TELKOM
NIKA
, Vol.15, No
.2, June 20
17
, pp. 903~9
1
1
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v15i2.6143
903
Re
cei
v
ed
De
cem
ber 5, 20
16; Re
vised
Ap
ril 28, 201
7; Acce
pted
May 13, 20
17
Autism Spectrum Disorders Gait Identification Using
Ground Reaction Forces
Che Za
w
i
y
a
h Che Hasan*,
Rozita Jaila
ni, Noorita
w
ati Md Tahir,
Rohilah Sah
a
k
F
a
cult
y
of Elec
trical Eng
i
ne
eri
ng, Univ
ersiti T
e
kno
l
og
i MAR
A
404
50 Sh
ah Al
am, Selan
gor, Mala
ysi
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: za
w
i
ya
h.has
an@
gmai
l.com
,
rozita@ie
ee.o
r
g, noorita
w
a
t
i
@
iee
e
.org,
i
l
a
h1
0@y
a
ho
o
.
co
m
A
b
st
r
a
ct
Autis
m
spectr
um d
i
sor
ders (
ASD) are a p
e
rman
ent ne
ur
odev
elo
p
m
e
n
tal dis
o
rder th
a
t
can be
ide
n
tified
d
u
rin
g
the
first few
years
of life
an
d ar
e
curr
ently
associ
ated
w
i
th the
ab
nor
mal
w
a
lkin
g p
a
ttern
.
Earlier
id
entific
ation
of this
p
e
rvas
iv
e dis
o
r
der co
ul
d pr
ovi
de ass
i
stanc
e
in d
i
a
gnos
is a
nd est
abl
ish r
a
pid
qua
ntitative
cli
n
ical
j
u
d
g
m
ent.
T
h
is
pa
per pre
s
ents
a
n
auto
m
ate
d
ap
proac
h
w
h
ich
ca
n
be
ap
pli
e
d
to
ide
n
t
ify
ASD g
a
it
patterns
usin
g thr
ee-d
i
mens
ion
a
l
(3D)
gr
o
u
n
d
reacti
on f
o
rc
es (GRF
). T
h
e study
i
n
volv
ed
classificati
on
o
f
gait p
a
tterns
of chi
l
dre
n
w
i
th ASD
and
typical
h
ealthy
chil
dren. T
h
e
GRF
data w
e
r
e
obtai
ne
d usin
g
tw
o
force pla
t
es durin
g self
-deter
mi
ned b
a
refoot w
a
lkin
g. T
i
me
-ser
ies
para
m
eter
i
z
a
t
ion
techni
qu
es w
e
re app
lie
d to th
e GRF
w
a
veforms to extr
act the i
m
p
o
rtant g
a
it feat
ures. T
h
e most do
min
a
nt
and c
o
rrect fe
atures for ch
ar
acteri
z
i
n
g
ASD
gait w
e
re
se
le
cted usi
ng stat
istical b
e
tw
een
-grou
p
tests a
n
d
stepw
ise discr
i
m
i
n
a
n
t ana
lysi
s (SW
D
A).
T
he selecte
d
featu
r
es w
e
re gro
u
p
ed i
n
to tw
o gro
ups w
h
ich s
e
rv
ed
as tw
o input
da
tasets to the k-
near
est nei
gh
b
o
r (KNN) cl
assi
fier. T
h
is study
de
mo
ns
trates
that the 3D
GR
F
gait fe
atures
s
e
lecte
d
usin
g
SW
DA are
rel
i
abl
e to
be
us
e
d
i
n
th
e i
d
e
n
tificatio
n
of
ASD
ga
it usi
n
g
KN
N
classifier w
i
th 8
3
.33% p
e
rfor
mance acc
u
racy.
Ke
y
w
ords
: au
tism sp
ectru
m
disor
ders, g
a
i
t classifi
c
a
tion
, k-nearest n
e
i
gh
bor, stepw
i
s
e discri
m
in
an
t
ana
lysis, grou
n
d
reactio
n
force
Copy
right
©
2017 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
A
u
t
i
sm sp
ect
r
um di
so
rde
r
s (A
S
D
) ar
e
cha
r
a
c
teri
ze
d by a con
s
tant deficit in so
cial
comm
uni
cati
on, social
int
e
ra
ction, a
n
d
the p
r
e
s
en
ce of restri
cte
d
and
repetiti
v
e behavio
rs. This
perva
sive an
d perma
nent
neuro
develo
p
mental di
so
rde
r
can b
e
recogni
ze
d d
u
ring the ea
rly
stage of the developm
ent
al perio
d of a child. One
o
f
the possibl
e
signs that could be u
s
ed
to
identify ASD
is the existen
c
e of motor
deficit
s, which includ
es a
b
norm
a
l gait, clum
sine
ss, and
irre
gula
r
m
o
tor
sign
s [1].
An abn
orm
a
l
gait i
s
d
e
fin
ed a
s
an i
r
re
gular style
of wal
k
in
g a
n
d
this
unu
sual con
d
ition coul
d cau
s
e dete
r
i
o
ration in
o
c
cupatio
nal and othe
r d
a
ily activities of
individual
s wi
th ASD. Previous
studie
s
h
a
ve re
p
o
rted
a wide
ran
g
e
of abnormal
gait pattern
s i
n
temporal and
spatial mea
s
ureme
n
ts, kinematic joi
n
t angles, ki
ne
tic joint moments an
d join
t
powers du
rin
g
wal
k
ing in i
ndividual
s wit
h
ASD [2, 3].
The id
entifica
t
ion of g
a
it a
bnormalitie
s
coul
d b
e
be
n
e
ficial fo
r the
early
dete
c
tion a
n
d
better treatm
ent planni
ng f
o
r children
with ASD
[4]. Current gait a
s
se
ssm
ent me
thods a
r
e oft
en
time-con
sumi
ng an
d hig
h
l
y
depen
dent
on the
clini
c
ian j
udgm
e
n
t, which lea
d
s to
su
bje
c
tive
interp
retation
s. With the
curre
n
t adva
n
ce
s in
g
a
it analysi
s
an
d instrument
ation, it not only
provide
ne
w i
n
sig
h
ts in
un
derstandi
ng
all aspe
cts
of
movement p
a
tterns, but
a
l
so sup
p
o
r
t
the
evolution of a
u
tomated dia
gno
sis of pat
hologi
cal di
so
rde
r
s.
Grou
nd
rea
c
tion force (GRF
) is
one
of the kin
e
t
ic mea
s
u
r
e
m
ents th
at has
bee
n
effectively used for the a
s
se
ssm
ent of
norm
a
l
an
d patholo
g
ical
movement
s and
al
so
for the
comp
ari
s
o
n
s
betwe
en pati
ents an
d no
rmal gro
u
p
s
[5]. In routine
gait analysi
s
,
force
plate
s
are
use
d
to mea
s
ure the G
R
F in three
di
mensi
o
n
s
,
na
mely medial
-l
ateral, ante
r
i
o
r-po
sterio
r, and
vertical
dire
ct
ions. T
he th
ree comp
onen
ts of G
R
F p
r
ovide a
com
p
lete interpret
a
tion on
ho
w
the
body weig
ht drop
s and
m
o
ves acro
ss the
supp
or
tin
g
foot d
u
rin
g
wal
k
in
g [6]. The
r
efo
r
e,
by
investigatin
g the
whole G
R
F com
pon
e
n
ts
i
s
exp
e
ct
ed to
be m
o
re effective to
identify spe
c
ific
locom
o
tion
chara
c
te
risti
c
s that ca
n b
e
used fo
r au
tomated id
en
tification. To
the be
st of o
u
r
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 15, No. 2, June 20
17 : 791 – 79
x
904
kno
w
le
dge, the only study
that investigates G
R
F co
mpone
nts in child
ren
with ASD is done
by
Ambro
s
ini
et
al [7]. The
stu
d
y re
porte
d
a
de
crea
se i
n
t
he
se
con
d
p
e
a
k
of
v
e
rt
i
c
al
f
o
rce
in
mo
st
of
their su
bje
c
ts.
In orde
r to re
cog
n
ize gait abno
rmalitie
s, machin
e lea
r
ning m
odel
s are u
s
ed to
cla
ssify
and di
scover
unde
rlying p
a
tterns of the
kinem
atic
a
n
d
kin
e
tic me
a
s
ureme
n
ts. T
he ap
plicatio
n of
machi
ne lea
r
ning cl
assifie
r
s for a
u
tom
a
ted re
cog
n
ition of gait pattern deviati
ons a
nd oth
e
r
variou
s biom
edical fields
has g
r
o
w
n e
norm
o
u
s
ly in
the last decade
s. Artificia
l
neural n
e
tworks
(ANN) and s
u
pport vec
t
or
mac
h
ines (SVM) have been
employed for rec
o
gnition and
cla
ssifi
cation
of Parkin
son’
s dise
ase [8], young-ol
d
ga
it patterns [9], cere
bral p
a
lsy childre
n [10],
and p
a
tients
with ne
urol
og
ical di
so
rde
r
s [11]. ANN
was al
so
su
ccessfully used
for cl
assifica
tion
of gende
r in
children [12]
and po
st-st
r
oke p
a
tient
s
[13]. Apart from that, k-n
eare
s
t neig
h
bor
(KNN) was
also
u
s
ed
a
s
a
p
a
ttern
cla
ssifie
r
fo
r gait
pattern
identificatio
n [14]
and
brain
balan
cing
cla
ssifi
cation
[15
]. KNN i
s
a
su
pervised
ma
chine
l
earning
cla
ssifie
r
whi
c
h
i
s
simpl
e
b
u
t
robu
st to
be
used in
st
atistical
e
s
timation a
nd
pattern
re
co
gnition. Thi
s
non
-pa
r
am
e
t
ric
cla
ssifi
cation
method p
r
edi
cted a
cla
ss l
abel to ea
ch
membe
r
of the test sa
mple
base
d
on vot
i
ng
rights of its
k-nearest nei
gh
bors dete
r
min
ed by a dista
n
ce met
r
ic [1
6].
Re
cently, stat
istical
featu
r
e
sel
e
ctio
n te
chni
qu
es such
as in
depe
nd
ent t-te
st [17]
, Mann
-
Whitney U [1
8], and step
wise meth
od of
discrimin
ant
analysi
s
(S
WDA) [13] we
re use
d
to sel
e
ct
signifi
cant fe
ature
s
in g
a
it resea
r
ch. Th
e i
nde
pend
en
t t-test and M
ann-Whitn
e
y U test (TM
W
U)
are the types of
between-group
tests that have t
he ability to
se
lect signifi
cant
features by
examining th
e mean
score of gait feature
s
a
c
ro
ss
two sepa
rate
grou
ps. Me
a
n
whil
e, SWDA is
freque
ntly co
ndu
cted to d
e
termin
e the
optimum se
t of input features fo
r gro
up memb
ership
predi
ction
an
d elimin
ate t
he le
ast
sig
n
ificant
and
unrel
ated fe
a
t
ures fro
m
t
he d
a
taset [19].
Previou
s
stu
d
i
es in
gait a
n
a
lysis
have v
a
lidated th
at
SWDA i
s
a
b
l
e
to ide
n
tify specifi
c
individ
ual
feature
s
that best dete
r
min
ed
gro
up pla
c
ement [13, 20
, 21].
The sca
r
city of research a
nd insufficien
t in
formation
rega
rdi
ng the
3D GRF in child
re
n
with ASD a
r
e demo
n
st
rat
ed glo
bally. Until no
w, th
ere i
s
n
o
pu
blish
ed literature d
ealin
g
with
automated
re
cog
n
ition of
ASD gait pattern
s ba
se
d
o
n
3D G
R
F d
a
t
a. Thus, this study pro
p
o
s
es
an automate
d
identificatio
n of ASD children u
s
in
g machi
ne lea
r
ni
ng cla
s
sifier
based on the
3D
GRF inp
u
t feature
s
. The
s
e features
were firs
t ex
tracted u
s
in
g
time-se
r
ie
s para
m
eteri
z
at
ion
method
s an
d
then we
re
sele
cted u
s
in
g two stati
s
tical featu
r
e selectio
n tech
nique
s. KNN is
employed to
model b
o
th input f
eatures and their
cl
assificatio
n
p
e
rform
a
n
c
e
s
with ea
ch in
put
dataset we
re
comp
ared. T
he re
st
of thi
s
pa
per
ha
s
been
org
ani
zed a
s
follow.
The next sect
ion
explain
s
the
prop
osed
met
hod fo
r th
e
st
udy. Sect
io
n
3 p
r
e
s
ent
s th
e expe
rime
ntal results
and
the
discu
ssi
on. Finally, Section
4 con
c
lud
e
s
the study.
2. Rese
arch
Metho
d
The ASD g
a
it identification is prim
a
r
ily generate
d
based on
the automatic gait
cla
ssif
i
cat
i
on
sy
st
em u
s
in
g
st
at
ist
i
c
a
l an
aly
s
is
a
nd m
a
chi
ne lea
r
ni
ng app
ro
ach
e
s. The p
r
o
p
o
se
d
sy
st
em
con
s
ist
s
of
f
i
v
e
seq
uen
ce p
r
oce
s
se
s of data a
c
qui
sition, pre
p
ro
cessing, featu
r
e
extraction, fe
ature sele
ctio
n, and gait pa
ttern
cla
s
sification as illu
strated in Figu
re
1.
Figure 1. The
overall proce
ss
of ASD gai
t identification
.
Feature
selection
Gait pat
t
ern
classification
Fea
t
ure
ex
tra
c
tio
n
I
nput dataset
Data
acquisition
3D GRF
(Fx
,
F
y
,
Fz)
Between-
g
r
oup test
K-
near
est neighbo
r
(KNN
)
Per
f
orm
a
nce
evaluation
Autis
m
s
p
ectru
m
d
i
so
rd
e
r
Preprocessing
Stepwise discr
i
m
i
nant analy
s
is
T
yp
icall
y
develo
p
in
g
Evaluation Warning : The document was created with Spire.PDF for Python.
TEL
K
from
perfo
and t
in 3
D
tec
h
n
A
SD
The
s
cla
s
s
accu
r
st
at
i
s
2.1.
D
labor
appr
o
The
p
st
ud
y
(NA
S
throu
fac
u
l
t
inde
p
injuri
e
emb
e
dire
c
t
even
t
perp
e
refer
e
parti
c
to pe
A
n a
v
Fi
g
2.2.
D
from
K
OM
NIKA
Autis
m
Sp
e
T
h
e
pr
oc
ASD and t
y
rm self-dete
hen were p
a
D
spa
c
e,
na
n
i
que
s. Sub
s
gait we
re s
e
s
el
ecte
d
fe
a
ifier. The
K
r
acy fro
m
1
s
tical feature
D
ata Ac
qui
s
The 3D
G
atory at
the
o
val for t
h
is
p
a
r
ent or gu
a
Thirty ch
i
y
. The child
r
S
OM
) ce
nter
g
h
so
cia
l
m
e
t
y mem
ber
s
p
ende
ntly wi
e
s o
r
mu
sc
u
l
Two for
c
e
d
ded i
n
the
t
ion
s
at
100
0
t
s f
o
r ea
ch l
e
ndi
cul
a
rly t
o
e
nc
e
pur
p
o
s
c
i
pant
s.
A
ll partici
rfo
r
m a
strai
v
erag
e o
f
te
n
g
ure 2. A ma
D
ata Pre
pro
A
suc
c
e
s
every p
a
rti
c
e
ct
rum
Diso
r
d
e
s
s of
identi
y
pically de
v
rmi
ned walk
i
a
ssed to the
me
l
y
F
x
,
F
y
s
e
que
ntly, t
h
e
le
cted by
e
a
tures
were
u
K
NN
w
a
s
th
e
0-fold cros
s
selection te
c
s
ition
G
RF
s data
a
Univ
e
r
si
ti
T
st
udy
w
a
s
a
a
rdi
an of e
a
c
i
ld
r
e
n
w
i
th
A
r
en
w
i
th
A
S
in Klan
g, S
e
e
dia networ
k
s
an
d n
e
ig
h
thout an
y
a
l
oskeletal di
s
c
e p
l
ates (
A
middle
o
f
a
0
Hz
. The fo
r
im
b of th
e
p
o
one anot
h
e
s
es
.
A
di
git
a
p
ant
s, dre
s
s
gh
t b
a
r
e
fo
o
t
n
walking tri
a
le parti
cipa
n
ces
sing an
d
s
sful valid
w
a
c
ip
ant. Th
e
3
I
S
d
ers
Gait Id
e
fying ASD
g
elopin
g
(TD
i
ng. Next, th
e
f
eature ex
tr
a
y
, and
Fz
w
h
e most
si
g
n
e
mploying t
w
u
se
d as th
e
e
n trai
ned
t
s
validatio
n
c
hni
que
s a
n
d
a
cqui
sition
w
T
eknol
ogi M
A
a
pp
roved b
y
c
h child sign
e
A
SD a
nd
thi
r
S
D w
e
re
e
n
r
e
lango
r, M
a
l
a
k
. The
typi
c
a
h
b
o
rh
ood
s
n
a
ssi
st
iv
e d
e
v
s
or
d
e
rs
.
A
dva
n
c
ed
M
6.5-metre
w
r
ce plate
s
w
e
p
ar
tic
i
pa
n
t
s
d
e
r
to record
t
a
l wei
ght
s
c
s
ed in
tight s
h
t
wal
k
in
g al
o
n
a
ls wa
s
re
co
r
t walkin
g on
d
Fea
t
ure
s
E
a
lki
ng trial
wi
3
D GRF dat
a
S
SN: 1693-6
9
e
ntification
U
ait pattern s
D
) children
u
e
3D
GRF
s
d
a
ction stage
.
w
ere extract
e
n
ificant and
w
o type
s of
s
input featu
r
t
o cla
s
sif
y
t
wa
s co
mp
u
d
the KNN
c
l
w
a
s
co
ndu
c
t
A
RA (
U
iTM
)
the Resea
r
c
e
d an inform
r
ty TD
childr
e
r
olle
d from
t
a
ysia a
nd l
o
c
a
l h
ealthy c
h
nearby. All
v
ice
s
an
d h
a
M
echani
cal
T
w
al
kway we
r
e
re al
so utili
z
d
uri
ng wal
k
i
n
t
he wal
k
ing
t
c
ale wa
s
u
s
hirt
s (fe
m
al
e
ng the
wal
k
w
r
ded fo
r ea
c
h
the wal
k
wa
y
E
xtrac
t
ion
th a c
l
ear fo
o
a from the
c
9
30
U
si
ng Grou
n
d
tarts
wit
h
th
e
u
si
ng t
w
o f
o
d
ata in thre
e
In this
s
t
ag
e
e
d u
s
in
g th
e
domina
n
t g
a
s
tatis
t
ical m
e
r
es for cl
as
s
h
e ASD
ga
u
ted to eval
l
assifier.
t
ed in the H
)
Shah Ala
m
c
h Ethics C
o
ed
c
o
ns
en
t
f
e
n
age
d
4 t
o
t
he Nationa
c
al com
m
un
i
h
ildren were
grou
p pa
r
t
a
d no m
e
di
c
T
ec
hn
o
l
og
y
r
e used to
m
z
ed to
detec
t
n
g trials
. Tw
t
rials from th
ed
to obtai
n
e
) and tight p
w
ay at thei
r
s
h
particip
ant
.
y
d
u
ring a g
a
o
t c
o
ntact o
n
c
ho
se
n trial
d
… (Che Za
w
e
acqui
sition
rce pl
ates
w
dire
ction
s
w
e
, gait featu
r
e
time-s
erie
s
a
it feature
s
e
thod as
de
p
s
if
ying gait
p
i
t. The
ave
r
u
ate the p
e
u
man Motio
n
m
, Selango
r,
o
mmittee of
f
orm prior to
o
12 yea
r
s
p
l
Autis
m
S
o
ty by appro
a
recruite
d fr
o
t
icipants w
e
c
al histo
r
y
o
Inc
., MA,
U
m
e
a
su
re the
3
t
initial foot c
o
video
ca
m
e frontal an
d
n
anth
r
op
o
m
ants p
r
ovi
d
e
s
elf
-
sel
e
ct
e
d
a
it motion c
a
p
n
each force
were filtere
d
w
i
y
a
h
Che
H
n
of
3D G
R
F
s
w
hile the
c
h
w
e
r
e prep
ro
c
r
es con
s
ist o
s
para
m
eteri
for charac
t
e
p
i
c
ted in Fi
g
p
attern us
in
g
r
ag
e cl
assifi
e
rforman
c
e
o
n & Gait A
n
Malaysia.
E
UiTM Shah
participatio
n
p
ar
tic
i
pa
te
d
o
cie
t
y
of
M
a
a
chin
g the p
a
o
m the fami
e
re able to
o
f lower ext
U
SA) which
3D GRF
s
in
c
on
tact and f
o
m
era
s
were
p
d
sagittal vie
w
m
etric data
o
e
d, we
re in
st
r
d
speed (Fig
u
pturin
g se
ss
plate wa
s c
d
usin
g a s
e
H
as
an
)
905
s
da
ta
h
i
l
dren
e
s
sed
f GRF
zat
i
on
e
r
i
z
i
ng
ure 1.
g
KNN
cation
o
f
the
n
alysi
s
E
th
i
c
a
l
Alam.
n
.
in this
a
laysi
a
a
rents
lies of
walk
rem
i
ty
we
re
three
o
ot off
p
la
c
ed
w
s f
o
r
o
f the
r
ucted
u
re
2).
i
on.
hosen
e
co
nd-
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 15, No. 2, June 20
17 : 791 – 79
x
906
orde
r lo
w-pa
ss Butterwort
h
filter
with
cutoff frequenc
y of 30 Hz
t
o
reduc
e
nois
e
.
Next, the 3D
GRF data were extrac
ted into
the ASCII text format
for data a
nal
ysis. The
s
e
p
r
ocesse
s
were
comp
uted u
s
i
ng the Vicon
Nexu
s softwa
r
e
version 1.8
.
5 (Vicon, Oxford, UK
).
In this
study, the 3D G
R
F
data from
si
n
g
le left limb
stance f
r
om th
e sel
e
cte
d
va
lid trial
wa
s analyzed
to repre
s
ent the gait attributes of
each partici
pant [2
2, 23]. The time comp
one
nts
of the 3D GRF were norm
a
lize
d
to the percent
a
ge o
f
stance p
hase time, where
a
s the 3D G
R
F
amplitude
s
were n
o
rm
ali
z
ed
to the
percenta
ge
of the p
a
rtici
pant’s bo
dy weig
ht [24,
25].
Normali
z
ation
steps we
re essentially
p
e
rform
ed
to
eliminate vari
ations a
m
on
g
the partici
pa
nts
with differe
nt height, body
mass, and d
u
r
ation of
sta
n
c
e ph
ase [26,
27]. After normali
zation, the
initial foot con
t
act co
rre
sp
o
nds to 0% an
d the
foot off
event co
rre
sp
onde
d to 100
% of the stance
pha
se.
In routin
e gai
t analysi
s
, th
e GRF du
rin
g
no
rmal
wal
k
ing i
s
g
ene
rally mea
s
ure
d
in three
dire
ction
s
(Fx
:
medial-late
r
al, Fy: anterior-p
oste
rio
r
, and Fz: vertica
l
). The 3D G
R
F patterns f
o
r a
TD pa
rticip
an
t are sh
own in Figure 3(a),
(b), and
(c). These graph
s also
sh
ow 17
ch
ara
c
te
ri
stic
points th
at were extracte
d
from the
cu
rves. Fy
2 was exclud
ed d
u
e
to ze
ro fo
rce valu
e du
ri
ng
mid-stan
ce. T
i
me-seri
e
s
pa
ramete
rizatio
n
tech
nique
s
were ap
plied
to each wave
form to extra
c
t
the in
stantan
eou
s valu
es
of amplitu
de
and it
s
relati
v
e
time [2
4, 28
]. This te
ch
ni
que i
s
co
nsi
d
ered
one of the most commo
n method
s o
f
gait data analysi
s
, whi
c
h is prefe
r
a
b
l
e, and clini
c
ally
accepta
b
le [2
9, 30].
The followi
ng
twenty GRF
gait features were
extract
ed: the local
pea
ks a
nd m
i
nimum
values of
the three GRF co
mpone
nts (F
x1,
Fx2,
Fx
3; Fy1, Fy3; and Fz1, Fz2, F
z
3); th
e rel
a
tive
time of occurren
ce
s (Tx1,
Tx2, Tx3; Ty1, Ty2,
Ty3;
and Tz1, Tz2
,
Tz3); loadi
n
g
rate, pu
sh-off
rate, an
d pe
a
k
ratio (T
able
1) [5, 24,
28,
31]. Loadi
ng
rate i
s
defin
e
d
as th
e am
pl
itude of the first
vertical
pea
k force
divide
d by it
s tim
e
o
c
curren
ce
. The
pu
sh
-off rate i
s
computed
a
s
the
amplitude of the second p
eak of
vertical force divid
ed by the time from the se
con
d
pea
k of
vertical fo
rce
until the e
n
d
of the
sta
n
ce
pha
se
[32]. The p
e
a
k
ratio i
s
calcul
ated a
s
the
amplitude
of the first p
e
a
k
of vertical force divi
de
d by
the amplitud
e of the se
co
nd vertical force
pea
k [28].
Table 1. The
extracted 3
D
GRF g
a
it
feature
s
and its a
bbreviatio
n
s.
Direction Gait
featu
r
e
Abbreviation
Medial-lateral Maximu
m supina
tion force
Fx1
Foot flat force
Fx2
Maximum pr
onation force
Fx3
Relative time to maximum supina
tion
Tx1
Relative time to foot flat
Tx2
Relative time to maximum pr
onation
Tx3
Anteri
or-p
os
teri
o
r
Maxi
mum br
aking force
F
y
1
Maximum pr
opulsion force
F
y
3
Relative time to maximum braking force
T
y
1
Relative time to zero force du
ring
midstance
T
y
2
Relative time to maximum pr
opulsion force
T
y
3
Vertical First
vert
ical peak force
Fz1
Vertical minimum
force
Fz2
Second vertical peak force
Fz3
Relative time to firs
t vertical peak force
Tz1
Relative time to vertical minimum force
Tz2
Relative time to sec
ond vertical peak force
Tz3
Loading rate
Loading rate
Push-off rate
Push-off rate
Peak ratio
Peak ratio
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Autism
Spectrum
Diso
rde
r
s Gait Identificat
ion
Usi
ng
Grou
nd… (Che Zawi
ya
h Che Ha
sa
n)
907
Figure 3. The
three groun
d
reactio
n
force co
mp
one
nts of a singl
e left limb stance durin
g the
stan
ce ph
ase
of a typically
developin
g
fe
male
pa
rticip
ant. (a) Me
dial-late
r
al dire
ction; (b
)
anterio
r-po
sterio
r dire
ction
;
and (c) verti
c
al directio
n.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 15, No. 2, June 20
17 : 791 – 79
x
908
2.3. Featur
es
Selection
Gene
rally, some extracte
d
featu
r
e
s
ma
y contain
re
dun
dant
and l
e
a
s
t si
gnifica
nce
informatio
n which
ca
n lea
d
to poo
r p
e
r
forma
n
ce in
the cla
s
sification sta
ge.
Hen
c
e, a fe
a
t
ure
sele
ction m
e
thod
wa
s em
ployed in o
r
d
e
r to sele
ct
the mo
st sig
n
i
ficant features a
s
well a
s
to
enha
nce the
cla
s
sifier p
e
rform
a
n
c
e.
In this
study, betwee
n
-g
rou
p
te
sts a
nd
step
wise
discrimi
nant
analysi
s
(S
WDA) were u
s
e
d
to
sele
ct the most si
gnificant gait feat
ure
s
.
Initially before cond
ucting
betwe
en-grou
p test, the
extracte
d g
a
it fe
ature
s
we
re e
x
plored
for normality usin
g the Sh
apiro
-Wilk (S
W) te
st sin
c
e
the sampl
e
size in ea
ch g
r
oup is le
ss than
50 [33]. Feat
ure
s
were
no
rmally dist
rib
u
ted if t
he ou
tcome of SW test (p-val
ue
) is g
r
eate
r
th
an
0.05. For n
o
rmally distrib
u
t
ed feature
s
,
the
mean score
s
were examined usi
n
g
indepe
nde
nt
t-
tests (T
), wh
erea
s Man
n
-Whitney U tests (M
WU) for non-n
o
rmal feature
s
.
The significant
differen
c
e be
tween the two grou
ps fo
r both tests
wa
s define
d
as
p < 0.05. Fe
ature
s
that were
s
t
atis
tically s
i
gnific
a
nt were c
h
os
en to be
as inp
u
t features in
cla
ssi
fication sta
g
e
.
Another stati
s
tical method
,
SWDA
was use
d
to ide
n
tify dominant f
eature
s
th
at made a
signifi
cant
co
ntribution
for
grou
p
sep
a
ra
tion a
c
ro
ss th
e two
group
s.
SWDA
was
perfo
rmed
u
s
i
n
g
the Wilks’ la
mbda meth
o
d
with the def
ault setting crit
eria of the F
value to enter is at lea
s
t 0.05
and
F valu
e t
o
remove
is l
e
ss tha
n
0.1
0
. Featu
r
e
s
within th
e
ran
ge of
F val
u
e
s
a
r
e
stati
s
tically
signifi
can
c
e
o
f
gro
ups di
scrimination [2
9, 34]. Bo
th sta
t
istical analy
s
es we
re
pe
rfo
r
med
u
s
ing
th
e
IBM SPSS
St
atis
tics
vers
ion 21.0 (IBM, New York
, USA).
2.4. Classific
a
tion Model
Cla
ssifi
cation
is
a p
r
o
c
e
s
s of
assig
n
in
g ea
ch
elem
ent in
a set
of data
into
target
c
a
te
go
r
i
es
o
r
c
l
ass
e
s
.
T
h
e u
l
tima
te
go
al o
f
th
is p
r
o
c
e
s
s
is
to
p
r
ed
ic
t th
e ta
r
g
et c
l
as
s fo
r e
a
ch
ca
se in the
dataset accu
rately
. The classificatio
n
stage
wa
s p
e
rform
ed u
s
i
ng Statistics
and
Machi
ne Le
arning Tool
box in Matlab version
R20
15a
(The M
a
thWo
rks Inc., USA
)
. The sel
e
cte
d
feature
s
, na
mely 3DG
R
F
-
TM
WU a
nd
3DG
R
F
-
SWDA
were fed into the KNN
c
l
as
s
i
fier. In this
study, the
cl
assificatio
n
t
a
sks were
e
x
plored
u
s
in
g four type
s
of dista
n
ce
met
r
ic
s:
cit
y
bloc
k,
correl
ation, cosin
e
, and Eu
clide
an, while
the k value was varie
d
fro
m
1 to 12.
In ord
e
r to
find the
be
st
model th
at chara
c
te
rizes
the input
dat
aset, it is im
portant to
impleme
n
t cross-valid
ation
method fo
r
model
eval
ua
tion. This
met
hod u
s
e
s
an i
ndep
ende
nt test
set which ha
s not b
een
u
s
ed
duri
ng th
e trainin
g
pro
c
e
ss to
evalu
a
te the mod
e
l perfo
rman
ce
[35]. Due
to
small
sampl
e
sizes u
s
ed
in
this stu
d
y, 1
0
-fold
cro
s
s v
a
lidation
met
hod i
s
cho
s
e
n
to
estimate the
gene
rali
zatio
n
ability of KNN
cla
ssifi
e
r
[25, 35]. In 10-fold
cross validation, the
dataset is
ran
domly divide
d into 1
0
eq
u
a
l or
nea
rly e
qual-si
z
ed
su
bset
s o
r
fold
s. Nine fol
d
s
a
r
e
use
d
for trai
n
i
ng a
nd th
e remainin
g o
n
e
fold i
s
u
s
e
d
for te
sting. 1
0
-cro
ss valid
ation i
s
repe
a
t
ed
for ten ite
r
ati
ons
so
that for e
a
ch num
ber
of iterat
io
ns, a
differen
t
fold is h
e
ld
out for eval
u
a
tion
and the oth
e
r nine folds
are use
d
for tra
i
ning. The
n
, the cla
s
sificati
on accu
ra
cy is cal
c
ul
ated
by
averagi
ng the
accura
cy for
the ten folds [35].
The mod
e
l perfo
rman
ce
with two types of
input
dataset an
d variation
s
of model
para
m
eters
was m
e
a
s
ured
usin
g conf
usion matrix
with two
c
l
as
ses,
TD and ASD. In this
s
t
udy,
true po
sitive (TP) is the nu
mber of ASD cases
co
rre
ctly classifie
d
and true
neg
ative (TN) is the
numbe
r of T
D
cases
co
rrectly cla
s
sifie
d
. The
effe
ctiveness of th
e TMWU
an
d SWDA feat
ure
sele
ction
were me
asure
d
by cal
c
ul
atin
g the
cla
s
sif
i
cat
i
on
a
c
c
u
ra
cy
w
h
ic
h
wa
s d
e
f
i
ned
a
s
t
h
e
corre
c
t cla
s
sification
s (TP a
nd TN) ra
te
made by the model ove
r
a dataset.
3. Results a
nd Discu
ssi
on
After complet
i
ng the feature extraction
usin
g parame
t
erizatio
n techniqu
es, twe
n
ty GRF
gait feature
s
were extra
c
ted as g
a
it pattern to
rep
r
ese
n
t the gai
t profiles of each parti
cip
ant.
Table 2 ta
bul
ates the m
e
a
n
s, sta
nda
rd
deviation
s (S
D), an
d the
p-value
distri
bution of ea
ch
extracted g
a
i
t
feature. Fro
m
the twenty gait f
eatures, only six significant featu
r
es h
a
ve bee
n
cho
s
e
n
u
s
in
g the
inde
p
ende
nt t-te
st and
Ma
nn-Whitney
U test
(TM
W
U) for AS
D
g
a
it
cla
ssifi
cation.
The
s
e
si
gnifi
cant
gait fe
ature
s
whi
c
h
h
a
ve a
p-val
u
e
less than
0.0
5
a
r
e
made
b
o
ld
in Tabl
e 2.
The d
o
mina
n
t
feature
s
a
r
e Fy3, Ty
2,
Fz3, T
z
3, p
u
s
h-off rate, a
nd pe
ak rati
o.
Pertainin
g
to
the mea
n
val
ues of the
do
minant f
eatu
r
es,
child
ren
with ASD were foun
d to
exhibit
signifi
cantly lowe
r Fy3, Ty2,
Fz3, Tz3,
and pu
sh
-off rate, but the peak ratio was si
gnifica
ntly
greate
r
in ASD as
comp
ared to TD.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Autism
Spectrum
Diso
rde
r
s Gait Identificat
ion
Usi
ng
Grou
nd… (Che Zawi
ya
h Che Ha
sa
n)
909
Based
on th
e SWDA a
p
p
roa
c
h, it is found
that only three d
o
minant g
a
it feature
s
,
namely F
z
3,
Ty2, and
Tx1
have the
ab
ility to di
scri
minate the
ASD gait f
r
om
the no
rmal
gait
pattern
s. All t
hese th
ree
d
o
minant
featu
r
es h
a
ve
a
hi
gh im
pa
ct on
the cl
assification p
r
o
c
e
s
s
with
its p-value di
stribution le
ss
than 0.05.
Table 2. Mea
n
, standa
rd d
e
viation (SD), and p-va
lu
e of the extract
ed 3D G
R
F g
a
it features.
Feature
ASD TD
p-value
Mean
SD
Mean
SD
Fx1
-2.37
2.66
-2.40
1.94
.584
Fx2
7.59
3.79
6.77
2.50
.301
Fx3
7.24
2.68
5.95
2.28
.051
Tx1
6.11
3.75
4.47
1.60
.130
Tx2
23.88
7.08
23.74
7.25
.944
Tx3
75.21
14.45
77.98
7.48
.717
F
y
1
-22.63
8.81
-19.48
3.75
.079
F
y
3
19.35
4.66
22.33
3.42
.007
T
y
1
12.15
4.60
14.10
4.02
.086
T
y
2
48.49
6.75
53.64
5.30
.
002
T
y
3
87.31
2.53
87.97
1.65
.294
Fz1 115.78
18.22
112.96
15.26
.595
Fz2 78.20
14.41
77.87
7.73
.209
Fz3 103.12
7.64
109.45
6.88
.001
Tz1 23.85
6.37
22.79
2.76
.304
Tz2 46.89
9.34
45.19
5.37
.391
Tz3 73.81
9.14
77.43
3.72
.048
Loading rate
5.32
2.09
5.05
1.09
.894
Push-off rate
4.28
1.17
4.97
0.80
.010
Peak ratio
1.12
0.16
1.03
0.15
.012
Table 3 sum
m
ari
z
e
s
the classificatio
n
accura
cy attained u
s
ing K
NN
cla
ssifie
r
with four
distan
ce
met
r
ics and
its optimize
d
k val
u
e
s
fo
r ea
c
h
3D
GR
F
-
T
M
WU
an
d
3
D
GR
F-
SWDA
datasets. It was ob
se
rved
t
hat the rates of correct
cl
assificatio
n
were
within th
e ran
ge 77%
to
83%. For the
3DGRF
-
TM
WU data
s
et
with six i
nput
feature
s
, th
e cityblo
c
k di
stan
ce
with
k=9
prod
uces 8
1
.67% accu
ra
cy as co
m
pare
d
to the other distan
ce.
Table 3. Cla
s
sificatio
n
accura
cy of KNN classifi
er
with four dista
n
ce metrics an
d
its optimized
k
values for the
TMWU a
nd
SWDA data
s
ets.
Distance
3DGR
F-
TMWU (
6
features)
3DGR
F-SWDA
(
3
features)
k Accur
a
cy
k
Accur
a
cy
Cit
y
block 9
81.67%
11
76.67%
Correlation
1 76.67%
8 76.67%
Cosine 9 80.00%
5 78.33%
Euclidean 7
76.67%
11
83.33%
Mean
while, t
he
combi
nati
on of th
ree
d
o
minant fe
ature
s
of
3DG
R
F-S
W
DA d
a
taset
and
KNN
cla
s
sifie
r
with
Eu
clid
ean
dista
n
ce
and
k=1
1
d
e
mon
s
trate
d
an imp
r
ove
d
perfo
rman
ce
for
ASD gait identific
a
tion
wit
h
83.33% acc
u
rac
y
.
Re
su
lts indi
cate
the p
o
tential
of usi
ng
bot
h
statistical fea
t
ure sel
e
ctio
n techniq
u
e
s
for t
he determinatio
n of signifi
cant an
d domina
n
t gait
feature
s
p
r
io
r to
perfo
rmi
ng id
entificat
ion of
AS
D
gait. In this
particular
c
a
s
e
, the SWDA
approa
ch p
r
o
duces a m
u
ch better
set o
f
predi
ctors.
This
study al
so hig
h
light
s
the releva
nce
of
the 3D
GRF
measureme
n
ts in ASD
g
a
it pattern
id
entification. F
u
ture
studie
s
sho
u
ld expl
ore
anothe
r type of possibl
e
gait feature
s
and ma
ch
ine cla
ssifie
r
s to enhan
ce classification
ac
cur
a
cy
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 15, No. 2, June 20
17 : 791 – 79
x
910
4. Conclusio
n
In this pap
er,
an identification sy
stem of
ASD gait ba
sed
on the 3
D
G
R
F is
pre
s
ente
d
.
This
study
h
a
s
evaluated
that the 3
D
G
R
F
gai
t
feature
s
extracted
u
s
ing
the time-se
r
ie
s
para
m
eteri
z
at
ion tech
niqu
e
s
and th
en se
lected u
s
in
g statistical feature
se
le
ction method
s
coul
d
be utilized f
o
r identificati
on of ab
normalities in ASD gait. Apart
from that, this study also
introdu
ce
s th
e impo
rtan
ce
of feature selectio
n
tech
nique
s fo
r se
lecting
domi
n
ant gait featu
r
es
prio
r t
o
cla
s
sif
i
cat
i
o
n
.
Ov
erall,
t
he sel
e
ct
ed 3
D
G
R
F gait feature
s
usi
ng
SWDA an
d the
opt
imize
d
K
N
N cla
s
sif
i
er
wer
e
su
c
c
e
s
sf
ully
discrimi
nated the 3D GRF gait pa
tterns into A
S
D
and T
D
gro
ups
with 83.
33% accu
ra
cy. These
fin
d
ing
s
wo
uld
be benefi
c
i
a
l for autom
atic
scree
n
ing
an
d diagn
osi
s
o
f
ASD and al
so for th
e de
tection of gai
t abnormalitie
s in individ
u
a
l
s
with ASD or o
t
her neu
rol
ogi
cal gait di
sorders.
Ackn
o
w
l
e
dg
ements
The auth
o
rs
woul
d like to
thank th
e Min
i
stry
of Hig
h
e
r
Edu
c
ation
(MOHE
)
, Mala
ysia for
the fund
s
re
ceived
thro
u
gh the
Ni
ch
e Res
earch
Grant S
c
h
e
m
e (NRGS),
proj
ect file:
600-
RMI/NRGS 5
/
3 (8/2013
), the Hum
an M
o
tion Gait
Analysis, Institut
e of Research Manag
eme
n
t
and In
novatio
n Unit (I
RMI) Premi
e
r
Lab
orato
r
y, Un
iv
ersiti
Te
knol
o
g
i MARA
(Ui
T
M) Sh
ah
Alam,
Selango
r. Th
e autho
rs al
so wi
sh to a
cknowl
edge
a
ll
contri
bution
s
esp
e
ci
ally the
Nation
al Auti
sm
Society of
Malaysia
(NASOM), vol
unteered
pa
rticipa
n
ts
an
d their fami
lies. Th
e
study
spo
n
sorship
wa
s funde
d b
y
the MOHE unde
r the Fe
deral T
r
aini
ng
(HLP)
schem
e 2014.
Referen
ces
[1]
America
n
Ps
ychiatric Ass
o
ci
ation. D
i
ag
nos
tic
and Statisti
cal Ma
nua
l of
M
ental
Disor
ders. 5th e
d
.
Arlingt
on, VA: America
n
Ps
y
c
hi
atric Assoc
i
at
ion; 20
13.
[2]
Kindr
ega
n D,
Galla
gher
L, G
o
rmle
y
J. Ga
it
Devi
ati
ons
in
C
h
ildr
e
n
w
i
th
Au
tism Spectrum
Disord
e
rs
: A
Revie
w
.
A
u
tis
m
Res
earc
h
an
d T
r
eatment
. 2015; 20
15: 8 p
ages.
[3]
Calhoun M, Long
w
orth M, Chester VL.
Ga
i
t
Pa
tte
rn
s in
Ch
il
dren w
i
t
h
Autism.
Clin
ical
Bi
omech
anics
.
201
1; 26(2): 20
0–2
06.
[4]
Chester VL, C
a
lh
oun M. Gait S
y
mmetr
y
i
n
Chil
dre
n
w
i
t
h
Autism.
Autism Research a
n
d
T
r
eatment
.
201
2; 201
2: 5 pag
es.
[5]
Giakas G, Baltzopo
ulos V. T
i
me and F
r
eq
u
enc
y
Dom
a
in
Anal
ys
is of Groun
d Reacti
on
F
o
rces Durin
g
W
a
lking: An In
vestigati
on of
Varia
b
il
it
y
a
nd
S
y
mmetr
y.
Gait & Posture
. 1997; 5(3): 18
9–1
97.
[6]
Perr
y
J. Gait Anal
ysis: N
o
rma
l and Path
ol
ogi
ca
l F
unctio
n
. MA, USA: SLACK Incorporate
d
.
1992.
[7]
Ambrosi
n
i D, Courc
hesn
e
E, Kaufman K. Mo
tion An
al
ysi
s
of Patients w
i
t
h
Infantil
e A
u
tism.
Gait &
Posture
. 19
98; 7(2): 188.
[8]
T
ahir NM, Man
ap H
H
. Parki
n
son D
i
se
ase G
a
it Cl
assificati
o
n
Bas
ed
on M
a
chi
ne
Lear
ni
n
g
Ap
proac
h.
Journ
a
l of App
l
ied Sci
ences
. 2
012; 12(
2): 180
–18
5.
[9]
Begg
RK, Pala
nis
w
am
i M, O
w
en B. Sup
port
Vect
or Mach
in
es for Automat
ed Gait Cl
assifi
cation.
IEEE
T
r
ansactio
n
s o
n
Bio
m
e
d
ica
l
Engi
neer
in
g
. 20
05; 52(5): 8
28–
838.
[10]
Kamruzzam
an
J, Begg RK. Supp
ort Vector Mach
i
nes an
d Other Pattern Reco
gniti
on A
ppro
a
ches to
the Di
ag
nosis
of Cere
bral
Pa
ls
y
G
a
it.
IEEE Transactio
n
s o
n
Bio
m
edic
a
l
Engi
neer
in
g
. 2
006; 5
3
(1
2)
:
247
9–
249
0.
[11]
Pradh
an
C, W
uehr
M, Akrami
F
,
Neu
h
a
euss
e
r M,
H
u
th S,
Brandt T
,
et al.
Automate
d
Cl
assificati
on
of
Neur
olo
g
ica
l
Di
sorders
of Ga
it Usi
ng S
pati
o
-
T
emporal Gait
Parameters.
J
o
urna
l of
Electro
m
yo
gra
phy
and Ki
nesi
o
l
o
g
y
. 2015; 25(
2): 413
–4
22.
[12]
Z
a
karia NK, J
a
ila
ni R, T
ahir NM. Applicati
on of ANN in
Gait F
eatur
es of Childr
en
for Gende
r
Classification.
Proced
ia Co
mputer Scie
nce
.
201
5; 76: 235
–
242.
[13]
Kaczmarcz
y
k
K, Wit A
,
Kraw
c
z
y
k M, Zaborski
J. Gait Classification in
Post-Stroke Patients Usi
n
g
Artificial Neural Net
w
orks.
Gait & Posture
. 2009; 30(2): 2
07–
210.
[14]
Mezgh
ani
N, H
u
sse S, B
o
ivi
n
K, T
u
rcot K, Ai
ssaou
i
R, Hag
e
meister N,
et al.
Autom
a
tic Classific
a
tio
n
of As
y
m
ptoma
t
ic an
d Osteo
a
rthritis K
nee
Gait
Patterns
Using
Ki
nemat
ic Data
F
eatur
es a
nd th
e
Near
est Neig
h
bor Cl
assifier.
IEEE Transactions on Bi
om
edical Engineering
. 2008; 5
5
(3): 123
0–
123
2.
[15]
Mustafa M,
T
a
ib M, Murat
Z,
Sula
iman N. C
o
mparis
on Bet
w
e
e
n
KNN an
d
ANN Classific
a
tion i
n
Brain
Bala
ncin
g A
p
p
licatio
n vi
a S
pectrogr
am Image.
J
ourn
a
l of
Co
mp
uter Sc
ienc
e & C
o
mp
utatio
n
a
l
Mathem
atics
. 2
012; 2(4): 1
7–2
2.
[16]
Coom
ans D,
D.L. Massart. Alternative
K-
Nearest
Neig
hb
our R
u
les in S
uper
vised Patter
n
Reco
gniti
on
:
Part 1. K-Near
est Nei
ghb
our
Classific
a
tio
n
b
y
Usi
ng Alte
rnative V
o
tin
g
Rul
e
s.
Anal
Chim
Acta
. 19
82; 136: 1
5–2
7
.
[17]
Ilias S, T
ahir N
M
, Jaila
ni R,
H
a
san
CZC.
Cl
a
ssificatio
n
of A
u
tism
Ch
ildr
en
Gait Patterns
Using
Ne
ural
Netw
ork and
Supp
ort Vecto
r
Machin
e
. 2016 IEEE S
y
m
posium on Computer Applications and
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Autism
Spectrum
Diso
rde
r
s Gait Identificat
ion
Usi
ng
Grou
nd… (Che Zawi
ya
h Che Ha
sa
n)
911
Industria
l Elect
r
onics (ISCAIE
)
.
Penang, Mal
a
y
s
ia. 20
16: 52
–56.
[18]
Alliso
n
K, Wri
g
le
y T
V., Vicen
z
ino
B, Ben
n
e
l
l KL, Grima
l
d
i
A, Hod
ges PW
. Kinem
atics a
nd K
i
netics
Durin
g
W
a
lki
n
g
in Indivi
du
als w
i
t
h
Glutea
l T
e
ndi
nop
ath
y
.
Cli
nical B
i
o
m
ec
ha
nics
. 201
6; 32: 56–
63.
[19]
Hub
e
rt
y
CJ, Olejnik S. Appl
ie
d MANOVA and Di
scrimi
nant
Anal
ys
is. Ne
w
Jerse
y
: Jo
hn
W
ile
y
& So
ns.
200
6.
[20]
Mulro
y
S, Gronle
y
J, W
e
iss W
,
Ne
w
s
am
C,
Perr
y
J. Use of Cluster Anal
ys
is for Gait Patter
n
Classific
a
tio
n
of Patients in t
he Earl
y a
nd L
a
te Recov
e
r
y
Phases F
o
l
l
o
w
ing Stroke.
Gait & Posture
.
200
3; 18(1): 11
4–1
25.
[21]
Ha
w
o
rt
h JL, H
a
rbo
u
rne
RT
,
Vall
abh
aj
osul
a
S, St
ergiou N.
Center of
Pre
ssu
re a
nd the
Projecti
on of
the T
i
me-Cour
se of Sitting Skill Acq
u
isiti
on.
Gait & Posture
. 2013; 3
8
(4): 8
06–
81
1.
[22]
Massaa
d
A, A
ssi A, Skal
li
W
,
Ghanem I. Rep
eat
a
b
il
it
y
and
Val
i
d
a
tion
of Gait D
e
vi
a
t
ion In
de
x
i
n
Chil
dre
n
: T
y
pic
a
ll
y Dev
e
l
opi
ng
and Cer
ebr
al Pals
y
.
Gait & Posture
. 201
4; 3
9
(1): 354
–3
58.
[23]
Di
xon
PC, B
o
w
t
ell M
V., Steb
bins
J. T
he Us
e of
Re
gressi
o
n
a
n
d
Norm
alis
ation
for th
e C
o
mparis
on
o
f
Spatio-T
empor
al Gait Data in
Chil
dre
n
.
Gait & Posture
. 201
4; 40(4): 52
1–5
25.
[24]
W
h
ite R, Agou
ris I, Selbie
RD, Kirkpatrick M.
T
he Variabi
lit
y
of
Force Platform Data in Normal and
Cere
bral Pa
ls
y.
Clinica
l
Bio
m
echa
nics
. 19
99
; 14(3): 185
–19
2.
[25]
Mana
p HH, T
ahir NM, Y
a
ssin
AIM.
Statistical Analys
is of P
a
rkinso
n Dis
eas
e Gait C
l
assific
a
tion
Usi
n
g
Artificial Neural Network
. IEEE International S
y
m
posium
on
Signal Pr
ocessing
and
Information
T
e
chnolog
y (IS
SPIT
)
. Bilbao,
Spai
n. 201
1: 60–6
5.
[26]
W
h
ittle MW
.
Gait Anal
ysis:
An Introducti
on.
4th ed. P
h
ila
de
lph
i
a, U
SA: Butter
w
ort
h
-Hei
nem
ann
;
200
7.
[27]
Stansfiel
d BW
, Hillm
an
SJ, H
a
zle
w
o
o
d
ME,
La
w
s
on AM, M
ann
AM, Lo
ud
on IR, et
al. N
o
rmalis
atio
n
of Gait Data in Chil
dre
n
.
Gait & Posture
. 200
3; 17(1): 81
–87
.
[28]
Greer NL, H
a
mill J, C
a
mpb
e
ll K
R
. D
y
n
a
m
i
cs of Ch
il
dren
’
s
Gait.
Hu
ma
n
Move
me
nt Sc
ienc
e
. 19
89;
8(5): 465
–4
80.
[29]
Reid SM, Gra
ham RB, Costi
gan PA. Differ
ent
iati
on of Yo
ung a
nd Old
e
r
Adult Stair Cli
mbin
g Gait
Using Pri
n
ci
pal
Compo
nent A
nal
ysis.
Gait & Posture
. 20
10; 31(2): 19
7–
203
.
[30]
W
o
lf S, Loose
T
,
Schablo
w
ski
M, Döderl
e
in L, Ru
pp R, Gerner
HJ, et al. Automated F
eatur
e
Assessment
in Instrumented Gait
Anal
ysis.
Gait & Posture
. 2006; 2
3
(3): 3
31–
33
8.
[31]
Su BL, Son
g
R, Guo LY,
Yen CW
. Ch
a
r
acterizin
g
Gai
t
As
y
mmetr
y
Via F
r
eq
uenc
y Sub-Ba
nd
Comp
one
nts o
f
the Ground R
eactio
n
F
o
rce.
Bio
m
e
dal S
i
gn
al Process
i
n
g
and C
ontro
l
. 2
015; 1
8
: 56
–
60.
[32]
McCror
y
J
L
, W
h
ite SC, Lifeso RM. Vertical
Groun
d Re
action F
o
rces:
Objective Me
asures of Gait
Follo
w
i
n
g
Hip
Arthropl
ast
y
.
Gait & Posture
. 200
1; 14(2): 10
4–1
09.
[33]
Ma
yers A. Introducti
on to Statistics and SP
SS in
Ps
y
c
hol
og
y. Harl
o
w
, Engl
and: Pe
arso
n Educati
o
n
;
201
3.
[34]
Deluz
i
o KJ, A
s
tephe
n JL. Bi
omech
anic
a
l F
eatur
es of Gai
t
W
a
veform Data Associat
ed
w
i
th K
n
e
e
Osteoarthritis. An Appl
icati
on
of
Princip
a
l Co
mpon
ent Ana
l
ysis.
Gait & Posture
. 2007; 2
5
(
1
): 86–9
3.
[35]
Kohav
i R, Provost F
.
Glossar
y
of
T
e
rms.
Machin
e Le
arni
ng
. 199
8; 30: 271
–
274.
Evaluation Warning : The document was created with Spire.PDF for Python.