This is a sel -a chi ed e sion o an o iginal a icle. This e sion
may di e om he o iginal in pagina ion and ypog aphic de ails.
Au ho (s):
Ti le:
Yea :
Ve sion:
Copy igh :
Righ s:
Righ s u l:
Please ci e he o iginal e sion:
CC BY-NC-ND 4.0
h ps://c ea i ecommons.o g/licenses/by-nc-nd/4.0/
P edic ing Child en's Myopia Risk : A Mon e Ca lo App oach o Compa e he
Pe o mance o Machine Lea ning Models
© 2024 SCITEPRESS
Published e sion
A iemjew, Pio ; Cybulski, Radosław; Emamian, Mohammad; G zybowski,
And zej; Jankowski, And zej; Lanca, Ca la; Meh a a an, Shi a; Młyński, Ma cin;
Mo awski, Ceza y; No dhausen, Klaus; Pä ssinen, Ola i; Ropiak, K zysz o
A iemjew, P., Cybulski, R., Emamian, M., G zybowski, A., Jankowski, A., Lanca, C., Meh a a an,
S., Młyński, M., Mo awski, C., No dhausen, K., Pä ssinen, O., & Ropiak, K. (2024). P edic ing
Child en's Myopia Risk : A Mon e Ca lo App oach o Compa e he Pe o mance o Machine
Lea ning Models. In A. P. Rocha, L. S eels, & J. V. D. He ik (Eds.), ICAART 2024 : P oceedings o
he 16 h In e na ional Con e ence on Agen s and A i icial In elligence, Volume 3 (pp. 1092-
1099). SCITEPRESS Science and Technology Publica ions.
h ps://doi.o g/10.5220/0012435500003636
2024
P edic ing Child en’s Myopia Risk: A Mon e Ca lo App oach o
Compa e he Pe o mance o Machine Lea ning Models
Pio A iemjew1 a, Radosław Cybulski1 b, Mohammad Hassan Emamian2 c,
And zej G zybowski3 d, And zej Jankowski1, Ca la Lanca4,8 e, Shi a Meh a a an5 ,
Ma cin Mły´
nski1 g, Ceza y Mo awski1, Klaus No dhausen6 h, Ola i P¨
a ssinen7 i
and K zysz o Ropiak1 j
1Uni e si y o Wa mia and Mazu y in Olsz yn, Poland
2Oph halmic Epidemiology Resea ch Cen e , Shah oud Uni e si y o Medical Sciences, Shah oud, I an
3Founda ion o Oph halmology De elopmen & Uni e si y o Wa mia and Mazu y, Poland
4Lisbon School o Heal h Technology, Lisbon, Po ugal
5Depa men o Biology, School o Compu e , Ma hema ical, and Na u al Sciences, Mo gan S a e Uni e si y, U.S.A.
6Depa men o Ma hema ics and S a is ics, Uni e si y o Jy ¨
askyl¨
a, Finland
7Ge on ology Resea ch Cen e and Facul y o Spo and Heal h Sciences, Uni e si y o Jy ¨
askyl¨
a, Finland
8Comp ehensi e Heal h Resea ch Cen e (CHRC), Escola Nacional de Sa´
ude P´
ublica, Uni e sidade No a de Lisboa,
Lisboa, Po ugal
[email p o ec ed], adoslaw[email p o ec ed], [email p o ec ed], ae[email p o ec ed],
[email p o ec ed], [email p o ec ed], shi a.meh a a an@mo gan.edu,
ma [email p o ec ed], ceza y[email p o ec ed], klaus.k.no [email p o ec ed], [email p o ec ed],
k [email p o ec ed]
Keywo ds: Myopia P edic ion, Machine Lea ning, Da a Analysis, Mon e Ca lo Simula ions, Lasso Reg ession.
Abs ac : This s udy p esen s he ini ial esul s o he Myopia Risk Calcula o (MRC) Conso ium, in oducing an inno-
a i e app oach o p edic myopia isk by using us wo hy machine-lea ning models. The da ase included
app oxima ely 7,945 eco ds (eyes) om 3,989 child en. We de eloped a myopia isk calcula o and an ac-
companying web in e ace. Cen al o ou esea ch is he challenge o model us wo hiness, speci ically
e alua ing he e ec i eness and obus ness o AI (A i icial In elligence)/ML (Machine Lea ning)/NLP (Na -
u al Language P ocessing) models. We adop ed a obus me hodology combining Mon e Ca lo simula ions
wi h c oss- alida ion echniques o assess model pe o mance. Ou expe imen s e ealed ha an ensemble
o classi ie s and eg ession models wi h Lasso eg ession echniques p o ided he bes ou comes o p edic -
ing myopia isk. Fu u e esea ch aims o enhance model accu acy by in eg a ing image and syn he ic da a,
including ad anced Mon e Ca lo simula ions.
ah ps://o cid.o g/0000-0001-5508-9856
bh ps://o cid.o g/0000-0003-1289-5318
ch ps://o cid.o g/0000-0002-1994-1105
dh ps://o cid.o g/0000-0002-3724-2391
eh ps://o cid.o g/0000-0001-9918-787X
h ps://o cid.o g/0000-0002-4249-3304
gh ps://o cid.o g/0000-0001-6869-078X
hh ps://o cid.o g/0000-0002-3758-8501
ih ps://o cid.o g/0000-0002-2976-8174
jh ps://o cid.o g/0000-0001-8314-0276
1 INTRODUCTION
1.1 The Role o AI/ML/NLP in
Diagnosing he Risk o Myopia in
Child en
Global Pe spec i e: The inc easing p e alence o
myopia, pa icula ly among child en, ep esen s a sig-
ni ican public heal h challenge (Wo ld Heal h O ga-
niza ion, 2015). Cha ac e ized by he eye’s inabili y
o ocus on dis an objec s, myopia no only comp o-
mises he quali y o li e bu also p edisposes indi idu-
1092
A iemjew, P., Cybulski, R., Emamian, M., G zybowski, A., Jankowski, A., Lanca, C., Meh a a an, S., Mły´
nski, M., Mo awski, C., No dhausen, K., Pä ssinen, O. and Ropiak, K.
P edic ing Child en’s Myopia Risk: A Mon e Ca lo App oach o Compa e he Pe o mance o Machine Lea ning Models.
Pape published unde CC license (CC BY-NC-ND 4.0)
In P oceedings o he 16 h In e na ional Con e ence on Agen s and A i icial In elligence (ICAART 2024) - Volume 3, pages 1092-1099
ISBN: 978-989-758-680-4; ISSN: 2184-433X
P oceedings Copy igh ©2024 by SCITEPRESS – Science and Technology Publica ions, Lda.
als o se ious ocula condi ions and isual impai men
la e in li e (Haa man e al., 2020). O e he pas ew
decades, he p e alence o myopia has inc eased on
a la ge scale, especially in Eas and Sou heas Asian
coun ies, whe e he p e alence o myopia in young
adul s is 80-90%, and an accompanying high p e a-
lence o high myopia in young adul s (10-20%) (Mo -
gan e al., 2012; Mo gan e al., 2018). Se e al s udies
ha e shown connec ions o myopia wi h pa en al my-
opia, longe educa ion, mo e nea wo k ime , and less
ime spen ou doo s (Huang e al., 2015; P¨
a ssinen
and Kauppinen, 2022). In addi ion o he abo e ac-
o s, nume ous o he ac o s can in luence he onse
and p og ession o myopia. Due o hose ac o s, he e
is g ea indi idual a ia ion in he de elopmen o my-
opia and i s p og ession. Fo he p e en ion o myopia
and slowing i s p og ession, i is o g ea in e es o
ha e p edic ion ools o know in ad ance which chil-
d en a e a isk o de eloping myopia. The me hods
de eloped in his manusc ip aim o imp o e he p e-
dic abili y o he de elopmen o myopia.
Applica ion o Myopia Diagnosis: In myopia
managemen , hese echnologies may ha e a c ucial
ole in he de elopmen o sc eening ools ha ac-
cu a ely p edic he onse and p og ession o my-
opia in child en. AI algo i hms can analyze as
da ase s, iden i ying pa e ns ha p ecede myopia de-
elopmen . Fo ins ance, machine lea ning models
ha e been ained o p edic myopia based on bio-
me ic da a, en i onmen al ac o s, and gene ic da a.
These models a e inc easingly being u ilized o ale
heal hca e p o ide s and pa en s o ea ly signs o
myopia, enabling imely in e en ion.
Challenges in Implemen a ion: Despi e hei po-
en ial, he deploymen o AI/ML/NLP in myopia di-
agnosis is no wi hou challenges. Issues such as
da a p i acy, he need o la ge and di e se da ase s
o model aining, and he in eg a ion o hese ech-
nologies in o exis ing heal hca e in as uc u es a e
ongoing conce ns. Mo eo e , ensu ing hese ad-
anced ools a e accessible ac oss a ious socio-
economic backg ounds emains a hu dle o achie ing
widesp ead bene i .
1.2 Selec ed Ini ia i es
Uni ed S a es. In esponse o he myopia su ge
among child en, he Uni ed S a es has launched AI-
powe ed ini ia i es ha in e wine esea ch wi h p ac-
ical applica ions. The Na ional Eye Ins i u e (NEI)
has ca alyzed his mo emen by unding esea ch in o
AI models capable o p edic ing myopia p og ession,
wi h ecen s udies demons a ing a 30% imp o e-
men in ea ly de ec ion accu acy. School-based p o-
g ams ha e seen an in usion o AI, pa icula ly du -
ing he COVID-19 pandemic, whe e inc eased sc een
ime has been linked o a ma ked ise in myopia cases
(Kuehn, 2021; Ma e al., 2022). These p og ams
bene i om collabo a ions such as he one be ween
he Ame ican Oph halmological Socie y and ech gi-
an s, aiming o c ea e s anda dized sc eening p o o-
cols ac oss a ious s a es.
China, Singapo e, Japan, and Sou h Ko ea. The
Fa Eas e n coun ies ha e es ablished a collabo a-
i e ne wo k, he Eas Asian Oph halmology Alliance
(EAOA), o acili a e he exchange o AI esea ch and
echnologies. G oundb eaking s udies, like China’s
AI-based analysis o e inal images, ha e epo ed an
80% accu acy in p edic ing myopia, showcasing he
powe o collabo a i e da a sha ing and algo i hm de-
elopmen . Fo example he (Foo e al., 2023). Sin-
gapo e’s na ional p og am, in eg a ing gene ic da a,
has seen a 20% inc ease in p edic i e p ecision pos -
COVID-19, add essing he li es yle changes ha ha e
po en ially accele a ed myopia a es in child en. This
collec i e e o unde sco es he po en ial o an in e -
na ional s anda d in myopia isk assessmen .
Eu ope. Eu ope’s in eg a ion o AI in myopia de-
ec ion is e ol ing, wi h he Eu opean Vision In-
s i u e leading mul i-coun y s udies ha emphasize
ea ly bioma ke iden i ica ion. Despi e challenges
wi h da a agmen a ion, ecen EU di ec i es ha e
sough o uni y heal h da a s anda ds, p omo ing e-
sea ch like he Pan-Eu opean Myopia S udy (PEMS),
which epo ed a 15% inc ease in myopia de ec ion
since he pandemic began. The GDPR (Gene al Da a
P o ec ion Regula ion), while s ingen , is adap ing o
os e secu e da a sha ing o AI applica ions, wi h
he ecen es ablishmen o he Eu opean Heal h Da a
Space aiming o acili a e his shi .
1.3 Machine Lea ning Models
T us wo hiness as an Impe a i e in
Heal hca e Applica ions
T us wo hiness in machine lea ning models an-
scends a me e desi able quali y, becoming impe a i e
in sensi i e applica ions like heal hca e. Use s, pa ic-
ula ly medical p o essionals, mus be con iden in he
models’ p edic i e capabili ies o make c ucial clin-
ical decisions. T us wo hiness is a composi e mea-
su e, including bu no limi ed o, e ec i eness, o-
bus ness, ai ness, in e p e abili y, eliabili y, ans-
P edic ing Child en’s Myopia Risk: A Mon e Ca lo App oach o Compa e he Pe o mance o Machine Lea ning Models
1093
pa ency, secu i y, eplicabili y, scalabili y, and com-
pliance wi h egula o y, en i onmen al, and social
s anda ds. In medical con ex s, e ec i eness equa es
o p edic i e accu acy and clinical ele ance, while
obus ness e lec s he model’s consis ency ac oss
di e se pa ien demog aphics and da ase s. These
a e he bed ock o ope a ional eliabili y o clinical
AI/ML applica ions. Unde his di ec i e, he ML
models should aim o elimina e bias, ensu ing equi-
able heal h ou comes. We implemen ed p i acy sa e-
gua ds, including ad anced enc yp ion and da a man-
agemen p o ocols, aligning wi h he HIPAA (Heal h
Insu ance Po abili y and Accoun abili y Ac ) s an-
da ds. This comp ehensi e app oach o us wo hi-
ness, g ounded in he highes go e nmen guidelines,
aims o ede ine he applica ion o AI in heal hca e,
ensu ing machine lea ning models ul ill public se -
ice while adhe ing o sa e y and e hical no ms. Rec-
ognized au ho i ies such as he Ame ican Medical In-
o ma ics Associa ion AMIA and Ins i u e o Elec-
ical and Elec onics Enginee s (IEEE) ha e long
s essed he signi icance o us wo hiness in clini-
cal AI/ML implemen a ions. The FDA is delinea ing
egula ions o AI in diagnos ics, emphasizing he im-
po ance o model alida ion and con inuous moni o -
ing. The NIH p omo es open science o model epli-
cabili y and independen e i ica ion, highligh ing he
necessi y o anspa ency and secu i y in p o ec ing
pa ien da a and p e en ing sys em misuse.
2 METHODOLOGY
2.1 MRC App oach o T us wo hiness
o Machine Lea ning Models
The Myopia Risk Calcula o Conso ium (MRC) ec-
ognizes he complex impe a i e o us wo hiness in
machine lea ning models, especially wi hin he deli-
ca e con ex o pedia ic oph halmology. Ou sys em-
a ic, phased app oach ha nesses he collec i e expe -
ise o in e na ional oph halmology expe s, ensu ing
ha he p ojec embodies a uly in e disciplina y col-
labo a ion. Da a scien is s, clinicians, and oph hal-
mologis s con ibu e uniquely, combining clinical in-
sigh s wi h ad anced compu a ional me hods o shape
ou me hodology. The main assessmen ML model
me hods p oposed in he MRC me hodology a e he
ollowing:
2.1.1 Applica ions o Mon e Ca lo Me hods
Ou use o Mon e Ca lo (Robe and Casella, 2013)
simula ions ex ends beyond da a augmen a ion; i en-
compasses he c ea ion o syn he ic da ase s ha mi -
o complex eal-wo ld a ia ions, hus suppo ing o-
bus model aining and alida ion. The applica ion
o Mon e Ca lo C oss-Valida ion (MCCV) and boo -
s apping echniques (Koha i, 1995) unde pins ou
models’ eliabili y, p o iding anspa en and s a is-
ically signi ican measu es o pe o mance.
2.1.2 Con idence In e al E alua ion
Pe o mance e alua ion o ou models anscends
poin es ima es, wi h con idence in e als d awn om
MCCV-de i ed s a is ics, ensu ing a eplicable and
us wo hy assessmen o model eliabili y. I is
wo h o emphasize ha :
•The applica ion o Mon e Ca lo me hods coupled
wi h con idence in e al e alua ions spea headed
by MRC se s a p oposal o new s anda ds in he
e alua ion o ML models’ e ec i eness and o-
bus ness o medical applica ions, championing
hei eliabili y and anspa ency.
•The echniques based on MCCV showcased
he ein ha e a i ma i ely passed he ini ial li mus
es in he ongoing pu sui o a comp ehensi e
us wo hiness assessmen amewo k o medi-
cal ML applica ions.
As we ha e na iga ed he ini ial phase, ou a-
di ional ML models, applied o alphanume ic da a,
ha e been me iculously c a ed o p edic myopia isk
in child en. This g oundwo k pa es he way o in-
co po a ing mo e in ica e da a ypes, such as image
da a, and sophis ica ed me hods. In pa icula , he
WisTech app oach o In e ac i e G anula Compu ing
(IG C) (Jankowski, 2017; Polkowski and A iemjew,
2015; Lin e al., 2023) will be explo ed o i s po en-
ial o dissec and u ilize causal ela ionships, enhanc-
ing ou model’s p edic i e p ecision h ough nuanced
‘wha -i ’ scena io analyses. In he c i ical domain o
da a p i acy, he conso ium adhe es o s ingen p o-
ocols aligned wi h global s anda ds such as GDPR
and HIPAA, ensu ing ha ou syn he ic da a gene a-
ion p ocess upholds he u mos pa ien con iden iali y
and secu i y. Ou me hodology’s in eg i y is unde -
pinned by an e hical amewo k ha guides syn he ic
da a applica ion, wi h o e sigh om an ins i u ional
e iew boa d dedica ed o main aining medical e hics
a he o e on o ou e o s. While he cu en o-
cus is on e ec i eness and obus ness, we a e laying
a comp ehensi e ounda ion o a mul i-dimensional
us wo hiness amewo k. This amewo k, adap -
able ac oss he us spec um, is c a ed o mee he
apex o medical p ac ice s anda ds. An icipa ing u-
u e clinical alida ion ials, we a e p epa ing o he
c i ical phase o p ac ical applica ion, aiming o in e-
ICAART 2024 - 16 h In e na ional Con e ence on Agen s and A i icial In elligence
1094
g a e ou models seamlessly in o he medical commu-
ni y o he bene i o pedia ic pa ien ca e.
3 EXPERIMENTAL PART
3.1 Pe o mance o he MRC (Ve sion
1.0)
Da a and Va iables. Da a o 3989 child en (7945
eyes) we e used o he analysis o his s udy. Chil-
d en om he Shah oud Schoolchild en Eye Coho
S udy we e included in his s udy. I is a p ospec-
i e coho s udy, conduc ed in Shah oud, no heas
I an, ha ec ui ed 5620 child en aged 6 o 12 in 2015
(baseline), wi h a ollow-up in 2018 (Emamian e al.,
2019). Cycloplegic e ac ions we e conduc ed o de-
ec myopia a baseline and i s p og ession a e h ee
yea s. A ques ionnai e was adminis e ed o collec
da a such as age, gende , nea wo k ime, ou doo
ime, li ing place, pa en al myopia, and mo he ’s ed-
uca ion. Ocula biome ics we e measu ed using he
Alleg o Biog aph. These a e c ucial isk ac o s o
unde s and and p edic myopia.
In ou s udy, we ca ego ized a iables o ML
models in o ou a ibu e classes:
1. E o less A ibu es: These a e a ibu es ha
can be calcula ed wi hou ad anced medical
knowledge o medical ins umen s. In ui i ely,
he alues o hese a ibu es should be eadily
a ailable.
2. Ad anced A ibu es: These a e a ibu es ha
a e no conside ed e o less, equi ing mo e spe-
cialized knowledge o equipmen .
3. Non-Cycloplegic A ibu es: These a e a -
ibu es ha do no equi e cycloplegic e ac ion.
4. Cycloplegic A ibu es: These a ibu es equi e
cycloplegic e ac ion, indica ing ha hey ela e
o an eye examina ion p ocedu e whe e he eye’s
cilia y muscle is empo a ily pa alyzed o de e -
mine e ac i e e o .
Dependen Va iables and Decision Classes. The
main ou come a iable o his s udy was he h ee-
yea sphe ical equi alen (SE2). The alue o SE2 is
equal o he inal SE (a e 3 yea s). Machine lea ning
echniques we e employed o c ea e bina y classi ie s
o myopia isk p edic ion and eg ession models o
SE2 p edic ion. We ocus on he ollowing decision
classes o bina y classi ie s:
•M01: The SE2 alue will be ≤-0.5 D.
•M02: The SE2 alue will be ≤-4.0 D.
•M05: The SE2 alue will be ≤-1.0 D.
In hese a ian s, he alues -0.5, -1 and -4 a e he
h esholds o c ea ing a bina y decision. In his way,
bina y decision sys ems a e o med o model he dis-
ease p edic ion p ocess.
Fo he eg ession models, we use SE2 as he de-
penden a iable. In he conside ed cases, he a i-
able SE2 depends on he independen a iables de-
i ed om he ini ial examina ion (baseline).
Model E alua ion Me hodology. We used an ag-
g ega ion o Mon e Ca lo c oss- alida ion and boo -
s ap me hods o assessing models’ e ec i eness and
obus ness. We calcula ed boo s ap con idence in-
e als o he ollowing s anda d s a is ics: classi ie
quali y measu es such as Sensi i i y, Speci ici y, P e-
cision, F1, Gm, AUC (Tha wa , 2021), and weigh ed
a e ages and eg ession quali y measu es like MSA,
MAE, and R2(Has ie e al., 2009).
Le us deno e by S one o he abo e s a is ics. In
ou p ojec , we use he ollowing calcula ion me hod-
ology o he alue o he s a is ic S measu ing he
model pe o mance (e.g., Sensi i i y, Speci ici y, P e-
cision, AUC, F1, Gm, MSA, MAE, and R2(Has ie
e al., 2009), e c.):
1. Gene a e 20 Mon e Ca lo-simula ed c oss-
alida ions using 10 olds o model e alua ion.
In o he wo ds, we ob ain 200 e alua ion olds
o model e alua ion (i.e., en imes wen y). This
leads o 200 alues o he s a is ic S measu ing
he model pe o mance (i.e., o any e alua ion
model om 200).
2. As a esul o he abo e i s s ep, we ha e 200
alues o S. I i is necessa y, hen you may use
mo e i e a ions o Mon e Ca lo-gene a ed c oss-
alida ions and employ Ma ko Chain Mon e
Ca lo (MCMC) me hods o be e app oxima -
ing he empi ical dis ibu ion o S (Robe and
Casella, 2000).
3. Calcula e he pe cen iles (Eubank, 2006) o he
gene a ed empi ical dis ibu ion o S.
4. Fo each pe cen ile p, epo he Qp(S) alue
(e.g., Q02(Sensi i i y)). Fo example, i S is he
measu e o AUC (i.e., A ea Unde he Cu e),
hen Q25(AUC) is he i s qua ile o AUC,
Q50(AUC) is he median, and Q75(AUC) is he
hi d qua ile.
5. P ima ily, ocus on Q05(S), Q50(S), and Q95(S).
6. Apply pe cen iles om he empi ical dis ibu ion
o de e mine con idence in e als o he s a is ic
S o he equi ed signi icance le el.
P edic ing Child en’s Myopia Risk: A Mon e Ca lo App oach o Compa e he Pe o mance o Machine Lea ning Models
1095

The quali y o classi ie s planned o deploymen
in a speci ic wo ld egion should e lec he le el o
medical ca e in a pa icula a ea. Gene ally, he mos
impo an ea u e is always he sensi i i y, which is as
high as possible on an accep able le el o p ecision
(and some imes o he pa ame e s). Condi ions may
be imposed on P ecision based on physician a ailabil-
i y. Fo example, we may p e e classi ie s wi h e y
high P ecision i doc o s a e una ailable. A e he
discussion, we concluded ha he p ima y assessmen
o he classi ie s should be using a weigh ed a e -
age, which we will symbolically deno e as SimWA ,
p e e ably acco ding o he o mula below.
SimWA =1
2Q05(Sensi i i y)+ 1
4Q05(Speci ici y)
+1
4Q05(P ecision).
(1)
Below, we p esen he classi ica ion esul s o
classi ie s so ed by SimWA alues.
Resul s. The e ec i eness o adi ional ML me h-
ods, such as Logis ic Reg ession (LR), k-Nea es
Neighbo (kNN), Suppo Vec o Machines (SVM),
Random Fo es (RF), G adien Boos ing (GB), and
A i icial Neu al Ne wo k (ANN), was e i ied. In ad-
di ion, he models we e cons uc ed as an agg ega ion
o adi ional models, i.e., ensemble models.
The bes esul s o he bina y classi ie s o he
de ined decision classes M01, M02, and M05 a e as
ollows.
Table 1: Selec ed bes esul s - a classi ica ion p oblem.
Using Cycloplegic A ibu es; Bes Me = Bes ML model
me hod, RF = Random Fo es , SVM = Suppo Vec o Ma-
chine, LR+RF=Ensemble model o Lasso Reg ession and
Random Fo es . See isualisa ion in Fig. 1.
S a is ic M01 M02 M05
Bes Me RF SVM LR+RF
SimWA 73% 64% 74%
Q05 (AUC) 86% 86% 87%
Q50 (AUC) 90% 94% 92%
Q05 (Sens) 79% 72% 79%
Q50 (Sens) 87% 91% 89%
Q05 (Spec) 91% 97% 94%
Q50 (Spec) 93% 98% 96%
Q05 (P ec) 44% 16% 44%
Q50 (P ec) 51% 22% 53%
In addi ion o bina y classi ie s, we ha e examined
eg ession models. We used SE2 as he dependen
a iable o eg ession models in his case. The a i-
able SE2 depends on he independen a iables de-
Figu e 1: Visualisa ion o he esul s o Table 1: Using Cy-
cloplegic A ibu es; Bes Me = Bes ML model me hod,
RF = Random Fo es , SVM = Suppo Vec o Machine,
LR+RF=Ensemble model o Lasso Reg ession and Random
Fo es .
Table 2: Selec ed bes esul s - a classi ica ion p oblem.
Using Non-Cycloplegic A ibu es; Bes Me = Bes ML
model me hod, LR+RF=Ensemble model o Lasso Reg es-
sion and Random Fo es , SVM = Suppo Vec o Machine,
RF = Random Fo es . See isualisa ion in Fig. 2.
S a is ic M01 M02 M05
Bes Me LR+RF SVM RF
SimWA 66% 72% 68%
Q05 (AUC) 80% 91% 85%
Q50 (AUC) 85% 99% 90%
Q05 (Sens) 71% 84% 81%
Q50 (Sens) 82% 99% 90%
Q05 (Spec) 87% 98% 86%
Q50 (Spec) 89% 99% 88%
Q05 (P ec) 33% 20% 24%
Q50 (P ec) 38% 28% 29%
i ed om he ini ial examina ion. Analogously o
bina y classi ie s, we analysed wo ypes o models:
1. Type 1: Using cycloplegic independen a iables.
2. Type 2: Using only non-cycloplegic independen
a iables.
ICAART 2024 - 16 h In e na ional Con e ence on Agen s and A i icial In elligence
1096
Figu e 2: Visualisa ion o he esul s o Table 2: Using
Non-Cycloplegic A ibu es; Bes Me = Bes ML model
me hod, LR+RF=Ensemble model o Lasso Reg ession and
Random Fo es , SVM = Suppo Vec o Machine, RF =
Random Fo es .
We used ypical eg ession model pe o mance in-
dica o s: Q05(R2 sco e), Q50(R2 sco e) and mea-
su ed diop e s: Q95(MSE), Q50(MSE), Q95(RMSE),
Q50(RMSE), Q95(MAE) and Q50(MAE). The mos
in e es ing esul s a e p esen in he ollowing ables:
Table 3: Selec ed bes esul s - a eg ession p oblem. Cy-
cloplegic A ibu es; Bes Me = Bes ML model me hod,
GB = G adien boos ing, RF = Random Fo es . See isual-
isa ion in Fig. 3.
S a is ic Model 1 Model 2
Bes Me GB RF
Q95 (MSE) 0.28 0.29
Q50 (MSE) 0.18 0.19
Q95 (RMSE) 0.52 0.54
Q50 (RMSE) 0.42 0.44
Q95 (MAE) 0.33 0.36
Q50 (MAE) 0.30 0.32
Q05 (R2 sco e) 0.69 0.68
Q50 (R2 sco e) 0.79 0.78
Figu e 3: Visualisa ion o he esul s o Table 3: Cyclo-
plegic A ibu es; Bes Me = Bes ML model me hod, GB
= G adien Boos ing, RF = Random Fo es .
Table 4: Selec ed bes esul s - a eg ession p oblem.
Non-Cycloplegic A ibu es; Bes Me = Bes ML model
me hod, GB = G adien Boos ing, ANN = A i icial Neu al
Ne wo k. See isualisa ion in Fig. 4.
S a is ic Model 1 Model 2
Bes Me GB ANN
Q95 (MSE) 0.50 0.63
Q50 (MSE) 0.40 0.48
Q95 (RMSE) 0.71 0.79
Q50 (RMSE) 0.63 0.69
Q95 (MAE) 0.53 0.59
Q50 (MAE) 0.46 0.52
Q05 (R2 sco e) 0.39 0.29
Q50 (R2 sco e) 0.53 0.45
4 CONCLUSIONS
This s udy ep esen s a key ad ancemen in us-
ing machine lea ning o p edic he isk o myopia
among child en. The Myopia Risk Calcula o (MRC)
Conso ium has success ully used classical machine
lea ning me hodologies wi h cu ing-edge AI inno a-
ions, o ging a ailblazing pa h in oph halmic heal h
P edic ing Child en’s Myopia Risk: A Mon e Ca lo App oach o Compa e he Pe o mance o Machine Lea ning Models
1097
Figu e 4: Visualisa ion o he esul s o Table 4: Non-
Cycloplegic A ibu es; Bes Me = Bes ML model
me hod, GB = G adien Boos ing, ANN = A i icial Neu-
al Ne wo k.
solu ions.
Key Findings:
•Valida ion o adi ional ML me hods o c ea -
ing classi ie s and eg ession models has been
achie ed, wi h p omising ou comes wi hin he I a-
nian pedia ic coho , hin ing a he po en ial o
c oss-popula ion applicabili y subjec o u he
empi ical in es iga ion.
•The applica ion o Mon e Ca lo me hods coupled
wi h con idence in e al e alua ions spea headed
by MRC se s a p oposal o new s anda ds in he
e alua ion o ML models’ e ec i eness and o-
bus ness o medical applica ions, championing
hei eliabili y and anspa ency.
•The echniques based on MCCV showcased
he ein ha e a i ma i ely passed he ini ial li mus
es in he ongoing pu sui o a comp ehensi e
us wo hiness assessmen amewo k o medi-
cal ML applica ions.
Now le s p esen ou u u e plans and inal hough s.
The MRC Conso ium is s a egically expanding i s
in es iga i e pu iew o include ans e lea ning
me hodologies, seeking o augmen he p ecision and
adap abili y o ou models o a a ie y o popula ion
da ase s, which will en ail he gene a ion o special-
ized syn he ic da a. Explo ing image da a, he conso -
ium an icipa es unlocking ad anced diagnos ic po-
en ial, he eby enhancing he u ili y o he models.
Wi h a nod o he u u e, he applica ion o he Wis-
Tech me hodology o In e ac i e G anula Compu -
ing (IG C) is an icipa ed o un a el he complexi ies
o causa ion, o e ing e ined ’wha -i ’ analy ical sce-
na ios ha could e olu ionize p edic i e accu acy.
The MRC is a he angua d o an e ol ing heal hca e
pa adigm whe ein AI and ML anscend hei oles as
me e compu a ional ools o become in eg al allies in
he deli e y o ad anced medical ca e. Cen al o his
pa adigm is he us wo hiness o ML models, a c i i-
cal componen ha ou esea ch add esses wi h a pio-
nee ing assessmen me hodology. As we s and on he
b ink o a ans o ma i e e a in medical echnology,
he conso ium is p opelling his mo emen wi h un-
wa e ing commi men and a ision o a u u e whe e
heal hca e is bo h inno a i e and eliable.
ACKNOWLEDGEMENTS
Shah oud School Child en Eye Coho S udy is
unded by he Noo Oph halmology Resea ch Cen-
e and Shah oud Uni e si y o Medical Sciences.
(G an numbe s: 9329, 960351). The expe imen s
we e ca ied ou , among o he s, by UWM MSc
s uden s Ma eusz ´
Sliwi´
nski, And zej S zeszewski,
Adam Jankowiak, Michał Domian, Paweł Budzi´
nski,
Ba osz ´
Cwiek, Jakub P zybo owski, Jakub Kasja-
niuk.
REFERENCES
Emamian, M. H., Hashemi, H., Khabazkhoob, M., Malihi,
S., and Fo ouhi, A. (2019). Coho p o ile: Shah oud
schoolchild en eye coho s udy (SSCECS). In e na-
ional Jou nal o Epidemiology, 48(1):27–27 .
Eubank, R. L. (2006). Quan ile. John Wiley & Sons, L d.
Foo, L. L., Lim, G. Y. S., Lanca, C., Wong, C. W., Hoang,
Q. V., Zhang, X. J., Yam, J. C., Schme e e , L., Chia,
A., Wong, T. Y., Ting, D. S. W., Saw, S.-M., and Ang,
M. (2023). Deep lea ning sys em o p edic he 5-yea
isk o high myopia using undus imaging in child en.
npj Digi al Medicine, 6(10).
Haa man, A. E. G., En ho en, C. A., Tideman, J. W. L.,
Tedja, M. S., Ve hoe en, V. J. M., and Kla e , C. C. W.
(2020). The Complica ions o Myopia: A Re iew and
Me a-Analysis. In es iga i e Oph halmology & Visual
Science, 61(4):49–49.
Has ie, T., Tibshi ani, R., and F iedman, J. (2009). The Ele-
men s o S a is ical Lea ning: Da a Mining, In e ence,
and P edic ion. Sp inge Se ies in S a is ics. Sp inge
New Yo k, NY, 2 edi ion.
Huang, H.-M., Chang, D. S.-T., and Wu, P.-C. (2015). The
associa ion be ween nea wo k ac i i ies and myopia
ICAART 2024 - 16 h In e na ional Con e ence on Agen s and A i icial In elligence
1098
in child en—a sys ema ic e iew and me a-analysis.
PLOS ONE, 10(10):1–15.
Jankowski, A. (2017). In e ac i e G anula Compu a ions
in Ne wo ks and Sys ems Enginee ing: A P ac ical
Pe spec i e. Lec u e No es in Ne wo ks and Sys ems.
Sp inge Cham, 1 edi ion. eBook Packages Enginee -
ing, Enginee ing (R0).
Koha i, R. (1995). A s udy o c oss- alida ion and boo -
s ap o accu acy es ima ion and model selec ion. In
IJCAI, olume 14, pages 1137–1145.
Kuehn, B. M. (2021). Inc ease in Myopia Repo ed Among
Child en Du ing COVID-19 Lockdown. JAMA,
326(11):999–999.
Lin, T.-Y., Liau, C.-J., and Kacp zyk, J., edi o s (2023).
G anula , Fuzzy, and So Compu ing. Encyclopedia
o Complexi y and Sys ems Science Se ies. Sp inge
New Yo k, NY, New Yo k, NY, 1 edi ion.
Ma, D., Wei, S., Li, S., Yang, X., Cao, K., Hu, J., Peng, X.,
Yan, R., Fu, J., G zybowski, A., Jin, Z., and Wang,
N. (2022). The impac o s udy-a -home du ing he
co id-19 pandemic on myopia p og ession in chinese
child en. F on ie s in Public Heal h, 9:720514.
Mo gan, I., F ench, A., Ashby, R., Guo, X., Ding, X., He,
M., and Rose, K. (2018). The epidemics o myopia:
Ae iology and p e en ion. P og ess in Re inal and Eye
Resea ch, 62.
Mo gan, I. G., Ohno-Ma sui, K., and Saw, S.-M. (2012).
Myopia. The Lance , 379(9827):1739–1748.
Polkowski, L. and A iemjew, P. (2015). G anula Com-
pu ing in Decision App oxima ion - An Applica ion o
Rough Me eology.
P¨
a ssinen, O. and Kauppinen, M. (2022). Associa ions
o nea wo k ime, wa ching , ou doo s ime, and
pa en s’ myopia wi h myopia among school child en
based on 38-yea -old his o ical da a. Ac a Oph hal-
mologica, 100(2):e430–e438.
Robe , C. and Casella, G. (2000). Mon e ca lo s a is ical
me hod. Technome ics, 42.
Robe , C. P. and Casella, G. (2013). Mon e Ca lo S a is i-
cal Me hods. Sp inge .
Tha wa , A. (2021). Classi ica ion assessmen me hods. Ap-
plied Compu ing and In o ma ics, 17(1):168–192.
Wo ld Heal h O ganiza ion (2015). The impac o my-
opia and high myopia. Repo , Uni e si y o New
Sou h Wales, Sydney, Aus alia. Repo o he Join
Wo ld Heal h O ganiza ion–B ien Holden Vision In-
s i u e Global Scien i ic Mee ing on Myopia.
P edic ing Child en’s Myopia Risk: A Mon e Ca lo App oach o Compa e he Pe o mance o Machine Lea ning Models
1099