Re ining p e-polysomnog aphy suspicion o
Obs uc i e Sleep Apnea Synd ome: Logis ic
and Bayesian analysis o clinical ac o s
Liliana Pa ícia Pin o Lei e
Oc obe |2012
5
h
ed.
i
Re ining p e-polysomnog aphy suspicion o
Obs uc i e Sleep Apnea Synd ome: Logis ic and
Bayesian analysis o clinical ac o s.
Liliana Pa ícia Pin o Lei e
Oc obe |2012
Ped o Pe ei a Rod igues, Faculdade de Medicina da Uni e sidade do Po o
C is ina San os, Faculdade de Medicina da Uni e sidade do Po o
5
h
ed
iii
Acknowledgemen s
This hesis is he culmina ion o one mo e academic s ep which means so much mo e han ha . This
Mas e s has gi en me he unique expe ience o mee ing unbelie able people and acqui ing knowledge ha
has allowed me o g ow as a human being.
I wan o hank all he people who we e wi h me, in some way, du ing his p ocess, ei he di ec ly o
indi ec ly.
To my pa en s, o all he lessons, lo e and sac i ices made h oughou my li e ha eally helped me o
always o e come he di icul ies.
To my men o s o ins an ly accep ing his p ojec wi h all he en husiasm and dedica ion.
To P o . Ped o Pe ei a Rod igues o all he pa ience and eaching, o se ing me objec i es and
challenges o which I hough I could no each.
To P o . C is ina San os o he en husiasm, in e es and a ailabili y in he a ea.
To he Sleep Labo a o y eam o Vila No a de Gaia Hospi al Cen e o helping me on he wo k
de elopmen , especially o my colleagues and iends o all hou s, D . Gisela Fon es and D . Fe nanda
Aze edo o ensu ing me he necessa y schedule shi s ha allowed me o a end he cou se and o
helping me in he da a collec ion.
To Ped o who encou aged me o a end he Mas e s, o all he lo e and pa ience and o making me
always belie e ha i is wo h igh ing o and lea ning.

Resumo
In odução: A Sínd ome da Apneia Obs u i a do Sono (SAOS) é uma doença que a ec a 2-4% da
população em odo o mundo. O mé odo pad ão pa a o diagnós ico de SAOS é a polissonog a ia (PSG),
um exame ca o, limi ado às á eas u banas e, consequen emen e, com g andes lis as de espe a.
Objec i o: De ini um mé odo auxilia de diagnós ico que p io iza os pacien es, du an e a consul a do sono,
pa a a ealização da PSG, de aco do com a p obabilidade de diagnós ico de SAOS.
Mé odos: Um es udo p ospec i o oi ealizado, incluindo pacien es adul os com suspei a de SAOS que
ealiza am PSG no labo a ó io do sono do Cen o Hospi ala de Vila No a de Gaia / Espinho. As
a iá eis es udadas o am de inidas a pa i de e isão de li e a u a e ecolhidas du an e a consul a. Fo am
colhidas duas amos as: uma coo e de eino pa a desen ol e os modelos e e i ica a sua alidade
in e na e ex e na a a és da alidação c uzada (VC), e uma coo e de alidação pa a e i ica a aplicação
dos modelos esul an es na p á ica clínica. Com as a iá eis signi ica i as ob idas na eg essão logís ica
(RL) uni a iada o am u ilizadas duas écnicas di e en es pa a cons ui os modelos: RL múl ipla e edes
Bayesianas usando os classi icado es Naï e Bayes (NB) e T ee Augmen ed Bayesian Ne wo k (TAN). A
sensibilidade e especi icidade o am analisadas pa a de e mina a espec i a pe o mance.
Resul ados: Fo am es udados 86 pacien es pa a cons ui os modelos, 52% dos quais com diagnós ico de
SAOS. A RL uni a iada mos ou seis a iá eis com in luência signi ica i a no esul ado: sexo masculino
(OR = 7,259, IC 95% = [1,096; 27,651]), índice de massa co po al (OR = 1,159, [1,030; 1,303])
ci cun e ência do pescoço, (OR = 1,341, [1,159; 1,550]), ci cun e ência abdominal (OR = 1,076, [1,025;
1,129]), apneias p esenciadas (OR = 4,725, [1,772; 12,599]) e álcool an es de do mi (OR = 3,307, [1,350;
8,100]). Fo am es ados dois di e en es limi es de sensibilidade. Com o in ui o de ob e 100% de
sensibilidade oi u ilizado um limi e de 10% na RL ob ido após a análise da cu a ROC (AUC = 80%
[70%, 89%]), 7% no NB e 2% no TAN, enquan o que se p e ende mos 95% de sensibilidade, os limi es
ob idos o am de 25% na RL, 10% pa a o NB e 22% no TAN. A VC da RL es ima que e é obus o pa a
ambos os limi es: 98% pa a a sensibilidade e 11% pa a a especi icidade e 89% -34%, espec i amen e. No
NB, usando 7% como limi e, os esul ados o am de 98% pa a a sensibilidade e de 18% pa a a
especi icidade, com o limi e mais ele ado (10%), os esul ados o am de 93% de sensibilidade e de 30% de
especi icidade. No TAN, usando 2% como limi e, os esul ados o am de 88% de sensibilidade e 23% de
especi icidade e com o limi e de 22% o am de 84% sensibilidade e 25% de especi icidade. Es es modelos
o am es ados numa segunda amos a compa á el de 33 pacien es pa a a alia o seu desempenho na
p á ica clínica. Os esul ados da RL apoiam as expec a i as de ambos os limi es (10% e 25%): 100% -0% e
88% -15%, espec i amen e. Os limi es de 7% e 2% usados pa a o NB e TAN espec i amen e, ob i e am
a mesma sensibilidade (94%), mas o TAN ob e e melho esul ado ela i amen e à especi icidade (7%) do
que o NB (0%). U ilizando os limi es mais ele ados, 10% pa a o NB e de 22% pa a o TAN, os dois
classi icado es ob i e am mais uma ez a mesma sensibilidade (89%), mas o TAN e elou melho
esul ado de especi icidade (13%) do que o NB (7%).
Discussão: A ci cun e ência do pescoço e apneias p esenciadas o necem in o mação su icien e pa a um
modelo clínico com base nos esul ados da RL. Se op a mos po edes Bayesianas de emos usa mais
a iá eis: sexo, índice de massa co po al, ci cun e ência abdominal e álcool an es de do mi . Pa a ambos
os modelos, u ilizando os limi es espec i os, podemos elabo a ês ní eis de p io idade dada a
p obabilidade de o pacien e e diagnós ico de SAOS: o g upo não p io i á io, um ní el in e médio e, po
im, um g upo de al a p io idade. Além des es esul ados, o uso das edes e ela duas p incipais an agens
que a RL adicional não pode esol e . P imei o, as edes Bayesianas podem lida com in o mações em
al a, e em segundo luga , pe mi em uma ep esen ação g á ica que pode se mais in e essan e pa a o
médico. Conside amos que o uso des es modelos na consul a do sono pode se uma e amen a ú il pa a a
iagem de pacien es que ealizem PSG e pode, e en ualmen e, ajuda a p io iza os doen es, pe mi indo
al ez eduzi o núme o de PSG com esul ado no mal.
Pala as-cha e
: ac o es de isco, Sínd oma da Apneia Obs u i a do Sono, diagnós ico, modelo clínico, edes Bayesianas,
sensibilidade e especi icidade.
Abs ac
In oduc ion: Obs uc i e Sleep Apnea (OSA) is a disease ha a ec s 2-4% o he popula ion a ound he
wo ld. The s anda d me hod o OSA diagnosis is polysomnog aphy (PSG), an expensi e exam, limi ed o
u ban a eas and, consequen ly, wi h long wai ing lis s.
Aim: To de ine an auxilia y diagnos ic me hod, ha p io i izes pa ien s du ing p e-polysomnog aphy
consul a ion, acco ding o hei p obabili y o OSA diagnosis.
Me hods: A p ospec i e s udy was conduc ed, including adul pa ien s wi h OSA suspicion ha pe o med
PSG a he Sleep Labo a o y o Vila No a de Gaia/Espinho Hospi al Cen e . The s udied a iables we e
de ined om li e a u e e iew and collec ed du ing consul a ion. Two samples we e collec ed: a aining
g oup o build he models and check in e nal and c oss- alida ion (CV) and a alida ion g oup o check
he esul an models in clinical p ac ice. Wi h he signi ican a iables achie ed wi h uni a ia e logis ic
eg ession (LR) we used wo di e en echniques, mul iple LR and Bayesian ne wo ks classi ie s- Naï e
Bayes (NB) and T ee Augmen ed Bayesian ne wo k (TAN) - o build models ha p edic s OSA diagnosis.
The sensi i i y and speci ici y was analyzed o de e mine hei pe o mance.
Resul s: We s udied 86 pa ien s in o de o build he models, 52% wi h OSA diagnosis. Uni a ia e LR
analysis showed six a iables wi h signi ican in luence on he ou come: male gende (OR=7.259, 95%
CI=[1.096;27.651]), body mass index (OR=1.159, [1.030;1.303]), neck ci cum e ence (OR=1.341,
[1.159;1.550]), abdominal ci cum e ence (OR=1.076, [1.025;1.129]), wi nessed apneas (OR=4.725,
[1.772;12.599]) and alcohol be o e sleep (OR=3.307, [1.350;8.100]). We es ed wo di e en cu o s o
sensi i i y. Aiming 100% o sensi i i y we used a 10% cu o on LR achie ed a e a ROC cu e analysis
(AUC=80% [70%;89%]), 7% on NB and 2% on TAN while aiming 95% o sensi i i y he cu o s we e
25% on LR, 10% o NB and 22% on TAN. The CV alida ion o LR model es ima es ha i was obus
o bo h cu o s (10% and 25%): 98%-11% and 89%-34%, espec i ely. On NB, using 7% as cu o , he
esul s we e 98% o sensi i i y and 18% o speci ici y and wi h he highe cu o (10%) he esul s we e
93% o sensi i i y and 30% o speci ici y. On TAN, using 2% as cu o , he esul s we e 88% o
sensi i i y and 23% o speci ici y and o 22% we e 84% o sensi i i y and 25% o speci ici y. These
models we e es ed on a sepa a e compa able 33-pa ien s coho o analyze hei pe o mance on clinical
p ac ice. Resul s o LR suppo ed he expec a ions o bo h h esholds: 100%-0% and 88%-15%,
ii
espec i ely. The 7% (NB) and 2% (TAN) cu o s ob ained he same sensi i i y (94%), bu TAN achie ed
be e esul s on speci ici y (7%) han NB (0%). Using he highe cu o s o 10% on NB and 22% o
TAN, he wo classi ie s ob ained once again he same sensi i i y (89%) bu be e esul s o speci ici y
on TAN (13%) han in NB (7%).
Discussion: Neck ci cum e ence and wi nessed apneas in o ma ion su ices o a clinical model based on he
LR esul s. I we use a BN we need mo e wo a iables: gende , body mass index, abdominal
ci cum e ence and alcohol be o e sleep. Fo bo h models using he espec i e cu o s we can p o ide
h ee le els o p io i y gi en he p obabili y o he pa ien ha ing OSA diagnosis, non-p io i y g oup, an
in e media e le el and, inally, a p io i y g oup. Besides hese esul s, he use o BN e eals wo main
ad an ages ha adi ional LR canno sol e. Fi s ly, BN can deal wi h missing in o ma ion; second, he
g aphical ep esen a ion can be mo e in e es ing o he physician. We conside ha he use o hese
models on sleep consul a ion can be a help ul ool o moni o pa ien s o pe o m PSG and e en ually
educe he numbe o no mal esul s PSG.
Key-wo ds:
isk ac o s, obs uc i e sleep apnea, diagnosis, clinical model, Bayesian ne wo k, sensi i i y and speci ici y.
P io Dissemina ion
The in es iga ion p o ocol o de elop his hesis wi h he p o isional i le ''Da a mining as an auxilia y
diagnos ic o he Synd ome o Obs uc i e Sleep Apnea: Is i possible o educe he numbe o
unnecessa y polysomnog aphies?'', was shown on 4 h Symposium on Medical In o ma ics, Oc obe 2011,
Po o, Po ugal
The p elimina y esul s was p esen ed on he In elligen Da a Analysis mee ing o ganized by he Heal h
In o ma ion and Decision Sciences depa men , Facul y o Medicine, Uni e si y o Po o, Po ugal on 25 h
o Janua y o 2012.

In oduc ion 1
1. In oduc ion
The synd ome o Obs uc i e Sleep Apnea (OSA) is a disease ha a ec s app oxima ely 4% o men and
2% o women wo ldwide bu is s ill unde es ima ed and unde diagnosed (Al Lawa i, Pa el, & Ayas, 2009;
Jennum & Riha, 2009; Madani & Madani, 2009; T. Young, E ans, Finn, & Pal a, 1997). I is cha ac e ized
by episodes o b ea hing cessa ion (apnea) o educ ion in ai low (hypopnea) du ing sleep o a leas 10
seconds as a esul o a uppe ai way collapse (Al Lawa i e al., 2009; Ibe C, 2007; Rech scha en, 1968;
Redline e al., 2007; Silbe e al., 2007). The se e i y o OSA is associa ed wi h he apnea-hypopnea index
(AHI), documen ed du ing sleep, which can be di ided in o mild (5 ≤ AHI <15), mode a e (15 ≤ AHI
<30) and se e e (AHI ≥ 30) ("Sleep- ela ed b ea hing diso de s in adul s: ecommenda ions o synd ome
de ini ion and measu emen echniques in clinical esea ch. The Repo o an Ame ican Academy o Sleep
Medicine Task Fo ce," 1999).The s anda d me hod o assessing his index, and he e o e de ining he
OSA diagnosis, is polysomnog aphy (PSG). Howe e , i is ime-consuming, expensi e and ela i i y
limi ed o u ban a eas which, consequen ly, o igina es high wai ing lis s (Sun, Chiu, Chuang, & Liu, 2010).
Fo a co ec diagnosis, i is impo an o i s de e mine he ac o s associa ed wi h he disease, and hen
use hem o calcula e he p obabili y o he p esence o OSA. Acco ding o he li e a u e, he isk ac o s
associa ed wi h OSA a e age, gende and body mass index (BMI). Howe e , he e is no consensus on he
weigh o hese ac o s in he p edic ion (Al Lawa i e al., 2009; Da ies, Ali, & S adling, 1992; Dogh amji,
2008; Ho s ein & Szalai, 1993; Kapu , 2010; Kohle , 2009; Manbe & A mi age, 1999; T. Young e al.,
1997; Te y Young e al., 2002; T. Young, Ska ud, & Peppa d, 2004). Some au ho s e e ed o he
ea u es such as neck ci cum e ence (NC), wi nessed apneas o diu nal somnolence as impo an isk
ac o s oo (Da ies e al., 1992; Poulio , Pe e s, Neu eld, & K yge , 1997). The pa h o a consensus is s ill
unde e mined.
In Po ugal, pa ien s a e e e ed by he p ima y ca e physician o a sleep consul , and hen he sleep
expe physicians decide he need o pe o m polysomnog aphy. Al hough pa ien s a e sc eened by he
physicians, based on clinical ac o s, he speci ici y o he en i e p ocess is a he low (48% o PSG
pe o med in 2010, in ou sleep labo a o y, esul ed nega i e o OSA, om which 75% had a comple ely
no mal esul o sleep diso de s) which, oge he wi h he limi ed a ailabili y o he se ice, yields long
wai ing lis s bo h o consul a ion and o pe o m PSG. This p oblem is also p e alen in o he sleep
labo a o ies and se e al s udies ha e been conduc ed o de ine he mos impo an ac o s o de e mine
2 In oduc ion
he p obabili y o ha ing OSA, and he eby educe and op imize he numbe o pa ien s ha ealize PSG,
assigning di e en p io i y o pa ien s (Al Lawa i e al., 2009; Da ies e al., 1992; Dixon, Schach e , &
O’B ien, 2003; Flemons, Whi elaw, B an , & Remme s, 1994; Gu ubhaga a ula, Maislin, & Pack, 2001;
Ho s ein & Szalai, 1993; Maislin e al., 1995; Poulio e al., 1997; Rodsu i, Hensley, Thakkins ian, D'Es e,
& A ia, 2004; Vine , Szalai, & Ho s ein, 1991; Te y Young e al., 2002).
To iden i y mo e quickly OSA pa ien s and possibly educe he numbe o PSGs, some p ocedu es ha e
been adop ed like he use o Po able Moni o s (PM) and p edic ion models, bu hese don´ shows
capable o s op he endency o inc eased wai ing lis s.
PM a e a use ul ool in cases o pa ien s wi hou como bid condi ions o medical diso de s, wi h a highe
p obabili y o mode a e o se e e OSA. O he wise, his me hod ends o unde es ima e se e i y o OSA,
because don´ allow de e mine sleep e iciency, and so, PSG ha e o be pe o med on he mos cases
(Collop e al., 2007).
P edic ion models we e buil based on ques ionnai es and p edic ion me hods o sc een pa ien s wi h a
highe p obabili y o OSA diagnosis (Flemons e al., 1994; Gu ubhaga a ula e al., 2001; Kaimakamis,
B a sas, Sichle idis, Ka ounis, & Magla e as, 2009; Kwia kowska, A kins, Ayas, & Ryan, 2007; Maislin e
al., 1995; Poulio e al., 1997; Sun e al., 2010; Vine e al., 1991). T adi ionally, hese models consis ed in
simple decision ules, he p ognos ic sco e and classi ica ion o pa ien s in o di e en isk ca ego ies. This
sco e is o en based on he combina ion o clinical a iables and has been buil o he gene al popula ion,
as well as o speci ic g oups. These models can be an al e na i e me hod in a sleep consul a ion o help in
he clinical decision o pe o m PSG. Bu , o cons uc clinical decision ules, we ha e o check some
cha ac e is ics o he models o alida e hei use in clinical p ac ice (Kononenko, 2001). The main
limi a ion is sensi i i y. These models need a high sensi i i y, as alse nega i es should be a oided, o
p e en excluding a pa ien wi h mode a e o se e e OSA om pe o ming PSG. No s udy ounded on
li e a u e in was i ed o 100% sensi i i y.
In Po ugal, we ound one s udy ha ied o implemen a sc eening ool o OSA. Vaz e al. used he
Be lin Ques ionnai e (BQ), one o he mos ecognized sc eening ool, o sc een pa ien s wi h OSA in a
sleep b ea hing clinic (Vaz e al., 2011). I includes 10 i ems o ganized in 3 ca ego ies conce ning sno ing
and wi nessed apneas (5 i ems), day ime sleepiness (4 i ems) and high blood p essu e /obesi y (1 i em).
Pa ien s a e also asked o p o ide in o ma ion on age, gende , weigh , heigh , neck ci cum e ence and
e hnici y. P ede e mina ion o high o lowe isk o OSA is based on esponses o each ca ego y o i ems.
The au ho s achie e a sensi i i y o 65.2 % and speci ici y o 80%, wha e eals a good disc imina ion bu
poo pe o mance in OSA iden i ica ion. These esul s a e simila o o he s udies ha ha e di e en
pe o mances and don´ e eal BQ as an al e na i e o sc een pa ien s (Ahmadi, Chung, Gibbs, & Shapi o,
2008; Chung e al., 2008; Gami e al., 2004; Gus e al., 2008; Ne ze , S oohs, Ne ze , Cla k, & S ohl, 1999;
Wein eich, Plein, Teschle , Resle , & Teschle , 2006).
Ano he equen p oblem in he applica ion o he models is he lack o in e nal and/o ex e nal
alida ion o he esul s (Da ies e al., 1992; Flemons e al., 1994; Ho s ein & Szalai, 1993; Vine e al.,
1991; Te y Young e al., 2002) which comp omises hei applica ion. Some s udies don´ show he
eg ession pa ame e s (Ho s ein & Szalai, 1993; Te y Young e al., 2002); o he use ac o s ha a e
subjec i e and may lead o lowe measu emen eliabili y (Ho s ein & Szalai, 1993). Ano he poin ha
limi s hei applica ion is he choice o a iables used in he cons uc ion o he models. As hey a e based
on clinical a iables, he missing o an impo an one comp omises hei esul s and, consequen ly, hei
alida ion. Neck ci cum e ence is one o he mos signi ican clinical a iable o iden i y pa ien s wi h
OSA acco ding o some s udies (Da ies e al., 1992) bu o he s udies don´ use his measu e o cons uc
hei models (Poulio e al., 1997; Rodsu i e al., 2004; Vine e al., 1991). Fu he mo e, he Ame ican
Academy o Sleep Medicine (AASM) ecommends he use o he 5 e en s pe hou cu o o dis inguish
pa ien s wi h OSA om hose whe e OSA is absen ("Sleep- ela ed b ea hing diso de s in adul s:
ecommenda ions o synd ome de ini ion and measu emen echniques in clinical esea ch. The Repo
o an Ame ican Academy o Sleep Medicine Task Fo ce," 1999). Some s udies used di e en cu o s o
desc ibe OSA (10, 15, 20 o 30) which is ques ionable (Flemons e al., 1994; Ho s ein & Szalai, 1993;
Maislin e al., 1995; Poulio e al., 1997; Vine e al., 1991) and can in oduce a la ge numbe o alse
nega i es and di icul ies in hei compa ison.
Today, hese models a e gene a ed by a i icial in elligence, using decision ees, neu al ne wo ks, suppo
ec o machines and Bayesian ne wo ks (BN) (Lee & Abbo , 2003; an Ge en, Taal, & Lucas, 2008). To
be use ul he models mus ha e ce ain cha ac e is ics, such as good pe o mance, good abili y o handle
da a en y e o s o omissions, anspa ency o diagnos ic knowledge, abili y o explain decisions, and he
algo i hm is able o educe he numbe o es s needed o make a eliable diagnosis (Kononenko, 2001)
Da a mining in ol es he ex ac ion o in o ma ion, whose goal is o disco e ac s and / o unknown o
hidden pa e ns in a da abase and ex ensi e in e ence ules o p edic ends. This p ocess is based on
combina ions o machine lea ning and s a is ical analysis (La ac, 2001; Lee & Abbo , 2003; P. Lucas,
2004; Mi chell, 1997; an Ge en e al., 2008). The da a mining ools ha e been used in sleep medicine o
c ea e models al e na i e o hose based in logis ic eg ession like decision ees applied o PSG signals,
gene ic algo i hm o sc een pa ien s wi h mode a e o se e e OSA.
Bayesian ne wo ks, in pa icula , ha e been used in medical domain in some a eas wi h high pe o mance
like in diagnoses o pneumonia and b eas cance , classi ica ion o cy ological indings, classi ica ion,
p edic ion o pa ien compliance o medica ion, p edic ion o clinician compliance o medical p ac ice
guidelines, p ognosis o head inju ies, de e mina ion o he isk ac o s o obesi y, and pa e n ecogni ion
4 In oduc ion
in na a i e clinical epo s (A onsky & Haug, 2000; E. Bu nside, Rubin, & Shach e , 2000; E. S. Bu nside,
2005; Hamil on e al., 1995; Lee & Abbo , 2003; Mon i oni, Ba els, Thompson, Sca pelli, & Hamil on,
1995; Sakella opoulos & Niki o idis, 1999; Tak ak, Ajmi Nabli, Ben O hmen, M i aoui, & Ben Hadj
Hamida, 2011). As BN a e a powe da a mining ool, we choose his me hod o c ea e a model o sc een
pa ien s wi h OSA.
Aim 5
2. Aim
The main objec i e o his wo k is o de ine an auxilia y diagnos ic me hod ha can suppo he decision
o pe o m polysomnog aphy, o p io i ize he wai ing lis o polysomnog aphy, in pa ien s ecommended
by a gene al p ac i ione , suspec ed o ha ing OSA, speci ically:
• P io i ize pa ien s ecommended o PSG;
• Reduce he numbe o ‘’unnecessa y polysomnog aphies’’ (inc ease speci ici y) o gi e a highe
p io i y in he wai ing lis o polysomnog aphy o pa ien s wi h mo e chances o OSA diagnosis;
• A oid he ecommenda ion “unnecessa y polysomnog aphy” o OSA cases (a oid alse
nega i es);
• P oduce e ec i e models o use in clinical p ac ice;
• Expand he g aphical in e p e a ion o esul s.

Backg ound 6
3. Backg ound
The uppe ai way includes he ex a ho acic achea, la ynx, pha ynx, nose and is sepa a ed in o h ee
egions: he nasopha ynx, wich is de ined om he nasal u bina es o he ha d pala e; he o opha ynx,
subdi ided in o he e o pala al egion ; and he hypopha ynx (K yge , 2005).
To s udy obs uc i e sleep apnea (OSA) we will ocus on he pha yngeal ai way, speci ically he
e opala al e oglossal egions because is he si e o uppe ai way closu e o na owing du ing sleep in
he majo i y o pa ien s wi h OSA. This pa is a condui o ai low connec ing he nose wi h he la ynx,
pha yngeal pa ency is c i ical. Wi h he excep ion o he wo ends o he espi a o y ai way ac ( he na es
and he small in apulmona y ai ways), he pha ynx is he only collapsible segmen o he espi a o y ac
(K yge , 2005).
The sleep s a e is associa ed wi h a dec ease in mo o ou pu o pha yngeal muscles. When his occu s
agains he backg ound o uppe ai way ana omic abno mali ies, se e e na owing o closu e o pha yngeal
ai way can occu .
3.1. Pa hogenesis o OSA
The pa hogenesis o OSA may be explained by some ac o s: al e a ions o uppe -ai way dila o muscle
ac i i y du ing sleep and his ana omy, lung olume, en ila o y con ol s abili y, sleep s a e s abili y and
os al luid shi s (Kapu , 2010; Yaggi & S ohl, 2010). The ela ionship o his ac o s in luence b ea h
and depends on a balance o o ces: o ces ha p omo e ai way collapse and opposing o ces ha
main ain uppe ai way pa ency (Yaggi & S ohl, 2010).
The balance o o ces p omo ing ai way collapse, like nega i e p essu e o en ila ion and ex aluminal
posi i e p essu e, and o ces o oppose hese collapsing, ac i i y o he pha yngeal dila o muscles (eg,
genioglossus) and enso palli ine dila o muscles a e onically ac i e, a e usually main ained du ing sleep
(Yaggi & S ohl, 2010) bu , in pa ien s wi h OSA, some o his con ols a e los . Dila o muscle ac i i y is
con olled by geniouglossus, he muscle ha o ms he majo i y o he body o he ongue (Kapu , 2010;
Yaggi & S ohl, 2010), and is esponsible o s i en and dila e a ious egions o he ai way. Any
al e a ions in muscle ac i i y o lowe end-expi a o y lung olume, inc eases he endency o he uppe
ai way collapse (Kapu , 2010).
3.1.1. Ven ila o y con ol s abili y
OSA causes g ea ins abili y o en ila o y con ol sys em and, consequen ly, a highe loop gain (measu e
o he s abili y o a nega i e- eedback con ol sys em) because au onomous sys em ha e o esponse o he
inpu s gene a ed by he uppe -ai way muscles (Kapu , 2010).
3.1.2. S abili y o sleep
The sleep s abili y is a ec ed by he numbe inc eased o a ousals ha occu s as a biological esponse o
he hypoxemia, causing an inc ease in espi a o y e o and accen ua e he changes in en ila ion. Neu al
espi a o y con ol cen e s esponse changing he le el o PaO2 and PaCO2 augmen ing and pe pe ua ing
espi a o y cycling (Kapu , 2010; Yaggi & S ohl, 2010).
3.1.3. Ros al luid shi s
Fluid displacemen om he legs caused by lowe body posi i e p essu e, by in la ion o an ishock
ouse s inc eases neck ci cum e ence, na ows he pha ynx, and inc eases collapsibili y in awake heal hy
subjec s has been shown o educe uppe -ai way size and inc ease collapsi y (Kapu , 2010; Yaggi & S ohl,
2010).
3.2. Sleep s ages and e en sco ing
The Ame ican Academy o Sleep Medicine (AASM) ecommend he c i e ia sco ing o sleep s ages and
he use o he e minology di ision in o wake ulness, Non Rapid Eyes Mo emen (NREM) wi h 3 s ages
N1, N2 and N3, and REM (Rapid Eyes Mo emen )(Ibe C, 2007; Silbe e al., 2007). Table 1 summa izes
he main cha ac e is ics o he i e s ages.
8 Backg ound
Table 1: C i e ia o sco e sleep s ages
S ages
Rules
Wake ulness
Eye blinks a a equency o 0.52Hz
Reading eye mo emen s
I egula conjuga e apid eye mo emen s associa ed wi h no mal o high chin muscle one
Epochs wi hou disce nible alpha hy hm
S age N1
A. In
subjec s who gene a e alpha hy hm, sco e s age N1 i alpha hy hm is a enua ed and
eplaced by low ampli ude, mixed equency ac i i y o mo e han 50% o he epoch.
B. In subjec s who do no gene a e alpha hy hm, sco e s age N1 commencing wi h he ea lies o
any o he ollowing phenomena:
1) Ac i i y in ange o 4-7Hz wi h slowing o backg ound equencies by zl Hz om hose o s age
W.
2) Ve ex sha p wa es.
3) Slow eye mo emen s.
S age N2:
A. One o mo e K complexes unassocia ed wi h a ousals
B. One o mo e ains o sleep spindles
S age N3
20% o mo e
o an epoch consis s o wa es o 0.5
-
2 Hz equencies wi h peak
-
o
-
peak ampli ude
o >75 µV in he on al de i a ion.
S age R
(REM sleep)
P esence o eye mo emen s a e and s age 2 absen o low ampli ude mixed equency EEG and
pe sis en ly low chin EMG one
REM: Rapid Eyes Mo emen .
Acco ding o he AASM, espi a o y e en s can be di ided in o apnea o hypopnea ( able 2).
Table 2: Sco ing o espi a o y e en s
E en
C i e ia
Apnea
D op in he peak he mal sens
o excu sion o a leas 90% o baseline
Du ing a leas 10 seconds
A leas 90% o e en ’s mee s he ampli ude educ ion c i e ia o apnea
Obs u
c
i e
:
Con inued o inc eased inspi a o y e o du ing he apnea
Cen al
:
Absence o inspi a o y e o
du ing he apnea
Mixed
:
Absence o inspi a o y e o in he ini ial po ion o he e en ollowed by esump ion o
inspi a o y e o in he second po ion
Hypopnea
(Recommended
ules)
The nasal p essu e excu sions d op by
≥
3
0% o baseline
The du a ion o his d op is a leas 10 seconds
The e is a ≥4% desa u a ion om p e-e en baseline o he e en is associa ed wi h a ousal
A leas 90% o e en ’s mee s he ampli ude educ ion c i e ia o hypopnea
Hypopnea
(Al e na i e
ules)
The nasal p es
su e excu sions d op by
≥50% o baseline
The du a ion o his d op is a leas 1o seconds
The e is a ≥3% desa u a ion om p e-e en baseline o he e en is associa ed wi h a ousal
A leas 90% o e en ’s mee s he ampli ude educ ion c i e ia o hypopnea
RERA
Sequence o b ea hs du ing a leas 10 seconds cha ac e ized by inc easing espi a o y e o o
la ening o he nasal p essu e wa e o m leading o an a ousal om sleep when does no mee
he c i e ia o apnea o hypopnea
RERA: Respi a o y e o - ela ed a ousal
3.3. OSA diagnosis
The diagnosis o OSA has o ollow he c i e ia A o B plus C (Sleep- ela ed b ea hing diso de s in adul s:
ecommenda ions o synd ome de ini ion and measu emen echniques in clinical esea ch. The Repo o an Ame ican
Academy o Sleep Medicine Task Fo ce, 1999):
A. Excessi e day ime sleepiness ha is no be e explained by o he ac o s;
B. Two o mo e o he ollowing ha a e no be e explained by o he ac o s:
-choking o gasping du ing sleep,
- ecu en awakenings om sleep,
16 Backg ound
Magni ude o day ime sleepiness associa ed wi h OSA was co ela ed wi h c ash isk. Un ea ed OSA is a
public isk because inc ease he isk o a ic acciden s and hei consequences (Al Lawa i e al., 2009;
Jennum & Riha, 2009).
3.6.8. Gene ics/Family his o y
The isk ac o s lis ed ea lie a e also ‘‘complex’’ ai s, and isk ac o s can ope a e ei he alone o in
combina ion o many, gene ic and amily his o y a e no an excep ion.
The gene ic in luence is mul i ac o ial a he han due o a single mu a ion o p o ein ac ion and p ohibi
de ini i e conclusions on gene ic unde pinnings o OSA and ha addi ional s udies a e needed o u he
de ine whe he he diso de uly has a gene ic componen
Some impo an cha ac e is ics such as c anio acial mo phology, cephalome ic abno mali ies, including
e oposi ion o he maxilla and mandible and a la ge so pala e, olume o he la e al pa apha yngeal
walls, ongue, so issue s uc u es and o he ac o s like sel - epo ed sleepiness, en ila o y con ol and
sleep cycles/ a chi ec u e ope a ing du ing sleep a e in pa he esul o a ious gene ic and
en i onmen al ac o s ha ac and in e ac o p oduce disease (Madani & Madani, 2009; Punjabi, 2008;
S ie e & Punjabi, 2005; Yaggi & S ohl, 2010). Al hough he di icul y o de ine he gene ic basis o
obs uc i e sleep apnea, he a ailable da a sugges s ha inqui ies abou amily his o y can ce ainly aid in
iden i ying possible pa ien s due o he amilial suscep ibili y o sleep apnea seems o inc ease di ec ly
wi h he numbe o a ec ed ela i es (Punjabi, 2008).
3.6.9. Smoking
Ai way in lamma ion and damage due o ciga e e smoke could al e he mechanical and neu al p ope ies
o uppe ai way and inc ease i s collapsibili y du ing sleep (Lam e al., 2010; Punjabi, 2008; Yaggi & S ohl,
2010). Sleep ins abili y, which has been linked o OSA, may be inc eased by o e nigh educ ions in
nico ine blood le els (Madani & Madani, 2009). Al hough he associa ion wi h OSA is ela i ely weak,
smoking may in e ac wi h and add o he ca dio ascula isk associa ed wi h OSA (Yaggi & S ohl, 2010)
3.6.10. Alcohol and seda i es
Alcohol and seda i es inges ion can induce apneic ac i i y in no mal o asymp oma ic indi iduals
p ecipi a e obs uc i e apneas and hypopneas du ing sleep (Punjabi, 2008) because elaxes uppe ai way
dila o muscles and so inc eases uppe ai way esis ance esul ing in hypo onia o he o opha yngeal
muscles (Dogh amji, 2008).

The e o e, alcohol in ake can p olong apnea du a ion, supp ess a ousals, inc ease equency o occlusi e
episodes and wo sen he se e i y o hypoxemia (Lam e al., 2010; Madani & Madani, 2009).
3.6.11. Como bid condi ions
Obs uc i e sleep apnea also has been implica ed in he e iology o como bid and ca dio ascula
condi ions, including hype ension, co ona y a e y disease, conges i e hea ailu e, and s oke (Punjabi,
2008).
OSA is high in pa ien s wi h hype ension and a casual ole o OSA in hype ension has been sugges ed in
se e al s udies (T. Young e al., 2004). Some s udies sugges s ha he e a e a po en ial ela ionship
be ween OSA and s oke, howe e , his implica ion needs o be p o ed (Punjabi, 2008).
The ea men o OSA can imp o e he condi ions ela ed abo e, and so, con i m he ela ionship
be ween hese condi ions. Some medical condi ions such as uncon olled hype ension, co ona y a e y
disease, conges i e hea ailu e, s oke, and diabe es melli us, undiagnosed obs uc i e sleep apnea should
be conside ed as a possible concomi an p oblem. The eason maybe ha in e mi en hypoxemia and
sleep dis up ion o obs uc i e sleep apnea a e dele e ious o glucose homeos asis and alle ia ing
obs uc i e b ea hing du ing sleep wi h con inuous posi i e ai way p essu e he apy has di ec e ec s in
imp o ing hype glycemia and imp o e he me abolic con ol (Punjabi, 2008).
3.7. Diagnos ic decision suppo
The de ini ion o clinical decision suppo sys ems is now a majo opic since i may help he diagnosis,
p ognosis, and ea men selec ion. Howe e , he complica ed na u e o eal-wo ld biomedical da a has
made i necessa y o look beyond adi ional bios a is ics wi hou loosing he necessa y o mali y (P.
Lucas, 2004). New compu a ional echniques a e be e a de ec ing pa e ns hidden in biomedical da a,
and can be e ep esen and manipula e unce ain ies. Fo example, nai e Bayesian app oaches a e closely
ela ed o logis ic eg ession (Schu ink e al., 2007). Bayesian app oaches ha e an ex eme impo ance in
hese p oblems as hey p o ide a quan i a i e pe spec i e, and allow aking in o accoun p io knowledge
when analyzing da a, o e ing a gene al and e sa ile app oach o cap u ing and easoning wi h unce ain y
in medicine and heal h ca e (P. J. F. Lucas, an de Gaag, & Abu-Hanna, 2004).
3.7.1. T adi ional clinical models
When we ha e a dicho omous ou come, he echnique o choice o s a is ical modeling is logis ic
eg ession (LR) (Tu, 1996). The ela ionship is achie e h ough he logis ic eg ession equa ion, which
18 Backg ound
allows de e mine which explana o y a iables in luence he ou come and, consequen ly e alua e he
p obabili y ha indi idual’s alues o he explana o y a iables, will ha e a pa icula ou come (Pe ie &
Sabin, 2009). The usual assump ion is ha hese p edic o a iables a e ela ed in a linea manne o he
log odds o he ou come o in e es (Tu, 1996).
Gi en he widely spend use o hese models, mo e de ails a e no p esen he e.
3.7.2. Beyond adi ional s a is ics
Ac ually, he heal h ca e se ices p oduce da a ha is inc easing e e y day as a consequence o new
echniques and he in eg a ion o da a om di e en sou ces. The adi ional s a is ics me hods, like
logis ic eg ession, ha e been unable o deal wi h g ea da abases, and so, he u iliza ion o new me hods,
pa icula ly machine lea ning ones, has been inc easing. Da a mining (DM) allows me hods o da a
p ep ocessing and isualiza ion, non-s a is ical me hods and new me hods based on p obabili ies and
s a is ics ha suppo s clinical decisions on models (P. Lucas, 2004).
A i icial in elligence is a b anch o compu e science and machine lea ning is one o i s subdi isions and,
om he beginning, is used in medical da abases. Tom Mi chell de ines machine lea ning as ‘’ he s udy o
compu e algo i hms ha imp o e au oma ically h ough expe ience. Success ul applica ions ange om
da a mining p og ams ha disco e gene al ules om la ge da abases, o in o ma ion il e ing sys ems
ha lea n use s ‘ eading p e e ences, o au onomous ehicles ha lea n o d i e on public highways”
(Mi chell, 1997). The h ee main b anches o machine lea ning a e s a is ical and pa e n ecogni ion
me hods like k-neighbo s and Bayesian classi ie s, induc i e lea ning o symbolic ules like decision ees,
decision ules o induc ion o logic p og ams and, inally, a i icial neu al ne wo ks (Kononenko, 2001).
The Knowledge Disco e y in la ge Da abases p ocess consis s o i e basic s eps: (1) p oblem
iden i ica ion; (2) da a ex ac ion; (3) da a p ep ocessing; (4) da a mining, and; (5) pa e n in e p e a ion
and p esen a ion. The main asks o da a mining in heal hca e may include (1) disco e ing associa ions, (2)
clus e ing, o (3) c ea ing p edic i e (classi ica ion/ eg ession) models (Lee & Abbo , 2003). DM is
conce ned wi h inding pa e ns in la ge da abases which a e in e es ing and alid. The e a e nume ous
da a mining algo i hms ha can be used in classi ica ion o p edic ion- p edic i e da a mining algo i hms,
o inds associa ions, clus e s- desc ip i e da a mining algo i hms (La ac, 2001).
Decision suppo me hods a e buil based in model selec ion ha uses he bes algo i hm o a gi en
da ase , and model in eg a ion/combina ion. They can p o ide an op imal solu ion and can be applied o
build ules o decision ees p oposing he bes classi ie o a gi en classi ica ion ask.
In he heal hca e/medical domain DM ools commonly used include neu al ne wo ks, decision ees and
Bayesian ne wo ks. Neu al ne wo ks a e designed o mimic he pa allel p ocessing abili y o he human
b ain. Decision ees use a epea ing se ies o b anches ha desc ibes associa ions be ween a ibu es and a
a ge a iable. Bayesian ne wo ks p o ides a p obabilis ic app oach (Lee & Abbo , 2003).
To be applied in medical diagnos ic asks, machine lea ning based sys ems ha e o some equi emen s
(Kononenko, 2001):
1) Good pe o mance: high alues o diagnos ic accu acy is c ucial. The pe o mance o mos
algo i hms is a leas equal o physicians.
2) Dealing wi h missing da a: In some pa ien s eco ds ha e lack o in o ma ion, so, dealing wi h
incomple e desc ip ions o pa ien s is impo an .
3) Dealing wi h noisy da a: Unce ain y and e o s a e common in medical da a. The e o e, ML
applica ions mus ha e e ec i e means o handling noisy da a.
4) T anspa ency o diagnos ic knowledge: he p oblem can be p esen ed by he sys em in a di e en
and new poin o iew, anspa en o physician, no see be o e in an explici o m.
5) Explana ion abili y: diagnosis mus be p esen ed in a clea way when diagnosis new pa ien s.
6) Reduc ion he numbe o es s: pa ien his o y has a la ge amoun o da a. The classi ie mus be
able o eliably diagnose wi h a small amoun o da a abou he pa ien s.
3.8. Bayesian Ne wo ks
Bayesian ne wo ks (BN) a e g aph-based o malisms o he ep esen a ion and manipula ion o unce ain
knowledge, based on p obabili y heo y. They p o ide a p obabilis ic app oach o in e ence which allow
aking in o accoun p io knowledge when analysing da a (P. Lucas, 2004; Mi chell, 1997). They a e
impo an o machine lea ning because hey p o ide a quan i a i e app oach o weighing he e idence
suppo ing al e na i e hypo hesis and ep esen a join p obabili y dis ibu ion and domain (o expe )
knowledge in a compac way (Lee & Abbo , 2003; Mi chell, 1997). They consis o a quali a i e and
quan i a i e pa . The quali a i e pa encodes, in a di ec ed g aph, he a iables unde s udy wi h hei
p obabilis ic in e ela ionships. The quan i a i e pa is a se o condi ional p obabili ies desc ibing he
s eng hs o he dependences be ween a iables ep esen ed in he quali a i e pa . This pa s oge he a e
su icien o de ine a join p obabili y dis ibu ion on he s a is ical a iables unde s udy (Coupé & an
de Gaag, 2002).
BN a e one o he mos popula unce ain y o malisms because: hey can handle noise, missing
in o ma ion and e eal p obabilis ic ela ions; possibili ies lea n om da a and inco po a e domain
knowledge and p o ide a good in e ace h ough hei compac g aphical ep esen a ion. Ano he
ad an age is ha ne wo ks a e lexible and he lea ned models can be used o many asks like p edic ion
20 Backg ound
o diagnosis. One p ac ical di icul y in applying Bayesian me hods is ha hey ypically equi e ini ial
knowledge o many p obabili ies. A second p ac ical di icul y is he signi ican compu a ional cos
equi ed o de e mine he Bayes op imal hypo hesis in he gene al case (Mi chell, 1997).
Fo all he cha ac e is ics e e ed abo e and also ha hey a e a powe ul da a mining echnique o
handling unce ain y in complex domains and a undamen al echnique o pa e n ecogni ion and
classi ica ion, he use o BN a e inc easing in he choice o da a mining echniques applied in medical
domain (Lee & Abbo , 2003). They allow he s epwise combina ion o p ognos ic e idence and p o ide
a quan i a i e measu e in e ms o p obabili ies (Sakella opoulos & Niki o idis, 2000).
3.8.1. P obabilis ic easoning
To unde s and BN some basic p obabilis ic concep s ha e o be lea ned.
I
X
and
Y
a e andom a iables, wi h p obabili y dis ibu ions and , he join p obabili y
dis ibu ion ep esen s he dis ibu ion o bo h a iables ela ed. This way, o a gi en e en ,
he ma ginal p obabili y o , can be calcula ed:
(
)
(
)
.,
∑
===
Y
YxXPxXP
The condi ional p obabili y is a concep e y impo an in medicine and is de ined as an e en
gi en he occu ence o
( )
.
)(
),(
|YP
YXP
yYxXP ===
Gi en he p ope ies o he join dis ibu ion, he o me equa ion can be ew i e as a amous heo em.
The Bayes heo em p o ides a di ec me hod o calcula ing he pos e io p obabili ies o he a ious
hypo heses gi en obse a ions. Mo e p ecisely, Bayes heo em p o ides a way o calcula e he p obabili y
o a hypo hesis
(
)
DhP | based on i s p io p obabili y )(hP , he p obabili y o obse ing a ious da a gi en
he hypo hesis
(
)
hDP |, and he obse ed da a i sel (Mi chell, 1997), making i he co ne s one o Bayesian
lea ning me hods:
( )
.
)(
)()|(
|DP
hPhDP
DhP =
)(XP )(YP
),( YXP xX
=
xX
=
xX
=
:yY
=
In many scena ios, he lea ne is in e es ed in inding he mos p obable hypo heses h o a se o
candida e hypo heses H gi en he obse ed da a. The maximally p obable hypo hesis is called maximum a
pos e io i (MAP) hypo hesis:
(
)
(
)
.|maxa g hPhDPh
Hh
MAP
∈
=
In o he cases, we assume ha e e y hypo hesis in H is equally p obable a p io i. So, we jus need conside
he e m
(
)
hDP |, called he likelihood o he da a
D
gi en h, o ind he mos p obable hypo hesis, and
any hypo hesis ha maximizes
(
)
hDP | is called a maximum likelihood (ML) hypo hesis, ML
h:
(
)
.|maxa g hDPh
Hh
ML ∈
=
The mos p obable classi ica ion o he new ins ance is ob ained by combining he p edic ions o all
hypo heses, weigh ed by hei pos e io p obabili ies. The Bayes op imal classi ica ion o he new
ins ances j
is desc ibed by (Mi chell, 1997):
(
)
(
)
.||maxa g
∑
∈
DhPh P iij
V j
The Bayes op imal classi ie o Bayes op imal lea ne is e e y sys em ha classi ies new ins ances based in he
equa ion abo e.
When mul iple dependences a e s ake, condi ional independence is de ined as a logic ha suppo s
symbolic easoning abou dependence and independence in o ma ion, making i possible o abs ac away
om he nume ical de ail o p obabili y dis ibu ions and he p ocess o assessing p obabili y
dis ibu ions.
Le X, Y, Z be se s o a iables, X is condi ionally independen o Y gi en Z i :
(
)
).|(,| ZXPZYXP =
This ea u e is he co ne s one o BN lea ning.
3.8.2. BN: De ini ion, ep esen a ion
Bayesian ne wo k is a g aphical ep esen a ion o s ochas ic (s a is ical) dependences and independences
among a iables. The ype o dependence and independence we a e dealing wi h is de e mined by he
di ec ion o a cs and whe he o no pa icula a iables a e ins an ia ed. I is based on he assump ion ha
he classi ica ion o pa e ns is exp essed in p obabilis ic e ms be ween p edic o s and ou come a iables-

22 Backg ound
condi ional assump ion (Lee & Abbo , 2003). Clea ly, he assump ions ha a e made in o mula ing his
p io knowledge a e c ucial, and a e a opic o much deba e (P. Lucas, 2004).
On he BN ep esen a ion, g aphical in o ma ion is quali a i e, nodes ep esen s a iables and a cs speci y
he (in) dependence be ween a iables (Sakella opoulos & Niki o idis, 2000). In his way, each node is
independen o all i s non-descenden nodes, gi en i s pa en s. This causes he join p obabili y
dis ibu ion P is equi alen o he p oduc o he (condi ional) dis ibu ions (Sakella opoulos & Niki o idis,
2000):
( )
( )
.|,,,
1
21 ∏
=
=
n
i
xin i
xPxxxP
π
K
Whe e i
x
π
is he se o pa en s o he o ex co esponding o he a iable i
X.
So, A Bayesian ne wo k
Β
is de ined as a pai
(
)
PG,
=
Β
, whe e
(
)
(
)
(
)
GAGVG ,
=
is an acyclic di ec ed
g aph wi h a se o e ices (o nodes)
(
)
{
}
n
XXXGV ,,, 21 K
=
and a se o a cs
(
)
(
)
(
)
GVGVGA
×
⊆
, and whe e P is a join p obabili y dis ibu ion de ined on he a iables
co esponding o he e ices
(
)
GV . The basic p ope y o a Bayesian ne wo k is ha he join
dis ibu ion
(
)
n
XXXP ,,, 21 K is equi alen o he p oduc o he (condi ional) p obabili ies:
( )
( )
.|,,,
1
21 ∏
=
=
i
Xin i
XPXXXP
π
K
Thus, a e he (condi ional) p obabili y dis ibu ions wi ch a e speci ied o he a iable i
X,
o ,,,1 ni K
=
in c ea ing a Bayesian ne wo k (Coupé & an de Gaag, 2002).
3.8.3. Building BN
The cons uc ion o a BN has wo phases(Lee & Abbo , 2003):
1. C ea ion o a BN s uc u e: a acyclic g aph which encodes p obabilis ic ela ionships among
a iables;
2. Assessmen o he p io and local condi ional p obabili ies: aining and es ing he ne wo k
s uc u e.
In con as wi h logis ic eg ession, whe e dependence and independence is hidden in app oxima ing
weigh s, in BN s uc u e hese a e explici ly ep esen ed (Lee & Abbo , 2003).
(
)
i
Xi
XP
π
|
We can cons uc a BN manually o lea n om da a. Manual cons uc ion in p ac ice a e ime consuming
because equi es access o human expe s. Lea ning om da a in nowadays a e much mo e a ac i e,
consequence o he inc easing o clinical and biological da a (P. J. F. Lucas e al., 2004).
The manual way comp ehends a ious s ages using he expe knowledge, ele an li e a u e and analysis
o a ailable pa ien da a (P. J. F. Lucas e al., 2004). The i e main s ages a e:
1. Selec ion o ele an a iables: is gene ally based on in e iews wi h expe s, desc ip ions o he
domain and an ex ensi e analysis o he pu pose o he ne wo k unde cons uc ion.
2. Iden i ica ion o he ela ionships among he a iables: de e mine how hose ac o s a e ela ed o
each o he . Dependence and independence ela ionships be ween hem ha e o be analysed and
exp essed in a g aphical s uc u e. Causal g aph, common e ec s and causes.
3. Quali a i e p obabilis ic and logical cons ains: quali a i e p obabilis ic de i ed om p ope ies o
s ochas ic dominance o dis ibu ions. Logical cons ains a e de i ed om unc ional ela ionship
be ween he a iables.
4. Assessmen o p obabili ies: Local condi ional p obabili y dis ibu ions P (Xi|pi(Xi)) o each
a iable Xi a e illed in. The equi ed p obabili ies can be ob ained om domain expe s o ,
al e na i ely, om da a.
5. Sensi i i y analysis and e alua ion: o be used in eal-li e p ac ice, BN a e es ed and e alua ed.
One way o assess ne wo k’s quali y is o pe o m a sensi i i y analysis wi h pa ien da a. The e
a e a ious ways o e alua e BN like measu ing classi ica ion pe o mance on a gi en se o eal
pa ien da a and measu ing simila i y o s uc u e o p obabili y dis ibu ion o a gold-s anda d
ne wo k o o he p obabilis ic model.
3.8.4. Lea ning BN om da a
BN can be lea n om da a wi hou explici access o knowledge o human expe s by explo ing a ious
issues such as compa ison o lea ning algo i hms, dealing wi h missing da a and e alua ion o he
ne wo ks lea ned (P. J. F. Lucas e al., 2004).
To c ea e BN om da a wi h lea ning pu pose his ha e o sa is y some equisi es: da a collec ion is e y
impo an o a oid bias ha in e e e in he BN impac and pu pose; da a’s a iables and alues should
ma ch cha ac e is ics o be modelled in he ne wo k o should a leas admi easy ansla ion; he size o
he da a ha e o allow eliable in o ma ion o p obabilis ic ela ionships among a iables disce ned; mus
ha e p ope ies ha allows he use o he mos lea ning algo i hms.
One impo an aspec is ha many s a is ical and lea ning me hods canno deal wi h missing alues and
he absence o missing alues has o be ensu ed by wo ways: emo ing he cases wi h missing da a o
24 Backg ound
illing (impu ing) missing da a. The i s me hod has o be applied wi h cau ion as i can esul in he loss
o a la ge amoun o aluable da a, hus leading o a dec ease in he obus ness o he models lea ned. The
second one means ha he missing alue is eplaced wi h an es ima e o he ac ual alue (P. J. F. Lucas e
al., 2004).
Lea ning Bayesian ne wo ks in ol es bo h s uc u e lea ning, i.e., lea ning he g aph opology om da a,
and pa ame e lea ning, i.e., lea ning he ac ual, local p obabili y dis ibu ions om da a. The e a e
basically wo app oaches o s uc u e lea ning: sea ch and sco e s uc u e lea ning, and cons ain -based
s uc u e lea ning.
Sea ch-and-sco e algo i hms sea ch o a BN s uc u e ha i s he da a bes (in some sense). These
me hods sea ch he space o all possible acyclic dig aphs by gene a ing a ious di e en g aphs in a
heu is ic way and compa ing hese o hei abili y o explain ha a hand. They s a wi h an ini ial ne wo k
s uc u e (o en a g aph wi hou a cs o a comple e g aph), and hen a e se he sea ch space o ne wo k
s uc u es by in each s ep emo ing an a c, adding an a c, o e e sing an a c. Recen sea ch-and-sco e-
algo i hms ake Ma ko equi alence in o accoun , i.e., hey sea ch in he space o equi alence classes o
Bayesian ne wo ks and he sco ing me hod hey use gi e he same sco e o equi alen ne wo ks. Bayesian
ne wo ks wi h di e en g aph opologies ha a e included in he same Ma ko equi alence class ep esen
exac ly he same condi ional-independence in o ma ion by d-sepa a ion. Examples o sea ch and sco e
algo i hms a e K2 and inclusion-d i en lea ning. They usually a e based on hill climbing (g eedy) sea ch (P.
J. F. Lucas e al., 2004). K2 pe o ms a g eedy sea ch ha ades o ne wo k complexi y o accu acy o e
he aining da a (Mi chell, 1997).
Cons ain -based algo i hms ca y ou a condi ional (in) dependence analysis on he da a and allow o he
easy inco po a ion o backg ound knowledge, i.e., p io knowledge on dependences o independences ha
hold o he domain unde conside a ion. Examples o cons ain -based lea ning algo i hms a e PC, NPC,
g owsh ink, and inc emen al associa ion (P. J. F. Lucas e al., 2004).
3.8.5. Naï e Bayes classi ie
The naï e Bayes (NB) classi ie is based on he simpli ying assump ion ha he a ibu e alues a e
condi ionally independen gi en he a ge alue (Mi chell, 1997):
(
)
(
)
∏
∈
=
i
iij
V
NB aP P
j
|maxa g
One in e es ing di e ence be ween he naï e Bayes lea ning me hods and o he lea ning me hods is ha
he e is no explici sea ch h ough he space o possible hypo hesis (Mi chell, 1997). Naï e Bayes is obus
o he p esence o i ele an a ibu es(bu edundan a iables mus be aken in o accoun , as hey ha e
impac on pe o mance), he a iabili y o a da a se is summa ized in con ingency ables and he
dimension o he decision model is independen o he numbe o examples.
The gene al s uc u e o a naï e Bayesian ne wo k is shown o igu e 1.
Figu e 1: Nai e Bayes ne wo ks
F ep esen s he ea u es a iables and C he class a iable
3.8.6. T ee augmen ed Bayesian ne wo k
T ee augmen ed Bayesian ne wo k (TAN) is an ex ension o naï e Bayes: educing he numbe o
independen assump ions, each node has a mos wo dependences, one condi ionally om he class and
o he condi ionally om o he a ibu e (P. J. F. Lucas e al., 2004) ( igu e 2).
Figu e 2: T ee augmen ed Bayesian ne wo k
F ep esen s he ea u es a iables and C he class a iable
32 Ma e ial and Me hods
4.6. Models alida ion
We use he esul s o sensi i i y and speci ici y o de e mine he pe o mance o ou models and choose
di e en h esholds, acco dingly.
The LR model was e alua ed wi h sensi i i y and speci ici y es ima es on ain da a, and using 10 imes 2-
old c oss alida ion o check o ex e nal alida ion. All he analysis was pe o med wi h SPSS s a is ical
so wa e (SPSS, Inc, Chicago, IL, USA). To e alua e he BN pe o mance models we used lea e-one-ou
CV.
An ex e nal alida ion was made applying he inal models, LR and BN, on a second compa able coho .

Resul s 33
5. Resul s
5.1. Pa ien s
On he i s da a collec ion, used o c ea e he logis ic eg ession (LR) and Bayesian ne wo ks (BN)
models, om he 113 pa ien s conside ed o inclusion, 27 we e excluded o se e al easons depic ed in
igu e 3. We collec ed da a om 86 pa ien s, 69 (80%) o which we e male and mean age was 56 yea s.
Fo y one pa ien s (48%) had no mal esul wi h age mean o 54 yea s; o he 45 pa ien s wi h obs uc i e
sleep apnea (OSA) (52%), 17 (37%) we e ca ego ized in o mild, 15 (33%) we e mode a e and 13 (30%)
we e se e e, and he mean age was 57 yea s ( able 6).
Figu e 3: Flow diag am o inclusion o pa ien s in he s udy
The analysis o uni a ia e eg ession showed 6 a iables ( able 6) wi h signi ican odds a io (OR): male
gende (OR=7.259, 95% CI=[1.096; 27.651]), body mass index (OR=1.159, [1.030; 1.303]), neck
ci cum e ence (OR=1.341, [1.159; 1.550]), abdominal ci cum e ence (OR=1.076, [1.025; 1.129]),
wi nessed apneas (OR=4.725, [1.772; 12.599]) and alcohol be o e sleep (OR=3.307, [1.350; 8.100]).
34 Resul s
Table 6: Desc ip ion and odds a ios o he 33 s udied a iables
No mal (
n=41
)
OSA (
n=45
)
Simple
OR
95%CI
Gende , n (%)
Male 27 (66) 42 (93) Re .
Female 14 (34) 3 (6) 7.259 [1.096;27.651]
E hnici y, n (%)
Eu opean 40 (98) 44 (98) - -
A ican 1 (2) 1 (2) - -
Age, mean (sd) 54 (14) 57 (13) 1.020 [0.988;1.052]
Sno e, n (%) 41 (100) 45 (100) - -
Wi nessed apneas, n (%) 20 (49) 36 (82) 4.725 [1.772;12.599]
Gasping/Shocking, n (%) 7 (17) 14 (31) 2.194 [0.783;6.142]
Mo o Vehicle C ashes, n (%) 3 (8) 3 (7) 0.872 [0.165;4.608]
Re eshing Sleep, n (%) 17 (41) 22 (49) 1.350 [0.575;3.169]
Humo al e a ions, n (%) 2 (5) 2 (4) 0.907 [0.122;6.751]
Noc u ia, n (%) 16 (39) 16 (36) 0.862 [0.359;2.069]
Res less Sleep, n (%) 4 (10) 9 (20) 2.312 [0.653;8.185]
Dec eased libido, n (%) 0 (0) 1 (2) - -
Mo ning headaches, n (%) 10 (24) 8 (18) 0.670 [0.236;1.906]
Alcohol be o e sleep, n (%) 12 (29) 26 (58) 3.307 [1.350;8.100]
Smoke , n (%)
No 22 (54) 25 (56) Re .
Yes 10 (24) 7 (16) 0.616 [0.200;1.894]
Ex-smoke 9 (22) 13 (29) 1.27 [0.456;3.543]
Seda i e use, n (%) 8 (18) 9 (20) 1.031 [0.356;2.986]
ESS, median ( ange) 8 (19) 8 (24) 0.980 [0.908;1.050]
Concen a ion dec ease, n (%) 8 (19) 3 (7) 0.295 [0.072;1.198]
BMI, mean (sd) 28 (4) 30 (5) 1.159 [1.030;1.303]
NC, mean (sd) 39 (3.4) 43 (3.8) 1.341 [1.159;1.550]
AC, mean (sd) 100 (10) 108 (12) 1.076 [1.025;1.129]
C anio acial and uppe ai way abno mali ies, n (%) 17 (41) 28 (62) 2.325 [0.979;5.526]
A ial ib illa ion, n (%) 1 (2) 1 (2) 0.909 [0.055;15.020]
S oke, n (%) 1 (2) 2 (4) 1.860 [0.162;21.319]
Myoca dial in a c ion, n (%) 4 (9) 2 (4) 0.430 [0.075;2.484]
Pulmona y hype ension, n (%) 0 (0) 0 (0) - -
Conges i e hea ailu e, n (%) 1 (2) 0 (0) - -
Diabe es, n (%) 8 (19) 9 (20) 1.031 [0.356;2.986]
Me abolic Synd ome, n (%) 0 (0) 0 (0) - -
Renal ailu e, n (%) 0 0 - -
Hypo hy oidism, n (%) 3 (7) 0 - -
Gas oesophageal e lux disease, n (%) 2 (5) 3 (7) 1.393 [0.221;8.783]
Hype ension, n (%) 21 (51) 22 (49) 0.911 [0.391;2.124]
BMI: Body Mass Index; NC: Neck ci cum e ence; AC: Abdominal ci cum e ence; ESS: Epwo h Somnolence Scale; OR: Odds
Ra io; CI: Con idence In e al
5.2. Logis ic Reg ession Model
A mul iple o wa d condi ional logis ic eg ession analysis was c ea ed wi h neck ci cum e ence (NC),
gende , wi nessed apneas (WA) and consume o alcohol be o e sleep. Abdominal ci cum e ence (AC) and
body mass index (BMI) a iables we e no conside ed o his model gi en hei high co-linea i y wi h he
s onges a iable, NC. A e wo s eps, NC and WA we e he inal a iables p esen in he equa ion o
he mul i a ia e eg ession, wi h in e cep -11.147 and coe icien s 0.256 (OR=1.292) and 1.134
(OR=3.108), espec i ely.
The ROC cu e ( ig. 4) analysis demons a ed an AUC o 80%, wi h a con idence in e al (CI) o [70%;
89%]. Gi en he good disc imina i e powe o he model, a cu o alue o 10% (pa ien s wi h p obabili y
o OSA highe han 10% we e ecommended PSG) was chosen o achie e a sensi i i y o 100% [92%;
100%] and a speci ici y o 5% [1%; 15%]. Howe e , aiming a a no so s ic alue o sensi i i y, we could
ge be e esul s o speci ici y. Ac ually, wi h a cu o alue o 25%, a sensi i i y o 95% [86%; 99%] and
speci ici y o 35% [22%; 50%] we e achie ed. Inspec ing e oneous classi ica ions wi h he cu o alue o
25%, he wo misclassi ied OSA pa ien s we e ac ually diagnosed wi h mild OSA ( ep esen ing 12% o
o al mild OSA pa ien s). Fo he 10% cu o alue he c oss- alida ion esul s we e 98±3% sensi i i y and
11±3.5% speci ici y, while o he 25% cu o alue he esul ing sensi i i y was 89±4% and 34±7% o
speci ici y.
36 Resul s
Figu e 4: ROC cu e o he model
5.3. Bayesian Ne wo ks
Figu e 5 ep esen s he Bayesian ne wo k and he condi ional p obabili ies o he 6 signi ican a iables
wi h signi ican OR achie ed on uni a ia e LR, male gende , NC; AC, WA and alcohol be o e sleep,
c ea ed wi h no class in o ma ion.
Figu e 5: Bayesian ne wo k and he condi ional p obabili ies
(WA: Wi nessed Apneas, BMI: Body Mass Index; NC: Neck ci cum e ence; AC: Abdominal ci cum e ence)
As we ound on mul iple LR, he e is a high associa ion be ween AC, BMI and NC. BMI in luences NC
and AC, wi h ce ain ha an obese ha e AC inc eased and a highe p obabili y o ha e a NC inc eased,
P(NC|Obese)=0.84, when compa ed wi h no obese, P(AC|¬Obese)=0.85 and P(NC|¬Obese)=0.19.
NC in luences OSA and alcohol be o e sleep. The e is a high p obabili y ha a pa ien has OSA gi en
ha ing NC inc eased, P(OSA|NC)=0.75, when compa ed wi h an NC no mal, P(OSA|¬NC)=0.34.
OSA in luences gende , wi h a high p obabili y o a male ha ing OSA diagnosis, P(Male|OSA)=0.93 and
lowe o be no mal, P(Male|¬OSA)=0.65. O cou se cau ion is ad ised in in e p e ing such dependences
as causa ion.
Gende in luences WA and alcohol be o e sleep, wi h he p esence o hese wo a iables mo e likely in
men, P(WA|Male)=0.74 and P(Alcohol|Male)=0.36 han in women P(WA|Female)=0.29 and
P(Alcohol|Female)=0.09
P(Male|OSA)=0.93
P(Male|¬OSA)=0.65
P(NC|Obese)=0.84
P(NC|¬Obese)=0.19
P(Alcohol| Male)=0.36
P(Alcohol|Female)=0.09
P(WA|Male)=0.74
P(WA|Female)=0.29
P(OSA|NC)=0.75
P(OSA|¬NC)=0.34
P(Obese)=0.37
P(AC|Obese)=1.00
P(AC|¬Obese)=0.85

38 Resul s
5.3.1. Naï e Bayes ne wo k classi ie
Figu e 6 shows he NB based model wi h he ou come o PSG as label a ibu e.
Figu e 6: Nai e Bayes classi ie
(BMI: Body Mass Index; NC: Neck ci cum e ence; AC: Abdominal ci cum e ence)
Aiming 100% o sensi i i y, we achie e 7% as cu o o NB, ob aining a sensi i i y o 100% and 25% o
speci ici y while using a highe cu o o 10%, o 95% o sensi i i y, he esul s we e 98% and 33% o
sensi i i y and speci ici y espec i ely, o in e nal alida ion ( able 8).
The same cu o s we e used on lea e-one-ou c oss- alida ion (CV) o check he ex e nal alida ion o
NB. Using 7% as cu o , he esul s we e 98% o sensi i i y and 18% o speci ici y, while using he 10%
cu o he esul s we e 93% o sensi i i y and 30% o speci ici y ( able 8).
The ma ginal p obabili ies i no in o ma ion is gi en o he ne wo k a e shown on igu e 7.
Figu e 7: Ma ginal p obabili ies o Naï e Bayes classi ie o he p esence o OSA
(BMI: Body Mass Index; NC: Neck ci cum e ence; AC: Abdominal ci cum e ence)
5.3.2. T ee Augmen ed Bayesian ne wo k
The ee augmen ed Bayesian ne wo k (TAN) model is shown in igu e 8.
Figu e 8: T ee augmen ed Bayesian ne wo k
(WA: Wi nessed Apneas, BMI: Body Mass Index; NC: Neck ci cum e ence; AC: Abdominal ci cum e ence)
One mo e ime, as we saw on NB and LR, we e i y high associa ion be ween AC, NC and BMI. Only
h ee a iables ha e one mo e dependence beyond he class a iable ou come: alcohol and WA a e
in luenced also by gende , as we saw in he ne wo k wi hou class in o ma ion, and gende is in luenced by
AC.
40 Resul s
As we did o NB, we es ed di e en cu o s o achie e di e en le els o sensi i i y and speci ici y. Fo a
100% o sensi i i y we choose 2% as cu o and o 95% o sensi i i y we achie e a cu o o 22%. As
in e nal alida ion esul s, he model wi h 2% cu o had a sensi i i y o 100% and 28% o speci ici y and
using a highe cu o o 22%, he esul s we e 95% and 38% o sensi i i y and speci ici y espec i ely
( able 8).
On lea e-one-ou CV, o check he ex e nal alida ion, using he same cu o s desc ibed abo e, he 2%
cu o has 88% o sensi i i y and 23% o speci ici y while he highe cu o had 84% o sensi i i y and
25% o speci ici y ( able 8).
The ma ginal p obabili ies o TAN a e p esen ed on igu e 9.
Figu e 9: Ma ginal p obabili ies o TAN
(WA: Wi nessed Apneas, BMI: Body Mass Index; NC: Neck ci cum e ence; AC: Abdominal ci cum e ence)
5.4. Valida ion on a second compa able coho
To es he pe o mance o he model in clinical p ac ice, we collec ed a second coho wi h 33 pa ien s
( ig. 10).
Figu e 10: Flow diag am o inclusion o pa ien s in he second compa able s udy
As we expec ed, he e was a highe alue o no mal esul s (45%), mainly male gende (76%), wi h mean
age o 53 yea s ( able 7). O he 18 pa ien s wi h OSA (54%), 6 (33%) we e ca ego ized in o mild, 7 (39%)
we e mode a e and 5 (28%) we e se e e. S a is ical es ing ( able 7) showed ha he wo coho s a e
compa able wi h espec o he main s udied a iables (OSA, gende , wi nessed apneas, alcohol be o e
sleep, BMI, NC and AC).
Table 7: Cha ac e is ics o wo samples
*-Chi-squa e es **-Fische es ……***- es
OSA: Obs uc i e Sleep Apnea; BMI: Body Mass Index; NC: Neck ci cum e ence; AC: Abdominal ci cum e ence; BMI : Body
Mass Index ecoded; NC : Neck ci cum e ence ecoded; AC : Abdominal ci cum e ence ecoded.
Tes
T ain
p
N=33
N=86
OSA,
n (%)
18(55)
45(52)
0.833*
Gende ,
n (%)
0.634*
Male
25(76)
69(80)
Female
8(24)
17(20)
Age,
mean (
sd)
53(14)
56(13)
0.456***
Wi nessed apneas,
n (%)
21(64)
56(65)
0.469*
Alcohol be o e sleep,
n (%)
17(52)
38(44)
0.466*
BMI,
mean (sd)
29(5)
29(4)
0.839***
NC,
mean (sd)
41(3)
41(4)
0.879***
AC,
mean (sd)
105(12)
105(12)
0.744***
AC inc eased, n (%
)
31(94)
75(90)
0.723**
NC inc eased, n (%)
16(49)
36(43)
0.664*
BMI Obese, n (%)
12(37)
31(36)
0.956*
48 Discussion
Figu e 14: In e ence using TAN wi h missing in o ma ion o WA
(WA: Wi nessed Apneas, BMI: Body Mass Index; NC: Neck ci cum e ence; AC: Abdominal ci cum e ence).
One mo e ime, NB esul s we e close o he eal esul and e en wi hou in o ma ion o WA, he
p obabili y o OSA was low.
Suppose ha he same pa ien is ecommended o he sleep labo a o y o pe o m PSG by elemedicine o
we ha e access o he elec onic egis e s, bu no o he pa ien , and we wan o p io i ize based on he
in o ma ion gi en: only ha is a emale, wi h no mal BMI and alcohol consump ion be o e sleep. This
si ua ion shows he impo ance o hese models, dealing wi hou in o ma ion o mo e han one a iable,
when we ha e he impossibili y o measu ing some missing a iables ( ig. 15 and 16).
Figu e 15: In e ence using NB wi h missing in o ma ion o WA, AC and NC
(WA: Wi nessed Apneas, BMI: Body Mass Index; NC: Neck ci cum e ence; AC: Abdominal ci cum e ence).

Figu e 16: In e ence using TAN wi h missing in o ma ion o WA, AC and NC
(WA: Wi nessed Apneas, BMI: Body Mass Index; NC: Neck ci cum e ence; AC: Abdominal ci cum e ence).
Compa ing hese p obabili ies o he esul s o igu e 13 and 14, we e i y ha p obabili y o OSA
diagnosis dec ease using TAN (17.41%) and inc ease using NB (27.22%) wi h he lack o in o ma ion o
h ee a iables. I we used he p io i iza ion sugges ed abo e using TAN, his pa ien would be classi ied
as a non-p io i y g oup e en wi h only h ee pa ame e s (gende , alcohol and BMI). These examples
shows he bias o NB o classi y new cases wi h missing in o ma ion and he capabili y o TAN o deal
wi h hese si ua ions as i uses one mo e a iable dependence o classi y.
Besides he ad an ages desc ibed abo e on dealing wi h missing in o ma ion, he g aphical ep esen a ion
mus be seen as ano he capi al gain, mainly he TAN model, ha shows mo e han one dependence
be ween he a iables. This is an ad an age compa a i ely o LR based models ha don´ ha e his
capabili y.
6.3. Limi a ions
Some ac o s we e no possible o assess due o he lack o ep esen a i eness in he sample (e hnici y,
sno e, dec eased libido, pulmona y hype ension, conges i e hea ailu e, me abolic synd ome, enal
ailu e and hypo hy oidism) which may ha e led o a somewha biased model. Also, because ou s udy was
conduc ed on pa ien s e e ed by p ima y ca e physicians o he sleep consul , he p e alence o OSA in
ou sample (52%) was highe han o gene al popula ions. Hence, no OSA p e alence es ima e can be
in e ed.
To ecode he con inuous a iables, NC, AC and BMI, we used measu es ha we ound on li e a u e, bu
he e a e no s anda d alues o ca ego ized alues in o no mal o al e ed, so hese may lead o some e o s
in bo de line cha ac e is ics. He e we used alue highe han 30 o ecode in o obese, some au ho s can
50 Discussion
conside ed 25 (p e-obese). To ecode NC and AC we chose he mos e e ed and consensual alues
c i e ia on li e a u e bu o he could exis .
O he ques ion is he me ic used o classi y OSA. As we explained on backg ound some au ho s
ques ioned i AHI is he mo e accu a e measu e o classi y OSA se e i y. Some sugges he use o RDI in
al e na i e o AHI.
Conclusions and ecommenda ions 51
7. Conclusions and ecommenda ions
In his s udy he main cha ac e is ics o obs uc i e sleep apnea (OSA) we e body mass index, neck
ci cum e ence, abdominal ci cum e ence, gende , wi nessed apneas and consume o alcohol be o e sleep.
We used wo di e en echniques o cons uc he models, one based on logis ic eg ession (LR) and o he
on Bayesian ne wo ks (BN). Using hese wo echniques, LR and BN, we did no aim o compa e he wo
models di ec ly, bu a he show hei esul s o acili a e choices and possibly, complemen he wo
me hods. They mus be seen has me hods ha suppo decisions bu do no subs i u e he physician ha
has always, acco ding o he clinical his o y o he pa ien , he las decision.
Wi h LR, he inal model used only wo a iables on he eg ession equa ion: neck ci cum e ence and
wi nessed apneas. The g ea limi a ion on he applica ion o his app oach is he use o WA, since i is
subjec i e and in some cases impossible o measu e.
The g ea ad an ages o BN a e he ac ha hey can deal wi h missing in o ma ion and he g aphical
ep esen a ion ha shows no only he alues o p obabili ies gi en he pa ien cha ac e is ics, bu also
ep esen s he ela ionship be ween a iables. This can be an al e na i e o he adi ional s a is ical
measu e, odds a io (OR), ha can be in e p e ed as a ela i e isk o disease in exposed o no exposed
pa ien s.
As we did no ind a alida ed model, es ed in Po uguese sleep labo a o ies, we hink ha ou models
consis in a alid me hod o sc een pa ien s wi h suspicion o OSA, be o e pe o ming PSG.
The g ea ad an age o ou solu ions is p io i izing OSA suspicion pa ien s in o di e en le els o p io i y
acco ding o hei cha ac e is ics, and consequen ly, hei p obabili y o con i ming he OSA diagnosis.
O he ad an age is ha he sys em could manage wai ing lis s au oma ically in consul a ion when he
physician inse s pa ien da a.
E en ually, we can educe he numbe o no mal esul exams, op imizing he a ailable esou ces and
making su e ha no se e e case wai s much o ime and, consequen ly, ea men .
Fu u e wo k 52
8. Fu u e wo k
Would be in e es ing o es hese models in a mul i-cen e sleep labo a o ies s udy o compa e ou esul s
o o he s be o e implemen ing a decision suppo sys em ha can be used du ing p e-polysomnog aphy
consul a ion. This clinical decision suppo sys em could be based on mul iple models, es ed in his wo k,
like logis ic eg ession and Bayesian ne wo ks using naï e Bayes and ee augmen ed Bayesian ne wo k.
As he sleep labo a o ies ha e also a high numbe o pa ien s e e ed by he p ima y ca e, we could
pe o m ano he s udy o implemen a decision suppo sys em o help p ima y ca e physicians wi h he
decision o send pa ien s o sleep consul a ion and in his way, p io i ize wai ing lis s o consul a ion.
Re e ences 53
9. Re e ences
Ahmadi, N., Chung, S. A., Gibbs, A., & Shapi o, C. M. (2008). The Be lin ques ionnai e o sleep apnea in
a sleep clinic popula ion: ela ionship o polysomnog aphic measu emen o espi a o y
dis u bance. [Compa a i e S udy]. Sleep B ea h, 12(1), 39-45.
Al Lawa i, N. M., Pa el, S. R., & Ayas, N. T. (2009). Epidemiology, Risk Fac o s, and Consequences o
Obs uc i e Sleep Apnea and Sho Sleep Du a ion. P og ess in Ca dio ascula Diseases, 51(4), 285-
293.
A onsky, D., & Haug, P. J. (2000). Au oma ic iden i ica ion o pa ien s eligible o a pneumonia guideline.
[E alua ion S udies Resea ch Suppo , Non-U.S. Go ' ]. P oc AMIA Symp, 12-16.
Be y, D. (2008). Case s udy: obs uc i e sleep apnea. [Case Repo s Re iew]. Medsu g Nu s, 17(1), 11-16;
quiz 17.
Boneka , H. W., & Ha din, K. A. (2003). Se e e uppe ai way obs uc ion du ing sleep. [Re iew]. Clin Re
Alle gy Immunol, 25(2), 191-210.
Bu nside, E., Rubin, D., & Shach e , R. (2000). A Bayesian ne wo k o mammog aphy. [Resea ch
Suppo , U.S. Go ' , P.H.S.]. P oc AMIA Symp, 106-110.
Bu nside, E. S. (2005). Bayesian ne wo ks: compu e -assis ed diagnosis suppo in adiology. [Resea ch
Suppo , Non-U.S. Go ' Re iew]. Acad Radiol, 12(4), 422-430.
Chung, F., Yegneswa an, B., Liao, P., Chung, S. A., Vai a ana han, S., Islam, S., Shapi o, C. M. (2008).
Valida ion o he Be lin ques ionnai e and Ame ican Socie y o Anes hesiologis s checklis as
sc eening ools o obs uc i e sleep apnea in su gical pa ien s. [Resea ch Suppo , Non-U.S.
Go ' ]. Anes hesiology, 108(5), 822-830.
Collop, N. A., Ande son, W. M., Boehlecke, B., Claman, D., Goldbe g, R., Go lieb, D. J., Schwab, R.
(2007). Clinical guidelines o he use o una ended po able moni o s in he diagnosis o
obs uc i e sleep apnea in adul pa ien s. Po able Moni o ing Task Fo ce o he Ame ican
Academy o Sleep Medicine. J Clin Sleep Med, 3(7), 737-747.
Coupé, V. M. H., & an de Gaag, L. C. (2002). P ope ies o Sensi i i y Analysis o Bayesian Belie
Ne wo ks. Annals o Ma hema ics and A i icial In elligence, 36(4), 323-356.
Da ies, R. J., Ali, N. J., & S adling, J. R. (1992). Neck ci cum e ence and o he clinical ea u es in he
diagnosis o he obs uc i e sleep apnoea synd ome. Tho ax, 47(2), 101-105.
Dixon, J. B., Schach e , L. M., & O’B ien, P. E. (2003). P edic ing Sleep Apnea and Excessi e Day
Sleepiness in he Se e ely Obese*. Ches , 123(4), 1134-1141.

54 Re e ences
Di ecção Ge al da Saúde ( Pla a o ma con a a obesidade). A ailable:
(h p://www.pla a o macon aaobesidade.dgs.p /P esen a ionLaye /calculado es.aspx?menuid=1
61&exmenuid=-1&calc=2.). [Accessed on 27 o May, 2012].
Dogh amji, P. P. (2008). Recogni ion o obs uc i e sleep apnea and associa ed excessi e sleepiness in
p ima y ca e. J Fam P ac , 57(8 Suppl), S17-23.
D eisei l, S., & Ohno-Machado, L. (2002). Logis ic eg ession and a i icial neu al ne wo k classi ica ion
models: a me hodology e iew. [Compa a i e S udy Re iew]. J Biomed In o m, 35(5-6), 352-359.
Eps ein, L. J., K is o, D., S ollo, P. J., J ., F iedman, N., Malho a, A., Pa il, S. P., Adul Obs uc i e Sleep
Apnea Task Fo ce o he Ame ican Academy o Sleep, M. (2009). Clinical guideline o he
e alua ion, managemen and long- e m ca e o obs uc i e sleep apnea in adul s. [P ac ice
Guideline]. J Clin Sleep Med, 5(3), 263-276.
Fe be , R., Millman, R., Coppola, M., Flee ham, J., Mu ay, C. F., Ibe , C. (1994). Po able eco ding in he
assessmen o obs uc i e sleep apnea. ASDA s anda ds o p ac ice. Sleep, 17(4), 378-392.
Flemons, W. W., Li ne , M. R., Rowley, J. A., Gay, P., Ande son, W. M., Hudgel, D. W., Loube, D. I.
(2003). Home diagnosis o sleep apnea: a sys ema ic e iew o he li e a u e. An e idence e iew
cosponso ed by he Ame ican Academy o Sleep Medicine, he Ame ican College o Ches
Physicians, and he Ame ican Tho acic Socie y. Ches , 124(4), 1543-1579.
Flemons, W. W., Whi elaw, W. A., B an , R., & Remme s, J. E. (1994). Likelihood a ios o a sleep apnea
clinical p edic ion ule. Ame ican Jou nal o Respi a o y and C i ical Ca e Medicine, 150(5 I), 1279-1285.
Fogel, R. B., & Whi e, D. P. (2000). Obs uc i e sleep apnea. [Resea ch Suppo , Non-U.S. Go ' Re iew].
Ad In e n Med, 45, 351-389.
Gami, A. S., P essman, G., Caples, S. M., Kanagala, R., Ga d, J. J., Da ison, D. E., Some s, V. K. (2004).
Associa ion o a ial ib illa ion and obs uc i e sleep apnea. [Compa a i e S udy Resea ch
Suppo , Non-U.S. Go ' Resea ch Suppo , U.S. Go ' , P.H.S.]. Ci cula ion, 110(4), 364-367.
Gu ubhaga a ula, I., Maislin, G., & Pack, A. I. (2001). An Algo i hm o S a i y Sleep Apnea Risk in a
Sleep Diso de s Clinic Popula ion. Am. J. Respi . C i . Ca e Med., 164(10), 1904-1909.
Gus, M., Goncal es, S. C., Ma inez, D., de Ab eu Sil a, E. O., Mo ei a, L. B., Fuchs, S. C., & Fuchs, F.
D. (2008). Risk o Obs uc i e Sleep Apnea by Be lin Ques ionnai e, bu no day ime sleepiness,
is associa ed wi h esis an hype ension: a case-con ol s udy. [Resea ch Suppo , Non-U.S.
Go ' ]. Am J Hype ens, 21(7), 832-835.
Hamil on, P. W., Mon i oni, R., Abmay , W., Bibbo, M., Ande son, N., Thompson, D., & Ba els, P. H.
(1995). Clinical applica ions o Bayesian belie ne wo ks in pa hology. [Resea ch Suppo , U.S.
Go ' , Non-P.H.S. Resea ch Suppo , U.S. Go ' , P.H.S. Re iew]. Pa hologica, 87(3), 237-245.
Ho s ein, V., & Szalai, J. P. (1993). P edic i e alue o clinical ea u es in diagnosing obs uc i e sleep
apnea. Sleep, 16(2), 118-122.
Ibe C, A.-I. S., Chesson AL, Quan SF. . (2007). The AASM manual o he sco ing o sleep and associa ed e en s:
ules, e minology and echnical speci ica ions. . Wes ches e : IL: Ame ican Academy o Sleep Medicine.
Jennum, P., & Riha, R. L. (2009). Epidemiology o sleep apnoea/hypopnoea synd ome and sleep-
diso de ed b ea hing. Eu Respi J, 33(4), 907-914.
Johns, M. W. (1991). A new me hod o measu ing day ime sleepiness: he Epwo h sleepiness scale. Sleep,
14(6), 540-545.
Kaimakamis, E., B a sas, C., Sichle idis, L., Ka ounis, C., & Magla e as, N. (2009, 3-6 Sep . 2009).
Sc eening o pa ien s wi h obs uc i e sleep Apnea synd ome using C4.5 algo i hm based on non linea analysis o
espi a o y signals du ing sleep. Pape p esen ed a he Enginee ing in Medicine and Biology Socie y,
2009. EMBC 2009. Annual In e na ional Con e ence o he IEEE.
Kapu , V. K. (2010). Obs uc i e sleep apnea: diagnosis, epidemiology, and economics. Respi Ca e, 55(9),
1155-1167.
Kohle , M. (2009). Risk ac o s and ea men o obs uc i e sleep apnea amongs obese child en and
adul s. Cu Opin Alle gy Clin Immunol, 9(1), 4-9.
Kononenko, I. (2001). Machine lea ning o medical diagnosis: his o y, s a e o he a and pe spec i e.
A i icial In elligence in Medicine, 23(1), 89-109.
K iege , A. C., & Redeke , N. S. (2002). Obs uc i e sleep apnea synd ome: i s ela ionship wi h
hype ension. [Re iew]. J Ca dio asc Nu s, 17(1), 1-11.
K ug, P. (1999). Sno ing and obs uc i e sleep apnea. [Re iew]. AORN J, 69(4), 792-801.
K yge , M. R. T., Demen W. (2005). P inciples and P a ices o Sleep Medicine (Fou h ed.): Elsi ie Saunde s.
Kuhlmann, U., Bo mann, F. G., & Becke , H. F. (2009). Obs uc i e sleep apnoea: clinical signs, diagnosis
and ea men . [Commen Edi o ial Re iew]. Neph ol Dial T ansplan , 24(1), 8-14.
Kwia kowska, M., A kins, M. S., Ayas, N. T., & Ryan, C. F. (2007). Knowledge-Based Da a Analysis: Fi s
S ep Towa d he C ea ion o Clinical P edic ion Rules Using a New Typicali y Measu e.
In o ma ion Technology in Biomedicine, IEEE T ansac ions on, 11(6), 651-660.
Lam, J. C., & Ip, M. S. (2010). Sleep & he me abolic synd ome. [Re iew]. Indian J Med Res, 131, 206-216.
Lam, J. C., Sha ma, S. K., & Lam, B. (2010). Obs uc i e sleep apnoea: de ini ions, epidemiology & na u al
his o y. Indian J Med Res, 131, 165-170.
La ac, N. (2001). Da a mining and decision suppo : a no e on he issues o hei in eg a ion and hei
ela ion o expe sys ems. In eg a ing aspec s o da a mining, decision suppo and me a-lea ning (IDDM-
2001), 1-8.
Lee, S. M., & Abbo , P. A. (2003). Bayesian ne wo ks o knowledge disco e y in la ge da ase s: basics o
nu se esea che s. [Re iew]. J Biomed In o m, 36(4-5), 389-399.
Lucas, P. (2004). Bayesian analysis, pa e n analysis, and da a mining in heal h ca e. [Re iew]. Cu Opin
C i Ca e, 10(5), 399-403.
Lucas, P. J. F., an de Gaag, L. C., & Abu-Hanna, A. (2004). Bayesian ne wo ks in biomedicine and
heal h-ca e. A i icial In elligence in Medicine, 30(3), 201-214.
Madani, M., & Madani, F. (2009). Epidemiology, pa hophysiology, and clinical ea u es o obs uc i e
sleep apnea. O al Maxillo ac Su g Clin No h Am, 21(4), 369-375.
56 Re e ences
Maislin, G., Pack, A. I., K ibbs, N. B., Smi h, P. L., Schwa z, A. R., Kline, L. R., Dinges, D. F. (1995). A
su ey sc een o p edic ion o apnea. Sleep, 18(3), 158-166.
Manbe , R., & A mi age, R. (1999). Sex, s e oids, and sleep: a e iew. Sleep, 22(5), 540-555.
Mi chell, T. (1997). Machine lea ning: McG aw Hill.
Mon i oni, R., Ba els, P. H., Thompson, D., Sca pelli, M., & Hamil on, P. W. (1995). P os a ic
in aepi helial neoplasia (PIN). Pe o mance o Bayesian belie ne wo k o diagnosis and g ading.
[Resea ch Suppo , Non-U.S. Go ' Resea ch Suppo , U.S. Go ' , P.H.S.]. J Pa hol, 177(2), 153-
162.
Ne ze , N. C., S oohs, R. A., Ne ze , C. M., Cla k, K., & S ohl, K. P. (1999). Using he Be lin
Ques ionnai e o iden i y pa ien s a isk o he sleep apnea synd ome. [Resea ch Suppo , Non-
U.S. Go ' Resea ch Suppo , U.S. Go ' , Non-P.H.S. Resea ch Suppo , U.S. Go ' , P.H.S.]. Ann
In e n Med, 131(7), 485-491.
Nuck on, T. J., Glidden, D. V., B owne , W. S., & Claman, D. M. (2006). Physical examina ion:
Mallampa i sco e as an independen p edic o o obs uc i e sleep apnea. Sleep, 29(7), 903-908.
Olson, E. J., Pa k, J. G., & Mo gen hale , T. I. (2005). Obs uc i e sleep apnea-hypopnea synd ome.
[Re iew]. P im Ca e, 32(2), 329-359.
Pagel, J. F. (2008). The bu den o obs uc i e sleep apnea and associa ed excessi e sleepiness. [Re iew]. J
Fam P ac , 57(8 Suppl), S3-8.
Palme , L. J., & Redline, S. (2003). Genomic app oaches o unde s anding obs uc i e sleep apnea.
[Resea ch Suppo , U.S. Go ' , P.H.S.
Re iew]. Respi Physiol Neu obiol, 135(2-3), 187-205.
Pa k, J. G., Rama , K., & Olson, E. J. (2011). Upda es on de ini ion, consequences, and managemen o
obs uc i e sleep apnea. [Re iew]. Mayo Clin P oc, 86(6), 549-554.
Pe ie, A., & Sabin, C. (2009). Medical S a is ics a a Glance: Wiley-Blackwell.
Pilla , G., & La ie, P. (2011). Obs uc i e sleep apnea: diagnosis, isk ac o s, and pa hophysiology.
[Re iew]. Handb Clin Neu ol, 98, 383-399.
Poulio , Z., Pe e s, M., Neu eld, H., & K yge , M. H. (1997). Using sel - epo ed ques ionnai e da a o
p io i ize OSA pa ien s o polysomnog aphy. Sleep, 20(3), 232-236.
P ac ice pa ame e s o he use o po able eco ding in he assessmen o obs uc i e sleep apnea.
S anda ds o P ac ice Commi ee o he Ame ican Sleep Diso de s Associa ion. (1994). Sleep,
17(4), 372-377.
Punjabi, N. M. (2008). The epidemiology o adul obs uc i e sleep apnea. P oc Am Tho ac Soc, 5(2), 136-
143.
Rech scha en, A. K. a. A. (1968). A Manual o s anda dized e minology, ecniques and sco ing sys em
o sleep s ages o human subjec s.
Redline, S., Budhi aja, R., Kapu , V., Ma cus, C. L., Ma eika, J. H., Meh a, R., Quan, S. F. (2007). The
sco ing o espi a o y e en s in sleep: eliabili y and alidi y. J Clin Sleep Med, 3(2), 169-200.
Rodsu i, J., Hensley, M., Thakkins ian, A., D'Es e, C., & A ia, J. (2004). A clinical decision ule o
p io i ize polysomnog aphy in pa ien s wi h suspec ed sleep apnea. Sleep, 27(4), 694-699.
Sakella opoulos, G. C., & Niki o idis, G. C. (1999). De elopmen o a Bayesian Ne wo k o he
p ognosis o head inju ies using g aphical model selec ion echniques. Me hods In Med, 38(1), 37-
42.
Sakella opoulos, G. C., & Niki o idis, G. C. (2000). P ognos ic pe o mance o wo expe sys ems based
on Bayesian belie ne wo ks. Decision Suppo Sys ems, 27(4), 431-442.
Schu ink, C. A., Vissche , S., Lucas, P. J., an Leeuwen, H. J., Buskens, E., Ho , R. G., Bon en, M. J.
(2007). A Bayesian decision-suppo sys em o diagnosing en ila o -associa ed pneumonia.
[Resea ch Suppo , Non-U.S. Go ' ]. In ensi e Ca e Med, 33(8), 1379-1386.
Schwab, R. J. (2005). Gene ic de e minan s o uppe ai way s uc u es ha p edispose o obs uc i e sleep
apnea. [Resea ch Suppo , N.I.H., Ex amu al Resea ch Suppo , U.S. Go ' , P.H.S. Re iew].
Respi Physiol Neu obiol, 147(2-3), 289-298. doi: 10.1016/j. esp.2005.06.006
Scu a i, M. (2010). Lea ning Bayesian ne wo ks wi h he bnlea n R package. Jou nal o S a is ical So wa e,
35(3), 1-20.
Sheldon, A., Belan, I., Neill, J., & Rowland, S. (2009). Nu sing assessmen o obs uc i e sleep apnea in
hospi alised adul s: a e iew o isk ac o s and sc eening ools. [Re iew]. Con emp Nu se, 34(1), 19-
33.
Silbe , M. H., Ancoli-Is ael, S., Bonne , M. H., Chok o e y, S., G igg-Dambe ge , M. M., Hi shkowi z,
M., Ibe , C. (2007). The isual sco ing o sleep in adul s. J Clin Sleep Med, 3(2), 121-131.
Sleep- ela ed b ea hing diso de s in adul s: ecommenda ions o synd ome de ini ion and measu emen
echniques in clinical esea ch. The Repo o an Ame ican Academy o Sleep Medicine Task
Fo ce. (1999). Sleep, 22(5), 667-689.
Sleep- ela ed b ea hing diso de s in adul s: ecommenda ions o synd ome de ini ion and measu emen echniques in clinical
esea ch. The Repo o an Ame ican Academy o Sleep Medicine Task Fo ce. (1999). (1999/08/18 ed. Vol.
22).
S ie e , T., & Punjabi, N. M. (2005). Demog aphics and diagnosis o obs uc i e sleep apnea. Anes hesiol
Clin No h Ame ica, 23(3), 405-420.
Sun, L., Chiu, H.-W., Chuang, C., & Liu, L. (2010). A p edic ion model based on an a i icial in elligence
sys em o mode a e o se e e obs uc i e sleep apnea. Sleep and B ea hing, 1-7.
Tak ak, J., Ajmi Nabli, T., Ben O hmen, H., M i aoui, A., & Ben Hadj Hamida, F. (2011). [The apeu ic
compliance in p ima y open angle glaucoma]. Tunis Med, 89(2), 142-146.
Tasali, E., Mokhlesi, B., & Van Cau e , E. (2008). Obs uc i e sleep apnea and ype 2 diabe es: in e ac ing
epidemics. [Re iew]. Ches , 133(2), 496-506.
Ta e, J., & Taso a, F. J. (2002). Mo e han a sno e: ecognizing he dange o sleep apnea. [Case Repo s
Re iew]. Nu sing, 32(8), 46-49.
Togei o, S. M., Cha es, C. M., J ., Palombini, L., Tu ik, S., Ho a, F., & Ne y, L. E. (2010). E alua ion o
he uppe ai way in obs uc i e sleep apnoea. [Re iew]. Indian J Med Res, 131, 230-235.
64 A achmen s
SPECIFICATIONS
Digi al Speci ica ions o Rou ine PSG Reco dings
Sampling a es
F equency
EEG
500Hz
EOG
500Hz
EMG
500Hz
ECG
500Hz
Ai low
100Hz
Oxime y
25Hz
Nasal p essu e
100Hz
Body posi ion
1Hz
Sno ing
500Hz
Rib Cage and Abdominal Mo emen s
100Hz
Fil e se ings
Low equency il e
High equency il e
EEG
0.3Hz
35Hz
EOG
0.3Hz
35Hz
EMG
10Hz
100Hz
ECG
0.3
Hz
70Hz
Respi a ion
0.1Hz
15Hz
Sno ing
10Hz
100Hz

Epwo h Somnolence Scale (ESS)
How likely a e you o doze o o all asleep in he ollowing si ua ions, in con as o eeling jus i ed?
This e e s o you usual way o li e in ecen imes.
E en i you ha e no done some o hese hings ecen ly, y o wo k ou how hey would ha e a ec ed
you. Use he ollowing scale o choose he mos app op ia e numbe o each si ua ion:
0 = Would NEVER doze
1 = SLIGHT chance o dozing
2 = MODERATE chance o dozing
3 = HIGH chance o dozing
Si ua ion
Change o dozing
Si ing and eading
Wa ching ele ision
Si ing, inac i e in a public place
( o example, a hea e o a mee ing)
As a passenge in a ca
o an hou wi hou a b eak
Lying down o es in he a e noon
when ci cums ances pe mi
Si ing and alking o someone
Si ing quie ly a e a lunch
wi hou alcohol
In a ca , while s opped o a ew
minu es in a ic
Each ques ion is sco ed om 0 o 3, gi ing a maximum sco e o 24.
66 A achmen s
C
CO
ON
NS
SE
EN
NT
TI
IM
ME
EN
NT
TO
O
I
IN
NF
FO
OR
RM
MA
AD
DO
O
Liliana Pa ícia Pin o Lei e, aluna de mes ado em In o má ica Médica da Faculdade de Medicina da
Uni e sidade do Po o p e ende ealiza in es igação, no âmbi o da sua disse ação, ecolhendo dados de
indi íduos suge idos pa a ealiza em polissonog a ia no labo a ó io de es udos do sono do Cen o
Hospi ala de Vila No a de Gaia/Espinho, EPE.
 O p esen e es udo em como objec i o a cons ução de á ios modelos, cons uídos que com
base em ca ac e ís icas dos doen es que a a és de e amen as de da a mining e a compa ação
dos seus esul ados de o ma a a alia qual o modelo que ob ém maio alidade, op imizando a
sensibilidade pa a a p edição dos casos mais p o á eis de doença.
 A ecolha de dados é e ec uada uma ez aquando a ealização da polissonog a ia;
 Não es ão p esen es bene ícios ou iscos pa a o sujei o;
 Se á man ida a con idencialidade de odos os dados ela i os ao sujei o;
 O sujei o pode á desis i da pa icipação na in es igação em qualque al u a, sem e de da
explicações, ap esen a desculpas ou eembolsa despesas;
 O in es igado usa á de anqueza du an e odo o p ocesso, limi ando o conhecimen o do sujei o
aos dados po ele designados como undamen ais ace ao objec i o da expe iência.
Eu, abaixo assinado_______________________________________________________
decla o que en endo os objec i os, ca ac e ís icas e du ação do es udo e que é de minha li e on ade que
pa icipo no mesmo.
_________________________________
____/____/_______
Au ho iza ion o s udy ealiza ion