This article appeared in a journal published by Elsevier.

The attached
copy is furnished to the author for internal non-commercial research
and education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elseviers archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/authorsrights

Author's personal copy

Applied Soft Computing 20 (2014) 127141

Contents lists available at ScienceDirect

Applied Soft Computing

journal homepage: www.elsevier.com/locate/asoc

A knowledge-based intelligent framework for anterior cruciate

ligament rehabilitation monitoring
S.M.N. Arosha Senanayake a, , Owais Ahmed Malik a ,
Pg. Mohammad Iskandar b , Dansih Zaheer c
a
b
c

Department of Computer Science, Faculty of Science, Universiti Brunei Darussalam (UBD), Gadong BE 1410, Brunei Darussalam
Faculty of Integrated Technologies, Universiti Brunei Darussalam (UBD), Gadong BE 1410, Brunei Darussalam
Sports Medicine & Research Center, Hassanal Bolkiah National Stadium, Berakas, Brunei Darussalam

a r t i c l e

i n f o

Article history:
Received 16 March 2013
Received in revised form 21 October 2013
Accepted 15 November 2013
Available online 26 November 2013
Keywords:
Case based reasoning (CBR)
Fuzzy systems
Support vector machine
Hybrid intelligent system
Anterior cruciate ligament (ACL)
Rehabilitation monitoring

a b s t r a c t
This study presents an integration of knowledge-based system and intelligent methods to develop a
recovery monitoring framework for post anterior cruciate ligament (ACL) injured/reconstructed subjects. The case based reasoning methodology has been combined with fuzzy clustering and intelligent
classication techniques in order to develop a knowledge base and a learning model for identifying
the recovery stage of ACL-reconstructed subjects and objectively monitoring the progress during the
convalescence regimen. The system records kinematics and neuromuscular signals from lower limbs of
healthy and ACL-reconstructed subjects using self adjusted non-invasive body-mounted wireless sensors. These bio-signals are synchronized and integrated, and a combined feature set is generated by
performing data transformation using wavelet decomposition and feature reduction techniques. The
knowledge base stores the subjects proles, their recovery sessions data and problem/solution pairs for
different activities monitored during the course of rehabilitation. Fuzzy clustering technique has been
employed to form the initial groups of subjects at similar stage of recovery. In order to classify the recovery stage of subjects (i.e. retrieval of similar cases), adaptive neuro-fuzzy inference system (ANFIS), fuzzy
unordered rule induction algorithm (FURIA) and support vector machine (SVM) have been applied and
compared. The system has been successfully tested on a group of healthy and post-operated athletes for
analyzing their performance in two activities (ambulation at various speeds and one leg balance testing)
selected from the rehabilitation protocol. The case adaptation and retention is a semi-automatic process
requiring input from the physiotherapists and physiatrists. This intelligent framework can be utilized by
physiatrists, physiotherapists, sports trainers and clinicians for multiple purposes including maintaining
athletes prole, monitoring progress of recovery, classifying recovery status, adapting recovery protocols and predicting/comparing athletes sports performance. Further, the knowledge base can easily be
extended and enhanced for monitoring different types of sports activities.
2013 Elsevier B.V. All rights reserved.

1. Introduction
The monitoring of recovery progress after anterior cruciate ligament (ACL) injury/reconstruction is crucial for athletes for not only
restoring their knee joint stability for dynamic activities and successful returning to sports but also minimizing the risks of re-injury,
cartilage degeneration and early onset of osteoarthritis [16]. ACL
injury causes deterioration in the sports performance or premature
end to the career for athletes as ACL is one of the critical ligament for

Corresponding author. Tel.: +673 2463001 1362; fax: +673 2461502.

E-mail addresses: arosha.senanayake@ubd.edu.bn, aroshas@ieee.org
(S.M.N. Arosha Senanayake), 11h1202@ubd.edu.bn (O.A. Malik),
iskandar.petra@ubd.edu.bn (Pg.M. Iskandar), smrc@brunet.bn (D. Zaheer).
1568-4946/$ see front matter 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.asoc.2013.11.010

knee joint stability, maintaining normal gait patterns, preventing

anterior tibial translation, and controlling knee axial rotation and
varus movements [7]. It can be completely or partially ruptured as
a result of deceleration due to sudden changes of direction, twisting and/or pivoting during sports, like soccer and basketball. The
tearing or rupture of ACL affects gait patterns and variability, and
causes changes in kinematics, kinetics and neuromuscular activities of athlete. The absence of ACL results in the loss of mechanical
control of the knee, loss of proprioception due to unavailability
of mechanoreceptors present in the ligament and neurophysiologic dysfunction, hindering the athletes to participate in sports
activities. Efcient and effective rehabilitation programs are essential for athletes following ACL reconstruction as well as for those
having ACL deciency. The rehabilitation programs are designed
to rebuild muscle strength, re-establish joint and neuromuscular

Author's personal copy

128

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141

control and to enable the athletes to return to pre-injury activity level. The goal based criteria (e.g. muscle strength, functional
and static knee stability, range of motion) are suggested to safe
return to sports rather than relying on some specic time period for
rehabilitation [79]. The conventional methods for post-ACL reconstruction recovery assessment test the static and dynamic knee
stability of subjects. However, the dynamic knee stability analysis
usually lacks quantiable constructs, and some subjective or partially objective evaluations and knee rating scores are performed in
order to evaluate the recovery progress of ACL reconstructed (ACLR) subjects [1015]. Moreover, physiological variations in subjects
and coopers/non-coopers phenomenon indicate that the rehabilitation process may differ for an individual or a group of subjects
[16]. Thus, in order to provide an effective cure and timely intervention during recovery process, a comprehensive rehabilitation
monitoring system is essential for ACL reconstructed (ACL-R) subjects.
Intelligent techniques have been found effective for different medical applications including medical diagnosis, pathology
classication, rehabilitation of lower limb movements after neurological disorders and implantation of articial limbs [1722].
These techniques assist clinicians in decision making process about
subjects health condition. Although, there has been number of
studies using intelligent mechanisms for assessment and classication of gait patterns for subjects having lower limb impairments
due to stroke or cerebral palsy but few efforts have been made for
gait and recovery analysis after ACL injury/reconstruction. Mostly
the assessment of injured/reconstructed subjects is still based on
the comparison of mean and peak values or some other statistics of kinematics, kinetics and neuromuscular parameters with
controlled subjects [6,2327]. Moreover, in most of the previous
studies individual parameters are observed and few efforts have
been made for developing an intelligent assistive tool for monitoring recovery of athletes after anterior cruciate ligament (ACL)
injury and reconstruction [28,29]. In order to differentiate between
ACL intact (ACL-I) and ACL decient/injured, an intelligent classication system has been developed using adaptive neuro-fuzzy
inference (ANFIS) system based on arthrometric data [30]. This
system recognizes the normal and ACL injured subjects with high
values of sensitivity and specicity but it does not consider the
knee dynamics changes in ACL-R subjects. A classication model
to test the gait normality of ACL-R subjects has been proposed
in [28] based on principal component analysis and regression
modeling techniques. This system used 3-D rotational kinematics
and the classication results indicated the presence of abnormal
gait patterns in frontal and transversal planes for ACL-R subjects
still 1 year after surgery with an accuracy of 93.75%. Recently,
an intelligent recovery classication model has been presented
for ACL-R subjects to monitor their rehabilitation progress during different ambulation and balance testing activities [19]. This
system integrates the kinematics and neuromuscular signals and
the accuracy of the classication results is approximately 95% for
different activities monitored. Although these studies used computational intelligent techniques for differentiating the normal and
impaired human motion but post-surgical monitoring and prognosis of rehabilitation and sports performance of ACL-R subjects
are fairly complex processes because of the involvement of multiple varying factors for each subject. The restoration of normal
gait patterns does not ensure the complete recovery and return
to high level sporting activities. In addition, keeping a record of
pre- and post-surgery treatments and their effectiveness, adjustments in individuals rehabilitation protocol and all the different
experiences during convalescence is a complicated phenomenon
for physiatrists and physiotherapists. These experiences can be
used to learn and adapt the rehabilitation procedure for ACL-R
subjects having similar parameters or patterns. In order to build

such an exhaustive system, a knowledge base of athletes proles

can be developed to store the pre-injury, pre-surgery and postsurgery (during rehabilitation period) information about athletes
knee dynamics and other relevant parameters. Based on this information and past experiences, a learning model can be developed for
improving the recovery after ACL injury/reconstruction. An effective method to build this learning system is to use case based
reasoning (CBR) paradigm which maintains a case repository of
past experiences (old problem-solution pairs) and solves the new
problems by using or adapting the existing solutions [31].
CBR has been used as a decision supporting technique in variety of healthcare applications [32]. Based on CBR approach, a gait
disorder analysis system has been proposed in [33] to help general
practitioners. This system records lateral and forward movements
of center of mass by using an accelerometer and retrieve similar
cases based on the experts knowledge. In order to provide diagnosis and prognosis for stroke patients, CBR based system has been
successfully developed using data collected through robotic tool
[34]. The system retrieves a potential solution for stroke diagnosis by nding similar cases from the repository of stroke patients
with explicit diagnosis and prognosis. Recently, different intelligent techniques have been combined with CBR approach to build
hybrid/soft CBR systems for complex medical domains. A hybrid
system combining Self Organizing Maps (SOM) and CBR has been
developed in [35] for evaluating postural control of subjects based
on trunk sway with a prediction accuracy of more than 90%. For
diagnosis and treatment of stress, a reliable hybrid CBR system (88%
accuracy) has been designed by using nger temperature [36]. The
hybrid models based on integration of CBR and fuzzy decision tree
provide accurate forecasting in the domains of breast cancer (98.4%
accuracy) and liver disorders (81.6% accuracy) [37]. Another application of combining CBR and cluster analysis is the development of
a health monitoring system for elderly subjects based on pulse rate
and blood oxygen saturation data [38]. This system classies the
pulse rate and blood oxygen saturation with an accuracy of 93%.
This research aims at developing an intelligent framework using
knowledge-based system and articial intelligence (AI) techniques
for monitoring and classication of recovery status of athletes,
and evaluating their sports performance after ACL reconstruction.
A knowledge base has been designed to store the subjects proles and problem/solution pairs by integrating adaptive learning
techniques and reasoning model to build an extensible recovery
progress monitoring system which provides efcient and effective
solution based on the existing cases. Non-invasive body-mounted
motion and electromyography (EMG) wireless sensors have been
used to collect the kinematics and neuromuscular data from ACL
reconstructed and healthy subjects during various ambulation and
balance testing activities. Different statistical and time/frequency
features have been extracted from kinematics and EMG data, and
a feature reduction method has been applied to reduce the large
number of features and generate a set of appropriate integrated
kinematics and neuromuscular patterns. In order to facilitate the
efcient selection and retrieval of cases and exible knowledge
base design, the fuzzy clustering technique has been used to
form the groups of subjects based on their current recovery stage
after ACL reconstruction. Different intelligent classication techniques (adaptive neuro-fuzzy inference system, fuzzy unordered
rule induction algorithm and support vector machine) have been
explored and compared for nding the similar case from the
knowledge base. Once relevant cases are selected, adaptation is performed by adjusting the recovery protocols for individuals based on
the previous known solutions and the performance evaluation can
be done. This framework will facilitate the clinicians, physiotherapists, physiatrists and sports trainers in determining the recovery
stage of ACL-R subjects based on the data collected during different rehabilitation activities and identifying the subjects lacking

Author's personal copy

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141

129

Fig. 1. A knowledge-based intelligent framework for monitoring recovery progress and sports performance of subjects after ACL reconstruction.

behind the desired level of recuperation. Additionally, the feedback/solution from the previous cases will assist them to take the
required therapy/training measures or accelerating their ongoing
activity level.
2. A knowledge-based intelligent framework for ACL
recovery monitoring
The knowledge-based intelligent framework for monitoring
rehabilitation progress and sports performance of ACL-R subjects is
shown in Fig. 1. The components and sequential operations of the
system are elaborated in the following sections.

Velcro straps to note the knee joint movements. For recording the
EMG signals, foam snap electrodes were placed on four different
knee extensor and exor muscles including vastus medialis (VM),
vastus lateralis (VL), semitendinosus (ST) and biceps femoris (BF) on
both legs of the subjects during ambulation activities. For balance
testing activity, EMG data were also collected from gastrocnemius
medialis (GM) muscle in addition to above four muscles. SENIAM
guidelines were followed to prepare the skin for better conductance
of signals and placing the electrodes on identied position on muscles [39]. The command module and the BioRadio were worn by the
subjects using a waste belt. These acquired kinematics and neuromuscular signals were the basis to derive raw pattern set for the
formation of knowledge base as described in Section 2.2.

2.1. System hardware and sensors placements

2.2. Data collection and processing
The knee joint dynamics are altered due to ACL
injury/reconstruction so the major focus of the rehabilitation
process is to restore the kinematics and neuromuscular control
of ACL-R subjects back to the normal or pre-injury level. In order
to measure these kinematics and neuromuscular signals, two
subsystems have been used in this study. The kinematics data
are recorded using wireless micro-electro-mechanical systems
(MEMS) motion sensors units with command module and a radio
for wireless data transmission (KinetiSense from ClevMed Inc.)
and the neuromuscular signals are acquired through a wireless
electromyography sensors unit (BioRadio) with electrodes and a
USB receiver for wireless data transmission (BioCapture System
from ClevMed. Inc.). The body-mounted MEMS motion sensors
measure the 3-D motion using three orthogonal gyroscopes and
three orthogonal accelerometers. The BioRadio, worn by the
subjects, records the EMG signals through surface electrodes
attached to the target muscles and then wirelessly transmits them
to the computer using USB receiver. These small, light-weight and
untethered sensors can be used without obstruction to monitor
required signals from lower limbs during dynamic activities (e.g.
walking, running, balance testing, one leg jumping) where the
use of conventional equipment (e.g. arthrormeter, goniometer,
wired-EMG) may be cumbersome or not feasible.
The motion sensors were attached to each leg (one on thigh and
one on shank as shown Fig. 1) of a subject using exible bulk and

2.2.1. Data acquisition

Both sensing units (KinetiSense and BioRadio) and system software were initialized and setup for recording signals from the
motion and EMG sensors. Each motion sensor recorded the 3-D
angular rates and 3-D linear acceleration during the motion performed by the subjects. The motion sensor data were sampled at
128 Hz (12 bit analog/digital resolution within a frequency range
020 Hz) and digitized in the sensor unit. The recorded angular rates and linear acceleration from all four sensors were sent
simultaneously in serial format to the Command Module which
transferred the digital data to the computer through the USB
Receiver connected to the computers USB port. The raw signals
(angular rates and accelerations) from motion sensors were viewed
and saved using KinetiSense software package for processing.
The surface EMG sensors were used to record the action
potential of skeletal muscles indicating the force of muscles. The
amplitude of the EMG measurements (usually in millivolts) was
related to the amount of force generated from the muscles contraction while performing the ambulation or balancing activities.
The sampling rate to collect EMG signals was set to 960 Hz at
12 bit Analog/Digital conversion and 2-D linear acceleration was
also recorded from BioRadio unit to synchronize both KinetiSense
and BioCapture devices. The EMG data recorded using BioRadio
were transferred to the computer wirelessly through USB receiver

Author's personal copy

130

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141

connected to the computers USB port. These signals were viewed

and stored using BioCapture software package for processing.
The data from both sensing units were simultaneously recorded
and transmitted to the same computer. These stored signals (angular rates, accelerations and raw EMG data) were then exported to
the les for further processing by custom-developed software.
2.2.2. Data processing
The raw signals from motion and EMG sensors were ltered and
transformed into the required format in order to prepare the data
for pattern set generation and extraction. For each motion sensor,
measurements for zero-referencing were obtained prior to starting
the experiment (actual motion) when the subjects were standing
upright position and these measurements were subtracted from
angular rate measurements of corresponding sensor during the
experiment. The measurements obtained from the MEMS gyroscopes were low-pass ltered using 6th order Butterworth lter
before computing the orientations in order to minimize the motion
artifacts. After ACL injury/reconstruction, the kinematics of knee
joint change and mainly the knee movements in sagittal plane are
affected. In order to monitor these changes, the exion/extension
measurements were computed for each gait cycle using angular
rates recorded through motion sensor units placed on the thigh
and shank segments of both legs. The motion sensors were aligned
to provide knee angle about the sagittal plane. Let  represents the
orientation of a lower limb segment (thigh or shank) then the orientations of both lower extremities ( RThigh ,  RShank ,  LThigh and  LShank )
can be estimated by applying trapezoidal integration method on
respective angular rates (RThigh , RShank , LThigh and LShank ) of
lower limbs. If (t) represents the angular rate of either thigh or
shank at time t and t is the sampling time, then the estimated
orientation ( (t) ) of thigh and shank at time t is computed using (1).
(t) =

(t) + (t+1)
2

t

(1)

The knee angle was computed by subtracting  Shank from  Thigh

for both legs. Knee exion/extension and corresponding angular
rates for thigh and shank are shown for a healthy (ACL intact) leg
in Fig. 2.
Similarly, the raw EMG data from all muscles were band-pass ltered (10450 Hz) using 4th order Butterworth lter and the mean
of each signal was subtracted from it. The processed data are stored
in the knowledge base and a preliminary prole of the subject was
generated.
2.2.3. Data segmentation and feature extraction
Important features were extracted from the processed kinematics and EMG data in order to design the case and knowledge base,
and perform the data analysis and classication tasks. The selection of right features minimizes the assessment error and helps in
achieving the high retrieval accuracy. The feature set consisted of
knee dynamics based on combined kinematics and EMG signals.
Both signals were synchronized before feature set generation due
to different sampling rates and recording delays and then based on
the ambulation or balance testing activity, the features set generation and selection steps were performed. The data segmentation for
ambulation activities was done by identifying the gait cycles during
each trial for all recording sessions of a subject. The knee kinematics and relevant muscles strength vary in each phase of a gait cycle
and these changes reect the progress of the recovery after the
ACL injury/reconstruction. The features were computed for different phases of a gait cycles for each subject (Fig. 3). The marking of
gait phases and, segmentation of data from multiple EMG channels
and motion sensors were based on the percentages dened in [40].
For balance testing activity, a window of four seconds was chosen

(based on experiments) as a data segment for kinematics and EMG

signals.
Three statistics values (root mean square, standard deviation
and maximum value for knee exion/extension) were computed
for each segment (i.e. seven phases of a gait cycle for ambulation
activities and a four seconds window for balance testing activity)
for kinematics data. Thus, a derived actual pattern set generated
consisted of a total of twenty-one kinematics features for each gait
cycle for ambulation activities and three kinematics features for
balance testing activity.
The EMG features represent the electrophysiological properties of the muscle bers during contraction. In different
applications, various time-domain, frequency-domain and timefrequency-domain features extracted from EMG signals have been
used [4143]. The wavelet transformation has been found more
appropriate for analyzing such bio-signals as compared to only
time/amplitude or frequency analysis due to stochastic and nonstationary nature of EMG signals. In this study, an EMG feature set
has been generated based on multilevel discrete wavelet decomposition analysis. By employing discrete wavelet transform (DWT),
EMG signal can be iteratively transformed into multi-resolution
subsets of coefcients using suitable wavelet basis function. The
time-domain EMG signal (EMG(t) ) is passed through various high
pass and low pass lters to obtain the approximation coefcient
subsets (e.g. cA1 , . . ., cA3 as shown in Fig. 4) and the detail coefcient subsets (e.g. cD1 , . . ., cD3 as shown in Fig. 4) where the
level of decomposition can be pre-dened (e.g. 3 in Fig. 4). Mathematically, after n level of decomposition, the original signal can be
represented as follows (2):
EMG(t) = cAn + cDn + cDn1 + cD1

(2)

The choices of level and mother wavelet depend on the domain

and applications, but for EMG analysis Daubechies 04/05 mother
wavelet with four/ve levels of decomposition has shown better performance results [4345]. In this research, Daubechies 05
wavelet basis function with ve levels of decomposition has been
used to compute EMG features. An example of EMG signal analyzed
by DWT with mother wavelet Daubechies 05 (db05) with 3-levels
of decomposition is shown in Fig. 4.
For various ambulation and balance testing activities, wavelet
coefcients (cD1 cD5 and cA5 ) were computed which represent
the energy distribution of the EMG signals from four/ve identied
muscles. From these coefcients, following ve statistical features
(maximum, minimum, mean absolute value, standard deviation
and average power of the six coefcients) were calculated for each
phase of gait cycle and balance testing segment. These features
were chosen based upon previous studies for bio-signal classication [43,46] and the experiments performed in this study for
recovery and performance assessment. Thus, a derived actual pattern set consisted of a total of 840 EMG features (6 coefcients 5
EMG features 4 muscles 7 phases) for each gait cycle for ambulation activities and 150 EMG features (6 coefcients 5 EMG
features 5 muscles) for balance testing activity.
2.2.4. Feature selection/projection
The lengths of a feature vector (combined kinematics and
EMG parameters) for each ambulation and balance activity were
861 (6 5 EMG features 4 muscles 7 phases + 3 kinematics
feature 7 phases) and 153 (6 5 EMG features 5 muscles + 3
kinematics feature), respectively. In order to reduce the length
of these feature vectors, different feature selection and reduction
algorithms were investigated. Instead of using the feature selection
algorithms which select only a subset of the features to represent the model, feature projection/reduction method was preferred
due to its efciency and effectiveness. The selection algorithms
(sequential forward selection/sequential backward selection) were

Knee Flexion/Extension (degrees)

Author's personal copy

60
40
20
0
1.5

2.5

3.5

4.5

3.5

4.5

3.5

4.5

Time (sec)

Angular Rate (rad/sec)

200
100
0
100
200
1

1.5

2.5

Angular Rate (rad/sec)

Time (sec)

400

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141

200
0
200
1

1.5

2.5

3
Time (sec)

Fig. 2. (a) Knee exion/extension (degrees), (b) Angular rate for thigh (rad/s) (c) Angular rate for shank (rad/s) for multiple gait cycles of an ACL intact leg.
131

Author's personal copy

132

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141

Knee Flexion/Extension (degrees)

60
50
40
30 Load
Response
20

Pre Swing

Terminal Stance

Mid Stance

Terminal
Swing

Mid Swing

Initial Swing

10
0

10

1.3

1.4

1.5

1.6

1.7

1.8

1.9

2.1

Time (sec)
Fig. 3. Variations in knee exion/extension during different gait phases.

found to be slow to test each subset for the prediction error for the
large the feature number of features and were unable to re-assess
the features importance as once a feature was added to the subset
it could not be removed. In this study, one of the widely used feature reduction/projection technique namely principal component
analysis (PCA) was applied to determine the best combination of
the original features to form a new and smaller feature set. PCA has
been successfully used reduce the dimension of bio-signals data set
by removing the redundancy in the data and replacing the group
of variables with a single variable while still not rejecting some of
the features completely from the data set [47,48]. PCA transformed
the original feature set of kinematics and EMG parameters f F RN
into a new feature set of variables v V RM of reduced dimension
by minimizing the mean-square error (MSE) between the original
set F and projected set V [49]. For a given set of kinematics and
EMG input vectors, fi (i = 1, . . ., n), where each fi is of dimension N
(N = 861 for ambulation and 153 for balance testing activity), PCA
linearly transformed each vector fi into new vector vi by using (3).

EMG (mV)

vi = AT fi

(3)

where A is the N N orthogonal matrix whose ith column ai is

the ith eigenvector of the sample covariance matrix in (4).
1 T
fi fi
n
n

Cov =

(4)

i=1

These new variables, called principal components, were the linear combination of original kinematics and EMG features, which
form an orthogonal basis for the space of data. The feature projection step can only be performed after collecting and extracting
features for multiple trials from a group of healthy and ACLR subjects so that PCs covers variations in kinematics and EMG
parameters. Hence, a nal pattern set was generated based on this
method which was used as the input to develop the knowledge
based system.
2.3. Fuzzy clustering
In order to make the case selection and retrieval process more
efcient and to avoid searching the whole case repository, the
knowledge base was designed using clustered based structure. The

0.5
0

0.5
0

0.5

1.5
Sampling Rate

1
cA1 0
1000

2000

3000

4000

5000

6000

0.05
0

7000

4
x 10
cD1
1000

2000

3000

4000

5000

6000

7000

0.2

1
cA2 0
1

2
1
cA3 0
1
0

2.5

1
0

2
0.05

cD2

0
500

1000

1500

2000

2500

3000

0.2
0

3500

500

1000

1500

2000

2500

3000

3500

0.5
cD3

0
0.5
500

1000

1500

500

1000

1500

Fig. 4. EMG signal for vastus lateralis of a healthy (ACL intact) subject walking at 5 km/h analyzed by DWT with Daubechies 05 mother wavelet at 3-levels of decomposition.

Author's personal copy

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141

clustering was performed by grouping the subjects according to

similarities in their kinematics and EMG parameters. There could
be various groups based on the health condition or stage of recovery of subjects so unsupervised fuzzy clustering has been adopted
to assign the subjects to a group due to the imprecise nature of
motion and neuromuscular parameters. In the domains of recovery
classication or gait analysis, variations in data are more common
and the clusters of cases may be overlapping such that one subject
may belong to more than one group with different degree of memberships. Fuzzy clustering has been applied for gait analysis and
identifying the effect of temporal patterns on the walking speed
[50,51].
Fuzzy clustering partitions the sample space and organizes the
data into approximate clusters [52,53]. In this study, fuzzy C-means
(FCM) algorithm was applied to the transformed feature set V
of kinematics and neuromuscular data collected during various
ambulation and balance testing activities. The fcm function in Fuzzy
Logic Toolbox from MATLAB starts with an initial guess for the cluster centers for marking the mean location of each cluster. Each
feature vector as a data point is then assigned a membership grade
for each cluster. fcm follows an iterative process to minimize the
objective function (5) and then decides the right cluster centers (6)
and membership grade (7) for each feature vector.
C
N 


um
||vi cj ||2
ij

(5)

i=1 j=1

N

um v
i=1 ij i

cj =

N

uij =

C

(6)

um
i=1 ij
1

(||vi cj ||/||vi ck ||)

k=1

2/m1

(7)

where vi is the ith vector of M -dimensional data from PCA, cj is

the M -dimension center of the cluster, uij is the degree of membership of vi in the cluster j, || || represents the similarity (Euclidian
distance) between any measured data and the center, and m is
the number of clusters. In the initial stage of populating the case
library, the variation in number of clusters indices is possible which
is reduced as more cases are added to the repository. In order to
identify the number of initial clusters, a modied cluster validity
measure has been used [54] which is mathematically illustrated as
follows (8).

K 
Validity =

i=1

2


 /KN
vV vci v ci


2

min(ci cj  )

(8)

where ci and cj are the cluster centers, N is the number of records

in the data set, k is the maximum number of clusters, v is a subset
(selected features) of V and ||v ci ||2 and ||ci cj ||2 represent the
intra and inter cluster distances, respectively.
In the proposed framework, fuzzy clustering step was an iterative process and as more cases would be generated and added to the
repository, the number of clusters would be rened. The clusters
were generated based on the data collected from two trials of each
activity from all subjects so that the variations in the gait/balance
patterns could be recorded and used to generate the classes. These
clusters were used to train the intelligent classier(s) for efcient
retrieval of cases.
2.4. Case generation
The generation of cases was based on two scenarios; initially
when the case repository was empty and subsequently when case
base was already populated with some cases. In the rst situation

133

when there was no case in the repository, data from a group of

healthy and ACL-R subjects at different recovery stages were collected and clustered using fuzzy clustering technique to form the
groups based on their current health condition and recovery stage.
The labels for these groups were identied by nding the distances
of their centers from the center of healthy subjects cluster and
were manually veried. Then new cases were generated based on
a semi-automatic process where both extracted features and recommendations from the clinicians were stored using the proposed
case representation (explained in Section 2.6.2). In the later scenario, a new case was added based on the outcome of CBR cycle
(retrieve, reuse, revise and retain) and re-grouping/re-training of
classier was performed if required. Each solved case, stored in
the repository, included the problem description, average value of
transformed attribute/value pairs for multiple trials, recovery stage,
class, recommendations and protocols followed during rehabilitation.
2.5. Intelligent classication mechanisms
The retrieval of a case from the case repository essentially means
matching a current input pattern with one or more stored patterns
or cases. Intelligent classication techniques have been proved very
efcient for performing pattern matching task. The variations in
kinematics signals, and non-stationary nature of EMG data lead
the recovery classication task challenging. The adaptive intelligent
techniques can be more useful for building models for such inputs.
These techniques can effectively identify the stochastic changes
in bio-signals, and can also deal with the impreciseness in measurements and variations due to subjects physiological conditions
[55,56]. More specically, fuzzy logic based techniques are very
useful when the data are incomplete, noisy, or imprecise. Three different intelligent classication techniques have been explored and
compared in this study in order to implement the case retrieval task
for recovery and performance analysis.
2.5.1. Adaptive neuro-fuzzy inference system (ANFIS)
ANFIS has been applied for addressing the problem of selecting
cases from overlapping cluster regions. The ANFIS is a fuzzy Sugeno
model that adapts the membership function parameters using neural network and learns from the given data set [57]. ANFIS combines
the learning capabilities of neural networks with fuzzy set theory
which handles the uncertainty arising due to overlapping cases. The
system adjusts the membership function () parameters based on
the given data, and the number of rules and output of fuzzy rules
is minimized. The overall output of an n-input system is given in
(9) where li is the previous layers output and wi is called the ring
strengths of the rules (10).
y=

n


i li =
w

i=1

i=1

wi =

n
wi li
i=1
n

n


ij (xj )

wi

(9)

(10)

j=1

The subtractive-clustering method was used to partition the

data due to large number of inputs, no requirement of setting the
number of clusters in advance and noise robustness. It is a onepass algorithm for estimating number of clusters by nding the high
density data point regions in feature space. The cluster center is the
point with the highest number of neighbors. The learning parameters of membership functions (premise parameters) and output
(consequent parameters) were tuned using the hybrid learning
algorithm. This algorithm combines the least square method and
gradient descent method that make the convergence faster in the

Author's personal copy

134

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141

large search space. The forward pass (least square method) and
backward pass (gradient descent method) are used to optimize the
consequent and premise parameters, respectively. After determining the consequent parameters, the output of ANFIS is calculated
and the premise parameters are adjusted based on output error by
using the back-propagation algorithm [57].
2.5.2. Fuzzy unordered rule induction algorithm (FURIA)
FURIA is a recent classication algorithm based on the RIPPER
algorithm with certain modications and extensions [58]. It learns
the fuzzy rules having soft boundaries as compared to conventional
strict rules and provides an unordered rule set using one vs. rest
strategy such that there is no default rule. A rule stretching mechanism is used in case when a new problem is not covered by any of
the existing rules. A fuzzy rule in FURIA is obtained through replacing the RIPPERs intervals with fuzzy intervals (IF ) of the form of
trapezoidal membership functions specied with four parameters
as follows:

IF =

1
bvc

va

ba a<v<b

(11)

dv

c<v<d

d
c

else

where b and c are the lower and upper bound of the core of the fuzzy
set, respectively, and a and d are the lower and upper bound of the
support, respectively. If a fuzzy selector (Ai IiF ) covers an instance
x = (x1 , . . ., xn ) to the degree IiF (xi ) then a fuzzy rule rF involving k
selectors (Ai IiF ), i = 1, . . ., k, covers x to the degree as follows:
n


r F (x) =

IiF (xi )

(12)

i=1..k

The fuzzication of a (Ai IiF ) is done by considering only the

relevant training data DTi and DTi is partitioned into positive and
negative instances. For a new query instance x, the support of this
class is dened by
sj (x) =

(j)

(13)

r (j) .CF(ri )
i

i=1...k
(j)

Different kernel mapping (quadratic, Gaussian radial basis function,

polynomial etc.) can be utilized other than the linear function to
improve the classiers performance based on the specic domain.
This study has compared the effect of three types of kernel functions
(linear, quadratic and radial basis function) on the classication
accuracy for ACL recovery analysis. In order to deal with the multiclass problem, the binary SVM has been modied using one vs. all
strategy.
2.6. Hybrid case based reasoning
A hybrid intelligent framework was developed by combining case-based reasoning (CBR) approach and adaptive intelligent
mechanisms for recovery stage diagnosis and prognosis of ACLR subjects. The framework utilizes the concept of solving new
problems by using/modifying the similar previous experiences
(problemsolution pairs). CBR problem-solving cycle consists of
four steps [31]:
Retrieve: Finding similar case(s) from the knowledge base whose
problem description best matches with the given problem.
Reuse: Reusing the solution of most similar case to solve the new
problem.
Revise: Adapting/modifying the chosen solution according to the
differences in new problem.
Retain: Storing the new problem-solution pair as a case once it
has been solved.
The details about the different components of the hybrid intelligent knowledge-based system are described in next sections.
2.6.1. Knowledge base (KB)
The overall structure of knowledge base is depicted in Fig. 5. The
knowledge base contains different types of information including
case library, raw and processed data, domain knowledge, historical data available for subjects (pre-injury, post-injury) and session
data during convalescence. In order to manage the knowledge base
repository, a relational database was used to reduce the storage
redundancy and provide exibility. The knowledge base evolves
with the time-period when new problems are presented and new
cases are added to the system. This evolution process makes it
more useful for domains where subjects specic monitoring and
prognosis mechanisms are required.

where CF (certainty factor of rule ri is dened as follows:

(j)

(j)

CF(ri ) =

(2DT /DT ) +

(j)
T

xD

(j)
T

xD

r (j) (x)

r (j) (x)

(14)

2.5.3. Support vector machines (SVMs)

SVMs combine the statistical learning theory with optimization
techniques and kernel mapping to solve the classication problem.
In case of binary support vector classication, the feature vectors
are separated based on optimal hyperplanes [59]. For linearly separable data, the hyperplane separating the data into classes can be
described using following equation:
xT w + w0 = 0

(15)

where x represents the feature vector set and w is normal to the

hyperplane. An optimal separating hyperplane maximizes the sum
of distances from the hyperplane to the closest data points on each
side. If yi represents the possible class labels (+1/1) then the optimal hyperplane can be found by minimizing ||w2 || subject to
yi .(xiT w + w0 ) 1

(16)

2.6.2. Case representation

Each case in the knowledge base is composed of two components: activity set and overall set. The activity set consists of subsets
of activity-based (walking at different speeds and balance testing)
problemsolution pair. The problem part of each subset is represented as attribute/value pairs for the selected kinematics and
EMG features and the corresponding solution part is made up of
the recovery classication and the treatments suggested for the
next stage. The overall set contains the performance evaluation
for the athletes, treatments given and followed at current stage of
recovery, link to the previous sessions outcomes (if any) and case
description. The case description contains the subjects biographical, sports and other relevant clinical information. This information
may provide assistance in rening the retrieval and adaption of
similar cases.
2.6.3. Retrieval of similar cases
In order to retrieve the most similar cases for the given scenario,
a two step process is performed. In the rst step, the best matching
class is identied for the new problem by using trained classier
as described in Section 2.5. In the second step, the most similar

Author's personal copy

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141

135

Fig. 5. Overall structure of knowledge base containing different types of data.

cases are retrieved from the identied cluster by using case density
function given in (11) [60].

3.1. Subjects


Case Density(e, C) =

3. Experimental setup

e C{e}

SMee

|C| 1

(17)

where SMee  represents the similarity between case e and e

(computed by cosine/Euclidian measure) and |C| is the number of
cases in class C. The retrieved cases are arranged in the descending
order of case densities and rst k cases are selected where k is the
users choice.

2.6.4. Reuse and repair of cases

After retrieving the most similar case(s), the next step is to
use and adapt the solution of this case to improve the recovery
process of the athlete for the next stage. This semi-automatic process requires the involvement of the clinicians/physiotherapist to
decide any changes in the rehabilitation protocol based on the
recommendations and indication of performance level from the
retrieved case(s). Additionally, modications may also be done in
the recommendation section of the previous stages of the new
problem.

2.6.5. Retaining the learned case

The new case may be retained in the library after formulating
a solution based on the adjustments of parameters for individuals.
The attribute/value pairs for the selected features are generated by
taking an average of their values for all sets. Whenever a new case
is added, the fuzzy clustering algorithm generates new clusters (if
required) and assigns different membership grades to the related
cases.

Nineteen subjects (6 healthy with no ACL or other knee injury

and 13 unilateral ACL reconstructed) were recruited from Performance Optimization Center at Ministry of Defense and Sports
Medicine and Research Center at the Ministry of Sports, Brunei
Darussalam. The healthy subjects were having an age (mean std)
of 29.5 6.90 years, height (mean std) 162.5 11.02 cm, and
weight (mean std) 64.25 17.10 kg. For ACL reconstructed
subjects the age (mean std), height (mean std) and weight
(mean std) were 31.6 3.52 years, 165.9 6.28 cm and
67.4 12.6 kg, respectively. The ACL-reconstructed subjects
were at different stages of recovery ranging from 3 months to
a year after surgery. All subjects read and signed an informed
consent form and ethical procedures were carried out according
to the guidelines approved by the Graduate Research Ofce and
Ethics Committee at Universiti Brunei Darussalam.

3.2. Activities monitored

During rehabilitation period of the subjects after ACL
surgery, different activities are assessed by the physiotherapists/physiatrists as per the rehabilitation protocol. Due to
variations in the post-surgery time periods of the subjects and
based on the guidelines from the physiatrist, two activities (walking at different speeds and single leg at surface open eye balance
testing) were selected from the rehabilitation protocol for the data
acquisition. All of the subjects walked at three different speeds:
4 km/h, 5 km/h and 6 km/h for 3035 s time on a motorized treadmill. The balance testing data were collected from the subjects
standing on single leg on a at surface for a time period of 1520 s.

Author's personal copy

136

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141

The balance test was performed for both legs of the ACL reconstructed subjects. The data were collected for multiple trials (2 trials
for ambulation activity and 4 trials for balance testing activity) for
each activity in order to take care of the variations in kinematics and
EMG parameters, and generating a large pattern set for performing
fuzzy clustering and training the classiers.

Table 1
Number of rules generated by ANFIS and FURIA classiers for different activities.
Activity\technique

ANFIS

FURIA

Walking at 4 km/h
Walking at 5 km/h
Walking at 6 km/h
Single leg balance test

8
5
7
15

4
5
3
3

4. Results and discussion

In order to generate the feature set and populate the case repository, data were collected from healthy and ACL-R subjects during
three ambulation and one balance testing activities. A large number of kinematics and EMG features were extracted from the human
motion during these activities to evaluate their effect on the recovery analysis after ACL reconstruction. However, the large feature
vector/set represented the data redundancy and made the classication process inefcient so PCA was applied to reduce the feature
sets for all activities monitored during this study. Before applying
PCA, the total data set was partitioned into two groups. The rst
group consisted of data from 14 subjects (4 healthy and 10 ACL-R)
and the second group contained data from 5 subjects (2 healthy and
3 ACL-R). The data from rst group were used to generate coefcient matrix from PCA, form clusters, train and test classiers, and
generate cases. For ambulation activities, the size of the pattern
set for rst group was 280 861 (i.e. 240 records from 14 subjects
for 20 gait cycles with 861 kinematics and EMG features). For balance testing activity, the size of the pattern set for the same group
was 56 153 (i.e. 56 records from 14 subjects for 4 segments with
153 kinematics and EMG features). The latter group was used to
test the performance of the framework. PCA was applied on the
data for rst group and the transformed features with respective
coefcient matrix were generated for each activity. The variations
explained by principal components (PCs) for these activities are
shown in Fig. 6. The number of PCs required for a cumulative distribution of variance of 90% or more vary among activities and were
found as follows: 42 PCs for walking at 4 km/h, 39 PCs for walking
at 5 km/h, 37 PCs for walking at 6 km/h and 6 PCs for balance testing. The coefcient matrix for each activity was stored to apply to
the data for new subjects. Fig. 7 shows the original data projected
on rst three PCs for walking at 4 km/h. Another advantage of PCA
is that it can also be useful in identifying the subjects who have
very different values of the input parameters (marked inside circles in Fig. 7). These data points can be treated as outliers/error
and removed from the data set before clustering/classication step
or further steps can be taken to investigate the subjects for whom
these data points belong to.
Before generating the cases, the transformed features were
clustered using FCM to form the groups of subjects who were
healthy or at similar stage of recovery after ACL reconstruction.
This step is part of the initial grouping and case generation process
and can be applied if re-grouping is required. Three clusters were
initially identied for each activity using proposed validity index
(8) which were manually veried and found to be appropriate.
Fig. 8 and Fig. 9 show 3-D scatter plot for the rst three PCs and the
cluster centers identied by FCM for walking at 5 km/h and 6 km/h
respectively, where clusters 1, 2 and 3 represent subjects less than
6 months after surgery, subjects between 6 months and 1 year
of surgery and subjects without ACL injury (i.e. healthy subjects),
respectively. Some of the data points lie on the boundary of
second and third clusters which depicts the overlapping of groups
such that some of the subjects belong to both clusters with certain
membership grades and cannot be completely categorized into one
group. These are the subjects who have recovered to a level where
they are similar to the subjects in healthy group. This is natural as
even after following the same rehabilitation protocol, the recovery

may depend on individuals other physical parameters. After forming the clusters, the case generation step was performed by storing
problemsolution pairs, corresponding treatments given/followed
by the subjects and the recommendations from the physiatrists.
In order to build an efcient case retrieval model, three different classiers (ANFIS, FURIA and SVM) were trained and tested for
each activity based on the respective clustered data for healthy and
ACL-R subjects, and their performance measures were compared.
The classiers models for ANFIS and SVM were developed using
MATLAB 7.0 functions and Weka (3.7.9) was used for generating
classier models for FURIA. The cross validations of all models were
done by partitioning the data into two groups: training data (70%
of the total data) and test data (30% of the total data). The partitions for ANFIS and SVM classiers were made by using cvpartition
function from MATLAB which randomly partitions the observations
into a training set and a test set with stratication such that both
training and test sets contain approximately the same class proportions as in the clusters groups (original groups). For FURIA,
the partitioning task was implemented using Weka Experimenter
where the randomized data split option was used for training and
testing the model. The training/testing phase for all classiers was
repeated 100 times and the average values of different performance
measures were computed.
The ANFIS networks were trained for 1000 epochs and the step
size was set to an initial value of 0.01. The networks were generated using gens2 with gaussmf input membership functions. The
parameters for initial membership functions were tuned using the
training/testing of the networks and the examinations of nal (after
training) membership functions indicated that there were signicant differences in the parameters for these membership functions
for all activities. Fig. 10 shows the initial (before training) and nal
(after training) membership functions of two inputs (PCs) for walking at 5 km/h speed. Similar trends were found for other activities
also. Both ANFIS and FURIA based classiers generated the rules
after training phase and these rules (models) were saved for further validation and retrieval of cases for new problems. However,
it was found that the number of rules generated by FURIA based
classiers were, in general, much less than the rules generated by
ANFIS models for different activities which indicates the efcient
rule learning/pruning mechanism in FURIA. The number of rules
generated by ANFIS and FURIA models for four different activities
is shown in Table 1. Multi-class SVM models were also developed
for different activities using three different kernel functions namely
linear, quadratic and Radial Basis Function (RBF). The overall classication accuracies of all three types of models were more than 95%
but linear-SVM model was found more accurate in classifying the
health status of subjects as compared for quadratic- and RBF-based
SVM models (Fig. 11).
The experiments showed that only rst few PCs were enough
to achieve high classication accuracy for all three types of classiers. In general, it was found that rst 37 PCs produced that
maximum classication accuracy for all techniques for different
activities. Fig. 11 shows a comparison of overall classication accuracy of three models for different activities. ANFIS based classiers
performed better than other two models for two walking activities
(4 km/h and 5 km/h) but it poorly classied the data for balance testing activity. For walking at 6 km/h and balance testing activities,

Author's personal copy

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141

137

50
Walking at 4 km/h
Walking at 5 km/h
Walking at 6 km/h
Single Leg Balance

45

Variations Explained (%)

40
35
30
25
20
15
10
5
0
0

10

15

20
25
30
Number of Principle Component

35

40

45

50

Fig. 6. Percentage of variance explained by rst 50 principal components for ambulation and balance testing activities.

Principal Component 3

30

20

10

10

20
20
10
0
10
20

Principal Component 2

30
40
50
60

20

10

20

10

40

30

50

60

Principal Component 1

Fig. 7. Original data projected on rst three principal components.

FURIA and SVM classiers were found more effective and these
models were able to classify the data for these two activities
without any error (100% accuracy). The sensitivity, specicity and
F-measure computed for each classication model for three groups
(class 1 subjects after less than 6 months of surgery, class 2
subjects between 6 months and 1 year of surgery, class 3 healthy

subjects without ACL injury) are shown in Tables 24. The high
values of these performance measures indicate that all three types
of classiers are suitable for classifying the recovery state/health
condition of subjects after ACL reconstruction (Fig. 12).
The trained models and cases for each activity were stored and
the second data set (data from 5 new subjects with known classes)

Principal Component 3

15

10

10
25
20
15
10
5
0

Principal Component 2

5
10
15

50

40

30

20

10

Principal Component 1

Fig. 8. Clusters centers identied by FCM for walking at 5 km/h cluster 1 () cluster 2 (o) cluster 3 (+).

10

20

Author's personal copy

138

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141

Principal Component 3

10

10

15
15
10

30

20
0

10
0

Principal Component 2

10

10

20
15

Principal Component 1

30

in2cluster1

in2cluster4 in2cluster2

in2cluster5 in2cluster3

Degree of membership

Degree of membership

Fig. 9. Clusters centers identied by FCM for walking at 6 km/h cluster 1 () cluster 2 (o) cluster 3 (+).

0.8
0.6
0.4
0.2
0
15

10

10

in3cluster1

0.6
0.4
0.2
0

15

10

in2cluster4in2cluster2

n2cluster5
i
in2cluster3

Degree of membership

Degree of membership

in2cluster1

0.8
0.6
0.4
0.2
0
15

10

5
0
5
10
c) Input 1 After Training

15

20

in3cluster4

0.8

0
5
10
b) Input 2 Before Training

a) Input 1 Before Training

in3cluster2

in3cluster1

in3cluster2

15

in3cluster4

0.8
0.6
0.4
0.2
0
10

0
5
10
d) Input 2 After Training

15

Fig. 10. Membership functions for two inputs of the ANFIS network before and after training for walking activity (5 km/h).

was used for validation of the retrieval process and the framework. PCA coefcient matrices were applied on the given data,
and three classication models were used to identify the classes
and retrieve the cases from the case repository. The classication
performances of three models were compared (Fig. 13) and it was

found that FURIA based classication model performed better than

the other two models for data from new subjects. After classifying the recovery stage, the relevant cases were selected using case
density function (17). The performance of each subject during each
Table 2
ANFIS classication performance for all activities.

Fig. 11. Overall classication accuracy using linear, quadratic and RBF kernel functions for SVM for different activities monitored during recovery analysis.

Activity

Class

Sensitivity (%)

Specicity (%)

F-measure (%)

Walking
4 km/h

1
2
3

99.66
98.98
100.00

99.37
99.75
100.00

99.54
99.19
100.00

Walking
5 km/h

1
2
3

99.54
95.99
100.00

97.31
99.62
100.00

98.85
97.42
100.00

Walking
6 km/h

1
2
3

99.32
100.00
100.00

100.00
99.70
99.90

99.65
99.64
99.88

Single leg
balance test

1
2
3

79.83
87.00
100.00

94.17
95.29
93.94

82.81
87.96
94.94

Author's personal copy

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141
Table 3
FURIA classication performance for all activities.
Activity

Class

Sensitivity (%)

Specicity (%)

F-measure (%)

Walking
4 km/h

1
2
3

96.50
94.10
98.60

97.80
98.50
100.00

96.50
91.40
99.30

Walking
5 km/h

1
2
3

98.50
96.60
100.00

97.80
98.80
100.00

98.50
96.60
100.00

Walking
6 km/h

1
2
3

100.00
100.00
100.00

100.00
100.00
100.00

100.00
100.00
100.00

Single leg
balance test

1
2
3

100.00
100.00
100.00

100.00
100.00
100.00

100.00
100.00
100.00

Table 4
SVM classication performance for all activities.
Activity

Class

Sensitivity (%)

Specicity (%)

F-measure (%)

Walking
4 km/h

1
2
3

98.22
100.00
100.00

100.00
98.91
100.00

99.08
99.51
100.00

Walking
5 km/h

1
2
3

89.79
100.00
100.00

100.00
93.09
100.00

94.44
97.69
100.00

Walking
6 km/h

1
2
3

100.00
100.00
100.00

100.00
100.00
100.00

100.00
100.00
100.00

Single leg
balance test

1
2
3

100.00
100.00
100.00

100.00
100.00
100.00

100.00
100.00
100.00

139

activity was compared with the most similar retrieved cases by

using their recommendations and next stage results.
This study demonstrates the successful development of a
knowledge-based framework for rehabilitation and performance
monitoring of ACL-R subjects using hybrid intelligent techniques
(CBR and intelligent clustering and classication algorithms).
Anterior cruciate ligament injury alters the human locomotion
and results in various short- and long-term affects including
loss of prioprioception, cartilage degeneration and early onset
of osteoarthritis. Although, through ACL reconstruction certain
improvements in the knee dynamics are achieved but the changes
in the kinematics and neuromuscular parameters have been
reported to persist in anterior cruciate ligament reconstructed
(ACL-R) subjects for several months after surgery. A comprehensive
system based on CBR methodology and hybrid intelligent techniques can be used to identify the need for timely intervention
in the current rehabilitation protocol of new subjects and making
adjustments as per the requirements, thus minimizing the risk of
re-injury. The knowledge-based system can store the pre-injury,
post-injury and recovery data in order to provide a detailed analysis about the recovery of subjects and a comparison of current
and past performance in any specic activity. The collection of relevant existing experiences for rehabilitation can help in designing a
personalized rehabilitation protocol as well as improving the generalized rehabilitation process.
Three different types of classiers (ANFIS, FURIA and SVM)
were investigated and the results show that for retrieving similar cases (i.e. differentiating the health/recovery stage of different
subjects), the performance of FURIA based model is slightly better than the other two models (Fig. 13). Various statistical features
were extracted from the kinematics and electromyography data for
ambulation and balance testing activities and these features provided useful information for developing accurate classiers. The
integration of kinematics and electromyography data has been
proved quite effective in identication of health condition of different subjects. The data processing of stochastic and non-stationary
EMG signals was done by using DWT analysis which provided reliable features in order to form data clusters and designing accurate
classiers for case selection and retrieval.

5. Conclusions and future work

Fig. 12. Overall classication accuracy using ANFIS, FURIA and SVM (linear) for
different activities monitored during recovery analysis.

Fig. 13. Overall classication accuracy for case retrieval using trained classiers
(ANFIS, FURIA and SVM) for new subjects to test the framework.

A knowledge-based framework has been implemented using

hybrid intelligent techniques which will help in diagnosis and prognosis of ACL-R subjects during recuperation. A case repository of
healthy and various ACL-R subjects at different stages of recovery
has been constructed using fuzzy clustering of integrated kinematics and EMG data for implementing CBR model. The framework
was validated for a group of healthy and ACL-R subjects with very
high retrieval accuracy. Three intelligent classiers were trained
and tested, and during training/testing phase FURIA and SVM models classied all instances of the data for ambulation at a speed of
6 km/h and balance testing activities with 100% accuracy. While
for other two activities (walking at speeds of 4 km/h and 5 km/h),
ANFIS classiers performed slightly better than the other models.
The trained classiers were then applied to a group of subjects not
included in the training and testing phase, and the results indicated that FURIA based classication model performed well for all
activities. This framework can provide a valuable feedback to clinicians and physiatrist in order to evaluate the rehabilitation status of
the subjects after ACL injury/surgery and determine the potential
solutions to the knee joint and neuromuscular problems during rehabilitation. Moreover, maintaining the athletes pre-injury,
post-injury and post surgery proles can provide a comprehensive platform for ACL injury prevention and sports performance

Author's personal copy

140

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141

evaluation. The system will be expanded in future by including data

for further activities from the rehabilitation protocol and collecting
pre- and post injury data from subjects in order the compare the
deviations/changes in the post-surgery sports performance. Additional parameters (e.g. gender, age, type of sports) will also be
considered for evaluating the recovery progress of subjects.
Acknowledgments
This work is supported by the University Research Council (URC)
grant scheme at the Universiti Brunei Darussalam under the grant
no. UBD/PNC2/2/RG/1(195) with the title Integrated Motion Analysis System (IMAS). The authors would like to acknowledge the
support provided by Ministry of Sports and Ministry of Defense,
Brunei Darussalam for providing national athletes as test subjects
for this study.

[21]

[22]

[23]

[24]

[25]

[26]

[27]

References
[1] G.A. Murrell, S. Maddali, L. Horovitz, S.P. Oakley, R.F. Warren, The effects of time
course after anterior cruciate ligament injury in correlation with meniscal and
cartilage loss, Am. J. Sports Med. 29 (2001) 914.
[2] B. Reider, M.A. Arcand, L.H. Diehl, K. Mroczek, A. Abulencia, C.C. Stroud, M. Palm,
J. Gilbertson, P. Staszak, Proprioception of the knee before and after anterior
cruciate ligament reconstruction, Arthroscopy 19 (2003) 212.
[3] H.G. Potter, S.K. Jain, Y. Ma, B.R. Black, S. Fung, S. Lyman, Cartilage injury after
acute, isolated anterior cruciate ligament tear: immediate and longitudinal
effect with clinical/MRI follow-up, Am. J. Sports Med. 40 (2011) 276285.
[4] M. Hall, C.A. Stevermer, J.C. Gillette, Gait analysis post anterior cruciate ligament
reconstruction: knee osteoarthritis perspective, Gait Posture 36 (2012) 5660.
[5] A. Jin, J. Lynch, C. Scholes, Q. Li, M. Coolican, D. Parker, Osteoarthritis following
anterior cruciate ligament reconstruction, J. Bone Joint Surg. Br. 94-B (2012)
125.
[6] B. Gao, N.N. Zheng, Alterations in three-dimensional joint kinematics of anterior cruciate ligament-decient and -reconstructed knees during walking, Clin.
Biomech. 25 (2010) 222229.
[7] G.D. Myer, M.V. Paterno, K.R. Ford, C.E. Quatman, T.E. Hewett, Rehabilitation after anterior cruciate ligament reconstruction: criteria-based progression
through the return-to-sport phase, J. Orthop. Sports Phys. Ther. 36 (2006)
385402.
[8] G.S. Roi, D. Creta, G. Nanni, M. Marcacci, S. Zaffagnini, L. Snyder-Mackler, Return
to ofcial Italian rst division soccer games within 90 days after anterior cruciate ligament reconstruction: a case report, J. Orthop. Sports Phys. Ther. 35
(2005) 2261, discussion 6166.
[9] D. Adams, D.S. Logerstedt, A. Hunter-Giordano, M.J. Axe, L. Snyder-Mackler,
Current concepts for anterior cruciate ligament reconstruction: a criterionbased rehabilitation progression, J. Orthop. Sports Phys. Ther. 42 (2012)
601614.
[10] T. Shaw, L.S. Chipchase, M.T. Williams, A users guide to outcome measurement
following ACL reconstruction, Phys. Ther. Sport 5 (2004) 5767.
[11] P.P. Gogia, J.H. Braatz, S.J. Rose, B.J. Norton, Reliability and validity of goniometric measurements at the knee, Phys. Ther. 67 (1987) 192195.
[12] P.J. Miller, Assessment of joint motion, in: J.M. Rothstein (Ed.), Measurement
in Physical Therapy, Churchill Livingstone, New York, 1985, pp. 103136.
[13] I. Fiebert, J. Gresley, S. Hoffman, K. Kunkel, Comparative measurements of
anterior tibial translation using the kt-1000 knee arthrometer with the leg in
neutral, internal rotation, and external rotation, J. Orthop. Sports Phys. Ther. 19
(1994) 331334.
[14] W.P. Hanten, M.B. Pace, Reliability of measuring anterior laxity of the knee joint
using a knee ligament arthrometer, Phys. Ther. 67 (1987) 357359.
[15] C.L. Highgenboten, A. Jackson, N.B. Meske, Genucom, kt-1000, and stryker knee
laxity measuring device comparisons. Device reproducibility and interdevice
comparison in asymptomatic subjects, Am. J. Sports Med. 17 (1989) 743746.
[16] S.L. Di Stasi, D. Logerstedt, E.S. Gardinier, L. Snyder-Mackler, Gait patterns differ between ACL-reconstructed athletes who pass return-to-sport criteria and
those who fail, Am. J. Sports Med. 41 (2013) 13101318.
[17] Y. Geng, L. Zhang, D. Tang, X. Zhang, G. Li, Pattern recognition based forearm
motion classication for patients with chronic hemiparesis, in: Proceeding of
the 35th Annual International Conference of the IEEE, Engineering in Medicine
and Biology Society (EMBC), 2013.
[18] H. Joutsijoki, K. Varpa, K. Iltanen, M. Juhola, Machine learning approach to an
otoneurological classication problem, in: Proceeding of the 35th Annual International Conference of the IEEE, Engineering in Medicine and Biology Society
(EMBC), 2013.
[19] S. Senanayake, O.A. Malik, M. Iskandar, D. Zaheer, 3-d kinematics and neuromuscular signals integration for post ACL reconstruction recovery assessment,
in: IEEE Proceedings of 35th Annual International Conference of the IEEE EMBS,
Osaka, Japan, 2013, ISBN 978-1-4577-0216-7, pp. 72217225.
[20] M. Vallejo, C.V. Isaza, J.D. Lopez, Articial neural networks as an alternative to traditional fall detection methods, in: Proceeding of the 35th Annual

[28]

[29]

[30]

[31]
[32]

[33]

[34]

[35]

[36]

[37]

[38]

[39]

[40]
[41]
[42]

[43]
[44]

[45]

[46]

[47]

International Conference of the IEEE, Engineering in Medicine and Biology

Society (EMBC), 2013.
W. You, A. Simalatsar, G. De Micheli, Parameterized SVM for personalized drug
concentration prediction, in: Proceeding of the 35th Annual International Conference of the IEEE, Engineering in Medicine and Biology Society (EMBC), 2013.
Y. Li, Y. Tian, W. Chen, SEMG signal classication for the motion pattern of
intelligent bionic articial limb, in: D. Liu, H. Zhang, M. Polycarpou, C. Alippi,
H.D. He (Eds.), Advances in Neural Networks ISNN 2011 se 58, Lecture notes
in Computer Science, 6675 ed., Springer, Berlin, Heidelberg, 2011, pp. 505513,
DA 2011/01/01.
C.O. Moraiti, N. Stergiou, H.S. Vasiliadis, E. Motsis, A. Georgoulis, Anterior cruciate ligament reconstruction results in alterations in gait variability, Gait Posture
32 (2010) 169175.
D.L. Shi, Y.B. Wang, Z.S. Ai, Effect of anterior cruciate ligament reconstruction
on biomechanical features of knee in level walking: a meta-analysis, Chin. Med.
J. (Engl.) 123 (2010) 31373142.
A.C. Thomas, M. Villwock, E.M. Wojtys, R.M. Palmieri-Smith, Lower extremity
muscle strength after anterior cruciate ligament injury and reconstruction, J.
Athl. Train. 48 (3) (2013), http://dx.doi.org/10.4085/1062-6050-48.2.14.
W.I. Drechsler, M.C. Cramp, O.M. Scott, Changes in muscle strength and EMG
median frequency after anterior cruciate ligament reconstruction, Eur. J. Appl.
Physiol. 98 (2006) 613623.
M. Lewek, K. Rudolph, M. Axe, L. Snyder-Mackler, The effect of insufcient
quadriceps strength on gait after anterior cruciate ligament reconstruction,
Clin. Biomech. (Bristol Avon) 17 (2002) 5663.
G. Leporace, L.A. Batista, A.M. Muniz, G. Zeitoune, T. Luciano, L. Metsavaht, J. Nadal, Classication of gait kinematics of anterior cruciate ligament
reconstructed subjects using principal component analysis and regressions
modelling, Conf. Proc. IEEE Eng. Med. Biol. Soc. 2012 (2012) 65146517.
S.M.N.A. Senanayake, O.A. Malik, P.M. Iskandar, Wireless multi-sensor integration for ACL rehabilitation using biofeedback mechanism, in: Proceedings
of ASME 2012 International Mechanical Engineering Congress & Exposition,
IMECE12, Houston, TX, USA, 2012.
Z. Heydari, F. Farahmand, H. Arabalibeik, M. Parnianpour, Adaptive neuro-fuzzy
inference system for classication of ACL-ruptured knees using arthrometric
data, Ann. Biomed. Eng. 36 (2008) 14491457.
A. Aamodt, E. Plaza, Case-based reasoning: foundational issues, methodological
variations, and system approaches, AI Commun. 7 (1994) 3959.
S. Begum, M.U. Ahmed, P. Funk, N. Xiong, M. Folke, Case-based reasoning systems in the health sciences: a survey of recent trends and
developments, Trans. Syst. Man Cybern. Part C 41 (2011) 421434,
http://dx.doi.org/10.1109/tsmcc.2010.2071862.
O.A. Kuchar, Y.L. Lazukova, D. Riordan, J.L. Leahey, A case-based reasoning system involving a quantized method for gait disorder diagnosis, in: Proceedings of
IASTED International Conference on Articial Intelligence and Soft Computing,
Cancun, Mexico, 1998, pp. 367370.
M. Baig, Case-based ReasoningAn Effective Paradigm For Providing Diagnostic
Support for Stroke Patients, Master of Science thesis, School of Computing,
Queens University, Canada, 2008.
A.U. Alahakone, S.M.N.A. Senanayake, A hybrid intelligent system combining
self organizing maps and case based reasoning for evaluating postural control, in: IEEE Conference on Multisensor Fusion and Integration for Intelligent
Systems (MFI), 2010.
M.U. Ahmed, S. Begum, P. Funk, A hybrid case-based system in clinical diagnosis
and treatment, in: Proceedings of the IEEE-EMBS International Conference on
Biomedical and Health Informatics (BHI), 2012, Hong Kong, 2012.
C.-Y. Fan, P.-C. Chang, J.-J. Lin, J.C. Hsieh, A hybrid model combining case-based
reasoning and fuzzy decision tree for medical data classication, Appl. Soft
Comput. 11 (2011) 632644.
M.U. Ahmed, H. Banaee, A. Lout, Health monitoring for elderly: an application
using case-based reasoning and cluster analysis, ISRN Artif. Intell. 2013 (2013)
11, http://dx.doi.org/10.1155/2013/380239.
H.J. Hermens, B. Freriks, C. Disselhorst-Klug, G. Rau, Development of recommendations for SEMG sensors and sensor placement procedures, J.
Electromyogr. Kinesiol. 10 (2000) 361374.
J. Perry, Gait Analysis: Normal and Pathological Function, SLACK Incorporated,
Thorofare, New Jersey, 1992.
R.N.A.-J. Khushaba, Adel, Fuzzy wavelet packet based feature extraction method
for multifunction myoelectric control, Int. J. Biomed. Sci. 2 (2007) 186194.
X. Zhang, X. Chen, Y. Li, V. Lantz, K. Wang, J. Yang, A framework for hand gesture
recognition based on accelerometer and EMG sensors, IEEE Trans. Syst. Man
Cybern. Part A: Syst. Hum. 41 (2011) 10641076.
A. Subasi, Classication of EMG signals using combined features and soft computing techniques, Appl. Soft Comput. 12 (2012) 21882198.
A. Phinyomark, A. Nuidod, P. Phukpattaranont, C. Limsakul, Feature extraction
and reduction of wavelet transform coefcients for EMG pattern classication,
Elektron. Elektrotech. 122 (2012) 2732.
A. Phinyomark, C. Limsakul, P. Phukpattaranont, Application of wavelet analysis
in EMG feature extraction for pattern classication, Measure. Sci. Rev. 11 (2011)
4552.
I. Gler, E.D.E.D.. beyli, Adaptive neuro-fuzzy inference system for classication of EEG signals using wavelet coefcients, J. Neurosci. Methods 148 (2005)
113121.
S.F.M.J. Sahin Ugur, Pattern recognition with surface EMG signal based wavelet
transformation, in: Proceedings of the IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2012.

Author's personal copy

S.M.N. Arosha Senanayake et al. / Applied Soft Computing 20 (2014) 127141
[48] L.Z.L. Zhang, H.P.H. Peng, C.Y.C. Yu, An approach for ECG classication based on
wavelet feature extraction and decision tree, in: International Conference on
Wireless Communications and Signal Processing (WCSP), Suzhou, China, 2123
October, 2010, pp. 14, http://dx.doi.org/10.1109/WCSP.2010.5633782.
[49] I.T. Jolliffe, Principal Component Analysis, 2nd ed., Springer-Verlag New York,
Inc., New York, 2002.
[50] M.J. OMalley, M.F. Abel, D.L. Damiano, C.L. Vaughan, Fuzzy clustering of children with cerebral palsy based on temporal-distance gait parameters, IEEE
Trans. Rehabil. Eng. 5 (1997) 300309.
[51] A. Manurung, J. Yoon, Fuzzy clustering of temporal parameters of gait during
stance phase for walking speed estimation, in: Proceedings of the International Conference on Control Automation and Systems (ICCAS), Gyeonggi-do,
2010.
[52] J. Dunn, A fuzzy relative of the isodata process and its use in detecting compact
well-separated clusters, J. Cybern. 3 (1973) 3257.
[53] J.C. Bezdek, Pattern Recognition with Fuzzy Objective Function Algorithms,
Kluwer Academic Publishers, Norwell, MA, USA, 1981.

141

[54] S. Ray, R.H. Turi, Determination of number of clusters in k-means clustering and
application in color image segmentation, in: Proceedings of the 4th International Conference on Advances in Pattern Recognition and Digital Techniques,
1999.
[55] M. Khezri, M. Jahed, A neuro-fuzzy inference system for SEMG-based identication of hand motion commands, IEEE Trans. Ind. Electron. 58 (2011) 19521960.
[56] S. Kocer, Classication of EMG signals using neuro-fuzzy system and diagnosis
of neuromuscular diseases, J. Med. Syst. 34 (2010) 321329.
[57] J.-S.R. Jang, ANFIS: adaptive-network-based fuzzy inference system, IEEE Trans.
Syst. Man Cybern. 23 (1993) 665685.
[58] J. Hhn, E. Hllermeier, Furia: an algorithm for unordered fuzzy rule induction
Data Mining and Knowledge Discovery, vol. 19, 2009, pp. 293319, LA English.
[59] J. Shawe-Taylor, N. Cristianini, An Introduction to Support Vector Machines
and Other Kernel-based Learning Methods, 1st ed., Cambridge University Press,
Cambridge, 2000.
[60] S. Shiu, S. Pal, Foundations of Soft Case-based Reasoning (Wiley Series on Intelligent Systems), Wiley-Interscience, NY, 2004.