In e na ional Jou nal o Da a Science and Analy ics
h ps://doi.o g/10.1007/s41060-025-00816-w
REGULAR PAPER
Syn he ic Da a Gene a ion o Heal hca e: Explo ing Gene a i e
Ad e sa ial Ne wo ks Va ian s o Medical Tabula Da a
Halal Abdul ahman Ahmed1·Juan A. Nepomuceno1·Belén Vega-Má quez1·Isabel A. Nepomuceno-Chamo o1
Recei ed: 7 Oc obe 2024 / Accep ed: 8 May 2025
© The Au ho (s) 2025
Abs ac
Recen ly, he medical and heal hca e ields ha e expe ienced signi ican imp o emen s. Howe e , he es ic ions o e hical
cons ain s, p i acy egula ions, and p ese a ion o sha ing sensi i e pe sonal in o ma ion limi access o eal pa ien da a.
Syn he ic da ase s wi h gene a i e models a e conside ed one o he mos eliable solu ions ha mee s ic da a p o ec ion
equi emen s. Syn he ic da a a e c ea ed in a con olled en i onmen bu possess he same s a is ical and s uc u al p ope ies
as eal da a. In his wo k, we gene a e syn he ic da a using six a ia ions o gene a i e ad e sa ial ne wo ks (GANs): GAN,
CGAN, CTGAN, CRAMER GAN, DRAGAN, and WGAN. We explo e he e icacy o syn he ic da a in h ee dis inc
heal hca e da ase s: B eas Cance Wisconsin (Diagnos ic), Lung Cance Pa ien , and Fe al Ca dio ocog aphy CTG. To
e alua e he pe o mance o hese gene a ed da ase s in classi ica ion asks, we employ wo di e se classi ie s, namely XGBoos
and SVM. In addi ion, we employ co ela ion and s a is ical analyses o sc u inise GAN models, iden i ying op imal a ian s
o speci ic da a gene a ion asks. Ou expe imen al amewo k encompasses he examina ion o o iginal ( eal), syn he ic,
and hyb id (o iginal and syn he ic) da ase s. Ou indings highligh a no able imp o emen in classi ica ion accu acy when
using ad anced GAN models such as CGAN and CTGAN o gene a e abula da a. This esea ch sheds ligh on he po en ial
o syn he ic da a in bols e ing da a p i acy while acili a ing meaning ul insigh s in he ealm o heal hca e analy ics.
Keywo ds Syn he ic da a gene a ion ·Gene a i e ad e sa ial ne wo ks ·P i acy-p ese ing da a ·Deep lea ning
1 In oduc ion
O e he pas decade, he medical ield has seen ema kable
ad ancemen s due o inc eased da a a ailabili y, machine
lea ning echniques, and a i icial in elligence. High-quali y
da ase s a e essen ial o aining and es ing machine lea n-
ing models in heal hca e, bu s ic p i acy es ic ions, da a
sca ci y, and e hical cons ain s limi access o such da a [1,
2]. To add ess his challenge, esea che s ha e de eloped
echniques such as Va ia ional Au oencode s (VAE) [3] and
Gene a i e Ad e sa ial Ne wo ks (GAN) o gene a e syn-
he ic da a. Syn he ic da a, c ea ed h ough algo i hms, is
c i ical when eal da a is challenging, expensi e, o limi ed
o acqui e. I p ese es he s a is ical and s uc u al cha ac-
e is ics o eal da a while main aining pa ien p i acy. In
heal hca e, syn he ic da a helps de elop accu a e machine
BHalal Abdul ahman Ahmed
[email p o ec ed]
1Depa men o Compu e Languages and Sys ems, Uni e si y
o Se ille, Se ille 41012, Spain
lea ning models, imp o e ea men s, and unde s and dis-
eases [4,5], enhancing he quali y o heal hca e and biome ic
[6] esea ch.
GANs, in pa icula , gene a e syn he ic da a esembling
eal medical eco ds wi hou con aining ac ual pa ien de ails,
hus o e coming da a sca ci y [7–9].
GANs ha e shown ema kable esul s in gene a ing eal-
is ic images and ideos o a ious applica ions. Han e al.
[10] success ully gene a ed b ain MR images. Jin e al. [11]
ocused on gene a ing CT images encompassing bo h nod-
ules and adjacen issues. Bhaga e al. [12] gene a ed ches
X- ay images o pneumonia pa ien s. Uzuno a e al. [13]
gene a ed ealis ic and high- esolu ion 2D and 3D medical
da a. Munia e al. [14] gene a ed syn he ic elec oca diog am
(ECG) da a. Mo eo e , GANs a e also used in o he ypes
o da a; o ins ance, Lei Xu e al. p esen ed Condi ional
Tabula GAN (CTGAN) [15], which aims o gene a e high-
quali y abula da ase s encompassing a ious da a ypes.
CTGAN employs a condi ional gene a o and an inno a i e
aining-by-sampling app oach o add ess he gene a ion o
imbalanced da a. The CTGAN model adop s a mode-speci ic
123
In e na ional Jou nal o Da a Science and Analy ics
no malisa ion echnique o e ec i ely handle he complexi y
o gene a ing mul i-modal nume ical columns. Li e al. [16]
p oposed EHR-M-GAN o gene a e mixed- ype ime-se ies
EHR da a. Mo ini e al. [17] employed CRAMER GAN o
gene a e syn he ic Pe sonal Name Reco ds (PNR) by ain-
ing he model on eal PNRs wi h nume ical and ca ego ical
ea u es. The expe imen al esul s indica e ha he gene a i e
model p oduces ealis ic syn he ic da a ha closely ma ches
he dis ibu ion o eal PNRs. Azman e al. [18]usedDRA-
GAN o gene a e syn he ic medical images; hese images a e
subsequen ly employed o classi y lung lesions in o benign
and malignan ca ego ies using Shu leNe . Chin-Cheong e
al. [19] employed WGAN o gene a e high-quali y nume i-
cal and ca ego ical he e ogeneous EHR da a in e ms o bo h
da a ideli y and da a u ili y by combining i wi h di e en-
ial p i acy (DP). Hussain e al. [20] p esen ed a solu ion
o add essing he da a de iciency in COVID-19 ches X-
ay images by using he Wasse s ein Gene a i e Ad e sa ial
Ne wo k (WGAN). The expe imen s demons a ed WGAN’s
e ec i eness in gene a ing syn he ic X- ay images. Se e al
esea ch s udies ha e u ilised WGAN [21–23].
S udies ha e explo ed a ious GAN-based models o
gene a ing syn he ic medical abula da a ha main ain s a-
is ical cha ac e is ics and model compa ibili y o o iginal
da a [24–26]. Al hough he gene a ion o syn he ic da a is
apidly inc easing, he gene a ion o syn he ic abula da a
s ill p esen s unique challenges compa ed o o he da a ypes
[27]. Rela i ely, he e a e less s udies on he gene a ion o
syn he ic abula da a compa ed o o he ypes o da a, such
as images, ex o speech [28–30].
This s udy makes se e al signi ican con ibu ions o he
ield o syn he ic da a gene a ion in heal hca e. Ou p ima y
con ibu ion lies in compa ing six di e en GAN a ian s -
GAN, CGAN, CTGAN, CRAMER GAN, DRAGAN, and
WGAN- o examine hei unc ionali ies and e icacy in
gene a ing high-quali y syn he ic abula da a o medical
p oblems. In pu sui o ou objec i e, we selec ed h ee di -
e en publicly a ailable medical da ase s, including B eas
Cance Wisconsin (Diagnos ic) [31], Lung Cance Pa ien
[32], and Fe al Ca dio ocog aphy (CTG) [33].
To access he simila i y and usabili y o syn he ic medical
da a and eal pa ien da a, we use wo widely ecognised clas-
si ica ion algo i hms, including XGBoos and Suppo Vec o
Machine (SVM). We will employ wo s a is ical me hods o
analyze he ela ionships be ween a iables in he da ase s
(Pea son and Spea man co ela ions). Fu he mo e, o e al-
ua e he pe o mance o he GAN models and making su e
he esul s a e accu a e and eliable, we will conduc s a is i-
cal e alua ions by using he S a is ical Tes s o Algo i hms
Compa ison (STAC) pla o m.
The emainde o his a icle is o ganised as ollows: Sec-
ion 2desc ibes he me hods o gene a ing syn he ic da a;
Sec ion 3desc ibes he ma e ials used; Sec ion 4 epo s and
discusses he esul s ob ained; and Sec ion 5p esen s he
conclusions and po en ial u u e wo k.
2 Me hods
This sec ion del es in o he me hods used o gene a e syn-
he ic medical da a using Gene a i e Ad e sa ial Ne wo ks
(GAN) me hods. We p o ide de ailed explana ions o he
selec ed GAN a ian s and classi ie s used o measu e he
pe o mance o he syn he ic da a. We ha e decided o check
whe he he beha iou o classi ying he gene a ed da a is
simila o he beha iou o eal da a. This aims o ensu e he
alidi y o he gene a ed da a h ough he pe o mance o
he classi ie s and he unde s anding o he limi a ions o he
GAN models.
2.1 Gene a i e Ad e sa ial Ne wo ks (GANs)
GAN is an a i icial in elligence model in oduced by Ian
Good ellow [34]. I is a obus amewo k o aining gene -
a i e models, capable o gene a ing new da a ha esembles a
gi en da ase . GANs a e composed o wo sub-ne wo ks: he
gene a o (G) and he disc imina o (D), based on he idea
o ze o-sum games [35]. These componen s a e wo neu-
al ne wo ks simul aneously ained h ough a compe i i e
p ocedu e in which one ne wo k a emp s o gain an ad an-
age while he o he ne wo k expe iences an equi alen loss.
Gene a o (G): The main ask o he gene a o is o gene a e
syn he ic da a ha esembles he eal da a o he aining se .
I akes andom noise as inpu and ans o ms i in o syn-
he ic da a. Disc imina o (D): The disc imina o is a bina y
classi ie designed o di e en ia e be ween eal da a om he
aining se and syn he ic da a gene a ed by he gene a o . I
ecei es bo h eal and gene a ed da a samples as inpu and
e u ns a p obabili y sco e indica ing he likelihood o he
inpu being eal.
These wo sub-ne wo ks compe e o imp o e hei p e-
dic i e accu acy by pi ing hemsel es agains one ano he .
When he gene a o is aining he disc imina o is inac-
i e, and when he disc imina o is aining, he gene a o is
inac i e. The gene a o becomes be e a gene a ing highe -
quali y syn he ic da a h ough his compe i i e p ocess, so
he disc imina o becomes be e a lagging a i icially gen-
e a ed da a. In Figu e 1, he a chi ec u e o GAN can be
obse ed.
2.1.1 Condi ional GAN (CGAN)
Condi ional GAN [36] is an ex ension o he s anda d GAN
model. In a s anda d GAN, he gene a o lea ns o gene a e
syn he ic da a om andom noise, whe eas he disc imina-
o a emp s o disc imina e be ween eal and ake samples.
123
In e na ional Jou nal o Da a Science and Analy ics
Fig. 1 Illus a ion o he GAN
a chi ec u e
Howe e , CGAN adds addi ional condi ional in o ma ion o
condi ional labels o he da a gene a o and he disc imina-
o o mo e a ge ed and con olled da a gene a ion. This
condi ional in o ma ion ins uc s he gene a o o gene a e
samples ha mee pa icula c i e ia o belong o a speci ic
class. The disc imina o conside s his addi ional condi ion
when disc imina ing be ween eal and ake da a. Typically,
condi ional da a is p o ided as an addi ional inpu o neu al
ne wo ks. The condi ional GAN a chi ec u e p o ides mo e
a ge ed and con olled da a gene a ion [37].
2.1.2 Condi ional Tabula GAN (CTGAN)
CTGAN [15] is a GAN-based model de eloped pa icula ly
o gene a e syn he ic abula da a wi h condi ional a ibu es.
Simila o CGAN, in CTGAN, he gene a o inpu s andom
noise and condi ional in o ma ion o gene a e syn he ic ab-
ula da a ha adhe es o he speci ied condi ions. CTGAN
can gene a e new ows o abula da a ha sa is y gi en
cons ain s o adhe e o speci ic c i e ia. CTGAN p o ides
a powe ul ool o gene a ing condi ional abula da a and
allows use s o con ol and in luence he cha ac e is ics o he
da a gene a ed h ough condi ional inpu s. The a chi ec u e
o GTGAN is adap ed o handle he cons ain s and depen-
dencies ound in abula da ase s.
2.1.3 CRAMER GAN
CRAMER GAN [38] uses he C ame dis ance, a mo e
s able me ic, o add ess aining ins abili ies and mode
collapse issues aced by s anda d GANs. C ame dis ance
p o ides a compu a ionally ac able measu e o he disc ep-
ancy be ween p obabili y dis ibu ions. By using he C ame
dis ance, CRAMER GAN imp o es he quali y o gene a ed
samples and encou ages di e si y in he gene a ed da a dis i-
bu ion, esul ing in a mo e ai h ul ep esen a ion o he ue
da a dis ibu ion. C ame GAN is an example o how di e -
en dis ance me ics can be used in GANs o achie e speci ic
objec i es and o e come aining challenges.
2.1.4 Deep Reg e Analy ic GAN (DRAGAN)
DRAGAN [39] is a p oposed egula isa ion echnique o ain
GANs, add ess he issue o mode collapse [40] and imp o e
aining s abili y. Mode collapse occu s when he GAN’s gen-
e a o ails o cap u e all he modes o di e se pa e ns in he
eal da a dis ibu ion, esul ing in a lack o di e si y in he
gene a ed samples. DRAGAN p oposes a no el app oach
o mi iga e his p oblem by cons aining he disc imina o
g adien s a ound eal da a poin s. By employing he egu-
la isa ion echnique o DRAGAN, he aining p ocess o
GANs achie es s abili y and a educed suscep ibili y o mode
collapse. enables as e aining, imp o ed s abili y, and be -
e modelling pe o mance compa ed o o he s able aining
p ocedu es like WGAN-GP (Wasse s ein GAN wi h G adi-
en Penal y) [41].
2.1.5 Wasse s ein GAN (WGAN)
Wasse s ein GAN [42] is an ad anced a ian o GANs. I
u ilises he Wasse s ein dis ance as he loss unc ion, o e ing
a mo e meaning ul measu e o da a dis ibu ion dissimi-
la i y compa ed o adi ional GANs. WGAN eplaces he
bina y disc imina o wi h a c i ic and en o ces a 1-Lipschi z
cons ain o s abili y du ing aining. As a esul , WGAN
achie es mo e s able aining, educes mode collapse issues
[40], and p o ides a be e e alua ion o he pe o mance o
he gene a o . I s e ec i eness has made WGAN popula in
gene a i e models and deep lea ning esea ch.
2.2 Classi ie s
2.2.1 Ex eme G adien Boos ing Classi ie
Ex eme G adien Boos ing (XGBoos ) [43] classi ie is
a machine lea ning algo i hm ha implemen s g adien -
boos ed decision ees, excelling in classi ica ion and eg es-
sion asks. Fu he mo e, he boos ing echnique employed
in his app oach is egula ised, allowing o he au oma ed
handling o missing alues. Addi ionally, he algo i hm has
been speci ically de eloped o show high e iciency, lexi-
123
In e na ional Jou nal o Da a Science and Analy ics
Fig. 2 P esen s a isual o e iew o he da a gene a ion p ocess using
GAN a ian s. As shown in s ep 1, he da a is di ided in o aining and
es se s, and hen XGBoos and SVM classi ie s a e used o e alua ion.
In s ep 2, da a is gene a ed om he aining se using a ious selec ed
GAN a ian s. S ep 3 in ol es he c ea ion o a mixed da ase h ough
he combina ion o o iginal and syn he ic da a. Finally, in s eps 4 and
5, bo h da ase s- he syn he ic and he mixed da ase s-a e subjec ed o
e alua ion using classi ie s
bili y, and po abili y, making i sui able o applica ion o
abula and s uc u ed da a. I can be used o classi ica ion
and eg ession asks.
2.2.2 Suppo Vec o Machines Classi ie
Suppo Vec o Machine (SVM) [44] is a commonly used
supe ised machine lea ning echnique o e icien ly han-
dling imbalanced da a. SVM is commonly applied o sol e
bo h classi ica ion and eg ession p oblems. No ably, i
demons a ed ha he SVM classi ie is una ec ed by da ase
class imbalance compa ed o o he algo i hms. The co e con-
cep o he SVM app oach is o ind an op imal sepa a ion o
‘hype plane’ using he ke nel o sepa a e wo o mo e a i-
ous classes and c ea e a igid bounda y be ween he samples,
which will assis in classi ica ion and eg ession.
3 Expe imen al Me hodology
The wo k low o his s udy, as shown in Figu e 2,in ol esa
se ies o well-de ined phases. The i s s ep is o di ide he
o iginal da ase s in o wo subse s: he aining se and he es
se . The aining se ge s 80% o he da a, and he es se ge s
20%. The ain da a a e used o ins uc he gene a i e mod-
els, speci ically o aining he gene a o and disc imina o
ne wo ks o each GAN a chi ec u e o gene a e syn he ic
da a. A e he gene a ion o da a by GAN a ia ions, he
quali y o he syn he ic da a is e alua ed by compa ing he
esul s ob ained wi h classi ica ion algo i hms agains hose
ob ained using he o iginal da a.
3.1 Hype pa ame e selec ion
This sec ion desc ibes he hype pa ame e uning p ocedu e
o he selec ed GAN models. Since GANs a e di icul o
ain and can be sensi i e o hype pa ame e s [45]. Se e al
ba ch sizes, noise dimensions, numbe s o epochs, and lea n-
ing a es a e es ed. Rega ding he alida ion schema, we ha e
used s a i ied K- old c oss- alida ion.
•Ba ch Size We conduc ed expe imen s wi h a ious al-
ues, including 16, 32, 64, 100, 128, and 264 o de e mine
he op imal ba ch size.
•Noise Dimension We e alua ed he e ec o a ying he
noise dimension alues o 50, 64, 100, 256, 264, 300,
and 350 on aining and es ing ac oss nume ous GAN
a ian s and da ase s. Based on aining and es ing pe -
o mance, we de e mined he op imal noise dimension
size o each GAN a ian and da ase .
•Numbe o Epochs Du ing aining, he numbe o
epochs indica es he numbe o imes he en i e da ase
is p esen ed o he GAN model. We expe imen ed wi h
epoch numbe s om 100 o 1000 o ensu e sa is ac o y
esul s wi hou o e i ing. Mo e in o ma ion abou he
numbe o epochs will be discussed in sec ion 5: Resul s
and Discussion.
•Lea ning Ra e The lea ning a e is a c ucial hype pa-
ame e ha a ec s he weigh adjus men s ep size o he
GAN models du ing aining. We es ed mul iple lea n-
ing a es o ine- une he models, including 0.01, 0.001,
1e-4, 1e-5, and 5e-6. These alues we e chosen based on
hei e ec on aining s abili y, con e gence speed, and
123
In e na ional Jou nal o Da a Science and Analy ics
Table 1 Hype pa ame e s Used o GAN Va ian s
Pa ame e s Range o Values
Ba ch Size 16, 32, 64, 100, 128, 264
Noise Dimension 50, 64, 100, 256, 264, 300, 350
Numbe o Epochs Range om 100 o 1000
Lea ning Ra e 0.01, 0.001, 1e-4, 1e-5, 5e-6
S a i ied K- old C oss-Valida ion 5- olds, 10- olds
o e all pe o mance. In gene al, 1e-5 and 5e-6 wo ked
well on ou da ase s. A e conduc ing a se ies o expe -
imen s ac oss a ange o lea ning a e alues, we ound
ha he op imal lea ning a e alue is 5e-6 o he BCW
and LC da ase s o selec ed GAN a ian s. Fo he CTG
da ase , bo h he lea ning a e alues o 1e-5 and 5e-6
pe o med well acco ding o di e en GAN a ian s.
Table 1p o ides a summa y o he expe imen al hype -
pa ame e alues. The hype pa ame e s include ba ch size,
noise dimension, numbe o epochs, lea ning a e, and s a -
i ied K- old c oss- alida ion. The alues co esponding o
each hype pa ame e s illus a e he ange and op ions in es-
iga ed du ing he expe imen s. Tables A1 and A2 in he
Supplemen a y Ma e ial sec ion show a de ailed summa y
o op imal hype pa ame e s o each GAN model, classi ie ,
and da ase .
3.2 E alua ion Me ics
Based on he con usion ma ix, we e alua ed he pe o mance
o ou model using a ious me ics, including accu acy, sen-
si i i y, speci ici y, which measu es he abili y o he model o
accu a ely iden i y nega i e ins ances, and F1-sco e, which
is he ha monic mean o p ecision and ecall. These measu es
a e shown in Equa ions 1-5.
Accu acy =(TP) +(TN)
(TP+TN+FP+FN) (1)
P ecision =(TP)
(TP + FP) (2)
Sensi i i y (Recall) =(TP)
(TP + FN) (3)
Speci ici y =(TN)
(TN + FP) (4)
F1-sco e =2×P ecision ×Recall
P ecision +Recall (5)
3.3 Da ase
The da ase comp ises a ious ypes o digi al da a, includ-
ing nume ical, ca ego ical, ime-se ies, and ex da a. The
quan i y o da a signi ican ly impac s he quali y o imple-
men ing e ec i e algo i hms o machine lea ning mod-
els. Fo his pape , we used he B eas Cance Wisconsin
(Diagnos ic) and he Fe al Ca dio ocog aphy(CTG) da ase s,
which a e publicly a ailable on he UCI Machine Lea ning
Reposi o y, and he Lung Cance Pa ien da ase is a ailable
on Kaggle. Table 2p o ides a de ailed summa y o he o igi-
nal da ase s be o e emo ing any ea u es, o e ing essen ial
in o ma ion such as he o al numbe o cases, ea u es, and
he dis ibu ion o he classes wi hin he da ase s. Fi s , when
p epa ing he da a, we checked i he e we e missing o dupli-
ca ed alues o handle. Mo eo e , elimina ing i ele an o
edundan ea u es can imp o e he pe o mance o he model
[46] and educing he dimensionali y o a da ase by d op-
ping less in o ma i e ea u es can imp o e compu a ional
e iciency and educe he isk o o e i ing [47]. To unde -
s and he co ela ion ela ionship be ween ea u es, we e e
o Sec ion 4.
•B eas Cance Wisconsin (BCW) B eas Cance Wis-
consin (Diagnos ic) consis s o 569 pa ien s wi h b eas
umou s, o which 212 cases a e malignan , and he
emaining 357 cases a e benign. Thi y- wo ea u es cha -
ac e ise he umou s. Th ee p ope ies ep esen each
ea u e: mean, s anda d e o , and wo s alue and he
ea u es a e speci ied by eal alues, excep o he label,
which is ca ego ical. We no iced ha he BCW da ase
had no missing alues, bu he las column was emp y,
and he Id column was edundan and no use ul, so we
had o d op hem. Also, we d opped ea u es ha a e
highly co ela ed wi h each o he ; edundan ea u es can
be candida es o emo al. Mo e de ails ela ed o he
d opping ea u es a e p o ided in Supplemen a y Ma e-
ial, and Table A3 in he Supplemen a y Ma e ial sec ion
shows a lis o ea u es ha we e d opped.
•Lung Cance Pa ien (LC) The Lung Cance da ase
con ains 1000 eco ds and 25 ea u es indica ing he
symp oms and isk le els (low, medium, and high) asso-
cia ed wi h ac o s ela ed o lung cance . We used 24
ea u es and d opped he Pa ien Id column. The ea u es
a e scaled on ei he a (1-7), (1-8), o (1-9) scale, whe e
1 ep esen s he minimum le el and 7, 8, and 9 ep e-
sen he maximum le el. Table A3 in he Supplemen a y
Ma e ial sec ion shows he lis o d opped ea u es.
•Fe al Ca dio ocog aphy(CTG) The Ca dio ocog aphy
(CTG) da ase consis s o Fe al Hea Ra e (FHR) and
U e ine Con ac ion (UC) da a classi ied by medical p o-
essionals. I encompasses 2,126 e al ca dio ocog am
samples ha we e au onomously p ocessed. These sam-
ples a e ca ego ised in o 1655 no mal, 295 suspicious,
and 176 pa hologic samples. Resea che s can use his
da ase o explo e bo h 10-class and 3-class classi ica-
ion p oblems. The o iginal aw da a o he CTG da ase
123
In e na ional Jou nal o Da a Science and Analy ics
Table 2 Da ase s O e iew
Name o Da ase To al A ailable Cases A ibu es Dis ibu ion o Classes
B eas Cance Wisconsin 569 Pa ien s 32 Two imbalanced classes
(Malignan (M)=357,
Benign(B)=212)
Lung Cance Pa ien 1000 Reco ds 25 Th ee classes (Low, Medium, and
High- isk le els); Low- isk
le el=303, Medium isk le el
class=332, High- isk le el
class=365
Ca dio ocog aphy (CTG) 2126 Samples 40 Th ee imbalanced classes; No mal
Cases Class(N)=1655,
Suspicious Cases Class(S)=295,
Pa hological Cases Class(P)=176
consis s o 40 ea u es. Two a ailable e sions o he
da ase exis publicly, wi h one con aining 21 ea u es
and ano he con aining 23. Di e en esea che s ha e
used di e en numbe s o ea u es; some used 21 [48–
50] and o he s 23 [51–53] al hough no all ea u es a e
deemed equally impo an , en ea u es a e essen ial ea-
u es [54]. In his s udy, we op ed o 23 ea u es. Table
A3 in he Supplemen a y Ma e ial sec ion p o ides he
lis o d opped ea u es. Table A4 shows a copa ission
be ween he sizes o each da ase acc oss di e en GAN
models.
4 Resul s and Discussion
This sec ion p esen s a compa a i e s udy and discusses he
expe imen s conduc ed p ima ily on di e en da ase s. The
expe imen s a e designed o in es iga e he gene al p ope ies
and pe o mance o a ious GAN models o gene a ing syn-
he ic da a in he clinical domain. The en i onmen al se up
o he expe imen s includes Py hon 3.7.12, Colab and Kag-
gle No ebook. We employed he YDa a-Syn he ic package
[55] o implemen GAN, CGAN, CRAMER GAN, DRA-
GAN, and WGAN. Fo CTGAN implemen a ion, we u ilised
he p e-exis ing CTGAN [56] lib a y. Addi ionally, we used
o he s anda d Py hon lib a ies such as Pandas, NumPy, Ran-
dom, Seabo n, Ma plo lib, and Sciki -lea n.
We calcula ed co ela ions be ween eal and syn he ic
da ase s o unde s and a iable ela ionships and iden i y ou -
lie s. Fo he BCW da ase , we u ilised Pea son co ela ion
[57] due o i s con inuous na u e. On he o he hand, o he
LC and CTG da ase s, we applied Spea man’s ank co ela-
ion [58] due o hei o dinal alues. This ailo ed app oach
allowed o a nuanced analysis, conside ing he dis inc cha -
ac e is ics o each da ase .
Finally, o implemen he s a is ical es , we used he S a is-
ical Tes s o Algo i hms Compa ison (STAC) [59] pla o m,
which pe o ms s a is ical analysis o compa e ou comes p o-
duced by compu a ional in elligence algo i hms. I is publicly
accessible om he STAC webpage. The implemen a ions o
he GAN models p esen ed in his pape a e eely acces-
sible in he Gi Hub eposi o y (h ps://gi hub.com/Halal-
Abdul ahman-Ahmed/MedSyn h_GANVa ian s).
4.1 B eas Cance Wisconsin Da ase
The expe imen al esul s gained om he GAN models on
he BCW da ase exhibi ed ou s anding pe o mance on syn-
he ic da ase s, wi h ema kably simila ou comes. Recen
s udies ha e highligh ed he po en ial o GAN models o
gene a e da a ha closely esembles he o iginal da a, bu
his does no always lead o imp o ed classi ie pe o mance
[60].
Th ough he conduc ed expe imen s, we obse ed in
Figu e 3 ha CRAMER GAN, DRAGAN, and WGAN ou -
comes signi ican ly d opped when implemen ed on mixed
da ase s. The accu acy o he classi ie s using mixed da ase s
dec eased as mo e aining da a was p o ided, indica ing
po en ial o e i ing. This may be because GAN models
a e ained o gene a e da a simila o he aining da a,
no da a ep esen a i e o he eal wo ld, his means ha
GAN-gene a ed da a can some imes be misleading. Classi-
ie pe o mance on mixed da ase s is lowe han on syn he ic
da ase s because GANs a e mo e suscep ible o o e i o pa -
icula pa e ns in he o iginal da a. This is due o he ac ha
he p esence o syn he ic da a can impede he abili y o GAN
models o dis inguish be ween eal and ake da a, which can
lead he model o s uggle o gene alise and lea n pa e ns
speci ic o he aining da a, leading o a d op in pe o mance.
In e es ingly, CGAN and CTGAN demons a ed excep-
ional pe o mance, e en on he mixed da ase , he eby
highligh ing hei e ec i eness in gene a ing syn he ic da a,
as illus a ed in Figu e 3. Mo eo e , CTGAN was mo e
sui able han o he GAN models o gene a ing BCW da a
123
In e na ional Jou nal o Da a Science and Analy ics
Fig. 3 Compa a i e Accu acy
o GAN Va ian s wi h XGBoos
and SVM classi ie s on BCW
Da ase s
Fig. 4 Compa a i e Analysis o F1-sco es o GAN Va ian s on BCW Da ase s
because i is designed o gene a e abula da a due o i s condi-
ional gene a ion app oach, unlike o he GAN a ian s wi h
mo e gene al-pu pose applica ions. The F1-sco e compa -
ison in Figu e 4p o ides a comp ehensi e o e iew o he
expe imen ou comes, o e ing insigh s in o he p ecision and
ecall ade-o . F1-sco e, which is a balanced measu e o a
model’s o e all pe o mance, is c i ical in machine lea ning
e alua ions.
Figu e 5shows he Pea son co ela ion calcula ed on
bo h he o iginal BCW da ase and he da a gene a ed by
CTGAN, p o iding aluable insigh s in o he ela ionships
be ween ea u es. The hea map o he o iginal da ase
demons a es nume ous s ong posi i e co ela ions among
ea u es. No ably, he e a e no s ong nega i e co ela ions,
sugges ing a lack o consis en in e se ela ionships be ween
ea u es. I we ake he diagnosis a iable as an example,
we obse e s ong co ela ions, pa icula ly wi h pa am-
e e s like pa ame e _ mean,compac ness_mean,
conca i y_mean, and conca e_poin _wo s .We
could say ha diagnosis is he mos c i ical a iable, as
his a iable shows whe he he pe son has b eas can-
ce o no , so we could say ha he bes GAN model is
he model ha can cap u e he same co ela ion be ween
diagnosis and hose ea u es. As we can obse e, he
diagnosis ea u e shows he same pa e n, i.e. we can
obse e in he syn he ic da ase gene a ed by CTGAN
hese s ong co ela ions men ioned be o e be ween diag-
nosis and pa ame e _mean,compac ness_mean,
conca i y_mean, and conca e_poin _wo s .
In he CTGAN-gene a ed da a, he diagnosis a iable
demons a es s ong posi i e co ela ions wi h he ollowing
ea u es: conca i y_mean,a ea_se,compac ness
_mean, and conca e_poin _wo s . Th ee ou o he
op ou co ela ions be ween he diagnosis a iable and hese
ea u es a e e ec i ely cap u ed by he CTGAN model. The
co ela ion coe icien s o CTGAN a e 0.74, 0.77, and 0.78,
whe eas he co esponding alues o he o iginal da ase a e
0.60, 0.70, and 0.79. Addi ionally, CTGAN eplica es he
123
In e na ional Jou nal o Da a Science and Analy ics
Fig. 5 Hea maps o BCW Da ase s (O iginal and Syn he ic Da ase s)
co ela ion in he a ea_se ea u e mo e accu a ely han
he o iginal da ase , wi h alues o 0.73 e sus 0.55, espec-
i ely.
The co ela ions in da a gene a ed by he CGAN a e p e-
sen ed in Figu e A1 in he Supplemen a y Ma e ial sec ion.
Among he op ou ea u es wi h he s onges co ela ions
o he diagnosis, CGAN appea s o eplica e all ea u es bu
wi h a no able inc ease in linea i y be ween hem. Finally,
GAN, DRAGAN, CRAMERGAN, and WGAN could no
cap u e he co ela ion be ween diagnosis and o he ea u es,
as shown in Figu e A1 in he Supplemen a y Ma e ial sec ion.
4.2 Lung Cance Pa ien Da ase
The expe imen al esul s showed ha CGAN and CTGAN
demons a ed good pe o mance, showcasing good esul s in
he classi ie s using syn he ic da ase s and e en ou pe o m-
ing o he models on he mixed da ase , see Figu e 6. CTGAN
eme ged as he mos success ul, exhibi ing consis en pe -
o mance ac oss syn he ic and mixed da ase s, as shown in
Figu e 6. On he o he hand, GAN, CRAMER GAN, DRA-
GAN, and WGAN showed ela i ely weake pe o mance,
mainly when applied solely o syn he ic da ase s. Su p is-
ingly, he e was a on ie imp o emen in pe o mance on he
mixed da ase ; howe e , hese models s ill needed o achie e
sa is ac o y esul s. These esul s can be a consequence o
hype pa ame e unning issues. Figu e 7 ep esen s compa -
a i e F1-sco e ou comes achie ed om e alua ing di e en
GAN models.
Figu e 8compa es he o iginal da ase wi h he syn he ic
da a om CRAMER GAN and WGAN-gene a ed da ase s.
I is e iden om Figu e 8 ha he gene a ed syn he ic da a
con ains a signi ican numbe o ou lie s. The da a gene -
a ed by CRAMER GAN does no exhibi s ong co ela ions,
indica ing ha changes in one a iable a e no consis en ly
associa ed wi h changes in he o he . In Spea man’s ank
co ela ion hea map o WGAN, we obse e no co ela ion
be ween he co esponding pai o ea u es and no mono-
onic ela ionship be ween he anks o he alues in ea u es.
Addi ionally, some columns in he gene a ed da a consis
o andom alues. This andomness migh esul in a lack
123
In e na ional Jou nal o Da a Science and Analy ics
Fig. 6 Compa a i e Accu acy
o GAN Va ian s on LC Da ase s
Fig. 7 Compa a i e Analysis o
F1-sco es o GAN Va ian s on
LC Da ase s
o s uc u ed ela ionships be ween he a iables in hose
columns.
In his s udy, some ea u es in he gene a ed syn he ic da a
do no ha e he same dis ibu ion as in he o iginal da ase .
This indica es a lack o ideli y in eplica ing he s a is i-
cal pa e ns o he o iginal ea u es. The men ioned issues,
such as ou lie s, misma ched ea u e dis ibu ions, and he
absence o co ela ion be ween ea u es, nega i ely impac ed
he pe o mance o he GAN, CRAMER GAN, DRAGAN
and WGAN on he LC da ase . Figu e A1 in he Supple-
men a y Ma e ial sec ion illus a es a g oup o Spea man’s
ank co ela ion hea maps ha compa e o iginal and syn-
he ic da ase s gene a ed by selec ed GAN models.
4.3 Fe al Ca dio ocog aphy(CTG) Da ase
Figu es 9and 10 isually ep esen he expe imen al esul s
on he CTG da ase . No ewo hy indings indica e sa is-
ac o y pe o mance in e ms o accu acy and F1-sco e
ac oss he classi ie ’s pe o mance using he da a gene a ed
by he mos GAN models. Rema kably, CGAN demon-
s a ed he mos excep ional le el o pe o mance. CGAN
e eals ema kably imp o ed accu acy, exceeding he o ig-
inal da ase compa ed o he o he selec ed GAN models.
This highligh s ha in oducing addi ional in o ma ion, such
as class labels, in o he CGAN can d ama ically inc ease he
accu acy o he classi ie s. We no ice a sligh d op in he pe -
o mance on mixed da ase s, p obably due o he o e i ing
o he GAN models. The CTG da ase is mo e complex han
he o he da ase s used in his s udy. The complexi y de i es
om he na u e o he ea u es wi hin he CTG da ase , which
has many bina y and o dinal ea u es, and his can con ibu e
o he isk o o e i ing in a model.
Howe e , i is possible o gene a e syn he ic da a wi h
sa is ac o y accu acy esul s while ob aining less a ou able
co ela ion esul s. In machine lea ning, pa icula ly in gen-
123
In e na ional Jou nal o Da a Science and Analy ics
wi h an ensemble app oach. In o ma ion Sciences 580, 221–242
(2021). h ps://doi.o g/10.1016/j.ins.2021.08.081
65. Bane jee, A., Chi nis, U., Jadha , S., Bhawalka , J., Chaudhu y, S.:
Hypo hesis es ing, ype i and ype ii e o s. Indus ial psychia y
jou nal 18(2), 127 (2009)
66. Bhano , K., Pede sen, J., Guyon, I., Benne , K.P.: In es iga ing
syn he ic medical ime-se ies esemblance. Neu ocompu ing 494,
368–378 (2022)
67. Gonçal es, A., Ma os, S., al.: Gene a ion and e alua ion o syn-
he ic pa ien da a. Jou nal o Biomedical In o ma ics 112, 103611
(2020) h ps://doi.o g/10.1016/j.jbi.2020.103611
68. So anzo, N., Bianconi, G., Al a ini, C.: Compa ing associa ion
ne wo k algo i hms o e e se enginee ing o la ge-scale gene eg-
ula o y ne wo ks: syn he ic e sus eal da a. Bioin o ma ics 23(13),
1640–1647 (2007)
69. Wol , M., T i sche , J., Landes, D., Ho ho, A., Schlö , D.: Bench-
ma king o syn he ic ne wo k da a: Re iewing challenges and
app oaches. Compu e s & Secu i y, 103993 (2024)
70. Vega-Má quez, B., Rubio-Escude o, C., Nepomuceno-Chamo o,
I.: Gene a ion o syn he ic da a wi h condi ional gene a i e ad e -
sa ial ne wo ks. Logic Jou nal o he IGPL 30(2), 252–262 (2022)
71. B eas Cance Machine Lea ning P edic ion — g askas.gi hub.io.
h ps://g askas.gi hub.io/pos /b eas _cance /. [Accessed 18-03-
2024]
72. Tumo Diagnosis (Explo a o y Da a Analysis) — kaggle.com.
h ps://www.kaggle.com/code/ha ik ishna9/ umo -diagnosis-
explo a o y-da a-analysis#Explo a o y-Da a-Analysis. [Accessed
18-03-2024]
73. A S udy o a B eas Cance Da ase — B eas Cance Da a S udy
— ucb-s a -159-s22.gi hub.io. h ps://ucb-s a -159-s22.gi hub.io/
hw07-G oup26/README.h ml. [Accessed 18-03-2024]
74. Sumb ia, S.: B eas Cance Diagnos ic Da ase - EDA
— medium.com. h ps://medium.com/analy ics- idhya/b eas -
cance -diagnos ic-da ase -eda- a0de80 15bd. [Accessed 18-03-
2024]
Publishe ’s No e Sp inge Na u e emains neu al wi h ega d o ju is-
dic ional claims in published maps and ins i u ional a ilia ions.
123
