 2015), lower limb
functional exercises (Gribble et al., 2009) and isolated
muscle fatigue (Gribble and Hertel, 2004, Gribble et
al., 2009) have on dynamic balance. The combined
physiological effects of central and peripheral fatigue
mechanisms may result in changes to the integration
of sensorimotor information from the balance
subsystems, leading to decreased balance
performance. Therefore, this research sets out to
evaluate the ability of a single lumbar mounted IMU
to objectively quantify dynamic balance
performance. It is hypothesised that a single IMU
system has the potential to accurately differentiate the
three reach directions (ANT, PM and PL) and
distinguish “normal” and “abnormal” balance as
influenced by fatigue.
icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support
16
2 METHODS
2.1 Subjects
Fifteen healthy participants aged between 18 and 40
(age: 23±4, weight: 67.5±8, height: 175±8, BMI:
22±2) who actively participate in sport were recruited
from the wider university population. Participants
were excluded from the study if they suffered from
chronic ankle instability, had sustained a lower limb
injury in the last six months, had vestibular, visual or
balance impairment, cardiovascular disease, any
neurological disease, or answered yes to any
questions on the PAR-Q (Warburton et al., 2011).
Ethical approval was obtained from the University
Human Research Ethics Committee and all
participants provided informed consent prior to
participating in the study.
2.2 Measures
2.2.1 Y-Balance Test
The YBT is an instrumented alternative to the SEBT,
capable of measuring dynamic postural control. The
YBT utilises three of the eight original SEBT reach
directions (ANT, PM and PL) and was developed in
order to provide a more objective reach distance
measurement, allowing for more accurate results,
collected in a less time consuming manner. The YBT
has been reported to demonstrate excellent intra-
tester (0.85-0.89) and inter-tester (0.97-1.00)
reliability (Plisky et al., 2009). The YBT requires
participants to stand on one leg, with their hands on
their hips, and slide a block as far as possible in the
three specified directions, with the contralateral limb,
before returning to bilateral stance. A fail is recorded
if the participant (1) uses the block for support, (2)
raises the stance heel from the platform, (3) makes
ground contact, (4) kicks the block forward to gain
extra distance or (5) removes one or both hands from
the hips during the task. The reach distances are then
normalised against the participant’s leg length using
the formula:
ℎ =
(ℎ) (ℎ)
(1)
Leg length is obtained by the same investigator
for each study participant by measuring the distance
from the anterior-superior iliac spine to the most
distal aspect of the medial malleolus (Gribble and
Hertel, 2003). The average YBT reach distances
scores used for analysis were obtained by finding the
mean of the three normalised maximal YBT scores in
each reach direction. The YBT testing protocol was
developed and conducted according to the guidelines
outlined by Gribble and colleagues (2012).
2.2.2 Modified Wingate Test
The Wingate test is traditionally used in the
measurement of peak anaerobic power and anaerobic
capacity (Smith and Hill, 1991). A modified version
of the extended Wingate test protocol employed by
Carey and Richardson (2003) was used during this
study in order to maximally fatigue participants. The
modified test requires participants to cycle at
maximal intensity for 60 seconds, rather than the
traditional 30 second protocol. The cycle ergometer
resistance is set to 0.075 g·kg-1 as per previously
published methods (Kraemer et al., 2000, Laurent et
al., 2007). Prior to commencement of the Wingate
test, participants completed a 5-minute warm-up
cycling at 50-60 RPM, which included 3 x 5 second
sprints. Following the 5-minute warm-up,
participants commenced cycling at a cadence of 50-
60 RPM for 30 seconds. At the end of the 30 second
period, the 60 second Wingate test commenced, and
participants were encouraged to maintain a maximal
effort for the duration of the 60 seconds in order to
ensure maximal fatigue.
2.2.3 Inertial Measurement Unit
A Shimmer3 IMU (Shimmer, Dublin, Ireland) was
mounted at the level of the fourth lumbar vertebra
(Figure 1). The IMU was calibrated and configured to
stream tri-axial accelerometer (±2 g), gyroscope
(±500 /s) and magnetometer (±1 gauss) data at 102.4
Hz via Bluetooth to an Android tablet, using Multi-
Shimmer sync software (Shimmer, Dublin, Ireland).
These data acquisition parameters were chosen based
on previous work carried out by our research group
Figure 1: Illustrates the mounting location of the Lumbar
IMU.
Objective Classification of Dynamic Balance Using a Single Wearable Sensor
17
investigating the use of IMUs in the evaluation of
exercise technique during similar movements, such as
the single leg squat (O'Reilly et al., 2015).
2.3 Procedure
On arrival to the performance laboratory, the
experimental protocol was explained to the
participants, and individuals completed 4 practice
trials in each direction, on their dominant leg (all right
leg dominant). Leg dominance was obtained by
asking the participants which leg they would use to
kick a ball (Wilkins et al., 2004). Following
completion of the practice trials, each participant was
fitted with the IMU as described above. Participants
then completed three recorded YBT in each direction
(randomised order) on the dominant limb. This was
repeated at 0, 10 and 20-minutes in order to provide a
pre-fatigue baseline measurement of dynamic
balance. YBT maximal reach distances and IMU data
were collected for each YBT attempt. If a participant
failed to complete the test as described above, the
individual reach direction was repeated, and an
annotation was recorded in the IMU data to denote a
failed and repeated reach direction.
Following the baseline assessment, participants
completed the modified Wingate protocol in order to
elicit maximal anaerobic fatigue. Immediately
following the Wingate protocol, participants
completed the YBT to capture the reduced dynamic
balance performance elicited by maximal anaerobic
fatigue.
2.4 Data Analysis
Nine signals were collected from the IMU;
accelerometer x, y, z, gyroscope x, y, z and
magnetometer x, y, z. Data were analysed using
MATLAB (2012, The MathWorks, Natwick, USA).
To ensure the data analysed applied to each
participant’s movement and in order to eliminate
unwanted high-frequency noise, the nine signals were
low pass filtered with an 8
th
order Butterworth filter
with a 20Hz cut-off. Nine additional signals were then
calculated. The 3-D orientation of the IMU was
computed using the gradient descent algorithm
developed by (Madgwick et al., 2011). The resulting
W, X, Y and Z quaternion values were also converted
to pitch, roll and yaw signals. The pitch, roll and yaw
signals describe the inclination, measured in radians,
of each IMU in the sagittal, frontal and transverse
planes respectively. The magnitude of acceleration
was also computed using the vector magnitude of
accelerometer x, y, z. The magnitude of acceleration
describes the total acceleration of the IMU in any
direction. This is the sum of the magnitude of inertial
acceleration of the lumbar spine and acceleration due
to gravity. Additionally, the magnitude of rotational
velocity was computed using the vector magnitude of
the gyroscopes x, y and z.
Each reach direction from each completed
YBT was extracted from the IMU data and resampled
to a length of 1000 samples; this was undertaken to
minimise the influence of the speed of repetition
performance on signal feature calculations. It ensures
the computed features related to differences in
movement patterns and not the participant’s exercise
tempo. Descriptive features were computed in order
to characterise the pattern of each of the eighteen
signals as the YBT was completed. These features
were namely 'Mean', 'RMS', 'Standard Deviation',
'Kurtosis', 'Median', 'Skewness', ‘ Range', ‘Variance',
'Max', ‘Index of Max’, 'Min', ‘Index of Min’,
'Energy', '25th Percentile', '75th Percentile', 'Level
Crossing Rate' and' Fractal Dimension' (Katz and
George, 1985) . This resulted in 17 features for each
of the 18 available signals producing a total of 306
features. These features were then used to develop
and evaluate a classifier for the automated detection
of reach direction in the YBT and a separate classifier
for the detection of pre-fatigue or fatigued YBT
performance. The random-forests method was
employed to perform classification of reach direction
and for the detection of fatigued YBT performance
(Breiman, 2001). This technique was chosen as it has
been shown to produce superior accuracy, sensitivity
and specificity scores in analysing exercise technique
with IMUs in comparison to the Naïve-Bayes and
Radial-basis function network techniques (Mitchell et
al., 2015). Four hundred decision trees were used in
each random-forest classifier.
The quality of the exercise classification
system was established using leave-one-subject-out-
cross-validation (LOSOCV) and the random-forests
classifier with four hundred trees (Fushiki, 2011).
Each participant’s data corresponds to one fold of the
cross validation. At each fold, one participant’s data
is held out as test data while the random forests
classifier is trained with all other participants’ data.
The held out data is used to assess the classifier’s
ability to correctly categorise unseen data. The use of
LOSOCV ensures that there is no biasing of the
classifiers, meaning the test subjects data is
completely unseen by the classifier prior to testing.
Previous research by Taylor et al (2010) has shown
that not employing this method of testing can skew
results significantly. In our system, each individual
reach direction was classified.
icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support
18
The scores used to measure the quality of
classification were total accuracy, average sensitivity
and average specificity. Accuracy is the number of
correctly classified observations divided by the total
number of observations completed; this is calculated
as the sum of the true positives (TP) and true
negatives (TN) divided by the sum of the true
positives, false positives (FP), true negatives and false
negatives (FN):
 =


(2)
The sensitivity and specificity were calculated for
each of the reach directions, sequentially treating
each label as the ‘positive’ class, and then the mean
and standard deviation across the five values was
taken. Sensitivity and specificity were computed
using formulas 3 and 4 below:
 =


(3)
 =


(4)
Sensitivity measures the effectiveness of a classifier
at identifying a desired label, while specificity
measures the classifiers ability to detect negative
labels. In the detection of fatigued balance, single
sensitivity and specificity scores were calculated,
treating pre-fatigued balance as the positive class and
fatigued balance as the negative class.
In reviewing the accuracy, sensitivity and
specificity scores produced by each classifier, 90% or
over was considered an excellent result, 80-89% was
considered a 'good' quality result, 60-79% was
considered a 'moderate' result and anything less than
59% was deemed a poor result. These values were
chosen by the authors after reviewing existing
literature on identifying exercises with IMUs. In
reviewing such literature, an existing accepted
standard for a good, moderate or poor classifier could
not be found. Therefore, the above system was agreed
on by the authors to facilitate interpretation of our
range of results.
3 RESULTS
ICC values for the three normalised reach directions
ranged from 0.976 – 0.986, indicating excellent test-
retest reliability across the pre-fatigue measures. Due
to the excellent ICC scores observed, the final pre-
fatigue measure was considered representative of the
pre-fatigue state, and was used in the comparison pre
and post-fatigue. The SEM ranged from 0.792-1.48
for the three YBT reach directions. The average
decrease in YBT reach distances following the fatigue
protocol was 2.65 ± 4.91 (ANT), 2.44 ± 3.06 (PM)
and 3.57 ± 4.27 (PL). Paired samples t-tests
demonstrated statistically significant differences (p <
0.05) between the final pre-fatigue YBT
measurement and the first post-fatigue measurement
in all reach directions (Table 1).
Table 1: Comparison of ICC, SEM and paired sample t-tests
for the YBT normalised Reach Direction for all three
directions. The level of significance was set to p < 0.05 and
statistically significant values were denoted with an*.
Reliability Analysis
Pre01, Pre02 & Pre03
Level of Significance
(p Values) for Post-hoc
Paired t-test
Reach
Distance
ICC SEM
Pre03 vs
Post01
ANT
0.986 0.792 0.049*
PM
0.976 1.482 0.008*
PM
0.978 1.134 0.006*
The classification algorithm for a single lumbar
mounted IMU was capable of differentiating the three
reach directions in the pre-fatigue baseline measures
with an accuracy of 97.80%, Sensitivity of 97.86 ±
0.89% and specificity of 98.90 ± 0.56%. Figure 2
presents a confusion matrix that illustrates the exact
percentage of reach direction repetitions that were
classified correctly and incorrectly. The rows
represent the actual reach direction recorded and the
columns show the classifier’s predicted reach
direction.
Figure 2: A confusion matrix showing multi-class
classification results for the three reach directions. The
percentage of reach direction attempts classified correctly
are marked in bold.
A single lumbar mounted IMU was capable of
discriminating pre and post-fatigue balance
performance with an accuracy of 61.90%-71.43%,
Objective Classification of Dynamic Balance Using a Single Wearable Sensor
19
sensitivity of 61.90%-69.04% and specificity of
61.90%-78.57% depending on which reach direction
was chosen (Table 3). When all reach directions were
considered together, balance performance was
classified with an accuracy of 70.24%, sensitivity of
64.28% and specificity of 76.19%.
Table 2: The accuracy, sensitivity and specificity results of
the classification algorithm in the detection of baseline and
fatigued dynamic balance.
ANT PM PL
All
Directions
Accuracy
61.90 71.43 70.24 70.40
Sensitivity
61.90 69.04 61.90 64.28
Specificity
61.90 73.80 78.57 76.19
4 DISCUSSION
The purpose of this study was to determine if data
derived from a single lumbar mounted IMU is
capable of accurately differentiating the individual
reach directions of the YBT, and classifying pre and
post-fatigue dynamic balance performance.
The traditional normalised YBT reach distance
results presented demonstrate that the modified
Wingate protocol had a detrimental effect on the
participant’s dynamic balance. The ICC values for the
pre-fatigue baseline assessments presented suggest
that the normalised YBT reach distance scores for
each reach direction possess excellent test-retest
reliability. The paired sampled t-test results (Table 1)
demonstrate that there was a statistically significant
difference between the final pre-fatigue measurement
and the post-fatigue measurements, suggesting that
the fatigue intervention had a detrimental effect on
the YBT reach distances for all three reach directions.
Additionally, the SEM results for all reach directions
was smaller than the average decrease in reach
distance between the final pre-fatigue and the post-
fatigue measurement, indicating that the fatigue
intervention had a negative effect on reach distance
scores. When the SEM is viewed in conjunction with
the ICC, it allows us to be sure that any deviation from
the baseline is as a result of the fatiguing intervention,
and not a consequence of natural biological variation.
The results presented in this paper support
previously published ones indicating that dynamic
balance is heavily influenced by isolated muscle
fatigue (Gribble et al., 2004, Gribble et al., 2009),
lower limb fatiguing exercises (Gribble et al., 2009),
treadmill running (Wright et al., 2013) and high
intensity intermittent exercise protocols (Whyte et al.,
2015). Whyte and colleagues (2015) investigated the
effect of high intensity intermittent exercise on
dynamic balance, as measured by the SEBT. It was
reported that the percentage reduction in SEBT reach
distance, for the ANT, PM and PL directions were
marginally lower than those presented in our study.
Importantly, these differences may be a result of the
different fatiguing interventions influencing the
sensorimotor system to different extents (Whyte et
al., 2015). Additionally, different methods of
dynamic balance assessments were utilised in the two
studies. Whyte and colleagues (2005) used the SEBT,
whereas the YBT was implemented in our study,
potentially explaining the difference in the magnitude
of change (Coughlan et al., 2012). These past
findings, combined with the results from this study,
demonstrate that at a group level, the fatigue
intervention had a negative effect on dynamic
balance.
The IMU classification system was capable of
differentiating individual YBT reach directions with
excellent levels of accuracy, sensitivity and
specificity. The confusion matrix (Figure 2)
illustrates the percentage of the reach directions
classified correctly and incorrectly, indicating where
the confusion occurred. The ANT reach direction was
classified with the greatest success rate of 99%,
followed by PM (98%), and then PL (97%). These
results may be expected as the three reach directions
utilise different strategies to complete a maximal
reach. The ANT reach direction involves a single
planar movement which incorporates a single leg
squat type movement, while the individual reaches
outside of their base of support. In contrast, the PM
and PL movements involve multi-planar movements,
requiring the individual to enter a single leg squat,
while rotating at the pelvis and trunk in order to
achieve a maximal reach distance. Indeed, previous
research conducted by Kang and colleagues (2015)
investigating trunk, pelvic and lower limb kinematic
strategies utilised during the YBT. The results
presented by their group demonstrate that the ANT
reach direction requires a largely different strategy to
the PM and PL directions. The ANT direction
requires minimal trunk and pelvic kinematic
movements, with 1° trunk extension, 4° trunk
ipsilateral flexion, 9° anterior pelvic tilt, and 1° of
pelvic ipsilateral rotation. The majority of the
movement strategies stem from sagittal plane
movements at the hip (30° flexion), knee (62°
flexion) and ankle (39° dorsiflexion). In contrast, the
PM and PL reach directions require large changes in
trunk and pelvic kinematics, with the PM reach
direction requiring 43° trunk flexion, 21° trunk
icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support
20
ipsilateral flexion, 39° anterior pelvic tilt and 0° of
pelvic contralateral rotation, and the PL reach
direction requiring 48° trunk flexion, 16° trunk
contralateral flexion, 38° anterior pelvic tilt and 11°
of pelvic contralateral rotation. These results clarify
the similarities and differences between the
movement strategies utilised during each reach
direction, contextualising how the classification
algorithm was capable of classifying the individual
reach directions with such high degrees of accuracy.
The YBT reach direction classification results
presented in this study are in line with previously
published IMU exercise identification results which
range between 85-95% depending on the exercises
and IMU setups (Giggins et al., 2014, Pernek et al.,
2015, Chang et al., 2007). Giggins and colleagues
(2014) demonstrated that a single IMU location could
differentiate between seven basic rehabilitation
exercises with an accuracy of between 93-95%
depending on the mounting location. Additionally,
Pernek and colleagues (2015) reported that a single
IMU system can correctly identity upper limb free
weight exercises with 85% accuracy. This is
significant as the excellent levels of accuracy (98%)
presented in this study were achieved using just 252
observations. In contrast, the exercise classification
work presented above used a greater number of
observations to train the classifiers, with Giggins et al
(2014) utilising 3940 observations and Pernek et al
(2015) using 440 observations per exercise.
The lumbar IMU classification algorithm was
capable of differentiating dynamic balance
performance, as influence by fatigue, with and
accuracy of between 62% and 71%, depending on the
reach direction (Table 3). The PM reach direction
demonstrated the highest classification accuracy
(72%), followed by the PL (70%) and then the ANT
(62%) reach direction. When all reach directions were
considered together, the classification algorithm was
able to differentiate normal and abnormal balance
with an accuracy of 70%.
These results would be expected, because as we
previously discussed above, the three reach directions
require different levels of movement strategy
complexity. The ANT reach direction presented with
the lowest degree of classification accuracy. The
ANT reach direction is the least complex movement,
predominantly requires sagittal plane movement of
the stance limb (Kang et al., 2015). It may be that the
ANT reach direction movement does not sufficiently
challenge the sensorimotor system in all individuals
to elicit a balance deficit large enough to be
consistently detected by the lumbar mounted IMU. In
contrast, the higher degree of accuracy observed in
the detection of abnormal balance during the PM and
PL reach directions are expected as these movements
require the individual to implement a more complex
multi-planar movement strategy. Both the PM and PL
reach directions require the individual to reach
outside of their base of support while utilising their
trunk as a mobile counter-lever, involving a
combination of complex multi-planar movements
occurring at the trunk, pelvis, hips, knee and ankle
(Kang et al., 2015, Fullam et al., 2014, Doherty et al.,
2016). This complex multi-planar movement may
more comprehensively challenge the integration of
the sensorimotor subsystems, resulting in more
pronounced strategy changes following the
introduction of a balance deficit, thus leading to
differences in the IMU data.
To the best of the authors knowledge this is the
first research study that has attempted to classify
dynamic balance performance using an IMU.
Previous research has investigated the ability of
single and multiple IMUs to detect technique
breakdown during compound lower limb exercises
such as the squat (O'Reilly et al., 2015) and single leg
squat (Whelan et al., 2015). Lower limb exercises
such as the single leg squat incorporate many of the
requirements involved during the YBT reach
directions, such as maintaining one’s balance while
executing a dynamic task on a single leg. Whelan and
colleagues (2015) reported that a single lumbar based
IMU mounted on the lumbar spine was capable of
classifying correct and incorrect single leg squat
technique with an accuracy of 92%. While the
classification accuracy presented by Whelan and
colleagues is higher than that of the YBT balance
performance classification presented in our study, it
is probable that the YBT classification performance
would be greatly improved by increasing the number
of observations used to train and test the classifier.
The results presented in this paper demonstrate
the potential of a single lumbar mounted IMU to
automatically classify YBT reach direction and
balance performance. This lays the groundwork for
the development of an accurate dynamic balance
performance classification system that can provide
accessible, in depth, clinically relevant information,
surrounding an individual’s dynamic balance, outside
of the constraints of a laboratory. Future work will
allow us to detect changes in movement and balance
strategy during the YBT, characterising the dynamic
balance defects. This would provide clinicians with
more in depth information which can be used to
comprehensively and objectively assess the
integration of the sensorimotor subsystems, in an
accessible manner. This has the potential to provide
Objective Classification of Dynamic Balance Using a Single Wearable Sensor
21
information in areas such as lower limb injuries,
identification of lower limb injury risk factors,
assessment of the motor function domain post-
concussion, as well as balance training in strength and
conditioning and rehabilitation.
A number of limitations to the study must be
acknowledged. The sample size and resultant number
of observations that could be used to train and
evaluate the classification algorithms were relatively
small, potentially resulting in decreased levels of
accuracy. It can be expected that as the number of
participants and observations are increased, there will
be a resultant increase in the accuracy of the balance
performance identification. Secondly, no gold
standard motion capture system was employed in this
study. However, YBT reach directions are commonly
accepted as the standard in clinical balance
assessments, and each participant was educated and
supervised by a Chartered Physiotherapist throughout
the duration of the study.
Extensive future work is required to improve the
classification results presented in this paper. Firstly, a
greater number of participants is required to increase
the size of the data set in order to establish a
normative dataset. Additionally, a classification
system with improved accuracy, sensitivity and
specificity will be developed. This may be achieved
through investigating the effectiveness of a single
IMU located at different anatomical positions,
collecting a larger data set to allow for more training
data for the classification algorithms and the
identification of new features to input into the
classifiers which enable further distinction of normal
and abnormal balance. Novel classification
techniques for IMU data may also be employed such
as the application of deep learning on the data. This
will also require a larger data set to be collected.
5 CONCLUSION
To conclude, the results presented in this paper
demonstrate that a lumbar mounted IMU is capable
of accurately distinguishing the three YBT reach
directions, as well as classifying balance performance
as influenced by a maximal anaerobic fatigue. This
work lays the foundations for the development of a
single IMU system, that can accurately differentiate
the YBT reach directions, as well as detect changes in
balance strategy, characterising and classifying
dynamic balance performance.
ACKNOWLEDGMENTS
This study was funded by the Science Foundation of
Ireland (12/RC/2289).
REFERENCES
Ahmadi, A., Mitchell, E., Destelle, F., Gowing, M.,
Oconnor, N. E., Richter, C. & Moran, K. Automatic
activity classification and movement assessment during
a sports training session using wearable inertial sensors.
2014 2014. IEEE, 98-103.
Alberts, J. L., Hirsch, J. R., Koop, M. M., Schindler, D. D.,
Kana, D. E., Linder, S. M., Campbell, S. & Thota, A.
K. 2015. Using Accelerometer and Gyroscopic
Measures to Quantify Postural Stability. J Athl Train,
50, 578-88.
Bonnechère, B., Jansen, B., Salvia, P., Bouzahouene, H.,
Omelina, L., Moiseev, F., Sholukha, V., Cornelis, J.,
Rooze, M. & Van Sint Jan, S. 2014. Validity and
reliability of the Kinect within functional assessment
activities: Comparison with standard stereophoto-
grammetry. Gait & Posture, 39, 593-598.
Breiman, L. 2001. Random Forests. Machine Learning, 45,
5-32.
Carey, D. G. & Richardson, M. T. 2003. Can Aerobic and
Anaerobic Power be Measured in a 60-Second Maximal
Test? J Sports Sci Med, 2, 151-7.
Chang, K.-H., Chen, M. Y. & Canny, J. 2007. Tracking
Free-Weight Exercises. In: Krumm, J., Abowd, G. D.,
Seneviratne, A. & Strang, T. (eds.) UbiComp 2007:
Ubiquitous Computing: 9th International Conference,
UbiComp 2007, Innsbruck, Austria, September 16-19,
2007. Proceedings. Berlin, Heidelberg: Springer Berlin
Heidelberg.
Coughlan, G. F., Fullam, K., Delahunt, E., Gissane, C. &
Caulfield, B. M. 2012. A comparison between
performance on selected directions of the star excursion
balance test and the Y balance test. Journal of athletic
training, 47, 366.
Doherty, C., Bleakley, C., Hertel, J., Caulfield, B., Ryan, J.
& Delahunt, E. 2016. Dynamic balance deficits in
individuals with chronic ankle instability compared to
ankle sprain copers 1 year after a first-time lateral ankle
sprain injury. Knee Surg Sports Traumatol Arthrosc,
24, 1086-95.
Doherty, C., Bleakley, C. M., Hertel, J., Caulfield, B.,
Ryan, J. & Delahunt, E. 2015. Laboratory Measures of
Postural Control During the Star Excursion Balance
Test After Acute First-Time Lateral Ankle Sprain. J
Athl Train, 50, 651-64.
Fullam, K., Caulfield, B., Coughlan, G. F. & Delahunt, E.
2014. Kinematic analysis of selected reach directions of
the Star Excursion Balance Test compared with the Y-
Balance Test. J Sport Rehabil, 23,
27-35.
Furman, G. R., Lin, C. C., Bellanca, J. L., Marchetti, G. F.,
Collins, M. W. & Whitney, S. L. 2013. Comparison of
the balance accelerometer measure and balance error
icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support
22
scoring system in adolescent concussions in sports. Am
J Sports Med, 41, 1404-10.
Fushiki, T. 2011. Estimation of prediction error by using K-
fold cross-validation. Statistics and Computing, 21,
137-146.
Giggins, Kelly, D. & Caulfield, B. Evaluating rehabilitation
exercise performance using a single inertial
measurement unit. 2013 7th International Conference
on Pervasive Computing Technologies for Healthcare
and Workshops, 5-8 May 2013 2013. 49-56.
Giggins, O., Sweeney, K. T. & B, C. The use of inertial
sensors for the classification of rehabilitation exercises.
2014 36th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society, 26-30
Aug. 2014 2014. 2965-2968.
Gribble, P. A. & Hertel, J. 2003. Considerations for
normalizing measures of the Star Excursion Balance
Test. Measurement in physical education and exercise
science, 7, 89-100.
Gribble, P. A. & Hertel, J. 2004. Effect of lower-extremity
muscle fatigue on postural control. Arch Phys Med
Rehabil, 85, 589-92.
Gribble, P. A., Hertel, J., Denegar, C. R. & Buckley, W. E.
2004. The effects of fatigue and chronic ankle
instability on dynamic postural control. Journal of
Athletic Training, 39, 321.
Gribble, P. A., Hertel, J. & Plisky, P. 2012. Using the Star
Excursion Balance Test to assess dynamic postural-
control deficits and outcomes in lower extremity injury:
a literature and systematic review. J Athl Train, 47, 339-
57.
Gribble, P. A., Robinson, R. H., Hertel, J. & Denegar, C. R.
2009. The effects of gender and fatigue on dynamic
postural control. Journal of sport rehabilitation, 18,
240.
Gribble, P. A., Terada, M., Beard, M. Q., Kosik, K. B.,
Lepley, A. S., Mccann, R. S., Pietrosimone, B. G. &
Thomas, A. C. 2015. Prediction of Lateral Ankle
Sprains in Football Players Based on Clinical Tests and
Body Mass Index. The American Journal of Sports
Medicine.
Herrington, L., Hatcher, J., Hatcher, A. & Mcnicholas, M.
2009. A comparison of Star Excursion Balance Test
reach distances between ACL deficient patients and
asymptomatic controls. The Knee, 16, 149-152.
Holden, S., Boreham, C., Doherty, C., Wang, D. &
Delahunt, E. 2016. A longitudinal investigation into the
progression of dynamic postural stability performance
in adolescents. Gait Posture, 48, 171-176.
Kang, M.-H., Kim, G.-M., Kwon, O.-Y., Weon, J.-H., Oh,
J.-S. & An, D.-H. 2015. Relationship Between the
Kinematics of the Trunk and Lower Extremity and
Performance on the Y-Balance Test. PM&R, 7, 1152-
1158.
Katz, M. J. & George, E. B. 1985. Fractals and the analysis
of growth paths. Bulletin of Mathematical Biology, 47,
273-286.
King, L. A., Horak, F. B., Mancini, M., Pierce, D., Priest,
K. C., Chesnutt, J., Sullivan, P. & Chapman, J. C. 2014.
Instrumenting the balance error scoring system for use
with patients reporting persistent balance problems
after mild traumatic brain injury. Archives of physical
medicine and rehabilitation, 95, 353-359.
Kraemer, W. J., Ratamess, N., Fry, A. C., Triplett-Mcbride,
T., Koziris, L. P., Bauer, J. A., Lynch, J. M. & Fleck, S.
J. 2000. Influence of resistance training volume and
periodization on physiological and performance
adaptations in collegiate women tennis players. Am J
Sports Med, 28, 626-33.
Laurent, C. M., Jr., Meyers, M. C., Robinson, C. A. &
Green, J. M. 2007. Cross-validation of the 20- versus
30-s Wingate anaerobic test. Eur J Appl Physiol, 100,
645-51.
Lee, J. B., Mellifont, R. B. & Burkett, B. J. 2010. The use
of a single inertial sensor to identify stride, step, and
stance durations of running gait. Journal of Science and
Medicine in Sport, 13, 270-273.
Madgwick, S. O., Harrison, A. J. & Vaidyanathan, A. 2011.
Estimation of IMU and MARG orientation using a
gradient descent algorithm. IEEE Int Conf Rehabil
Robot, 2011, 5975346.
Mitchell, E., Ahmadi, A., O'connor, N. E., Richter, C.,
Farrell, E., Kavanagh, J. & Moran, K. Automatically
detecting asymmetric running using time and frequency
domain features. 2015 2015. IEEE, 1-6.
Munro, A. G. & Herrington, L. C. 2010. Between-session
reliability of the star excursion balance test. Phys Ther
Sport, 11, 128-32.
O'reilly, M., Whelan, D., Chanialidis, C., Friel, N.,
Delahunt, E., Ward, T. & Caulfield, B. Evaluating squat
performance with a single inertial measurement unit.
2015 2015. IEEE, 1-6.
Pernek, I., Kurillo, G., Stiglic, G. & Bajcsy, R. 2015.
Recognizing the intensity of strength training exercises
with wearable sensors. Journal of Biomedical
Informatics, 58, 145-155.
Plisky, P. J., Gorman, P. P., Butler, R. J., Kiesel, K. B.,
Underwood, F. B. & Elkins, B. 2009. The Reliability of
an Instrumented Device for Measuring Components of
the Star Excursion Balance Test. N Am J Sports Phys
Ther, 4, 92-9.
Plisky, P. J., Rauh, M. J., Kaminski, T. W. & Underwood,
F. B. 2006. Star Excursion Balance Test as a predictor
of lower extremity injury in high school basketball
players. Journal of Orthopaedic & Sports Physical
Therapy, 36, 911-919.
Schaefer, J. L. & Sandrey, M. A. 2012. Effects of a 4-week
dynamic-balance-training program supplemented with
Graston instrument-assisted soft-tissue mobilization for
chronic ankle instability. J Sport Rehabil, 21, 313-26.
Smith, C. A., Chimera, N. J. & Warren, M. 2015. Associa-
tion of y balance test reach asymmetry and injury in
division I athletes. Med Sci Sports Exerc, 47, 136-41.
Smith, J. & Hill, D. W. 1991. Contribution of energy
systems during a Wingate power test. British Journal of
Sports Medicine, 25, 196-199.
Taylor, J. L. & Gandevia, S. C. 2008. A comparison of
central aspects of fatigue in submaximal and maximal
voluntary contractions. J Appl Physiol (1985), 104,
542-50.
Objective Classification of Dynamic Balance Using a Single Wearable Sensor
23
Taylor, P. E., Almeida, G. J., Kanade, T. & Hodgins, J. K.
2010. Engineering in Medicine and Biology Society,
EMBC. Classifying Human Motion Quality for Knee
Osteoarthritis using accelerometers.
Warburton, D. E., Jamnik, V. K., Bredin, S. S., Mckenzie,
D. C., Stone, J., Shephard, R. J. & Gledhill, N. 2011.
Evidence-based risk assessment and recommendations
for physical activity clearance: an introduction. Appl
Physiol Nutr Metab, 36 Suppl 1, S1-2.
Whelan, D., O'Reilly, M., Ward, T., Delahunt, E. &
Caulfield, B. Evaluating Performance of the Single Leg
Squat Exercise with a Single Inertial Measurement
Unit. 2015 2015 ACM, 144-147.
Whyte, E., Burke, A., White, E. & Moran, K. 2015. A high-
intensity, intermittent exercise protocol and dynamic
postural control in men and women. Journal of athletic
training, 50, 392-399.
Wilkins, J. C., Valovich Mcleod, T. C., Perrin, D. H. &
Gansneder, B. M. 2004. Performance on the Balance
Error Scoring System Decreases After Fatigue. Journal
of Athletic Training, 39, 156-161.
Wright, K. E., Lyons, T. S. & Navalta, J. W. 2013. Effects
of exercise-induced fatigue on postural balance: a
comparison of treadmill versus cycle fatiguing
protocols. Eur J Appl Physiol, 113, 1303-9.
Yang, S., Zhang, J.-T., Novak, A. C., Brouwer, B. & Li, Q.
2013. Estimation of spatio-temporal parameters for
post-stroke hemiparetic gait using inertial sensors. Gait
& Posture, 37, 354-358.
Zijlstra, W. & Hof, A. L. 2003. Assessment of spatio-
temporal gait parameters from trunk accelerations
during human walking. Gait Posture, 18, 1-10.
icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support
24