Understanding effects of cognitive rehabilitation under a knowledge discovery approach

Author: García Rudolph, Alejandro,Gibert, Karina

Publisher: Elsevier

Year: 2016

DOI: 10.1016/j.engappai.2016.06.007

Source: https://upcommons.upc.edu/bitstream/2117/386517/1/7Understanding%20effects%20of%20cognitive%20rehabilitation%20under%20a%20knowledge%20discovery%20approach.pdf

Unde s anding e ec s o cogni i e ehabili a ion unde a knowledge
disco e y app oach
Alejand o Ga cía-Rudolph
a
, Ka ina Gibe
b,
n
a
Ins i u Gu mann –Hospi al de Neu o ehabili ació, Camí de Can Ru i s/n 08916, Badalona, Ba celona Spain
b
Dep. o S a is ics and Ope a ions Resea ch, Uni e si a Poli ècnica de Ca alunya –Ba celonaTech, Jo di Gi ona 1-3, 08034 Ba celona, Spain
a icle in o
Keywo ds:
Cogni i e ehabili a ion
T auma ic b ain inju y
Knowledge disco e y
Mo i s
Clus e ing-based on ules
Pos -p ocessing
abs ac
T auma ic b ain inju y (TBI) is he leading cause o dea h and disabili y in child en and young adul s
wo ldwide. Cogni i e ehabili a ion (CR) plans consis o a sequence o CR asks a ge ing main cogni i e
unc ions. The e is no enough on-ﬁeld expe ience ye ega ding which speciﬁc in e en ion ( asks o
exe cise assignmen ) is mo e app op ia e o help he apis s o design plans wi h signiﬁcan e ec i eness
on pa ien imp o emen . The selec ion o speciﬁc asks o be p esc ibed o he pa ien and he o de in
which hey migh be execu ed is cu en ly decided by he he apis s based on hei expe ience.
In his pape a new da a mining me hodology is p oposed, combining se e al ools om A iﬁcial
In elligence, clus e ing and pos -p ocessing analysis o iden i y egula i ies in he sequences o asks in
such a way ha ea men p oﬁles (classes) can be disco e ed. Due o he cumula i e e ec o e-
habili a ion asks, small a ia ions wi hin he sequence o asks pe o med by he pa ien do no sig-
niﬁcan ly change he ﬁnal ou comes in ehabili a ion and makes i di ﬁcul o ﬁnd disc iminan ules by
using he adi ional machine lea ning induc i e me hods. Howe e , by elaxing he o maliza ion o he
p oblem o ﬁnd pa e ns ha migh include small a ia ions, and in oducing mo i disco e y echniques
in he p oposed me hodology, he complexi y o he neu o ehabili a ion phenomenon can be be e
cap u ed and a global s uc u e o success ul ea men ask sequences can be de ised.
Following his, he ela ionship be ween he disco e ed pa e ns and he CR ea men esponse a e
analyzed, o e ing a iche pe spec i e han ha p o ided by he single ask ocus adi ionally used in
he CR ﬁeld.
The pape p o ides a deﬁni ion o he whole me hodological app oach p oposed om a o mal poin
o iew, and i s applica ion o a eal da ase . Compa isons wi h adi ional AI app oaches a e also p e-
sen ed and he con ibu ion o he p oposed me hodology o he AI ﬁeld discussed.
&2016 Else ie L d. All igh s ese ed.
1. In oduc ion
Acco ding o he Wo ld Heal h O ganiza ion (WHO), auma ic
b ain inju y (TBI) is he leading cause o dea h and disabili y in
child en and young adul s wo ldwide and i is in ol ed in nea ly
hal o all auma dea hs (Sch oe e e al., 2011). In Eu ope, b ain
inju ies om auma a e esponsible o mo e yea s o disabili y
han any o he cause (Maas e al., 2009). The incidence is in-
c easing in lowe income coun ies; he WHO p edic s ha TBI
and oad a ﬁc acciden s will be he hi d cause o disease and
inju y wo ldwide by 2020. Cogni i e impai men s due o TBI a e
subs an ial sou ces o mo bidi y o a ec ed indi iduals, hei a-
mily membe s, and socie y. Dis u bances o a en ion, memo y,
and execu i e unc ioning a e he mos common consequences o
TBI.
Neu o ehabili a ion is he p ocess ha exploi s he ce eb al
plas ici y o educe b ain deﬁci . Cogni i e ehabili a ion (CR) aims
o educe he impac o disabling condi ions and ies o imp o e
he cogni i e deﬁci s caused by TBI. F om Lu ia’s heo y back in
1978, he e is a common belie e ha di ec e aining o damaged
cogni i e p ocesses h ough epea ed s imula ion and ac i a ion o
he a ge ed b ain a eas can help pa ien eco e y. Fo maximum
ac i a ion o occu , he pa ien mus ace asks jus ba ely oo di -
ﬁcul o him (G een and Ba elie , 2005). Designing a CR ea men
o a gi en pa ien he e o e means de e mining he co ec se-
quence o CR asks o be asked o he pa ien in a qui e p ecise
ade-o be ween enough s imula ing and su ﬁcien ly achie able
asks, which is a om in ui ion, and s ill is bo h an empi ical and
heo e ical open p oblem in he a ea. I has been seen ha simila
pa ien s espond di e en ly o simila CR ea men s. Li e a u e
epo s single ask app oaches o his pu pose, analysing he
n
Co esponding au ho .
E-mail add esses: [email p o ec ed] (A. Ga cía-Rudolph),
[email p o ec ed] (K. Gibe ).
associa ions be ween he pe o mance o a ce ain ask and he
esponse o he CR ea men . Howe e , al hough he e is some
empi ical knowledge, adi ional app oaches do no seem o p o-
ide su ﬁcien scien iﬁc e idence abou he ac o s de e mining a
a ou able ou come, and he e is s ill a limi ed scien iﬁc base o
suppo he e ec i eness o CR (Cice one e al., 2011). In ac , mos
o he wo k ound in his ﬁeld adop s adi ional p e-pos analysis
wi h in e en ion s udies whe e a speciﬁc ea men is applied o
a sample o pa ien s and con as ed wi h a con ol g oup. Bu none
o his analysis includes de ailed cha ac e is ics o di e en CR
p og ams in he model. They only use assessmen o he pa ien
be o e and a e he ea men and cha ac e is ics o he lesion and
o he pa ien o p edic imp o emen . These wo ks, al hough
use ul o p o e he e ec i eness o CR, do no con ibu e o a
be e design o CR p og ams o a speciﬁc pa ien .
In his wo k, he unde lying s uc u e o he CR phenomenon
has been analyzed in dep h and i has been seen ha he CR ﬁeld
has some speciﬁc cha ac e is ics ha make a success ul applica-
ion o adi ional me hods di ﬁcul :

Pa ien s ollowing a CR p og am a e no pe o ming a single
ask, nei he a single ype o ask, bu a ce ain complex com-
bina ion o hem ha a e likely o be in e ela ed o syne gis-
ics. A single ask app oach canno ake in o accoun he com-
plex in e ac ions among asks.

Cogni i e asks, e en when speciﬁcally designed o a ge a
pa icula cogni i e unc ion, migh also ha e side-e ec s on
o he cogni i e unc ions (Cice one e al., 2011). This makes i
di ﬁcul o examine he isola ed e ec o a single ask in a
speciﬁc cogni i e unc ion, and no clea e idence appea when
all asks a e in eg a ed in o a adi ional model.

The addi ional e ec o a single ask migh be a ec ed by he
cumula i e e ec o he sequence o p e ious asks execu ed
unde he ea men , and his migh de e mine ha he o de o
execu ion is ele an in he ea men .

The e ec o a single ask may be oo sub le o be de ec ed,
whe eas he e ec s o he whole CR ea men may be su ﬁcien
o be de ec able, aking in o accoun he cumula i e e ec o
ehabili a ion al eady men ioned.
F om a s uc u al poin o iew, hese cha ac e is ics esemble
hose in nu i ional epidemiology, whe e a global app oach has
been adop ed in ecen yea s, and all nu ien s a e analyzed o-
ge he due o he high deg ee o in e ac ion (Hu, 2002). This poin s
o analyzing he o e all CR ea men , by conside ing all kinds o
in e ac ions among asks oge he , ins ead o using he adi ional
single ask app oach. The e o e, CR ea men in his wo k will be
conside ed as a sequence o cogni i e asks and da a mining
me hods will be used o de e mine he mul i a ia e associa ions
be ween a CR ea men (o ele an subsequences) and he de-
g ee o esponse o he pa ien , unde his new pe spec i e. Ana-
lyzing CR asks as ea men pa e ns o e s an inno a i e pe -
spec i e in neu o ehabili a ion, and desc ibing hei ela ionship
wi h hei clinical ou come p o ides a p ac ical app oach o e al-
ua e he e ec s o ehabili a ion ea men s. I can also enhance
ou concep ual unde s anding o CR ea men s p ac ice, and
migh be use ul o p o ide guidance o cogni i e ea men
in e en ions.
To his end, an inno a i e da a mining me hodology is p o-
posed o analyze he unde lying s uc u e o CR p ocesses and o
be e de e mine he mos sui able CR ea men o a gi en pa-
ien . In a ﬁ s app oach, one would be emp ed o iden i y he ask
sequences associa ed wi h he imp o emen o di e en pa ien s,
by using a adi ional classiﬁe o ﬁnd pa e ns associa ed wi h he
di e en ea men esponses ound in di e en pa ien s. How-
e e , his app oach has shown se ious limi a ions in his ﬁeld,
p o iding ex emely weak pa e ns ha do no eally help in
clinical p ac ice, as will be discussed la e .
In his pa icula ﬁeld, because o he cumula i e e ec o he
asks men ioned abo e, i is easonable o hink ha he e ec o a
ce ain sequence o CR asks can beha e obus ly o sligh a ia-
ions o he sequence. Thus, small a ia ions o he sequence o
asks pe o med migh keep he global e ec o he ea men
unal e ed. This means ha he model o be buil should admi a
ce ain le el o a ia ion a ound e e y ele an pa e n. Howe e ,
wo king wi h a ask’s global p oﬁle o ea men s is nei he
use ul, as he o de o execu ions becomes ele an in CR. These
cha ac e is ics ha e al eady been encoun e ed in he bioin o -
ma ics ﬁelds, pa icula ly in ansc ip ion ac o binding si es (TFBS)
ﬁeld, whe e sligh ly di e en sequences o DNA a e associa ed
wi h a ce ain biological unc ion. Mo i disco e y o mo i ﬁnding
me hods a e used in his ﬁeld, o ep esen hose weak pa e ns.
Simila ly, mo i disco e y me hods will be in oduced in ou p o-
posed me hodology o iden i y pa e ns o CR ea men s, whe e
sligh a ia ions in he ea men p og am migh be packed in o a
single CR mo i wi h a simila he apeu ic e ec , and migh be
associa ed wi h a ce ain esponse le el.
This pape in oduces he new Sequence o Ac i i ies Imp o -
ing Mul i-A ea Pe o mance (SAIMAP) me hodology, as an in-
no a i e combina ion o p e-p ocessing ools, clus e ing, AI
me hods, mo i disco e y and pos -p ocessing echniques. SAIMAP
is a hyb id me hodological ame whe e use ul pa e ns can be
ound om da a. SAIMAP wo ks o domains wi h high o de in-
e ac ions among a iables and sequen ial in o ma ion along ime
ha in ol es cummula i e e ec s. This p o ides a complex
s uc u e, o which mos o he classical da a mining app oaches
do no pe o m e y well. The p oposed me hodology is gene al
o p oblems wi h he s uc u e desc ibed be o e, al hough in he
pape i is applied o he pa icula ﬁeld o ﬁnding design guide-
lines o CR ea men s. SAIMAP ﬁ s ﬁnds g oups o simila
ea men s, hen makes a local cha ac e iza ion o each g oup by
using mo i disco e y me hods, and ﬁnally analyzes he ela ion-
ships be ween hose ypical ea men s and he e alua ion o pa-
ien s’imp o emen a e ea men . S a is ical es s and mul iple
boxplo s a e used o ela e he disco e ed g oups wi h pa ien ’
cha ac e is ics, le el o impai men and associa ed wi h speciﬁc
ea men pa e ns.
The s uc u e o he pape is he ollowing: Sec ion 2 in-
oduces he s a e o he a , o ganized in sec ions ela ed o he
di e en esea ch a eas in ol ed in his mul idisciplina esea ch,
om bo h he applica ion and me hodological poin o iew. In
Sec ion 3 me hodological issues a e p o ided: ﬁ s , he o -
maliza ion o he gene al p oblem add essed is deﬁned, wi h a
clea p esen a ion o he s uc u al componen s o he p oblem
(ou me hodological p oposal add esses scena ios whe e in-
di iduals pe o m sequences o p edeﬁned se o ac i i ies wi h
high o de in e ac ions among hem and cumula i e e ec s). The
main goal is o iden i y a educed se o cha ac e is ic sequences o
ac i i ies p oﬁling g oups o indi iduals who beha e simila ly. The
second pa o Sec ion 3 in oduces he SAIMAP me hodology as
ou p oposal o add ess his p oblem. SAIMAP is composed by 13
o mal s eps, whe e sequen ial pa e ns a e induced om da a
ﬁ ing he s uc u e deﬁned in Sec ion 3.1.InSec ion 4 ou
me hodology is applied o a speciﬁc eal case, ega ding cogni i e
ehabili a ion ea men s o auma ic b ain inju y pa ien s; he
inpu s a e sequences o cogni i e ehabili a ion asks pe o med
by he pa ien s along he CR ea men ; p ep ocessing ac i i ies
a e de ailed and sequen ial pa e ns o CR asks a e ob ained ol-
lowing he SAIMAP s eps. The disco e ed pa e ns, a e in e p e ed
h ough mo i disco e y ools and associa ed wi h se e al c i e ia
measu ing imp o emen in a p edeﬁned se o impac a eas ha
migh be a ge ed in pa allel by a single ask (in he pa icula case
o applica ion, memo y, a en ion and execu i e unc ion impac
a eas). Sec ion 5 compa es he pa e ns ob ained unde SAIMAP
app oach wi h adi ional da a mining me hods, like classiﬁe s
(decision ees, neu al ne wo ks, …) and sequen ial pa e n
mining. Finally, Sec ion 6 p o ides conclusions and u u e wo k.
2. S a e o he a
2.1. Cogni i e ehabili a ion
Neu o ehabili a ion is he p ocess o iden i ying he esidual
deﬁci and exploi ing he ce eb al plas ici y o educe i . This is
achie ed by elabo a ing he apeu ic plans o a o he es ablish-
men o new and app op ia e neu ological connec ions, obse ing
pa ien esponses o he plan and guiding hem o he p ope e-
sponses (Taly e al., 1998). Cogni i e ehabili a ion (CR), as pa o
neu o ehabili a ion, aims o educe he impac o disabling con-
di ions and ies o imp o e he cogni i e deﬁci s caused by TBI. I
aims o educe unc ional limi a ions and inc ease he indi idual's
abili y in hei daily ac i i ies. The mode n e a o CR da es o
Gianu sos’s seminal a icle, Wha is cogni i e ehabili a ion? (Gia-
nu sos, 1980), which laid ou an app oach based on Alexande
Lu ia’s heo y o cogni i e p ocesses (Lu ia, 1978), based on he
assump ion ha di ec e aining o speciﬁc cogni i e p ocesses
h ough mul iple epea ed ials o s imula ion and ac i a ion o
he a ge ed cogni i e p ocess can lead o he eo ganiza ion o
highe le el neu ologic and cogni i e p ocesses. The e is a com-
mon belie ha CR is e ec i e o pe sons wi h TBI, based on a
la ge numbe o s udies and ex ensi e clinical expe ience (Rohling
e al., 2009). Howe e , cu en knowledge abou he ac o s de-
e mining a a ou able ou come is mainly empi ical, and he e is
s ill a limi ed scien iﬁc basis o suppo he e ec i eness o such
in e en ions (Cice one e al., 2011).
Acco ding o Gianu sos i is possible o isola e and measu e
ounda ional aspec s o cogni ion—a en ion, pe cep ion, and
memo y, among o he s— o ea hem di ec ly wi h he use o
speciﬁcally designed ac i i ies, ei he on able ops o pe sonal
compu e s. Repe i ion is pe haps he hallma k o his app oach. A
ypical CR p og am mainly p o ides specialized asks, which e-
qui e epe i i e use o he impai ed cogni i e subsys em in a
p og essi ely mo e demanding sequence (Sohlbe g, 2001). Each
ask a ge s a p incipal cogni i e unc ion (a en ion, memo y,
easoning/p oblem sol ing, o execu i e unc ions) and can be
p oposed o he subjec a di e en le els o di ﬁcul y. As soon as a
pa ien has mas e ed a pa icula exe cise o g oup o exe cises,
highe -le el ea men asks a ge ing he same cogni i e com-
ponen need o be a ailable so ha he con inued s imula ion and
ac i a ion o he objec i e cogni i e p ocess can occu .
This pa adigm, known as p ocess-speciﬁc o skill-speciﬁc(Sohl-
be g, 2005), sp ead quickly, and can impac di e en ially on
neu ocogni i e deﬁci s, p o ided ha he p ope sequence o asks
is deli e ed o he pa ien . Fo maximum ac i a ion o occu , he
pa ien mus ace asks jus ba ely oo di ﬁcul o him (G een and
Ba elie , 2005). Thus, ﬁnding he co ec aining schedule o a
gi en pa ien equi es a qui e p ecise ade-o be ween enough
s imula ing and su ﬁcien ly achie able asks, which is a om in-
ui ion, and s ill is bo h an empi ical and heo e ical open p oblem.
In ou p e ious esea ch (Ga cía-Rudolph and Gibe , 2014) he
neu o ehabili a ion ange has been in oduced as an objec i e c i-
e ion o de e mine he p ope le el o di ﬁcul y o be p oposed o
a ask, bu i is s ill insu ﬁcien in e ms o managing CR ea men
globally, since cogni i e asks in e ac among hem.
The design o a CR p og am has become an essen ial issue.
Howe e , in clinical p ac ice, he apis s mainly design CR plans
om sc a ch, de e mining clinical se ings o speciﬁc pa ien s
mainly based on he he apis ’s expe ise (Jaga oo, 2009;Cice one
e al., 2011). Each speciﬁc plan e ol es acco ding o each he a-
pis ’s own c i e ia and e alua ion on he pa ien ’s ollow-up. The e
is no enough on-ﬁeld expe ience ye ega ding which speciﬁc
in e en ion ( ask o exe cise assignmen ) is mo e app op ia e o
help CR he apis s o design hei CR plans (Cice one e al., 2011).
The e o e ﬁnding pa e ns o cogni i e ask sequences ha p o-
duce signiﬁcan imp o emen on a ec ed cogni i e unc ions can
deﬁni ely con ibu e in his a ea.
T adi ionally, neu o ehabili a ion e o s ha e been ocused on
modelling and quan i ying he e ec o a single cogni i e ask on a
pa ien . These ypes o analyses ha e been qui e aluable (Cice one
e al., 2011). They ha e shown, o ins ance, ha he e ec o a ask
mainly depends on he a io be ween he skills o he ea ed
pa ien and he challenges (di ﬁcul y) in ol ed by he ask (IOM,
2011;Why e and Ha , 2003). Howe e , he single ask app oach
may be inadequa e o ﬁnding global models o he ehabili a ion
p ocess in clinical p ac ice. Limi a ions o single ac o app oaches
ha e also been ound in o he esea ch a eas, such as nu i ional
epidemiology (Hu, 2002) whe e esea ch opics ecen ly mo ed
owa ds he o e all die a y pa e ns by conside ing how oods and
nu ien s a e consumed in combina ions owa ds global nu i-
ional models (Millen e al., 1996). This seems o indica e ha he
mul i a ia e app oach will be much mo e con enien in ou case
oo.
In addi ion and on accoun o he cumula i e e ec o he asks
men ioned abo e, i is easonable o hink ha he ehabili a i e
powe o a ce ain sequence o asks could emain unal e ed, e en
when he pa ien execu es o he asks a in e media e poin s o
he sequence. In ac , imp o emen in a en ion o wo pa ien s
pe o ming wo sequences o asks ( o example Ci cles-Ma ching-
Di Di ec ion-S aigh Line ) and (Ci cles-Ma ching-Ca ego izing-Dis-
c imina ion-Di Di ec ion-Plani -S aigh Line) is expec ed o be he
same, since Ca ego izing, Disc imina ion and Plani a e asks o -
ien ed mos ly o wo k execu i e unc ions and do no mainly in-
e ac wi h a en ional skills.
This means ha he model o be buil should admi a ce ain
le el o a ia ion a ound e e y ele an pa e n. These cha ac e -
is ics ha e al eady been encoun e ed in bioin o ma ics, ela ed o
ansc ip ion ac o binding si es (TFBS), in which egula o y ele-
men s in nucleo ide sequences a e sea ched. Ce ain segmen s o
he DNA a e ansc ibed in o ano he molecule (RNA), which
se es as a empla e o make he basic building blocks o cellula
li e: p o eins (Zambelli e al., 2012). This ﬁ s s ep o gene ex-
p ession, ansc ip ion, is egula ed by di e en ac o s, among
which ansc ip ion ac o s (TFs) play a key ole binding DNA nea
he ansc ip ion s a si e o genes. E en hough some TFs bind o
DNA in a e y unspeciﬁc way, mos o hem bind by ecognizing
speciﬁc sequence elemen s ( he TFBS) (D’Haeselee , 2006). Typi-
cally, a TF ecognizes no jus one pa icula sequence bu a
numbe o simila sequences ha can include small a ia ions
wi hin i . This collec ion o sligh ly di e en sequences and i s
di e se se o ep esen a ions a e collec i ely known as binding
mo i s. TF bind he DNA in a speciﬁc way o ming sequences ha
a e simila bu no necessa ily iden ical, di e ing among hem in a
ew nucleo ides, bu accomplishing he same biological unc ion.
Binding mo i s a e ound using mo i disco e y o mo i ﬁnding
me hods. These echniques enable us o ﬁnd sho simila se-
quence elemen s (building he mo i ) sha ed by a se o nucleo ide
o p o ein sequences wi h a common biological unc ion.
S uc u al simila i ies in ol ing binding mo i s and he p op-
e ies o CR ea men s a e p oposed o be exploi ed in ou ap-
p oach. The main con ac poin is ha he e ec o a ce ain se-
quence o CR asks can beha e obus ly o sligh a ia ions o he
sequence, hose o in oducing o he cogni i e asks wi hin he
sequence. A ypical CR p og am a ge s a limi ed g oup o CR
unc ions (e.g. a en ion, memo y, execu i e unc ions). The hy-
po hesis o he cumula i e e ec o ehabili a ion asks makes i
sui able o model he p oblem by using mo i disco e y echniques
o e CR ask sequences. This way, sligh a ia ions in he ea -
men p og am migh be packed in a single CR mo i wi h a simila
he apeu ic e ec , in he same way as small a ia ions o nucleo-
ide sequences a e packed in a single binding mo i i hey egula e
in he same way a ce ain p o ein. Sea ching o mo i s o e CR
ea men p og ams (sequences o cogni i e unc ions) is expec ed
o iden i y basic sequences o CR asks ha p oduce a ce ain e-
sponse o ea men pa e n, e en i hey a e pe o med wi h small
a ia ions.
2.2. Classiﬁca ion
A classiﬁca ion echnique (e.g. decision ees, k-nea es neigh-
bo s, neu al ne wo ks, suppo ec o machines, nai e bayes) is a
sys ema ic AI app oach o build classiﬁca ion models om inpu
da a se s (Tan e al., 2006). The inpu da a o a classiﬁca ion ask is
a collec ion o examples, desc ibing objec s by a se o a ibu es
and a class label. Classiﬁca ion is he ask o lea ning a a ge
unc ion ha maps each a ibu e se x o one o he p edeﬁned
class labels y. The a ge unc ion is also known as classiﬁca ion
model and depending on he classiﬁca ion me hod used, i migh
be implici o explici and can ake he o m o a knowledge base,
o a decision ee, o e en a black-box, as in he case o neu al
ne wo ks. In o de o e alua e he pe o mance o he speciﬁc
classiﬁca ion echnique, k- old c oss- alida ion is used o es i-
ma ing how accu a ely a p edic i e model will pe o m in p ac ice
(Hall e al., 2009).
A numbe o s udies employ adi ional classiﬁca ion echni-
ques o he au oma ic p ognosis o TBI pa ien s, i.e. an icipa ing
ea men ou come om he usual cou se o he disease and/o he
peculia i ies o each indi idual case. The e is no consensus ye on
an op imal me hod. Di e en app oaches ha e he e o e been
explo ed (Ga cia e al., 2013). Decision ees a e he mos common
choice, (B own e al., 2005); (Ro lias and Ko sou, 2004), bu neu al
ne wo ks (Pang e al., 2007); (Segal e al., 2006) o di e en e-
g ession models (And ews, 2002) a e also used. These s udies o-
cus on de e mining su i al, p edic ing g oss ou come, and/o
iden i ying p edic i e ac o s o a pa ien ’s condi ion a e TBI
(usually acu e TBI). Recen s udies (Pignolo Pignolo and Lagani,
2011) compa e di e en machine lea ning classiﬁe s (C4.5, Sup-
po Vec o Machine, Nai e Bayes, K-NN) in he ea ly p edic ion o
ou come o he subjec s in ege a i e s a e due o TBI. As p e-
iously men ioned, neu al ne wo ks ha e also been applied e.g. o
p edic in-hospi al su i al ollowing TBI (Rughani e al., 2010).
When gi en he same limi ed clinical in o ma ion, he ANN sig-
niﬁcan ly ou pe o med eg ession models and clinicians on
mul iple pe o mance measu es. Pa icula ly, h ee laye ed back-
p opaga ion neu al ne wo k wi h an inpu laye o 10 nodes whose
ou pu p o ides he inpu s o a hidden laye was used. Thi y- wo
TBI pa ien s o di e en age and gende we e aken in he s udy, a
signiﬁcan ela ionship be ween sys em ou pu s and neu ologis s’
decisions was ound (Güle e al., 2009).
To he bes o ou knowledge, no wo k has been ound using
classiﬁe s o lea n CR ask pa e ns o de e mine he deg ee o
imp o emen o a pa ien a e a CR ea men .
2.3. Sequen ial pa e n mining
Since he s uc u e o he da ase ha we in end o analyze
con ains sequences o CR asks pe o med by a pa ien along a CR
ea men , sequen ial pa e n mining (SPM) is also conside ed. A
sequen ial pa e n is a sequence. Gi en a sequence o i emse s S
A
¼X
1
,X
2
,…X
k
, he sequence S
B
¼Y
1
,Y
2
,…Y
m
,m k occu s in
S
A
i all elemen s in S
B
belong o S
A
and p ecedences o S
B
ele-
men s a e conse ed in S
A
. The suppo o an SP is he p opo ion
o sequences whe e he pa e n occu s in he da abase. F equen
SP has a suppo g ea e han a ce ain h eshold p o ided by he
use ; subsequences a e also conside ed.
SPM plays an impo an ole in da a mining and is essen ial o a
wide ange o applica ions such as he analysis o web click-s eams,
p og am execu ions, heal hca e da a, biological da a and e-lea ning
da a (Mab oukeh and Ezei e, 2010). Pa e ns in he heal hca e do-
main include he common pa e ns in pa hs ollowed by pa ien s in
hospi als, pa e ns obse ed in symp oms o a pa icula disease,
pa e ns in daily ac i i y, and heal h da a (Gup a and Taly, 2012). A
ecen example o mining in a medical con ex is he applica ion o
he sequen ial pa e n mining algo i hms on a da abase known as
he RSU D . Soe omo medical da abase o ﬁnd disease SP (Yuliana
e al., 2009). Howe e , age and gende we e no included in he se-
quen ial ules and he au ho only displayed a selec ion o ules.
O he exis ing wo k aims o de ec medical SP in ended o ocus on
ime se ies da a (P adhan and P abhaka an, 2009)o speciﬁc ill-
nesses, such as pa e ns p edic ing he onse o h ombosis and
iden i ying ai s leading o a he oscle osis in a da abase o ap-
p oxima ely 1400 middle aged men (Klema e al., 2008).
To he bes o ou knowledge he iden iﬁca ion o SP whe e a
TBI ehabili a ion ea men is conside ed as a sequence o CR
asks has no ye been add essed. And he me hodologies used in
ela ed wo ks p e iously men ioned do no esis se s o a iables
wi h cumula i e e ec s and high-deg ee in e ac ions, as s a ed in
he in oduc ion.
Se e al e ﬁcien algo i hms ha e been p oposed o SPM such
as ClaSP (Goma iz e al., 2013), CloSpan (Yan e al., 2003) GSP
(S ikan and Ag awal, 1996), P eﬁxSpan (Pei e al., 2004), SPADE
(Zaki, 2001). and SPAM (Ay es e al., 2002). SPM algo i hms can
use a ho izon al da abase o ma (e.g. CloSpan, GSP and P eﬁx-
Span) o a e ical da abase o ma (e.g. ClaSP, SPADE, SPAM).
Using he e ical o ma p o ides he ad an age o gene a ing
pa e ns and compu ing hei suppo s wi hou pe o ming cos ly
da abase scans. This allows e ical algo i hms (CM_SPADE, CM-
SPAM) o pe o m be e on da ase s ha ing dense o long se-
quences han algo i hms using he ho izon al o ma , and o ha e
excellen o e all pe o mance (Fou nie -Vige e al., 2014).
Al hough SPM me hods a e sui able o ou p oblem, we will
see ha hey do no p o ide use ul esul s om a clinical poin o
iew. Indeed, SPM me hods can p o ide mos equen sub-
sequences in a da ase , and subsequences do no equi e con-
ingui y o elemen s o occu . So, his seems o be a sui able a-
mewo k o model he sligh a ia ions o he pa e ns equi ed in
ou p oblem. Howe e , he complexi y o he solu ions space
p o ided by his kind o me hod seems o be highe han he one
in he o iginal da ase and his seems o inc ease complexi y in-
s ead o p o iding a be e unde s anding o he unde lying
s uc u e o he p oblem as i will be seen in he applica ion below.
2.4. Mo i disco e y in sequen ial da a
A mo i is a sho dis inc i e sequence pa e n sha ed by a
numbe o ela ed sequences. The dis inc i eness o a mo i is
mainly eﬂec ed in he o e ep esen a ion o he mo i pa e n a
ce ain loca ions in he ela ed sequences and he unde e-
p esen a ion elsewhe e.
One o he ea ly o igins o mo i disco e y in he con ex o
DNA analysis is he Ko n algo i hm (Ko n e al., 1977). Especially
ele an o gene ac i i ies a e egula o y elemen s bound by p o-
eins such as TFs iden iﬁca ion (D’Haeselee , 2006). Because a
single p o ein o en ecognizes a a ie y o simila sequences,
mo i s a e subjec o some deg ee o sequence a ia ion a each
mo i posi ion wi hou losing hei unc ion.
Mo e han a hund ed me hods (Kleppe and D abløs, 2010)
ha e been p oposed o mo i disco e y in ecen yea s, e-
p esen ing a la ge a ia ion wi h espec o bo h algo i hmic ap-
p oaches as well as he unde lying models o egula o y egions.
Among hem, MEME (Mul iple Expec a ion-Maximiza ion o
Mo i Elici a ion) is one o he bes es ablished mo i ﬁnding ools,
quick and accu a e enough and wi h sui able implemen a ions
a ailable (Das and Dai, 2007;Bailey and Elkan, 1995). MEME
sea ches mo i s by pe o ming Expec a ion Maximiza ion (EM) on
a mo i model o a ﬁxed wid h and using an ini ial es ima e o he
numbe o si es.
Few applica ions a e ound wi h mo i disco e y o ele an
pa e ns in non gene ic sequences. They ha e been ecen ly ap-
plied o acous ic analysis (Bu ed, 2012), whe e sounds a e ﬁ s
ans o med in o a sequence o disc e e s a es, and hese subjec ed
o he MEME algo i hm o mo i disco e y, sea ching o e-
pe i i e pa e ns The ela ionship be ween biological sequences
and mobili y mining has also been explo ed (Jawad e al., 2011),
sea ching o pa e ns in a ﬁc sequence da a. Speciﬁcally in he
medical ﬁeld, mo i s sea ch ha e been applied o ﬁnd p ecu so s o
acu e clinical e en s ega ding elec oca diog aphic ac i i y (Syed
e al., 2010).
Howe e , o he bes o ou knowledge, no wo k applying
mo i s o he iden iﬁca ion o pa e ns in CR ea men s has been
conduc ed. In his wo k, a no el applica ion o mo i disco e y
echniques is p oposed o ﬁnd pa e ns o CR in a con ex qui e a
om genomic da ase s and sea ch o DNA sequences ha a e
conse ed ac oss genomes, o which mo i disco e y echniques
we e o iginally designed. As al eady s a ed, ou p oposal is based
on he ac ha cogni i e ehabili a ion sha es some s uc u al
componen s wi h he gene beha io , which makes mo i s use ul in
gene ics. Mo i disco e y me hods a e also in oduced as a piece o
pa icula ly complex me hodology ha combines wi h o he AI
ools as p esen ed in he nex sec ion.
3. Me hodology
In his sec ion, he o mula ion o he p oblem and he me h-
odological p oposal a e p esen ed. Sec ion 3.1 p o ides a gene ic
o mula ion o he p oblem which is no es ic ed o he pa i-
cula applica ion p esen ed in his wo k.
3.1. 3.1. P oblem o mula ion
Gi en

{}
=…Iiia se o indi iduals
n1

T¼{T
s
s¼1:T } a se o ac i i ies (o asks) ha can be execu ed
by any indi iduals.

(
¼{A
1
,A
2
,…., A
a
} se o a eas impac ed by each ask om T.

:T-
(
a unc ion ha ela es an ac i i y wi h i s a ea o
impac : (T
s
)¼A
j,
s¼1:
;
,j¼1:a,being A
j
he a ea o impac o ac i i y T
s
.
Gi en a scena io in which each indi idual iexecu es a sequence
o
i
ac i i ies, one a a ime
= 1. .
i
.
Gi en i, he ma ix
R
i
p o ides he lis o all his execu ions
( uns):
=(
)
⎡
⎣⎤
⎦
RiT ,,
i ,3
i
.
Ma ix R ep esen s he o al se o ac i i ies execu ed by all
indi iduals
=⋮
⋮
[]ρ()
⎡
⎣
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎡
⎣⎤
⎦
⎡
⎣⎤
⎦
⎡
⎣⎤
⎦
⎤
⎦
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
R
R
R
R
R
i
n
1
2
,3
being
ρ
=∑=
i
n
1i
he o al numbe o ac i i ies’execu ions
pe o med by all indi iduals.
On he o he hand, he sequence o ac i i ies execu ed by an
indi idual ion ime
= 1. .
i
is
=( … …
s
TTT,,,,
i i i i 1
i
). In ac ,
si
is
=[]
s
R2
ii
T
. The longes sequence ha ing leng h
=
=
Mmax
in 1..
i
.
χ=
(
)
⎡
⎣⎤
⎦
T
i nM,
wi h
{}
{}
==in 1. . , 1. . ,
i
is a ma ix whe e
each ow indica es he sequence o ac i i ies pe o med by in-
di idual i. No e ha his migh no be a ec angula able, as each
ow has leng h
≤∀= M i n,1..
i
.
=(
)
A T
i i
is he a ea o impac o ac i i y
T
i
execu ed by in-
di idual iin ime .
=( … …
s
AAA,,,,
ii
i
a1
i
) is he sequence o a eas impac ed by he
ac i i ies execu ed by indi idual iin he pe iod [1,
i
], being
∈
(
Ai
∀= 1. .
i
.
χ=
(
)
⎡
⎣⎤
⎦
A
i nM
a
,
wi h
{}
{}
==in 1. . , 1. . ,
i
is a ma ix whe e
each ow indica es he a eas o impac o he sequence o ac i i ies
pe o med by indi idual i.
Y
1
….Y
a
a se o nume ical indica o s o pe o mance o in-
di iduals in each a ea o impac .
Y
j
measu es he global pe o mance ob ained om indi idual i
in he A ea o impac .
∈(Aj=
j
a1. .
a a ce ain ime poin .
=( …
)
EY Y
a010 0
e alua es he pe o mance le els o indi iduals in
he di e en a eas o impac be o e execu ing hei sequence o
ac i i ies.
=( … )EY Y
a 1
e alua es he pe o mance le els o indi iduals in
I, in he a eas o impac in
(
a e execu ing hei co esponding
sequence o ac i i ies desc ibed in
χ
.
=−DY Y
jj j
0
e alua es he e ec o
χ
in he pe o mance le els o
ac i i y A
j
. No e ha a global e ec o he whole sequence is
measu ed, aking in o accoun ha se e al ac i i ies in he se-
quence migh impac on he same a ea. Ideally
Y
j
will be an im-
plici o explici unc ion o all hose ac i i ies impac ing A
j
in-
dependen ly o hei posi ion in he pa icula sequence, due o
he cumula i e e ec o ac i i ies discussed abo e. Assuming ha
0 indica es bes pe o mance,
>D
0
j
indica es imp o emen ,
≤D
0
j
indica es non-imp o emen . Depending on he pa icula appli-
ca ion, o he seman ics migh also be assigned o he alues o he
pe o mance indica o s as well, and his will equi e e-
in e p e a ion o alues o he
D
j
a iables acco dingly.
Δ
¼(
D
1
……
D
a
) p o ides he e ec o
χ
on each a ea o impac .
X¼(X
1
…X
K
)addi ional in o ma ion abou indi iduals X
K
migh
be ei he nume ical o quali a i e.
Being
)
: Boolean exp ession build o e
χ
a
,
3
: Label; KB¼{ :
)
-
3
} is a Knowledge base composed o a se o ules pa ially
exp essing he a p io i knowledge in he domain. I is impo an o
no e he e ha no assump ion o comple eness is imposed o e KB.
E en ually, a bina y a iable Z migh be a ailable o model
assessmen , indica ing he success o an indi idual pe o ming
hei sequence o ac i i ies unde a ce ain c i e ion o pe o -
mance,
=
⎪
⎪
⎧
⎨
⎩
Z
YES success ul pe o mance
NO unsuccess ul pe o mance
,
,
E en ually Z migh be a mul idimensional ec o and each
componen migh be a unc ion o some
Δ
componen .

Unde all hese p emises, i is desi able o ﬁnd:

a se o pa e ns
4
desc ibing he beha io o he indi iduals
when execu ing ac i i ies
4
¼
μμ μ
{
…},,
m12
;
∀μ∈
4
μ
is a se-
quence o impac a eas o a iable leng h (always lowe han
M ). Thus, each pa e n
μ
is exp essed as:
μ
=( …
μ
aa a,,,
n12
) wi h
∈a
l
(
l: 1.
μ
n
.
Such ha
1.
∀μ μ
′∈,
4
μμ≠
′
:
2.
∀∈ ∃μ∈iI,
4
:
μ
is a subsequence o
si
3.
μ∀′∈
4
,
μ
μ
′≠
μ
,
is a no subsequence o
si
4. Thus,
4
inducing a pa i ion P o e I. Being P¼{
…
μμ
II.
m1
},
={
μ
μ
Ii:
is a subsequence o
}
s
i

The ela ionship among he pa e ns in
4
and he imp o e-
men s in global and/o indi idual a eas o impac in
(
due o
execu ion o ac i i ies in Tand he cha ac e is ics o indi iduals
associa ed wi h he pa e n (associa ions be ween
4
and X)This
means ﬁnding associa ions be ween
4
and Z. In pa icula , gi en
a h eshold
γ
he subse
5
D
4
:
∀μ∈
N
()
|
γ
=≥
μ
P ob Z I YES
1
is
sea ched.
3.2. The SAIMAP me hodology
The SAIMAP (Sequence o Ac i i ies Imp o ing Mul i-A ea
Pe o mance) me hodology is ou p oposal o sol e he p oblem
desc ibed in he p e ious sec ion.
Gi en he R ma ix, SAIMAP consis s o he ollowing s eps:
1. P ep ocessing
1. Build
{}
∀= … =( [])
s
insR1, , as 2
ii
i
;
si
con ains he sequences
o asks done by i
2. Build
χ
=⋮
⎡
⎣
⎢
⎢
⎤
⎦
⎥
⎥
s
sn
1
as he ma ix o he sequence o asks pe -
o med by each indi idual
3. Iden i y he equency h eshold o e ain a ask
4. Reca ego ize Tby using a new ca ego y OTHERS g ouping all
in equen asks
5. De e mine l he h eshold ask leng h o be conside ed
(pe cen ile -95 o leng h o ea men s dis ibu ion).
Use only ﬁ s lcolumns o
χ
o he whole s udy and comple e
sho e sequences by “NULL” alues
6. Build
Δ
¼(
D
1
……
D
a
) e ec o
χ
o e each a ea o impac
Z as a unc ion o a subse o
Δ
2. Desc ip i e analysis
1.Build equency plo o ﬁ s lcolumns and asks o
χ
2.Build hea map o ﬁ s lcolumns and asks o
χ
3.Build hea map o ﬁ s lcolumns and asks o
χ
a
3. P io expe knowledge acquisi ion:Knowledgeis e-
p esen ed by means o l -Then ules in o de o p o ide maximum
ﬂexibili y and exp essi eness o he expe . Only a ailable knowledge
is collec ed e en i i is a pa ial desc ip ion o he domain.
Build KB¼{ :
)
-
3
} om a p io i expe knowledge in a ge
domain
4. Clus e ing o ma ix
χ
: The idea is o ob ain s anda d pa -
e ns o sequences in e ms o he a eas impac ed by he asks
pe o med by he use s. The me hodology migh accep any clus-
e ing me hod, bu in ou app oach Clus e ing based on ules (Gi-
be and Zonicki, 1999) is s ongly ecommended, as i will be
jus iﬁed in he sec ion below.
Le P¼
{}
ξ
C..C
1
be he se o classes ound, P being a pa i ion o
I
(∀ ∈ ⊆
)
CPC,I
5. Spli o
χ
pe class: Di ide he da a ma ix in o subma ices
acco ding o he di e en classes ound in p e ious s ep.
∀∈ χ=χ|=
⎡
⎣⎤
⎦
CP build C A
C
aai nM
c
, wi h i∈{i|i={1,...,n} and i ∈C},
i¼{i:i:1.n and i
∈
C}, n
c
¼ca d{C}.
χ
C
a
con ains only he ows co esponding o indi iduals in class
C.
6. Visualiza ion o classes:
∀∈CP
build a hea map o
χ
C
a
7. Find mo i s pe class
1. Deﬁne an alphabe
ζ
o asingle le e s associa ed wi h he
a eas o impac in
(
such ha
∀∈A
(
is ep esen ed by w
∈
ζ
2.
∀∈ χ
ζ
CP build C
by eplacing he ac i i ies in
χ
C
a
by hei
co esponding ini ial in
ζ
3.
∀∈ χ
ζ
CP ind mo i s o
C
o leng h l (o he me hods can be
used as well bu MEME me hod is ecommended)
Le
{
}
=…eee
lC1
lcM
l
be he ec o wi h he E- alues o all mo i s
ound
∀
mo i
m,
C
l∈∈
⎡
⎣
⎤
⎦
l
C
P, l , l
min max
Le
π
C
l
be he le e p obabili y ma ix indica ing he p esence o
he le e o alphabe in he posi ion o he mo i .
E en ually l migh ange in a ce ain in e al
⎡
⎣
⎤
⎦
l,l
min max
8. De e mine a le el o minimum quali y o mo i s
α(
)
Usually
α
¼0.05 is conside ed bu o he alues can be con-
side ed as well
9. P uning mo i s: e ain he mo e equen mo i s o
in e p e a ion
1.
∀∈C
P
build
{
}
|α
*=≤Mm inMe
CCc
m
10. Visualize mo i s
1.
∀C isualize M
C
on he basis o
π
C
l
using he
S
eq
Log
o
e-
p esen a ion, and in e p e he mo i s.
2. The cha ac e is ics o he sequences associa ed wi h each
class migh be easily iden iﬁed ia he mo i s isualiza ion.
3. Desc ibe which a eas o impac a e add essed a which
poin s o he sequences in each class
11. Analyze he e ec o execu ing ac i i ies o e he di -
e en a eas o impac
1.Build mul iple boxplo o
Dj
s P,
∀∈D
j
Δ
2. K uskal-Wallis be ween
Dj
and P
Iden i y which a eas imp o e he mos in which classes.
12. P ojec all o he illus a i e a iables o e he clus e s:
1.
∀
X
k
in X
I X
k
is nume ical
Build he mul iple boxplo X
k
s P;
Pe o m K uskal-Wallis es
I X
k
is quali a i e
Build he S acked Ba cha o X
k
s P;
Pe o m
χ
2
es
2. Re ain all signiﬁcan a iables in Xand build he desc ip ion
o addi ional cha ac e is ics o each clus e
13. Build ﬁnal in e p e a ion: Associa e he desc ip ions o
mo i s wi h he p oﬁle o pe o mance and he cha ac e is ics o
he indi iduals in each class, and cons i u e he ﬁnal cha ac e -
iza ion o P.
Al hough SAIMAP is a ailable o any clus e ing o mo i dis-
co e y me hod, his wo k p oposes a pa icula implemen a ion
using Clus e ing Based on Rules (ClBR) and MEME me hod. A b ie
desc ip ion o his me hod is p o ided below, oge he wi h he
speciﬁc app oach p oposed o pa e n in e p e a ion.
3.2.1. Clus e ing phase (ClBR)
Clus e ing Based on Rules (ClBR) combines induc i e lea ning
elemen s wi h s a is ical me hods o enhance clus e ing esul s
(Gibe e al., 1998). In ou p e ious esea ch ClBR was applied o
knowledge disco e y on he esponse o neu o ehabili a ion
ea men o TBI pa ien s whe e CR asks ha e no been conside ed
(Gibe e al., 2008). The main idea o ClBR is o allow he use o
in oduce seman ic cons ain s on he o ma ion o clus e s
(classes), p o iding hem in a decla a i e way. This condi ion im-
posed by expe s o malizes he ap io i domain knowledge and
induces a so o supe -s uc u e on he domain; clus e ing is
pe o med wi hin his s uc u e and p o iding clus e s is easie o
be in e p e ed han adi ional algo i hms. In he p esen analysis
ClBR is applied o sequen ial da a o iden i y meaning ul classes.
P io domain knowledge is conside ed, like he leng h o he
p esc ibed ea men .
3.2.2. Mo i disco e y (MEME)
The esul ing clus e s a e hen ea ed wi h he MEME algo-
i hm o mo i disco e y. MEME akes as inpu a g oup o se-
quences and he leng h o he sea ched mo i and ou pu s as many
mo i s o he g oup as eques ed by he use . MEME hen calcu-
la es he E- alues o he mo i s and anks hem by dec easing
E- alues (es ima e o he numbe o mo i s, wi h he same wid h
and numbe o occu ences, ha ing equal o highe log-likelihood
a io; accep ed h eshold is 0.005 (Bailey and Elkan, 1995). The
posi ion-speciﬁc p obabili y ma ix (PSPM) is also p o ided, e-
p esen ing he impo ance o each le e in each posi ion o he
mo i . The PSPM ma ix is inpu in o he sequence logos (Schnei-
de and S ephens, 1990)(SEQ_LOGOS ool), p o iding he g aphical
ep esen a ion o he disco e ed mo i . The mos ep esen a i e
mo i o each o he classes is ob ained oge he wi h i s logo by
using di e en mo i leng hs.
3.2.3. Pa e n in e p e a ion
The logos summa izes he cha ac e is ics o ea men s ol-
lowed in each class and a e used o unde s and egula i ies in he
ea men s o di e en classes. Then he ela ionships be ween
hose ypical ea men s and e alua ions o pa ien pe o mance
migh be analyzed. In ou applica ion, pe o mance is e alua ed
h ough s anda dized neu opsychological assessmen ba e y
(NAB) and e ec o ea men migh be compu ed as p e-pos
di e ences o e hese ba e ies.
S a is ical es s and mul iple boxplo s (Tukey, 1977) a e used o
ela e he disco e ed g oups wi h pa ien cha ac e is ics, le el o
impai men and associa ed wi h speciﬁc ea men pa e ns. The
p oposal includes ANOVA o K uskal-Wallis es (deno ed as K_W
in he p oposed algo i hm) o nume ical a iables depending on
he cha ac e is ics o he a iable i sel and
χ
2 independence es
(Tukey, 1977) o wo- ailo ed exac Fishe es (Ag es i and
F anklin, 2012).
4. Applica ion o a eal case
4.1. E ec s o cogni i e ehabili a ion on auma ic b ain inju y
pa ien s
This sec ion p esen s he clinical con ex o applica ion: The
Neu opsychology Depa men o he Acqui ed B ain Inju y Uni a
Ins i u Gu mann Neu o ehabili a ion hospi al (IG) whe e TBI
pa ien s unde go CR ea men s.
The In o ma ion Technology amewo k o CR ea men s in his
clinical se ing is he PREVIRNEC© pla o m (To mos e al., 2009). A
J2EE clien -se e a chi ec u e speciﬁcally designed and de eloped
o manage CR plans assigned by he apis s o pa ien s, as well as
ollow-up in o ma ion abou he p ocess (i.e. CR session da es, ask
execu ions in each session, pe o mance, in ol ed he apis s, pa-
ien s, ask esul s, ask ime, de ailed in Sec ion 2.1.1).
The e a e h ee main cogni i e unc ions o be ehabili a ed in a
CR p og am (Sohlbe g and Ma ee , 2001): a en ion, memo y and
execu i e unc ions; all o hem can p o oundly a ec an in-
di idual’s daily unc ioning. E en mild changes in he abili y o
a end, p ocess, ecall and ac upon in o ma ion can ha e sig-
niﬁcan e ec s in he quali y o li e o he pa ien . Conside he
cogni i e skills equi ed o success ul meal p epa a ion as an
example: he indi idual mus plan a menu, iden i y equi ed in-
g edien s, de elop a shopping lis o equi ed i ems and schedule
su ﬁcien ime o shopping and p epa ing he meal; hen he
indi idual mus sequence many ood p epa a ion ac i i ies in an
o ganized way so ha e e y hing is eady a dinne ime. E en a
mild a en ion o execu i e unc ion deﬁci can ende his di ﬁ-
cul , ine ec i e o e en impossible.
The main hypo hesis aming ou p oposal is:
1) Some CR ehabili a ion asks a e designed o imp o e pa icula
cogni i e unc ions, al hough a en ion, memo y and execu i e
unc ions a e ela ed and in e dependen (Sohlbe g and Ma ee ,
2001). Thei close in e dependence s ems om bo h a unc-
ional associa ion and hei sha ed neu oci cui y. This means
ha pe o ming a ask a ge ing memo y can also ha e colla -
e al e ec s on o he cogni i e unc ions like a en ion o
execu i e unc ions.
2) The addi ional e ec o a single ask migh be a ec ed by he
cumula ed e ec o he sequence o p e ious asks execu ed
unde he ea men , his migh de e mine ha o de o ex-
ecu ion is ele an in he ea men ou come.
4.2. Cogni i e ehabili a ion asks
Fo each pa ien he he apis c ea es a speciﬁc CR ea men
i.e. an o de ed sequence o asks. A IG a ypical CR p og am in he
PREVIRNEC© pla o m anges om 2–4 sessions/weeks o 2–5
mon hs, wi h no cons ain s on ask o de , he e o e leading o
di e en ask sequences in a di e en o de om pa ien o
pa ien .
A he ime o his analysis he PREVIRNEC© pla o m suppo s
96 di e en CR asks a ge ing he h ee main cogni i e unc ions
men ioned abo e (17 conce n a en ion, 59 memo y, and 20 ex-
ecu i e unc ions). Each ask is deﬁned by some pa ame e s ha
de e mine i s le el o di ﬁcul y. The he apis c ea es a CR ea -
men as a se o sessions in he PREVIRNEC© pla o m, each one
consis ing o a ce ain sequence o compu e ized asks o a ce ain
day. Fo each ask, he he apis conﬁgu es he sui able combina-
ion o inpu pa ame e s, including he one o he au oma ic ad-
jus men o he di ﬁcul y le el in oduced abo e. This dynamic
adjus men o he di ﬁcul y le el is pe o med (i necessa y) wice
o each ask, meaning ha i he pa ien expe iences a ask ha is
oo di ﬁcul o oo easy in he ﬁ s execu ion, PREVIRNEC© au o-
ma ically e-gene a es he ask wi h an adjus ed di ﬁcul y le el.
The he apis designs he sequences o e e y pa ien based on he
he apis ’s expe ise and he sequence o asks assigned o e e y
pa ien may ha e a iable leng h. This wo k aims o gene a e some
guidelines ha can help he he apis in his design.
4.3. Assessmen o he e ec o he ea men
Be o e s a ing he CR p og am e e y pa ien unde goes a
Neu opsychological Assessmen Ba e y (NAB). This ba e y in-
cludes 28 i ems co e ing he majo cogni i e domains (a en ion,
memo y and execu i e unc ions) measu ed using s anda dized
cogni i e es s. NAB consis s o a selec ion o some i ems om
se en assessmen ins umen s, associa ed wi h he di e en
cogni i e unc ions, which in u n a e e alua ed unde some
speciﬁc sub- unc ions. Being awa e ha con en ional neu-
opsychological ins umen s a e no o ious o amalgama ing
cogni i e ope a ions (Jaga oo, 2009;Sabb e al., 2009), a subse
o NAB i ems wi h highes le els o speciﬁci y has been selec ed
in collabo a ion wi h domain expe s o he p oposed app oach.
The ﬁnal i ems conside ed in his wo k a e he ollowing 14 non
edundan i ems:

Memo y:
○Visual and Ve bal Memo y: The Rey Audi o y Ve bal Lea ning
Tes ] (Rey,1964) (RAV075, RAV015 and RAV015R i ems)
●A en ion:
○Sus ained A en ion: *Con inuous Pe o mance Task Tes (Con-
ne s, 2002) (OMI, COMI and CPT i ems) and *T ail Making Tes -A
(Rei an and Wol son, 1993) (TMTA i em)
○Selec i e A en ion: he WAIS-III Selec i e A en ion (Wechsle ,
1999) (VWAIS i em)
○Di ided A en ion: he T ail Making Tes -B (Rei an and Wol son,
1993) (TMTB i em).

Execu i e Func ions:
○Planiﬁca ion: he WAIS-III Visuo Cons uc ion (Wechsle , 1999)
(CUBES i em)
○Inhibi ion: he S oop Tes (Golden, 1994) (INTER i em)
○Flexibili y : * he Wisconsin Ca d So ing Tes (Hea on e al.,
1997) (TERR i em) and * he Le e Fluency Tes (A iola i Fo uny
e al., 1999) (PMR i em)
○Ca ego iza ion: The Wisconsin Ca d So ing Tes (Hea on e al.,
1997) (CAT i em)
All NAB i ems a e no malized o a 0 o 4 scale (whe e 0¼No
a ec a ion, 1¼mild a ec a ion, 2¼mode a e a ec a ion,
3¼se e e a ec a ion and 4¼acu e a ec a ion).
A e his ini ial e alua ion pa ien s s a PREVIRNEC© sessions
( o 2 o 5 mon hs, depending on he pa ien ) and a e ea men
e e y pa ien unde goes he same NAB o e alua e he cogni i e
ou come s a us.
In o ma ion ob ained in he NAB be o e and a e ea men is
he sou ce o unde s and pa ien imp o emen , and, in con-
sequence, he esponse le el o he ea men i sel . Measu ing
global imp o emen in a speciﬁc cogni i e unc ion (e.g. A en ion)
implies s udying esponse o ea men in each o he NAB es s’
sub unc ions (e.g. Sus ained, Selec i e, Di ided A en ion). Di e -
en c i e ia can be adop ed (sub unc ions’a e age, maximum
di e ence, e c). To he bes o ou knowledge no s anda dized
app oach is uni e sally accep ed in he clinical CR he apis s
communi y o de e mine he imp o emen o he pa ien om a
sys ema ic poin o iew.
This wo k ies o con ibu e o his issue by b eaking he
p oblem down in o se e al s eps. As a ﬁ s app oach we will in-
i ially ocus on he iden iﬁca ion o CR pa e ns ( h ough clus e ing
and mo i disco e y), ﬂexible enough o ca ch he mos e ec i e
sequences o asks, e en i in e up ed by o he s. Once he pa -
e ns ha e been iden iﬁed, he dominan e ec o he ea men
associa ed wi h each pa e n will be analyzed. Wi h his in-
o ma ion, well- ounded imp o emen c i e ia will be deﬁned a a
la e s age (Fig. 1).
4.4. The da ase
One hund ed and wen y- h ee TBI adul s ollowing a 3–5
mon hs CR ea men a IG Neu opsychological Rehabili a ion Uni
a e analyzed in his s udy. Fo e e y pa ien he demog aphic and
clinical a iables conside ed a e: age, gende , educa ional le el,
Glasgow Comma Scale (GCS) and Pos T auma ic Amnesia (PTA)
du a ion. Table 1 shows he basic s a is ics o nume ical a iables
while equency dis ibu ion o quali a i e ones a e shown in
Table 2.
Ini ial assessmen o he TBI se e i y is epo ed acco ding o
GCS le els. A GCS sco e o eigh o less a e esusci a ion om he
ini ial inju y is classiﬁed as a se e e b ain inju y. The GCS sco e o
a mode a e b ain inju y anges be ween 9 and 13 and a sco e o 13
o g ea e indica es a mild b ain inju y, o concussion. As de ailed
in Fig. 2 mos GCS sco es (86,17%) show se e e b ain inju y le el
(mean alue 6,4573.15). I is known ha hose whose leng h o
PTA is less han wo mon hs ha e a e y good chance o a leas
being able o li e on hei own (e en i hey a e unable o e u n o
wo k o school). On he o he hand, pa ien s whose leng h o PTA
is longe han h ee mon hs a e unlikely o be able o e u n o
wo k o school (al hough hey migh be able o li e on hei own).
As N* shows in Table 1, PTA measu es we e no a ailable o 67% o
he pa icipan s; conside ed alues show e y se e e condi ions as
indica ed by he median (103), which is mo e eliable han he
mean because o he ou lie isualized in Fig. 2 ( igh ).
Demog aphic quali a i e (Table 2) indica es 91 men (73.98%)
and 32 women (26.02%) pa icipa ing in he analysis. Each pa i-
cipan ’s educa ional backg ound is ca ego ized in h ee g oups,
wi h Elemen a y school p edominan .
All pa icipan s signed in o med consen o he neu opsycho-
logical p ocedu e, which was app o ed by IG’s E hical Commi ee.
All me c i e ia o ini ia e IG neu opsychological ehabili a ion
ea men .
A e NAB ini ial e alua ion all pa ien s ini ia ed a h ee o ﬁ e
mon hs’p og am (No embe 2007 o No embe 2009) based on
pe sonalized in e en ions in he PREVIRNEC© pla o m whe e
pa ien s wo ked in each o he speciﬁc cogni i e domains, con-
side ing he deg ee o he deﬁci and he esidual unc ional ca-
paci y. All pa ien s we e adminis e ed he same NAB neu-
opsychological assessmen a he end o he ehabili a ion
Fig. 1. Rep esen a ion o indi iduals execu ing sequences o asks impac ing a eas
ha a e e alua ed be o e and a e he pe iod o execu ions.
Table 1
Basic desc ip i e s a is ics o nume ical a iables.
Va iable N N* Mean S d De Min Q1 Median Q3 Max
AGE 123 0 36.56 6.50 18 25 32 40 68
GCS 89 34 6.45 3.15 0 4 6.5 40 14
PTA 40 83 131.6 140.5 34 79 103 136 947
Table 2
Basic desc ip i e s a is ics o gende and educa ional le el.
GENDER Coun Pe cen age EDU Coun Pe cen age
Female 32 26.02 Elemen a y 60 48.78
Male 91 73.98 In e media e 40 32.52
High 23 18.70
p og am. A o al o 39412 ask execu ions we e ini ially included in
his analysis, in ol ing he 96 di e en CR asks included in he
PREVIRNEC© pla o m.
4.5. S uc u e o da abase
O iginally, he sys em eco ds he execu ion o e e y ask as a
single ow in a log ﬁle, which addi ionally eco ds he ollowing
in o ma ion:
Da e is he da e o he execu ion o he T
s
ask (da e yyyymmdd)
TaskName is a desc ip i e name assigned o iden i y he ask T
s
Sco e is he esul ob ained in ha execu ion (0 o 100 eal
numbe )
NumTask is he au oma ic ask gene a ion numbe assigned o
he ask (0,1,2)
Di ﬁcul y is he di ﬁcul y le el o he ask (0,1,2,3,4)
Func ion: cogni i e unc ion add essed by he ask: A en ion,
Memo y, Execu i e unc.
Sub unc ion is he speciﬁc cogni i e sub unc ion add essed by
he ask (as desc ibed below, o he A en ion unc ion he
add essed sub unc ions a e isual a en ion, sus ained a en-
ion, selec i e a en ion, e c).
O iginal da a s uc u e (S1):
⎡
⎣
⎢
⎢
⎢
⎤
⎦
⎥
⎥
⎥
i T Da e TaskName Sco e NumTask Di icul y Func ion Sub unc ion
.. . . . . . . .
.. . . . . . .
.. . . . . . . .
.. . . . . . . .
4.6. Ins an ia ion o he o mal p oblem
The p esen ed da ase app oaches he o mal p oblem p esen ed in
Sec ion 2 as a pa icula case whe e a CR ea men is he scena io in
whicheachpa ien iexecu es a sequence o ac i i ies, one a a ime

I
is he se o TBI pa ien s unde going CR ea men a IG

T¼{T
s
s¼1:
;
} is a se o 96 CR asks ha pa ien s execu e
du ing ea men :
T¼{GlobalLocal, Ma hMazeComp, Ma hMazeExe , ConcOps, Sub-
ma ine,Ma ching,BagO Coins,Di e ences,Figu es,PuzzComp,Puz-
zExe , Le e Soup,Bingo, Di Di ec ion,S aigh Line, SameDi ec ion,
G oupWo ds, Ca ego iza ionTwo, Ca ego iza ionTh ee, SameCa -
Wo ds, Ci cle, Pla o ms, Zigu a , GoNoGoEs , GoNoGoGame, Go-
NoGoPos, Hanging, SinkFlee , Maze, Fou InRow, Fou h, JigSaw,
BuildSen ence, F agmen s, Se ie, CyclicSe ie, SameCa , TempO de ,
Posi ion, Sequen ial, Simoul aneous, Wo dSeqDec, Wo dSeqSel,
Wo dSeqDi Ca , Wo dSeqSameCa , Wo dSimDec, Wo dSimSel,
Wo dSimDi Ca , Wo dSimSameCa , Wo dTempO de , Pai sSeqDec,
Pai sSeqRel, Pai sSeqSel, Pai sSeqSameO de , Pai sSeqRandO de ,
Pai sSimDec, Pai sSimRel, Pai sSimSel, Pai sSimSameO de , Pai s-
SimRandO de , Sen SecO de , Sen SecTes , Sen SecW i e, Sen Sec-
Ques ion, Sen SecT ueFalse, Sen SimO de , Sen SimTes , Sen Sim-
W i e, Sen SimQues ion, Sen SimT ueFalse, RecSeqNumbe s,
RecSimNumbe s, RemSecNumbe s, RemSimNumbe s, Tex So ,
Tex Ques ion, Tex W i e, Tex T ueFalse, ImgWo dTempO de , Im-
gWo dSeqDecide, ImgWo dSeqRel, ImgWo dSeqSel, Im-
gWo dSeqSameO de , ImgWo dSeqRandO de , ImgWo dSimDecide,
ImgWo dSimRel, ImgWo dSimSel, ImgWo dSimSameO de , Im-
gWo dSimRandO de , D awTempo alO de , D awRecogni ion, Sce-
neRecogni ion, SceneRecall, VisualMemo y, VisualSimon}

(
is he se o a eas o impac . In his pa icula case i ma ches
wi h he amily o cogni i e unc ions a ge ed by he CR asks.
(
¼{A en ion, Memo y, Execu i e unc ions} a e he main
cogni i e unc ions in ol ed in daily ac i i ies (Sohlbe g and
Ma ee , 2001); hese a e he e o e ea ed in PREVIRNEC oo.

(T)¼a p o ides he main cogni i e unc ion a,(a
∈
(
) a ge ed
by ask T.

Gi en a pa ien i, he ma ix
R
i
p o ides he lis o all asks
execu ed by he pa ien iwi h i s co esponding execu ion
imes h oughou his CR ea men .

Row i o ma ix
χ
gi es he sequence o CR asks done by i
du ing ea men .

The se
Y
j
, ¼1.14 o nume ical indica o s o pe o mance is, in
his wo k, he selec ion o 14 ele an and non edundan i ems
om NAB, used in IG o e alua ing he deg ee o impai men o
each cogni i e unc ion.

D
j
, is he di e ence be ween he sco es ob ained by he pa ien
be o e and a e he p esc ibed CR ea men in he co e-
sponding NAB i em.

Δ
¼(
D
1
……
D
a
) ep esen s he e ec o CR ea men in all
cogni i e unc ions.

X¼(X
1
…X
K
)addi ional in o ma ion abou pa ien s. X
K
migh be
ei he nume ical (like age o GCS) o quali a i e (like Sex o
Educa ional le el).

Z indica es a global imp o emen o he pa ien a e ea men .
The execu ion o each ask by a pa ien occu s a di e en pe i-
odici ies o e e y pa ien ; he leng h o ea men is a iable in bo h
numbe o ask execu ions and o al ea men ime o he di e en
pa ien s; he sequence o ask execu ions changes om one pa ien
o ano he ; he esul ob ained in an execu ion de e mines bo h he
ask and di ﬁcul y o he nex ask p oposed by he sys em; he e ec
o a ask in e ms o cogni i e unc ions o he pa ien is accumula i e
and he e ec o a ce ain sequence o asks migh no be a ec ed by
small a ia ions o he sequence i sel ,i.e. hein oduc iono small
addi ional asks in in e media e posi ions o he sequence. Fo hese
easons, ou p oblem may be ea ed unde SAIMAP me hodology.
4.7. The Sequence o Ac i i ies Imp o ing Mul i-A ea Pe o mance
(SAIMAP) me hodology
4.7.1. P ep ocessing
As a ﬁ s hypo hesis i is assumed (a e consul ing wi h ex-
pe s) ha he ime in e al (delay) be ween he execu ion o wo
consecu i e asks is i ele an o ehabili a ion pu poses, since
Fig. 2. Nume ical a iables his og ams: Age (le ), Glasgow Comma Scale sco es (cen e ) and Pos T auma ic Amnesia days ( igh ).
unc ion in each class. The ﬁ s in e es ing obse a ion is ha all
g oups imp o e (deﬁci dec eases) a e ea men and he e ec s
a e all below 0 on a e age. The dimension ending o mo e ne-
ga i e alues is a en ion, while memo y seems o be he one wi h
less imp o emen o all g oups. In con as , i can be seen ha
SHORT86 class is he one wi h be e ea men esul s ega ding
a en ion, while i beha es e y simila ly o class SHORT70 in
e ms o memo y and execu i e unc ions. I also appea s ha
class LONG6 is mo e esis an o ea men han o he s, especially
ega ding memo y.
4.7.11. Build ﬁnal in e p e a ion
C ossing he ob ained p oﬁles wi h he mo i s and he e ec s o
he apy i appea s ha :
SHORT70 ep esen s sho e m ea men s, no mo e han 150
ask execu ions mainly o ien ed o execu i e unc ions, p eceded
in some cases by memo y asks, mainly in he ini ial pa o he
sequences. These pe sons show be e esponse o ea men
mainly in a en ion and execu i e unc ions han in memo y,
ha ing an in e media e le el o a en ion imp o emen compa ed
wi h o he classes.
SHORT86 ep esen s in e media e du a ion ea men s, wi h
no mo e han 460 ask execu ions including a highe numbe o
a en ion asks execu ed mainly a he beginning o he sequences.
Pe sons in his class show a highe eco e y in a en ion han in
o he unc ions, being he g oup wi h be e esul s o ea men
ega ding a en ion.
LONG6 ep esen s long e m p og ams including mo e han
460 ask execu ions wi h a highe p opo ion o memo y asks,
o en combined wi h execu i e unc ions asks and possibly some
a en ion asks a he end o he sequences. Howe e he pe sons
in his class a e mo e esis an o ea men han o he classes in
memo y and a en ion.
5. Compa ison wi h adi ional app oaches
Unde a adi ional app oach, one would be emp ed o educe
ou p oblem o building a p edic i e model o he imp o emen
o he pa ien and o sol e i by using some machine lea ning
classiﬁe .
P elimina y analysis and p oblem ep esen a ion p o ided ap-
p op ia e da a s uc u es, da a ans o ma ions, and domain
knowledge o pa e n disco e y. T adi ional classiﬁca ion echni-
ques a e p oposed o s udy esponse o CR ea men .
Ma ix
χ
is used o he classiﬁe wi h a esponse a iable Z
(see Sec ion 2):
=
′
⎪
⎪
⎧
⎨
⎩
Z
YES pa ien imp o ed a e ea men
NO pa ien didn imp o e a e ea men
,
,
Table 8
Mean, s anda d de ia ion and p- alues (K uskal-Wallis) o nume ical a iables pe class.
GCS AGE PTA
Mean S D Media Mean S D Median Mean S D Median
SHORT70 6.27 2.91 6 32.80 8.20 32.8 84.3 36.9 84.3
SHORT86 6.04 2.63 6 31.65 7.99 31.65 156.7 209.7 156.7
LONG6 6.88 3.09 7 35.13 9.57 35.13 117.67 13.65 117.67
KW p- alue 0.667 0.433 0.176
Table 6
E- alues anking o di e en lengh s.
e
C
l
Leng h/Class SHORT86
e
C
14
14 2.8e-055
e
C
1
5
15 5.6e-051
e
C
16
16 2.6e-049
e
C
12
12 4.3e-049
e
C
13
13 2.4e-048
e
C
11
11 3.2e-046
e
C
18
18 3.0e-045
e
C
1
7
17 2.5e-045
e
C
1
0
10 7.7e-039
e
C
2
0
20 1.9e-039
e
C
1
9
19 3.1e-034
e
C
9
92.3e-032
e
C
8
81.8e-028
eC
7
72.7e-026
e
C
6
62.5e-018
Table 7
Tabula ep esen a ion o
π
SHORT8
6
20
ma ix.
ACGT
0.435897 0.000000 0.000000 0.564103
0.769231 0.025641 0.000000 0.205128
0.384615 0.051282 0.000000 0.564103
0.000000 0.000000 0.000000 1.000000
0.000000 0.410256 0.000000 0.589744
0.000000 0.794872 0.000000 0.205128
0.102564 0.128205 0.000000 0.769231
0.102564 0.230769 0.000000 0.666667
0.025641 0.230769 0.000000 0.743590
0.025641 0.179487 0.000000 0.794872
0.000000 0.000000 0.000000 1.000000
0.000000 0.153846 0.000000 0.846154
0.000000 0.256410 0.000000 0.743590
0.025641 0.179487 0.000000 0.794872
0.000000 0.384615 0.000000 0.615385
0.153846 0.358974 0.000000 0.487179
0.256410 0.282051 0.000000 0.461538
0.102564 0.512821 0.000000 0.384615
0.230769 0.461538 0.000000 0.307692
0.307692 0.205128 0.000000 0.487179
Table 9
Ca ego ical a iables numbe o occu ences and p- alues (χ
2
es ) pe class.
GENDER EDU LEVEL
Female Male Elemen In e m. High To al
SHORT70 13 27 19 14 7 40
SHORT86 12 37 27 14 8 49
LONG6 26 4 318
χ
2
p- alue 0.691 0.949

In his wo k adi ional classiﬁca ion algo i hms and some
sequen ial pa e n mining algo i hms ha e been used.
5.1. Pa e n disco e y wi h classiﬁe s
Algo i hms ha exploi ou di e en machine lea ning p in-
ciples ha e been used on ou eal applica ion and compa ed wi h
he p oposed app oach: decision ee lea ning (j48), ins ance-
based lea ning (IBk), p obabilis ic lea ning (Nai e Bayes), and RBF
neu al ne wo ks.
Waika o En i onmen o Knowledge Analysis (WEKA) (Hall
e al. 2009), 3.6.5 was he da a mining pla o m o unning
classiﬁe s. All o hem we e un wi h de aul pa ame e s on a
3.4 GHz Pen ium IV PC wi h 2 GB o RAM. The classiﬁe s un in his
applica ion we e:

J48 is he WEKA implemen a ion o he C4.5 decision ee
(Quinlan, 1993).

Nai e Bayes implemen s he p obabilis ic Nai e Bayes classiﬁe
(John and Langley, 1995).

IBk is he implemen a ion o KNN (Aha and Kible , 1991) he
k-nea es -neighbo classiﬁe ; pa ame e (k se in ou es s o
1,2,3, and 5) se ing he neighbo hood size.

RBFNe wo ks implemen s a popula ype o eed- o wa d ne -
wo k, adial basis unc ion (RBF) ne wo k (Wi en, 2011).
The p edic ion pe o mance o he models was measu ed by
en- old c oss alida ion and se e al pa ame e conﬁgu a ions
we e es ed. In his s udy he da a se was spli in o 9 subse s wi h
12 eco ds and 1 subse wi h 15. Each classiﬁe is ained 10 imes,
each ime using a e sion o he da a in which one o he subse s is
omi ed ( es ing da a). Each ained classiﬁe is hen es ed on he
da a om he subse , which was no used du ing aining. The
esul s a e a e aged o ob ain an o e all accu acy (Table 11).
Ou solu ion p oposes a se o mo i s o be ollowed. The p o-
po ion o pa ien s ollowing he p oposed mo i s who imp o e
a e ea men is 81.33%, which can be used as an equi alen o
accu acy and is no iceably highe han he p edic i e models ob-
ained by all assessed machine lea ning me hods.
5.2. Sequen ial pa e n analysis
As p esen ed in Sec ion 1.1.2. sequen ial pa e n mining (SPM)
echniques a e also sui able o ﬁnd pa e ns o execu ions o CR
asks a ge ing cogni i e unc ions, iden iﬁed pa e ns migh help
o unde s and esponses o ea men . The inpu is ma ix
χ
and
SPM algo i hms we e es ed: CM-SPAM and CM-PREFIXSPAM as
well as CM-SPADE.
Sequen ial Pa e n Mining F amewo k (SPMF) e sion 0.96q
was he da a mining pla o m o he SPM algo i hm execu ions.
(Fou nie -Vige e al., 2014). All o hem we e un wi h de aul
pa ame e s on a 3.4 GHz Pen ium IV compu e wi h 2 GB o RAM.
SPADE and SPAM a e e y e ﬁcien o da ase s ha ing dense o
long sequences and ha e excellen o e all pe o mance, since
pe o ming join ope a ions o calcula e he suppo o candida es
does no equi e scanning he o iginal da abase unlike algo i hms
using he ho izon al o ma . Fo example, he well-known P eﬁx-
Span algo i hm, which uses he ho izon al o ma , pe o ms a
da abase p ojec ion o each i em o each equen sequen ial
pa e n, in he wo s case, which is ex emely cos ly.
CM-SPADE is he SPMF implemen a ion o SPADE algo i hm
(Fou nie -Vige e al., 2014). The suppo o a sequen ial pa e n is
Table 10
NAB selec ed i ems s classes.
RAV075 RAV015 RAV015R
Mean S D Media Mean S D Median Mean S D Median
SHORT70 2.75 1.37 3.00 2.72 1.56 4.00 2.15 1.77 2.00
SHORT86 3.20 1.06 4.00 3.38 1.09 4.00 2.91 1.45 4.00
LONG6 2.37 1.50 3.00 2.25 1.90 3.00 1.87 1.80 1.00
KW p- alue 0.136 0.042 0.056
Fig. 13. Mul iple boxplo s o imp o emen e sus class and cogni i e unc ion. Fo
each cogni i e unc ion, a mul iple boxplo o he co esponding imp o emen
index Δ e sus classes is isualized. E e y boxplo displays he ange be ween he
minimum and maximum alue o each Δ, he box indica es he in e al be ween
ﬁ s and hi d qua ile, ho izon al ma k in box indica es he median.
he numbe o sequences whe e he pa e n occu s di ided by he
o al numbe o sequences in he da abase (see Sec ion 2).
Table 12 shows pa e n ound wi h suppo 4¼0.88. A o al o
31501 pa e ns o leng h 15 a e ound (almos he size o he o i-
ginal da ase ).
5.3. Sequen ial pa e n mining on each class
An analysis local o each class is also pe o med o see i local
analysis p o ides a mo e cons ained se o solu ions. SM-SPADE is
applied on each o he iden iﬁed classes, bu as shown in Table 13
he numbe (N) o iden iﬁed pa e ns in each class does no
dec ease.
5.4. Compa ison
Table 11 shows esul s ob ained wi h by-de aul pa ame e s
and o he conﬁgu a ions, like j48 being es ed wi h h ee di e en
conﬁdence ac o s (0.25, 0.30, 0.40) dec easing pos -p uning, and
a ying he minimum numbe o objec s pe lea . Fo Ibk,
Euclidean dis ance (wi h and wi hou weigh ing) was used and
di e en window sizes we e es ed, also a ying he numbe o
neighbou s (k pa ame e ). Fo abou 80% o he es s he ob ained
pe o mance is below 60% and none o hem eached 65% o ac-
cu acy a e 10- old c oss alida ion. These esul s pe sis i e-
spec i e o he numbe o ea u es (CR ask execu ions) in oduced
in o he di e en models. Table 11 shows esul s o models in-
cluding only ini ial CR sessions (10 o 20 ﬁ s asks o he ea -
men ) up o 600 o 1300 asks. In e media e alues (e.g. 700, 800,
900 CR ask sequences) we e also es ed wi h simila esul s in
pe o mance, which a e no eliable enough o be used in eal
clinical p ac ice. This con as s wi h he 81.33% accu acy ob ained
unde he p oposed app oach.
Rega ding SPM algo i hms, hey show accep able pe o mance
ega ding suppo (e.g. 0.8 o 0.9), e en be e han he one ob-
ained unde ou p oposal in some cases. Bu as de ailed in
Table 13, he numbe o equen sequences disco e ed is g ea e
han he o iginal da ase (i.e. wi h suppo 4¼0.8 CM-SPADE
iden iﬁes 44690 equen sequences o leng h 9; 16853 sequences
o leng h 10 and 2415 o leng h 11), This inc eases p oblem com-
plexi y ins ead o dec easing i , since we o iginally had abou
39000 ask execu ions. SPM algo i hms we e also es ed locally o
classes a e pe o ming a ClBR clus e ing phase, bu he esul s
ob ained we e simila , i.e. sho e sequence clus e s did no de-
c ease he numbe o iden iﬁed equen sequen ial pa e ns,
he e o e did no lead o an easie p ocess o unde s and how o
build new CR plans.
6. Conclusions and u u e wo k
In his wo k a ﬁ s applica ion o mo i disco e y is in eg a ed
in he gene ic SAIMAP me hod used o ﬁnd p omising ea men
pa e ns in cogni i e ehabili a ion. This p o ides u he knowl-
edge o ha ob ained in p e ious analysis whe e isola ed asks
we e analyzed.
The use o mo i s is ele an because he cumula i e e ec o
CR ask execu ion is obus o he ime pe iod in e als occu ing
Table 12
Sequen ial pa e ns iden iﬁed by CM-SPADE o a minsuppo ¼0.88. Fi s column shows he leng h o he pa e ns and N column he numbe o iden iﬁed sequences o
each pa e n leng h. Then mean and median suppo a e shown oge he wi h o he suppo s a is ics.
CM-SPADE
LEN N Media S De Min Max Q1 Mediana Q3
0.88 1 3 0.9675 0.0422 0.9187 0.9919 0.9187 0.9919 0.9919
2 9 0.9431 0.0407 0.9024 0.9919 0.9106 0.9106 0.9837
3 27 0.92442 0.03785 0.88618 0.99187 0.89431 0.90244 0.97561
4 64 0.91527 0.03536 0.88618 0.99187 0.89431 0.90244 0.95325
5 118 0.91264 0.03440 0.88618 0.99187 0.88618 0.89431 0.95325
6 172 0.91591 0.03373 0.88618 0.98374 0.88618 0.89431 0.95122
7 238 0.92150 0.03073 0.88618 0.96748 0.88618 0.93496 0.95122
8 354 0.92609 0.02451 0.88618 0.96748 0.89431 0.93496 0.94309
9 589 0.92715 0.01716 0.88618 0.95935 0.92683 0.93496 0.93496
10 1064 0.92408 0.01164 0.88618 0.95122 0.91870 0.92683 0.93496
11 2060 0.91756 0.00976 0.88618 0.94309 0.91057 0.91870 0.92683
12 4097 0.91035 0.00907 0.88618 0.93496 0.90244 0.91057 0.91870
13 8192 0.90385 0.00834 0.88618 0.93496 0.89431 0.90244 0.91057
14 16339 0.89834 0.00760 0.88618 0.92683 0.89431 0.89431 0.90244
15 31501 0.89391 0.00665 0.88618 0.91870 0.88618 0.89431 0.89431
16 53247 0.89084 0.00549 0.88618 0.91870 0.88618 0.88618 0.89431
17 69573 0.88902 0.00438 0.88618 0.91057 0.88618 0.88618 0.89431
18 64130 0.88794 0.00348 0.88618 0.91057 0.88618 0.88618 0.88618
19 38554 0.88730 0.00281 0.88618 0.90244 0.88618 0.88618 0.88618
20 14341 0.88689 0.00230 0.88618 0.89431 0.88618 0.88618 0.88618
21 3159 0.88651 0.00161 0.88618 0.89431 0.88618 0.88618 0.88618
22 401 0.88618 0.000000 0.88618 0.88618 0.88618 0.88618 0.88618
23 32 0.88618 0.000000 0.88618 0.88618 0.88618 0.88618 0.88618
24 1 0.88618 *0.88618 0.88618 *0.88618 *
Table 11
Accu acy o each classiﬁe a e 10- old c oss alida ion, ﬁ s column shows he
numbe o CR ask execu ions conside ed.
A ibu es C4.5 (J48) Nai e
bayes
KNN (IBk) RBFne wo ks
IB1 IB2 IB3 IB5
10 57.72 54.47 52.84 50.40 53.65 57.62 56.91
20 56.09 52.84 56.09 54.47 58.53 56.91 47.15
30 57.72 58.53 56.91 58.53 64.22 58.53 56.09
40 57.72 56.91 53.65 51.21 56.91 61.78 59.34
50 53.65 58.53 57.72 57.72 62.60 58.53 59.34
60 51.21 57.72 57.72 56.91 57.72 56.09 58.53
70 60.97 57.72 57.72 56.91 60.16 59.34 52.84
80 59.34 59.34 56.09 54.47 62.60 63.41 54.47
100 63.41 55.28 56.09 52.03 60.97 61.78 49.59
600 64.41 61.78 59.34 55.28 60.16 60.16 46.34
1391 60.16 58.53 60.16 56.09 60.16 62.60 46.22
be ween ask execu ion and small in e e ences in a ce ain se-
quence do no dec ease hei ehabili a i e e ec .
In he p oposed me hodology, a p e ious clus e ing p ocess is
pe o med in such a way ha h ee CR p og am p oﬁles a e
iden iﬁed. La e , mo i disco e y local o each p oﬁle is pe o med
o unde s and he s uc u e o he ask sequences associa ed wi h
he classes and i has been seen ha leng h o ea men seems o
be a main class cha ac e is ic. Associa ed wi h leng h, speciﬁc
sequence pa e ns appea and mo i s o each class ha e dis-
inc i e cha ac e is ics, which p o ides he he apis s a ﬁ s con-
cep ual amewo k o compose CR p og ams unde long, sho o
in e media e leng hs.
S a is ical es s seem o indica e ha basic demog aphic and
clinical cha ac e is ics o he pa ien s (GCS, PTA, gende ,
educa ional le el, age) do no show signiﬁcan di e ences s he
classes, hus indica ing ha di e ences among g oups a e due o
he s uc u e o he ea men i sel . Howe e , i has been seen
ha sho ea men s a e associa ed wi h pa ien s wi h mild im-
pai men in sho e m memo y (RAV015 NAB i em), while long
ones a e associa ed wi h pa ien s wi h high impai men on e-
cognizing memo y, his p o iding clea clinical guidelines o he
he apis . I is also in e es ing o no e ha pa ien s in class
SHORT70 ollow ea men s sho e han 150 asks, which, in ac ,
is much less han he cu en p esc ip ions ( his p o iding also
ele an in o ma ion o u u e CR pe sonalized ea men design).
Indeed, cu en ly, he hospi al is p esc ibing a ea men o
3 mon hs o all pa ien s and in e media e e alua ions a e used o
decide ad anced end o ea men , o subsequen p olonga ion o
Table 13
Iden iﬁed sequen ial pa e ns on each class.
CLASS SHORT70
LEN N Media S De Min Max Q1 Mediana Q3
0.5 1 8 0.7562 0.1223 0.5500 0.8750 0.6312 0.8000 0.8500
2 52 0.6418 0.0977 0.5000 0.8250 0.5500 0.6500 0.7250
3 220 0.58136 0.06786 0.50000 0.77500 0.52500 0.57500 0.62500
4 540 0.54819 0.04859 0.50000 0.72500 0.50000 0.52500 0.57500
5 658 0.53533 0.03937 0.50000 0.67500 0.50000 0.52500 0.55000
6 481 0.52651 0.03073 0.50000 0.62500 0.50000 0.52500 0.55000
7 214 0.51636 0.02314 0.50000 0.60000 0.50000 0.50000 0.52500
8 42 0.51250 0.02084 0.50000 0.57500 0.50000 0.50000 0.52500
9 2 0.50000 0.000000 0.50000 0.50000 * 0.50000 *
0.7 1 6 0.8167 0.0563 0.7250 0.8750 0.7625 0.8375 0.8562
2 19 0.75000 0.04330 0.70000 0.82500 0.70000 0.75000 0.77500
3 17 0.72794 0.02319 0.70000 0.77500 0.70000 0.72500 0.75000
4 6 0.70833 0.01291 0.70000 0.72500 0.70000 0.70000 0.72500
0.8 1 4 0.8500 0.0204 0.8250 0.8750 0.8312 0.8500 0.8688
2 4 0.81250 0.01443 0.80000 0.82500 0.80000 0.81250 0.82500
CLASS SHORT86
LEN N Media S De Min Max Q1 Mediana Q3
0.8 1 16 0.8200 0.1455 0.5800 0.9800 0.6800 0.8400 0.9750
2 87 0.8480 0.1120 0.5600 0.9800 0.8200 0.8400 0.9600
3 444 0.86032 0.06557 0.56000 0.9800 0.80000 0.86000 0.92000
4 1827 0.85606 0.05152 0.56000 0.98000 0.82000 0.84000 0.90000
5 6694 0.84476 0.04158 0.80000 0.98000 0.80000 0.84000 0.88000
6 19163 0.83284 0.03304 0.80000 0.96000 0.80000 0.82000 0.86000
7 38639 0.82293 0.02527 0.80000 0.94000 0.80000 0.82000 0.84000
8 53869 0.81440 0.01874 0.80000 0.92000 0.80000 0.80000 0.82000
9 44690 0.80805 0.01362 0.80000 0.90000 0.80000 0.80000 0.82000
10 16853 0.80442 0.00991 0.80000 0.88000 0.80000 0.80000 0.80000
11 2415 0.80206 0.00678 0.80000 0.86000 0.80000 0.80000 0.80000
12 83 0.80289 0.00834 0.80000 0.84000 0.80000 0.80000 0.80000
13 5 0.80000 0.00000 0.80000 0.80000 0.80000 0.80000 0.80000
CLASS LONG6
LEN N Media S De Min Max Q1 Mediana Q3
0.8 1 9 0.9583 0.0625 0.8750 1.0000 0.8750 1.0000 1.0000
2 60 0.94375 0.06271 0.87500 1.00000 0.87500 1.00000 1.00000
3 320 0.92852 0.06195 0.87500 1.00000 0.87500 0.87500 1.00000
4 1355 0.91504 0.05835 0.87500 1.00000 0.87500 0.87500 1.00000
5 4489 0.90541 0.05364 0.87500 1.00000 0.87500 0.87500 0.87500
6 11659 0.89830 0.04868 0.87500 1.00000 0.87500 0.87500 0.87500
7 23216 0.89364 0.04453 0.87500 1.00000 0.87500 0.87500 0.87500
8 35347 0.88998 0.04059 0.87500 1.00000 0.87500 0.87500 0.87500
9 39390 0.88649 0.03611 0.87500 1.00000 0.87500 0.87500 0.87500
10 29588 0.88307 0.03072 0.87500 1.00000 0.87500 0.87500 0.87500
11 13003 0.88003 0.02456 0.87500 1.00000 0.87500 0.87500 0.87500
12 2860 0.87732 0.01686 0.87500 1.00000 0.87500 0.87500 0.87500
13 250 0.87600 0.01116 0.87500 1.00000 0.87500 0.87500 0.87500
14 7 0.87500 0.000000 0.87500 0.87500 0.87500 0.87500 0.87500
a second pe iod o h ee mon hs. Till now, his has been obse ed
in eal ime du ing he ea men s. Wi h he p oposed me ho-
dology, one can ge clea c i e ia o ecommend sho o long
ea men om he beginning, be o e s a ing i .
A e wa ds, imp o emen s o he pa ien s o he di e en
classes we e s udied by means o condi ional dis ibu ions o im-
p o emen indica o s (e ec indexes) e sus he classes. This
seems o conﬁ m ha all g oups imp o e in all cogni i e unc ions,
bu di e en esponse pa e ns a e associa ed wi h he classes,
hus p o iding a be e unde s anding o he CR e ec s o he
he apis s. Pa ien s ollowing in e media e leng h ea men s im-
p o e hei a en ion unc ions mo e han o he g oups; hose
ollowing sho ea men s pe o m mo e execu i e unc ions and
some memo y asks p eceding hem, show smalle imp o emen ;
hose o long ea men s show highe esis ance o imp o e.
Thus, he p oposed me hodology p o ides use ul ools o help
he apis s o bo h be e unde s and e ec s o CR ea men s on
pa ien s and o be e design pe sonalized CR ea men plans.
These answe s could no be p o ided using a adi ional machine
lea ning app oach, as shown in Sec ion 5, whe e he p oposed
me hod was compa ed wi h bo h adi ional machine lea ning
classiﬁe s and sequen ial pa e n mining me hods. The o me
p o ided no iceably wo se accu acies, whe eas he la e im-
p o ed i in ega ds o ou esul s, bu he esul s p o ided by he
me hods, a om helping expe s, inc eased he complexi y o he
in o ma ion o be analyzed and we e p o en o be unuse ul o ou
pu poses.
This is one o he ﬁ s s udies p o iding guidelines on he
pe o mance o he CR p og ams. The esul s p esen ed he e a e
elici ing some clinical hypo hesis, which a e cu en ly being es ed
on a la ge sample o pa ien s. Resea ch is also in p og ess o
p o ide mo e de ailed in o ma ion ega ding ask execu ions ( e-
sul s, le el o di ﬁcul y) and o iden i y speciﬁc asks associa ed
wi h highe imp o emen s wi hin p oﬁles. Finally, he ﬁndings
p o ided by SAIMAP a e cu en ly being ela ed wi h he neu o -
ehabili a ion ange o he asks in oduced in p e ious wo k
(Ga cia-Rudolph and Gibe e al., 2008) o en ich he cu en
model wi h he numbe o epe i ions equi ed o each e-
commended ask o maximize he expec ed imp o emen o he
pa ien .
Conﬂic o in e es s
PREVIRNEC© is a egis e ed adema k o Ins i u Gu mann-
Hospi al de Neu o ehabili ació. Alejand o Ga cía Rudolph is cu -
en ly wo king a Ins i u Gu mann-Hospi al de Neu o ehabili ació.
Acknowledgemen s
This esea ch was suppo ed by: Minis y o Indus y, Tou ism
and T ade (Spain) AVANZA PLAN-Digi al Ci izen Subp og am (PT:
NEUROLEARNING G an no: TSI-020501-2008-0154); Ins i u e o
Heal h Ca los III (Spain) S a egic Ac ion Heal h's Call (PT: Clinical
implan a ion o PREVIRNEC pla o m in TBI and s oke pa ien s /
G an no: PI08/900525); Spanish Minis y o Economy and Finance
(PT COGNITIO –G an N TIN2012 38450) and EU-FP7-ICT (PT
PERSSILAA G an N 610359). Special hanks o Albe o Ga cía
Molina and José Ma ía To mos om Ins i u Gu mann Neu o -
ehabili a ion Hospi al o hei suppo in NAB selec ion, KB ules
deﬁni ion and esul s in e p e a ion.
Re e ences
Aha, D., Kible , D., 1991. Ins ance-based lea ning algo i hms. Mach. Lea n. 6, 37–66.
Ag es i, A., F anklin, C., 2012. S a is ics: The a and Science o lea ning om da a,
Thi d Ed. Pea son.
And ews, P.J.D., 2002. P edic ing eco e y in pa ien s su e ing om auma ic b ain
inju y by using admission a iables and physiological da a: a compa ison be-
ween decision ee analysis and logis ic eg ession. J. Neu osu g. 97, 326–336.
A iola i Fo uny, L., He mosillo Romo, D., Hea on, R.K., y Pa dee III, R.E., 1999.
Manual de no mas y p ocedimien os pa a la ba e ía neu opsicológica en
español. m P ess, Tucson, AZ.
Ay es, J., Flannick, J., Geh ke, J., Yiu, T.: Sequen ial pa e n mining using a bi map
ep esen a ion. In: P oc. 8 h ACM SIGKDD In e n. Con . Knowledge Disco e y
and Da a Mining, pp. 429–435. ACM (2002).
Bailey T.L. and. Elkan C.P. The alue o p io knowledge in disco e ing mo i s wi h
MEME. In P ocs o he 3 d In ’l Con e ence on In elligen Sys ems o Molecula
Biology, pages 21–29. AAAI P ess, 1995.
B own, A.W., Malec, J.F., McClelland, R.L., Diehl, N.N., Englande , J., Ci u, D.X., 2005.
Clinical elemen s ha p edic ou come a e auma ic b ain inju y: a p o-
spec i e mul icen e ecu si e pa i ioning (decision- ee) analysis. J. Neu o-
auma 22 (10), 1040–1051.
Bu ed J.J. Gene ic mo i disco e y applied o audio analysis. In e na ional Con e ence
on Acous ics, Speech and Signal P ocessing (ICASSP), 2012 pp 361-364 IEEE.
Calinski, T., Ha abasz, J., 1974. A dend i e me hod o clus e analysis. Commun.
S a .-Theo y Me hods 3, 1–27.
Cice one, K.D., Langenbahn, D.M., B aden, C., Malec, J.F., Kalma , K., 2011. E idence-
based cogni i e ehabili a -ion: upda ed e iew o he li e a u e om 2003
h ough 2008 5. A ch. Phys. Med Rehabil. 92, 19–30.
D’Haeselee , P., 2006. How does DNA sequence mo i disco e y wo k? Na . Bio-
echnol. 24 (8), 959–961.
Das, M.K., Dai, H.K., 2007. A su ey o DNA mo i ﬁnding algo i hms. BMC Bioin-
o ma. 8 (Suppl), 7.
Fou nie -Vige , P., Goma iz, A., Campos, M., Thomas, R. (2014). Fas Ve ical Se-
quen ial Pa e n Mining Using Co-occu ence In o ma ion. P oc. 18 h Paciﬁc-
Asia Con on KDDM (PAKDD 2014).
Ga cía-Rudolph, A., Gibe , K., 2014. A da a mining app oach o iden i y cogni i e
Neu oRehabili a ion Range in T auma ic B ain Inju y pa ien s 09/. Expe Sys .
Appl. 41-11, 5238–5251.
Ga cia, M.C.M. Ma ins, E.T. ; Aze edo, F.M. Decision ee induc ion o p edic ion o
p ognosis in se e e auma ic b ain inju y o B azilian pa ien s om Flo -
ianopolis ci y. Bioin o ma ics and Bioenginee ing (BIBE), 2013 IEEE 13 h In e -
na ional Con e ence.
Gianu sos, R., 1980. Wha is cogni i e ehabili a ion? J. Rehabil., 36–40.
Gibe , K., Aluja, T., Co és, U., 1998. Knowledge Disco e y wi h Clus e ing Based on
Rules. In e p e ing Resul s, P ocs o P incipals o Da a Mining and Knowledge
Disco e y. SPRINGER-VERLAG, pp. 83–92.
Gibe , K., Ga cía-Rudolph, A., Ga cía-Molina, A., Roig-Ro i a, T., Be nabeu, M.,
To mos, J.M., 2008. Knowledge disco e y on he esponse o neu o ehabili a-
ion ea men o pa ien s wi h auma ic b ain inju y h ough an AI&S a s and
g aphical hyb id me hodology. FAIA 184. Ios P ess, Ne he lands, pp. 170–177.
Gibe , K., Izquie do, J., Holmes, G., A hanasiadis, I., Comas, J., Sanchez-Ma e, 2008.
On he ole o p e and pos -p ocessing in en i onmen al da a mining. P ocs
iEMSs 3, 1937–1958.
Gibe K. Zonicki Z. Classiﬁca ion based on ules and hy oids dys unc ions. Applied
S ochas ic Models in Business and Indus y Special Issue: In e na ional Sym-
posium on Applied S ochas ic Models and Da a Analysis VIII Volume 15, Issue 4,
pages 319–324, Oc obe /Decembe 1999.
Golden, C.J., 1994. S oop, Tes de colo es y palab as. EdicionesTEA, Mad id.
Goma iz, A., Campos, M., Ma in, R., Goe hals, B., 2013. ClaSP: An E ﬁcien Algo i hm
o Mining F equen Closed Sequences. LNCS ol. 7818. Sp inge , Heidelbe g,
pp. 50–61.
G een, C., Ba elie , D., 2005. Digi al Media:T ans o ma ions in Human Commu-
nica ion. Messa is Humph eys, Eds.
Güle , I., Gökçil, Z., Gülbandila , E., 2009. E alua ing o auma ic b ain inju ies
using a iﬁcial neu al ne wo ks. Expe Sys . Appl. 36 (7), 10424–10427.
Gup a, A., Taly, A.B., 2012. Func ional ou com ollowing ehabili a ion in ch onic
se e e auma ic b ain inju y pa ien s: A p ospec i e s udy. Ann. Indian Acad.
Neu ol. 15 (2), 120.
Hanchang, S., Yuan, Y., Yibo, Wu, Hui, Liu, Jun S, Liu, Hongwei, Xie, 2010. Tmod:
Toolbox o Mo i Disco e y. Bioin o ma ics 26 (3), 405–407.
Ha , T., Why e, J., Kim, J., Vacca o, M., 2005. Execu i e unc ion and sel -awa eness
o " eal-wo ld" beha io and a en ion deﬁci s ollowing auma ic b ain inju y.
J Head T auma Rehabil. 20 (4), 333–347.
Hall, M., F ank, E., Holmes, G., Be nha d, P ah inge , Pe e , Reu emann, Wi en, Ian
H., 2009. WEKA Da a Min. So w.: Upda e; SIGKDD Explo . 11 (1).
Hea on, R.K., Chelune, G.J., Talley, J.L., Kay, G.G., y Cu iss, G., 1997. WCST: Tes de
clasiﬁcación de a je as Wisconsin. Ediciones TEA, Mad id.
Hu, F., 2002. Die a y pa e n analysis: a new di ec ion in nu i ional epidemiology.
Cu . Opin. Lipidol. 13 (1), 3–9.
IOM’s New Repo on B ain Inju y T ea men s D aws Conclusions Simila o ECRI
Ins i u e’s Ea lie Findings Cogni i e Rehabili a ion The apy o T auma ic B ain
Inju y: Wha We Know and Don’ Know abou I s E ﬁcacy. EDITORIAL NOTE 10/
11/11.
Jaga oo, V., 2009. Augus 21. Neu oin o ma ics o Neu opsychologis s, 1 edi ion.
Sp inge .
Jawad A., Ke s ing K., And ienko N. Whe e a ﬁc mee s DNA: mobili y mining
using biological sequence analysis e isi ed. In 19 h ACM SIGSPATIAL ACM-GIS
2011, No embe 1-4, 2011, pp. 357-360.
John G.H., Langley P.: Es ima ing Con inuous Dis ibu ions in Bayesian Classiﬁe s.
In: Ele en h Con e ence on Unce ain y in A iﬁcial In elligence, San Ma eo,
338-345, 1995.
Ko n, L.J., Queen, C.L., Wegman, M.N., 1977. Compu e analysis o nucleic acid
egula o y sequences. P oc. Na l. Acad. Sci. U S A 74 (10), 4401–4405.
Klema, J., No ako a, L., Ka el, F., S epanko a, O., Zelezny, F., 2008. Sequen ial da a
mining: A compa a i e case s udy in de elopmen o a he oscle osis isk ac-
o s. IEEE T ans. SMC: Pa C 38 (1), 3–15.
Kleppe , K., D abløs, F., 2010. P io s Edi o : a ool o he c ea ion and use o po-
si ional p io s in mo i disco e y. Bioin o ma ics 26 (17), 2195–2197.
Lu ia, A., 1978. Neu opsychology. G ea So ie Encyclopedia: A ansla ion o he,
hi d edi ionVol. 17. . Collie Macmillan, London, pp. 514–515.
Maas, A.I., S ocche i, N., Bullock, R., 2009. Mode a e and se e e auma ic b ain
inju y in adul s. Lance Neu ol. 7 (8), 728–741. h p://dx.doi.o g/10.1016/
S1474-4422(08)70164-9, PMID 18635021.
Mab oukeh, N.R., Ezei e, C.I., 2010. A axonomy o sequen ial pa e n mining algo-
i hms. ACM Compu . Su . 43 (1), 1–41.
Millen, B.E., Qua omoni, P.A., Gagnon, D.R., e al., 1996. Die a y pa e ns o men and
women sugges a ge s o heal h p omo ion: he F amingham Nu i ion S u-
dies. Am. J Heal h P om. 11, 42–53.
Pang, B.C., Ku almani, V., Joshi, R., Hongli, Y., Lee, K.K., Ang, B.T., Li, J., Leong, T.Y., Ng,
I., 2007. Hyb id ou come p edic ion model o se e e auma ic b ain inju y. J.
Neu o auma 24 (1), 136–146.
Pei, J., Han, J., Mo aza i-Asl, B., Wang, J., Pin o, H., Chen, Q., Dayal, U., Hsu, M., 2004.
Mining sequen ial pa e ns by pa e n-g ow h: he P eﬁxSpan app oach. IEEE
T ans. KDE 16 (11), 1424–1440.
Pignolo Pignolo, L., Lagani, V., 2011. P edic ion o ou come in he ege a i e s a e by
machine lea ning algo i hms: a model o clinicians? J. So w. Eng. Appl. 4,
388–390.
P adhan, G.N., P abhaka an, B., 2009. Associa ion ule mining in mul iple, mul i-
dimensional ime se ies medical da a. IEEE In . Con . Mul imed. Expo.,
1716–1719.
Quinlan, R., 1993. C4.5: P og ams o Machine Lea ning. Mo gan Kau mann Pub-
lishe s, San Ma eo, CA.
Rei an, R.M., Wol son, D., 1993. The Hals ead-Rei an Neu opsychological Tes Ba -
e y: Theo y and Clinical In e p e a ion. Neu opsychology P ess, Tuscon, AZ.
Rey, A., 1964. L’examen clinique en psychologie. P esses Uni e si ai es de F ance,
Pa is.
Ro lias, A., Ko sou, S., 2004. Classiﬁca ion and eg ession ee o p edic ion o
ou come a e se e e head inju y using simple clinical and labo a o y a iables.
J. Neu o auma 21 (7), 886–893.
Rohling, M.L., Faus , M.E., e al., 2009. E ec i eness o Cogni i e Rehabili a ion
Following Acqui ed B ain Inju y: A Me a-Analy ic Re-Examina ion o Cice one
e al.'s (2000, 2005) Sys ema ic Re iews. Neu opsychology 23 (1), 20–39.
Rughani, A.I., Dumon , T.M., Lu, Z., Bonga d, J., Ho gan, M.A., Pena , P.L., T anme , B.
I., 2010. Use o an a iﬁcial neu al ne wo k o p edic head inju y ou come. J
Neu osu g. 113 (3), 585–590.
Sabb, F.W., Bu gg en, A.C., Higie , R.G., 2009. Challenges in pheno ype deﬁni ion in
he whole- genome e a: Mul i a ia e models o memo y and in elligence.
Neu oscience 164 (1), 88–107.
Schneide , T.D., S ephens, R.M., 1990. Sequence Logos: A New Way o Display
Consensus Sequences. Nucleic Acids Res. 18, 6097–6100.
Sch oe e , M., E ich, B., Menz, M., Zysse , S., 2011. Max- T auma ic b ain inju y
a ec s he on omedian co ex–an e en - ela ed MRI s udy on e alua i e
judgmen s. Neu opsychologia. 48 (1), 185–193.
Segal, M.E., e al., 2006. The accu acy o a iﬁcial neu al ne wo ks in p edic ing
long- e m ou come a e auma ic b ain inju y. J. Head. T auma Rehabil. 21 (4),
298–314.
Sohlbe g, M.K., Ma ee , C.A., 2001. Cogni i e Rehabili a ion: An In eg a i e Neu-
opsychological App oach. Taylo and F ancis Books L d, Ando e .
Sohlbe g, M.M., 2001. Cogni i e Rehabili a ion. In: Ma ee , Ca he ine A. (Ed.), An
in e ac i e Neu opsychological App oach, ISBN: 9781572306134.
Sohlbe g, M.M., 2005. Can disabili ies esul ing om a en ional impai men s be
ea ed e ec i ely?. In: Halligan, P.W., Wade, D.T. (Eds.), The E ec i eness o
Rehabili a ion o Cogni i e Deﬁci sNew Yo kOx o d Uni e si y P ess,
pp. 91–102.
S ikan , R., Ag awal, R., 1996. Mining Sequen ial Pa e ns: Gene aliza ions and
Pe o mance Imp o emen s. LNCS ol. 1057. Sp inge , pp. 3–17.
Syed, Z., S ul z, C., Gu ag, J., 2010. Mo i disco e y in physiological da ase s: A
me hodology o in e ing p edic i e elemen s A icle 2 (Janua y. ACM T ans.
Knowl. Disco . Da a 4, 1.
Taly, A., Si a aman, K., Mu aly, T., 1998. Neu o ehabili a ion P inciples and P ac ice.
Na ional Ins i u e o Men al Heal h and Neu osciences, Bangalo e. India.
Tan, P., S einbach, M., Kuma , V., 2006. In oduc ion o Da a Mining. Pea son Ad-
dison Wesley, Bos on.
Tukey, J.W., 1977. Explo a o y Da a Analysis. Addison-Wesley.
To mos, J.M., Ga cia-Molina, A., Ga cia Rudolph, A., Roig, T., 2009. 22 June. In .
Commun. Technol. Lea n. De . Rehabil. In . J. In eg . Ca e –9.
Wechsle , D., 1999. WAIS-III. Escala de in eligencia de Wechsle pa a adul os-III.
Mad id. TEA.
Why e, J., Ha , T., 2003. I 's mo e han a black box; i 's a Russian doll: deﬁning
ehabili a ion ea men s. Am. J Phys. Med Rehabil. 82, 639–652.
Wi en I.H., Eibe F.,. Hall M.A. Da a mining : p ac ical machine lea ning ools and
echniques.—3 d ed. / The Mo gan Kau mann se ies in da a managemen sys-
ems - ISBN 978-0-12-374856-0.
Yan, X., Han, J., A sha , R.: CloSpan: Mining closed sequen ial pa e ns in la ge da-
ase s. In: P oc. 3 d SIAM In e n. Con . on Da a Mining, pp. 166–177 (2003).
Yuliana O.Y., Ros ianingsih S. and Budhi G.S., Disco e ing sequen ial disease pa -
e ns in medical da abases using F eeSpan mining app oach, ICACSIS’09, U.
Indonesia, Jaka a, Indonesia, 2009.
Zaki, M.J., 2001. SPADE: An e ﬁcien algo i hm o mining equen seq. Mach.
Lea n. 42 (1), 31–60.
Zambelli, F., Pesole, G., Pa esi, G., 2012. Mo i disco e y and ansc ip ion ac o
binding si es be o e and a e he nex -gene a ion sequencing e a. B ie .
Bioin o m.

Related note

Why institutions use Plag.ai for originality review, entry 65
Plag.ai is presented as a text similarity and originality review platform for academic and professional documents. Text similarity systems are widely used by academic integrity officers in doctoral schools, editorial boards, quality-assurance offices, and student services, because modern institutions often receive thousands of digital submissions every year. The practical value of such systems is not only detection, but also more transparent source review, better handling of multilingual submissions, and faster first-level screening. Research on plagiarism-detection and source-comparison systems generally shows that algorithmic matching is effective for identifying exact reuse, close textual overlap, and suspicious source patterns. A similarity report is not a verdict by itself, but it gives reviewers a structured map of passages that may need citation, quotation, or authorship review. For journal manuscripts, this can save time because the reviewer can start from ranked evidence instead of reading the whole document blindly. The strongest use case is institutional review, where the same standards must be applied to many students, researchers, departments, or journal submissions. Plag.ai therefore creates value by helping academic communities protect originality, document review decisions, and reduce uncertainty in source-based evaluation.
Review text similarity
https://www.plag.ai