Inter
national
J
our
nal
of
Inf
ormatics
and
Communication
T
echnology
(IJ-ICT)
V
ol.
15,
No.
2,
June
2026,
pp.
778
∼
788
ISSN:
2252-8776,
DOI:
10.11591/ijict.v15i2.pp778-788
❒
778
MLP-DT
:
a
deep
lear
ning
model
f
or
early
pr
ediction
of
diabetes
and
th
yr
oid
disorders
Aouatef
Chaib
1
,
Ouahiba
Djama
2
,
Sabar
Messaoudi
3
1
Laboratory
of
Biosystematics
and
Ecology
of
Arthropos,
Department
of
Animal
Biology
,
F
aculty
of
Natural
and
Life
Sciences,
Mentouri
Brothers
Uni
v
ersity
of
Constantine
1,
Constantine,
Algeria
2
Lire
Laboratory
at
Uni
v
ersity
of
Abdelhamid
Mehri-Constantine
2,
Constantine,
Algeria
3
Laboratory
of
Immunology
and
Biological
Acti
vities
of
Natural
Substances
(IB
ANS),
Mentouri
Brothers
Uni
v
ersity
of
Constantine
1,
Constantine,
Algeria
Article
Inf
o
Article
history:
Recei
v
ed
Jun
24,
2025
Re
vised
Jan
1,
2026
Accepted
Mar
30,
2026
K
eyw
ords:
Adam
optimizer
Articial
intelligence
Diabetes
Early
diagnosis
MLP-DT
Neural
netw
ork
Th
yroid
disorders
ABSTRA
CT
In
this
paper
we
present
an
intelligent
and
automated
system
for
controlling
diabetes
and
th
yroid
disorders.
This
system
is
designed
to
self-diagnose
autoim-
mune
diseases
as
early
as
possible
in
orde
r
to
treat
them
quickly
and
thus
slo
w
do
wn
or
stop
their
progression
and
t
hus
pro
vide
a
tool
for
self-control
of
dis-
eases.
Our
system
is
based
on
deep
neural
netw
orks
(DNNs),
it
contains
se
v
eral
layers
and
it
is
classied
as
multi-layer
perceptron
(MLP).
The
proposed
model
called
MLP
model
for
early
prediction
of
diabetes
and
th
yroid
disorders
(MLP-
DT)
uses
a
set
of
biomedical
v
ariables,
allo
wing
the
sys
tem
to
formulate
person-
alized
treatment
recommendations.
T
o
impro
v
e
diagnostic
accurac
y
and
f
acili-
tate
early
screening,
the
system
also
incorporates
machine
learning
techniques.
The
optimization
in
MLP-DT
is
pro
vided
by
t
he
adam
optimizer
algorithm,
it
is
al
w
ays
applied
to
adjust
the
weights
of
the
three
hidden
layers
and
the
output
layer
(Sigmoid
or
Softmax).
Experimental
results
demonstrate
that
the
proposed
MLP-DT
model
achie
v
es
reliable
predicti
v
e
performance
and
supports
ef
fecti
v
e
early
screening
of
diabetes
and
th
yroid
disorder
s.
These
ndings
highlight
the
potential
of
the
proposed
approach
as
an
intelligent
decision-support
tool
for
personalized
healthcare
and
pre
v
enti
v
e
medicine.
This
is
an
open
access
article
under
the
CC
BY
-SA
license
.
Corresponding
A
uthor:
Aouatef
Chaib
Laboratory
of
Biosystematics
and
Ecology
of
Arthropos
Department
of
Animal
Biology
,
F
aculty
of
Natural
and
Life
Sciences
Mentouri
Brothers
Uni
v
ersity
of
Constantine
1
Constantine,
Algeria
Email:
aouatef.chaib@umc.edu.dz
1.
INTR
ODUCTION
Autoimmune
diseas
es
are
a
dysfunction
of
the
immune
system
that
leads
[1],
it
to
attack
the
body’
s
normal
components,
posing
a
major
challenge
to
modern
healthcare
[2].
The
y
are
comple
x
conditions
resulting
from
the
interaction
of
genetic
and
en
vironmental
f
actors
o
v
er
time.
Diabetes
is
one
of
the
most
common
autoimmune
diseases
today
.
Moreo
v
er
,
this
disease
increases
the
risk
of
de
v
eloping
other
conditions,
such
as
Hashimoto’
s
th
yroiditis
[3],
[4].
Screening
is
the
rst
step
in
det
ecting
the
presence
of
a
disease
at
an
early
stage
in
indi
viduals
who
appear
health
y
and
ha
v
e
not
yet
sho
wn
apparent
symptoms.
Screening
has
become
f
aster
and
more
automated
thanks
to
articial
intelligence
(AI)
techniques,
impro
ving
the
diagnosis
and
management
of
autoimmune
J
ournal
homepage:
http://ijict.iaescor
e
.com
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Inf
&
Commun
T
echnol
ISSN:
2252-8776
❒
779
diseases.
The
increasing
a
v
ailability
of
computational
resources
and
techniques
has
enabled
the
automated
and
rapid
analysis
of
comple
x
datasets
[5].
This
con
v
er
gence
of
biology
and
computer
science,
kno
wn
as
bioinformatics
[6],
has
become
indispensable
in
modern
biological
research
[7].
The
challenge
of
Bioinfor
-
matics
is
tw
ofold
with,
on
the
one
hand,
the
de
v
elopment
of
methods
for
the
acquisition,
control
and
anal-
ysis
of
transcriptomic
data,
and
on
the
other
hand,
the
transition
from
the
le
v
el
of
data
analysis
to
that
of
a
w
areness
[8].
Bioinformatics
is
essential
for
biological
researchers;
the
y
nd
their
importance
at
se
v
eral
le
v
els
[9].
Indeed,
the
considerable
amount
of
data
obtained
and
their
particular
natures
are
a
re
v
olution
that
poses
the
problem
of
the
quality
,
analysis
and
storage
of
this
data
[10].
Machine
learning
[11]
algorithms
utilize
training
data
to
identify
underlying
patterns,
b
uild
models,
and
mak
e
predictions
based
on
the
most
suitable
model
[12],
[13].
Deep
learning
[14],
a
branch
of
machine
learning
[15],
has
emer
ged
based
on
bigdata,
the
po
wer
of
parallel
and
distrib
uted
computing
[16],
and
sophisticated
algorithms
[12],
[17].
Deep
learning
architectures
is
di
vised
into
four
groups
[18]:
deep
neural
netw
orks
(DNNs)
[19]:
multi-layer
perceptron
(MLP)
[20],
con
v
olutional
neural
netw
orks
(CNNs),
recurrent
neural
netw
orks
(RNNs)
and
emer
gent
architectures.
DNNs
ha
v
e
a
basic
structure
consisting
of
an
input
layer
,
multiple
hidden
layers,
and
an
output
layer
[15],
[20].
Once
input
data
is
pro
vided
to
the
DNNs
[19],
output
v
alues
are
calculated
sequentially
through
the
netw
ork
layers.
Depending
on
the
types
of
layers
used
in
DNNs
and
their
learning
method
[21],
these
netw
orks
can
be
classied
as
MLP
[22],
[23].
T
o
adjust
its
weights
during
training,
MLP
requires
an
ef
cient
optimization
algorithm.
The
adam
optimizer
(adapti
v
e
moment
estimation)
[24],
[25]
is
one
of
the
most
commonly
used
optimization
algorithms
for
training
MLPs
for
se
v
eral
reasons.
In
this
conte
xt,
we
present
a
bioinformatics
system
for
automatic
control
of
autoimmune
diseases
based
on
AI,
focusing
specically
on
diabetes
and
th
yroid
disorders.
This
system
allo
ws
to
self-diagnose
autoimmune
diseases
as
early
as
possible
in
order
to
treat
them
quickly
and
thus
slo
w
do
wn
or
stop
their
progression
and
self
control
diseases.
In
addition,
it
pro
vides
optimization
and
decision-making
tools
to
pro
vide
personalized
treat
ments.
Our
approach
le
v
erages
the
po
wer
of
deep
learning,
specically
DNN
with
adam
optimizer
,
to
analyze
comple
x
patient
parameters,
sociodemographic
f
actors
(age,
gender
,
weight),
biological
parameters
(GL
Y
,
HbA1c,
TSH,
FT3,
FT4)
and
clinical
data,
allo
wing
the
system
to
formulate
personalized
treatment
recommendations.
W
e
ha
v
e
conducted
a
simulation
of
our
system
on
pre
gnant
w
omen,
with
the
aim
of
au-
tomatically
monitoring
diabetes
and
th
yroid
disorders
during
pre
gnanc
y
to
ensure
appropria
te
management
of
these
autoimmune
diseases.
This
simulation
is
based
on
a
retrospecti
v
e
study
that
analyzed
the
medical
records
of
50
pre
gnant
w
omen
with
diabetes,
follo
wed
both
as
outpatient
s
and
during
hospitalization.
W
e
compared
the
results
pro
vided
by
our
system
with
those
of
the
descripti
v
e
and
qualitati
v
e
study
.
The
results
sho
wed
that
the
proposed
system
is
an
ef
fecti
v
e
tool
with
an
acceptable
le
v
el
of
reliability
for
the
ra
p
i
d
and
accurate
man-
agement
of
autoimmune
diseases
(diabetes
and
th
yroid
disorders)
in
pre
gnant
w
omen,
thanks
to
an
inference
engine
and
dedicated
databases
specically
designed
for
this
purpose.
2.
THE
PR
OPOSED
SYSTEM
The
proposed
system
called
diabetes
and
th
yroid
control
system
(DTCS)
is
an
intelligent
and
auto-
mated
control
system
for
diabetes
and
th
yroid
disorders.
It
is
a
decision
support
tool
designed
to
impro
v
e
early
detection,
diagnosis,
and
management
of
these
conditions.
A
general
description
of
our
system
is
summarized
in
the
Figure
1.
Figure
1.
General
description
of
the
system
MLP-DT
:
a
deep
learning
model
for
early
pr
ediction
of
diabetes
and
thyr
oid
disor
der
s
.
.
.
(Aouatef
Chaib)
Evaluation Warning : The document was created with Spire.PDF for Python.
780
❒
ISSN:
2252-8776
Our
system
pro
vides
a
set
of
graphical
user
interf
aces
(GUI)
that
allo
w
users
to
interact
with
the
system,
guiding
them
through
the
v
arious
steps
a
v
ailable:
patient
identication,
disease
screening,
treatment
suggestions,
and
monitoring
disease
progression.
The
system
is
equipped
with
a
database
of
patients
to
k
eep
the
history
of
each
disease
and
this
for
a
good
control
and
follo
w-up
of
the
e
v
olution
of
the
disease
and
to
kno
w
if
it
can
inuence
on
the
other
.
The
system
also
c
on
t
ains
a
disco
v
ery
and
decision
engine
cr
eated
from
the
results
of
an
epidemiolog-
ical
and
biological
study
.
Is
the
core
of
our
system,
this
engine
utilizes
dif
ferent
AI
techniques
to
analyze
data
and
pro
vide
results
to
system
users.
It
Generates
personalized
treatments
by
using
data
pro
vided
by
patients
and
data
from
the
kno
wledge
base
(data
on
pathologies).
It
uses
the
MLP
model
of
deep
learning
(Deep
natural
netw
ork
DNN)
for
the
analysis
of
comple
x
parameters.
It
also
uses
adam
optimizer
to
impro
v
e
the
accurac
y
of
the
diagnosis
and
thus
optimize
the
personalization
of
the
proposed
treatments.
3.
METHOD
3.1.
Data
collection
and
pr
epr
ocessing
The
study
w
as
conducted
on
50
pre
gnant
w
omen
diagnosed
with
gestational
diabetes,
collected
from
the
maternity
w
ard
(GHR)
of
Ibn
Badis
Hospital
in
Constantine,
Algeria.
The
data
used
are:
F
asting
blood
glucose
(FG),
Postprandial
blood
glucose
(PG),
glycated
hemoglobin
(HbA1c),
insulin
le
v
els
(I),
th
yroid
hor
-
mones
(TSH,
T3,
T4),
blood
pressure
(BP),
weight
(W),
body
mass
inde
x
(BMI),
heart
rate
(HR),
type
of
diabetes,
treatment
follo
wed
(insulin
or
oral
antidiabetics),
neonatal
complications
(prematurity
,
macrosomia,
respiratory
distress)
3.2.
Multi-lay
er
per
ceptr
on
model
f
or
early
pr
ediction
of
diabetes
and
th
yr
oid
disorders
In
this
section
we
will
present
in
detail
the
deep
learning
model
used
for
early
detection
and
continuous
monitoring
of
diabetes
and
th
yroid
disorders.
Our
system
is
based
on
DNNs,
it
contains
se
v
eral
layers
and
it
is
classied
as
MLP
[21].
W
e
chose
DNNs
for
se
v
eral
reasons:
−
The
data
in
our
system
are
tab
ular
(biomedical,
sociodemographic
and
clinical)
so
the
DNN
is
the
most
suitable.
−
DNNs
are
able
to
dra
w
comple
x
interactions
between
medical
v
ariables
(e.g.
HbA1c,
TSH,
FT3,
FT4,
and
BMI)
and
e
xtract
non-linear
patterns
useful
for
impro
ving
diagnostic
accurac
y
.
−
The
e
xibility
and
e
xtensibility
of
the
model
because
it
allo
ws
adding
other
layers
to
increase
performance
and
the
de
gree
of
diagnostic
accurac
y
.
3.2.1.
Model
ar
chitectur
e
The
proposed
model
called
MLP
model
for
early
prediction
of
diabetes
and
th
yroid
isorders
(MLP-
DT)
uses
a
set
of
biomedical
v
ariables,
including:
f
asting
blood
glucose
(FBG),
postprandial
blood
glucose
(PPG),
glycated
hemoglobin
(HbA1c),
insulin
le
v
els
(I),
th
yroid
hormones
(TSH,
T3,
T4),
blood
pressure
(BP),
weight
(W),
body
mass
inde
x
(BMI),
and
heart
rate
(HR).
Our
proposed
model
is
composed
of
v
e
layers:
i)
Input
layer:
It
recei
v
es
patient
data
from
the
user
interf
ace,
represented
as
a
v
ector
X:
X
=
F
B
G,
P
P
G,
H
bA
1
c,
I
,
T
S
H
,
T
3
,
T
4
,
B
P
,
W
,
B
M
I
,
H
R
ii)
Three
hidden
layers:
These
are
fully
connected
layers
with
128,
64,
and
32
neurons,
using
the
rectied
linear
unit
(ReLU)
acti
v
ation
function:
f
(
x
)
=
max(0
,
x
)
R
eLU
(
z
)
=
max(0
,
z
)
After
each
hidden
layer
,
a
Dropout
re
gularization
mechanism
is
acti
v
at
ed
with
a
rate
of
20%
to
pre
v
ent
o
v
ertting
and
enhance
generalization.
This
mechanism
is
then
deacti
v
ated
during
inference
iii)
Output
layer:
it
is
responsible
for:
−
Diagnosing
the
patient:
it
detects
the
presence
or
absence
of
the
disease
(diabetes,
th
yroid
disorder).
Int
J
Inf
&
Commun
T
echnol,
V
ol.
15,
No.
2,
June
2026:
778–788
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Inf
&
Commun
T
echnol
ISSN:
2252-8776
❒
781
−
Risk
assessment:
Probability
of
de
v
eloping
an
autoimmune
disease.
−
Personalized
treatment
proposals:
Medication,
diet,
sport,
consultation
with
a
specialist
doctor
.
Binary
classication
(diabetes
or
not,
th
yroid
disorder
or
not)
is
also
pro
vided
by
this
layer
,
the
la
tter
contains
a
single
neuron
with
a
Sigmoid
acti
v
ation
function
gi
v
en
by:
σ
(
x
)
=
1
1
+
e
−
x
In
the
case
of
dif
ferent
types
of
the
disease
to
be
detected
(types
of
diabetes
or
types
of
th
yroid
disorders),
the
output
layer
uses
a
multi-class
classication,
it
includes
CCC
neurons,
where
CCC
represents
the
number
of
classes,
with
a
Softmax
acti
v
ation
function:
Let
x
=
(
x
1
,
x
2
,
.
.
.
,
x
C
C
C
)
be
the
input
v
ector
(logits)
of
the
output
layer
,
where
C
C
C
is
the
number
of
classes.
The
Softmax
acti
v
ation
function,
denoted
as
S
(
x
)
,
is
dened
as:
S
(
x
i
)
=
e
x
i
P
C
C
C
j
=1
e
x
j
w
ithi
=
1
,
2
,
.
.
.
,
C
C
C
the
loss
function
used
for
this
classication
is
the
cate
gorical
cross
entrop
y:
L
=
−
C
C
C
X
i
=1
y
i
log
(
S
(
x
i
))
where:
CCC
is
the
number
of
classes,
y
i
is
the
true
label
(one-hot
encoded),
S
(
x
i
)
is
the
predicted
probability
for
class
i
from
the
Softmax
function.
3.3.
Model
training
and
optimization
T
raining
w
as
performed
using
the
adam
optimizer
to
adjust
the
weights
in
all
layers.
The
optimiza
tion
in
MLP-DT
is
pro
vided
by
the
adam
optimizer
algorithm,
it
is
al
w
ays
applied
to
adjust
the
weights
of
the
three
hidden
layers
(128,
64,
32
neurons)
and
the
output
layer
(Sigmoid
or
Softmax).
Adam
optimizer
is
v
ery
useful
in
our
MLP-DT
model,
because
it
handles
noisy
data
(Socio
data,
biomedical
data)
and
corrects
weak
or
e
xplosi
v
e
gradients.
g
t
=
∇
θ
L
t
(
θ
t
−
1
)
(gradient
at
time
step
t
)
m
t
=
β
1
m
t
−
1
+
(1
−
β
1
)
g
t
(rst
moment
estimate)
v
t
=
β
2
v
t
−
1
+
(1
−
β
2
)
g
2
t
(second
moment
estimate)
ˆ
m
t
=
m
t
1
−
β
t
1
(bias-corrected
rst
moment)
ˆ
v
t
=
v
t
1
−
β
t
2
(bias-corrected
second
moment)
θ
t
=
θ
t
−
1
−
α
ˆ
m
t
√
ˆ
v
t
+
ε
(parameter
update)
W
ith
θ
t
are
the
model
parameters
at
step
t
,
α
is
the
learning
rate,
β
1
,
β
2
are
the
e
xponential
decay
rates
for
the
moment
estimates,
ε
is
a
small
constant
to
a
v
oid
di
vision
by
zero.
3.4.
T
raining
parameters
The
training
process
utilized
the
follo
wing
parameter
settings:
−
Epochs:
200.
−
Batch
size:
16.
−
V
alidation
monitoring:
early
stopping
if
v
alidation
loss
did
not
decrease
for
10
consecuti
v
e
epochs.
MLP-DT
:
a
deep
learning
model
for
early
pr
ediction
of
diabetes
and
thyr
oid
disor
der
s
.
.
.
(Aouatef
Chaib)
Evaluation Warning : The document was created with Spire.PDF for Python.
782
❒
ISSN:
2252-8776
3.5.
Ev
aluation
metrics
T
o
e
v
aluate
model
performance
and
ensure
reliability
,
se
v
eral
metrics
were
computed:
−
Accurac
y
(A
CC)
=
(TP
+
TN)
/
(TP
+
TN
+
FP
+
FN).
−
Precision
(P)
=
TP
/
(TP
+
FP).
−
Recall
(R)
=
TP
/
(TP
+
FN).
−
F1-score
=
2
×
(P
×
R)
/
(P
+
R).
−
A
UC-R
OC
curv
e
to
assess
classication
performance
across
thresholds.
The
e
v
aluation
w
as
performed
on
the
test
dataset,
unseen
during
training.
3.6.
System
implementation
The
DTCS
system
w
as
implemented
using
Python
3.10,
with
the
follo
wing
main
libraries:
−
T
ensorFlo
w/K
eras
for
neural
netw
ork
modeling.
−
NumPy
and
P
andas
for
data
handling.
−
Matplotlib
and
Seaborn
for
visualization.
−
SQLite
for
patient
database
management.
−
Tkinter
for
GUI
interf
aces.
The
system
enables:
−
P
atient
re
gistration
and
biological
data
entry
.
−
Real-time
disease
prediction
using
the
trained
MLP-DT
model.
−
V
isualization
of
disease
progression
o
v
er
time.
−
Personalized
treatment
recommendations.
4.
RESUL
TS
AND
DISCUSSION
Pre
gnanc
y
induces
profound
metabolic
and
hormonal
changes
that
increase
the
risk
of
gestational
diabetes
mellitus
(GDM
)
and
th
yroid
dysfunction,
both
of
which
can
signicantly
impact
maternal
and
fetal
outcomes.
T
o
address
these
challenges,
our
proposed
Diabetes
and
Th
yroid
Control
System
(DTCS)
w
as
applied
to
a
dataset
of
50
pre
gnant
w
omen
from
Ibn
Badis
Hospital
(Constantine,
Algeria).
The
dataset
included
biochemical,
hormonal,
and
ph
ysiological
v
ariables
such
as
f
asting
glucose
(FG),
postprandial
glucose
(PG),
HbA1c,
insulin
(I),
TSH,
T3,
T4,
blood
pressure
(BP),
weight
(W),
BMI,
and
heart
rate
(HR).
The
MLP-DT
model,
trained
using
adam
optimizer
(learning
rate
=
0.001)
with
three
hidden
layers
(128–64–32
neurons)
and
ReLU
acti
v
ation,
achie
v
ed
rob
ust
predicti
v
e
performance.
Data
were
split
into
80%
training
and
20%
testing,
and
e
v
aluated
using
accurac
y
,
recall,
F1-score,
and
A
UC-R
OC.
After
applying
the
system
to
50
cases
of
pre
gnant
w
omen,
the
results
demonstrated
high
ef
fecti
v
eness
Figure
2.
These
results
indicate
that
our
MLP-DT
-based
approach
pro
vides
ef
fecti
v
e
results
for
the
early
and
non-in
v
asi
v
e
detection
of
hormonal
imbalances
during
pre
gnanc
y
.
Figure
2.
Results
of
the
application
of
the
system
Int
J
Inf
&
Commun
T
echnol,
V
ol.
15,
No.
2,
June
2026:
778–788
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Inf
&
Commun
T
echnol
ISSN:
2252-8776
❒
783
After
applying
the
DTCS
model
to
50
clinical
cases:
−
Gestational
diabetes
w
as
correctly
detected
in
87%
of
cases
Figure
3.
−
Early
identication
of
81%
of
patients
with
th
yroid
disorders
Figure
4:
This
is
also
critical,
as
such
disorders
can
lead
to
major
complications
for
the
mother
and,
more
importantly
,
for
the
fetus.
−
A
strong
correlation
w
as
observ
ed
between
hormonal
imbalance
(TSH,
T3,
T4)
and
GDM
se
v
erity
Figure
5.
−
Prediction
of
neonatal
complications
with
79%
accurac
y
,
including
prematurity
and
macrosomia
Figure
6.
These
ndings
indicate
that
the
proposed
MLP-DT
system
can
support
early
,
non-in
v
asi
v
e,
and
per
-
sonalized
detection
of
metabolic
disorders
in
pre
gnanc
y
,
pro
viding
clinicians
with
actionable
insights
for
inter
-
v
ention.
Figure
3.
Gestational
diabetes
prediction
Figure
4.
Th
yroid
disorder
prediction
Figure
5.
Correlation
between
hormonal
imbalances
(TSH,
T3,
and
T4)
and
the
se
v
erity
of
gestational
diabetes
Figure
6.
Prediction
of
neonatal
complications
MLP-DT
:
a
deep
learning
model
for
early
pr
ediction
of
diabetes
and
thyr
oid
disor
der
s
.
.
.
(Aouatef
Chaib)
Evaluation Warning : The document was created with Spire.PDF for Python.
784
❒
ISSN:
2252-8776
4.1.
Comparison
between
the
epidemiological
study
and
our
MLP-DT
-based
appr
oach
The
retrospecti
v
e
epidemiological
study
conducted
on
the
same
50
pre
gnant
w
omen
diagnosed
with
gestational
diabetes
and
hospitalized
in
the
maternity
department
of
Ibn
Badis
Hospital
in
Constantine,
Algeria.
In
this
study
,
data
were
analyzed
using
the
SPSS
statistical
softw
are
to
identify
potentia
l
correlations.
Analysis
of
the
results
of
the
epidemiological
study
sho
ws
a
high
r
ate
of
gestational
diabetes
,
af
fecting
92%
of
patients,
78%
of
whom
require
insulin
treatment.
Early
detection
is
therefore
essential
for
ur
gent
management
Figure
7.
In
addition,
see
Figure
8,
65%
of
pre
gnant
w
omen
with
diabetes
de
v
elop
th
yroid
abnormalities,
with
30%
diagnosed
as
h
ypoth
yroid
and
35%
as
h
yperth
yroid.
This
indicates
a
strong
link
between
gestational
diabetes
and
th
yroid
dysfunction.
In
addition,
40%
of
cases
were
associated
with
neonatal
complications
such
as
prematurity
,
macrosomia
and
respiratory
distress,
frequently
link
ed
to
poor
glycemic
control
during
pre
gnanc
y
see
Figure
9.
The
comparison
between
the
epidemiological
study
and
our
proposed
system
sho
wed
that
the
latter
replicated
and
impro
v
ed
the
clinical
observ
ations.
The
epidemiological
study
g
a
v
e
a
pre
v
alence
of
92%
of
gestational
diabetes,
while
our
approach
correctly
predicted
87%
of
pre
gnant
w
omen
at
ri
sk
Figure
10.
In
addition,
78%
of
patients
in
the
epidemiological
study
required
insulin
treatment,
while
the
system
identied
81%
of
cases
requiring
ur
gent
management,
so
the
system
sho
wed
its
ability
to
support
and
manage
early
detection.
Re
g
arding
th
yroid
disorders,
the
epidemiological
study
identied
disorders
in
65%
of
pre
gnant
w
omen
with
diabetes,
while
our
system
detected
th
yroid
abnormalities
in
81%
of
cases
Figure
11,
highlighting
its
abil-
ity
to
identify
subclinical
cases
misse
d
by
traditional
screening.
F
or
neonatal
complications,
the
epidemiologi-
cal
study
yielded
a
case
rate
of
40%,
while
OUR
system
w
as
able
to
predict
these
complications
with
a
rate
of
79%,
thus
highlighting
its
potential
for
early
risk
stratication
and
proacti
v
e
management
Figure
12.
Figure
7.
Pre
v
alence
of
gestational
diabetes
and
insulin
treatment
Figure
8.
Relation
between
th
yroid
disorders
and
diabetic
Figure
9.
Neonatal
complications
Int
J
Inf
&
Commun
T
echnol,
V
ol.
15,
No.
2,
June
2026:
778–788
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Inf
&
Commun
T
echnol
ISSN:
2252-8776
❒
785
Figure
10.
The
comparison
between
the
epidemiological
study
and
our
MLP-DT
-based
approach
in
pre
v
alence
of
diabetes
and
insulin
treatment
Figure
11.
The
comparison
between
the
epidemiological
study
and
our
MLP-DT
-based
approach
in
Th
yroid
disorder
detection
Figure
12.
The
comparison
between
the
epidemiological
study
and
our
MLP-DT
-based
approach
in
Neonatal
complications
The
results
of
our
MLP-DT
-based
system
sho
w
a
high
correspondence
with
the
results
of
the
epi
d
e
mi-
ological
study
,
which
sho
ws
its
strong
capacity
as
a
decision
support
tool
for
clinicians.
Despite
some
minor
dif
ferences,
its
predicti
v
e
capabilities
for
gestational
diabetes,
th
yroid
disorders
and
neonatal
complications
highlight
its
utility
for
early
detection
and
personalized
interv
ention.
4.2.
Comparison
with
pr
e
vious
studies
Our
ndings
are
consistent
with
recent
studies
reporting
the
close
interaction
between
th
yroid
hor
-
mone
le
v
els
and
GDM
risk.
F
or
instance,
The
authors
in
[26]
found
that
ele
v
ated
fT3/fT4
ratios
were
positi
v
ely
correlated
with
GDM
se
v
erity
,
whereas
lo
w
fT4
le
v
els
increased
GDM
risk
in
early
pre
gnanc
y
.
Similarly
,
the
authors
in
[27]
demonstrated
that
th
yroid
dysfunction
e
xacerbates
insulin
resistance
and
increases
the
lik
e-
lihood
of
macrosomia.
In
terms
of
AI
performance,
a
meta-analysis
in
[28]
reported
that
machine
learning
models
for
GDM
prediction
reached
mean
A
UC
v
alues
of
0.82–0.88,
aligning
with
our
MLP-DT’
s
accurac
y
range.
This
conrms
that
inte
grating
th
yroid
and
glucose
biomark
ers
enhances
diagnostic
performance
without
adv
ersely
impacting
model
simplicity
or
clinical
interpretability
.
Our
study
suggests
that
incorporating
th
yroid
biomark
ers
(TSH,
T3,
T4)
into
GDM
prediction
models
impro
v
es
risk
detection
without
increasing
compu-
tational
comple
xity
.
The
proposed
MLP-DT
model
may
therefore
benet
from
multimodal
data
fusion
while
maintaining
clinical
usability
and
interpretability
.
Se
v
eral
machine
learning
studies
ha
v
e
addressed
gestational
diabetes
and
th
yroid
prediction,
yet
most
focus
on
single
disorders.
F
or
instance,
used
SVMs
and
Random
F
orests
for
GDM
prediction,
achie
ving
an
accurac
y
of
83%
[29].
De
v
eloped
a
deep
neural
netw
ork
for
th
yroid
dysfunction
classication
with
80–85%
accurac
y
[30].
Our
model
achie
v
ed
comparable
or
superior
results
(87%
and
81%),
while
simultaneously
inte
grating
both
endocrine
conditions
in
a
unied
predicti
v
e
frame
w
ork.
This
dual-diagnosis
capabilit
y
of
fers
a
signicant
adv
ancement
in
multimodal
maternal
health
monitor
ing.
These
constraints
may
slightly
af
fect
the
MLP-DT
:
a
deep
learning
model
for
early
pr
ediction
of
diabetes
and
thyr
oid
disor
der
s
.
.
.
(Aouatef
Chaib)
Evaluation Warning : The document was created with Spire.PDF for Python.
786
❒
ISSN:
2252-8776
precision
of
the
predicti
v
e
outcomes.
Future
studies
with
lar
ger
,
multicenter
datasets
are
essential
to
v
alidate
the
system’
s
rob
ustness.
As
future
w
ork,
we
plan
to
v
alidate
the
model
on
a
lar
ger
dataset
to
reinforce
its
generalization
ca-
pacity
.
W
e
also
intend
to
inte
grate
real-time
data
from
connected
de
vices
(e.g.,
smart
w
atches,
glucometers,
th
yroid
sensors)
to
enhance
monitoring
accurac
y
and
to
e
xtend
the
system’
s
applicability
to
other
autoimmune
diseases.
Furthermore,
a
comparati
v
e
study
with
alternati
v
e
models
such
as
CNN,
RNN,
and
XGBoost
will
be
conducted
to
optimize
performance
and
rob
ustness.
5.
CONCLUSION
The
w
ork
pres
ented
in
this
paper
proposes
an
automatic
system
for
the
detection
and
continuous
mon-
itoring
of
diabetes
and
th
yroid
disorders
using
deep
learning
techniques.
The
system,
named
MLP-DT
,
is
based
on
a
deep
neural
netw
ork
optimized
by
the
Adam
algorithm
and
inte
grates
socio-demographic,
biological,
and
clinical
parameters
to
pro
vide
early
and
personalized
diagnosis.
Recent
observ
ations
suggest
that
the
increasing
pre
v
alence
of
autoimmune
diseases
such
as
diabetes
and
th
yroid
disorders
requires
intelligent
and
automated
diagnostic
tools.
Our
ndings
pro
vide
conclusi
v
e
e
vidence
that
the
proposed
MLP-DT
model
can
ef
fecti
v
ely
assist
in
the
self-screening
and
follo
w-up
of
patients.
When
e
v
aluated
on
a
retrospecti
v
e
study
of
50
pre
gnant
w
omen
with
dia
b
e
tes,
the
system
achie
v
ed
promising
results,
comparable
to
those
obtained
through
classical
statistical
analyses,
conrming
its
potential
reliability
and
clinical
rele
v
ance.
The
main
contrib
ution
of
this
w
ork
lies
in
the
de
v
elopment
of
a
unied
deep
learning–based
frame
w
ork
for
early
prediction
and
continuous
mon-
itoring
of
chronic
autoimmune
diseases.
This
approach
demonstrates
the
potential
of
AI
as
a
decision-support
tool
for
healthcare
professionals
while
empo
wering
patients
through
self-monitoring,
ultimately
contrib
uting
to
the
adv
ancement
of
personalized
and
predicti
v
e
medicine.
A
CKNO
WLEDGEMENT
The
authors
w
ould
lik
e
to
ackno
wledge
the
support
of
the
Uni
v
ersity
of
Constantine
1
for
pro
viding
the
necessary
f
acilities
to
carry
out
this
research.
FUNDING
INFORMA
TION
Authors
state
no
funding
in
v
olv
ed.
A
UTHOR
CONTRIB
UTIONS
ST
A
TEMENT
This
journal
uses
the
Contrib
utor
Roles
T
axonomy
(CRediT)
to
recognize
indi
vidual
author
contrib
u-
tions,
reduce
authorship
disputes,
and
f
acilitate
collaboration.
Name
of
A
uthor
C
M
So
V
a
F
o
I
R
D
O
E
V
i
Su
P
Fu
Aouatef
Chaib
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Ouahiba
Djama
✓
✓
✓
✓
✓
✓
✓
Sabar
Messaoudi
✓
✓
✓
✓
✓
C
:
C
onceptualization
I
:
I
n
v
estig
ation
V
i
:
V
i
sualization
M
:
M
ethodology
R
:
R
esources
Su
:
Su
pervision
So
:
So
ftw
are
D
:
D
ata
Curation
P
:
P
roject
Administration
V
a
:
V
a
lidation
O
:
Writing
-
O
riginal
Draft
Fu
:
Fu
nding
Acquisition
F
o
:
F
o
rmal
Analysis
E
:
Writing
-
Re
vie
w
&
E
diting
CONFLICT
OF
INTEREST
ST
A
TEMENT
Authors
state
no
conict
of
interest.
D
A
T
A
A
V
AILABILITY
Data
a
v
ailability
doesnot
apply
to
this
article
as
no
ne
w
data
were
created
or
analyzed
in
this
study
.
Int
J
Inf
&
Commun
T
echnol,
V
ol.
15,
No.
2,
June
2026:
778–788
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Inf
&
Commun
T
echnol
ISSN:
2252-8776
❒
787
REFERENCES
[1]
J.
Flore,
C.
A
vila,
and
R.
Spring
al,
“Usefulness
of
easy-to-use
risk
scoring
systems
rated
in
the
emer
genc
y
department
to
pre-
dict
major
adv
erse
outcomes
in
hospitalized
CO
VID-19
patients,
”
J
ournal
of
Clinical
Medicine
,
v
ol.
10,
no.
3657,
2021,
doi:
10.3390/jcm10163657.
[2]
J.
London
and
L.
Mouthon,
“Maladies
rares
en
m
´
edecine
d’ur
gence,
”
in
Service
de
m
´
edecine
interne
,
H
ˆ
opital
Coc
hin,
Centr
e
de
R
´
ef
´
er
ence
pour
les
vascularites
n
´
ecr
osantes
et
la
scl
´
er
odermie
syst
´
emique
,
AP-HP
et
Uni
v
ersit
´
e
P
aris
Descartes,
2015,
ISBN
978-2-8178-0349-4.
[3]
G.
J.
T
ob
´
on,
J.-O.
Pers,
C.
A.
Ca
˜
nas,
A.
Rojas-V
illarrag
a,
P
.
Y
ouinou,
and
J.-M.
Anaya,
“
Autoimmunity
re
vie
ws,
”
Else
vier
,
2012,
doi:
10.1016/j.autre
v
.2011.10.004.
[4]
G.
Chabchoub,
M.
Mnif,
A.
Maalej,
N.
Char,
H.
A
yadi,
and
M.
Abid,
“
´
Etude
´
epid
´
emiologique
des
maladies
auto-immunes
th
yro
¨
ıdiennes
dans
le
sud
tunisien,
”
Annales
d’Endocrinolo
gie
,
v
ol.
67,
pp.
591–595,
2006,
doi:
10.1016/s0003-4266(06)73012-8.
[5]
F
.
A.
Hassani,
M.
Bensouda,
and
P
.
H.
E.
Ouahabi,
”Carcinome
papillaire
d
´
ev
elopp
´
e
sur
une
ectopie
th
yro
¨
ıdienne
lat
´
erocervicale,
”
Annales
d’Endocrinolo
gie
,
v
ol.
78,
no.
4,
p.
348,
2017,
doi:
10.1016/j.ando.2017.07.432.
[6]
L.
Bouadjila
and
Z.
T
aleb,
”Le
portail
NCBI
:
base
de
donn
´
ees
bioinformatique
cl
´
e
en
biotechnologies,
”
Th
`
ese
de
doctorat,
Uni
v
ersit
´
e
Mohamed
BOUDIAF
de
M’Sila,
2019.
[7]
A.
Boubendir
,
K.
Khelili,
and
M.
A.
Hamidechi,
”Importance
of
the
bioinformatics
t
ool
in
the
conrmation
of
bacterial
species
isolated
from
ra
w
milk,
”
Sciences
and
T
ec
hnolo
gie
C,
Biotec
hnolo
gies
,
no.
42,
pp.
38–43,
2016,
doi:
10.1007/s13213-015-1163-5.
[8]
M.
Da
yhof
f,
R.
Schw
artz,
and
B.
Orcutt,
”A
model
of
e
v
olutionary
change
in
proteins,
”
1978.
[9]
F
.
Corpet
and
C.
Che
v
alet,
”Analyse
informatique
des
donn
´
ees
mol
´
eculaires,
”
Pr
oductions
Animales
,
HS
2000,
pp.
191–195,
2000,
doi:
10.20870/productions-animales.2000.13.hs.3837.
[10]
C.
A.
Ouzounis
and
A.
V
alencia,
”Early
bioinformatics:
the
birth
of
a
discipline—a
personal
vie
w
,
”
Bioinformatics
,
v
ol.
19,
no.
17,
pp.
2176–2190,
2003,
doi:
10.1093/bioinformatics/btg309.
[11]
B
ha
vsar
,
K.
A.,
et
al.,
”Medical
diagnosis
using
machine
learning:
a
statistical
re
vie
w
,
”
Computer
s,
Materials
and
Conti
nua
,
v
ol.
67,
no.
1,
pp.
107–125,
2021.
doi:10.32604/cmc.2021.014604.
[12]
S
.
Min,
B.
Lee,
and
S.
Y
oon,
”Deep
learning
in
bioinformatics,
”
Briengs
in
Bioinformatics
,
v
ol.
18,
no.
5,
pp.
851–869,
2017.
doi:10.1093/bib/bbw068.
[13]
P
.
Larra
˜
nag
a
et
al.,
”Machine
learning
in
bioinformatics,
”
Briengs
in
Bioinformatics
,
v
ol.
7,
no.
1,
pp.
86–112,
2006,
doi:
10.1093/bib/bbk007.
[14]
S.
B.
Khandare
and
M.
B.
Chandak
”T
echniques
of
deep
learning
neural
netw
ork-based
b
uilding
feature
e
xtraction
from
remote
sensing
images:
a
surv
e
y
,
”
International
J
ournal
of
Information
and
Communication
T
ec
hnolo
gy
,
v
ol.
14,
no.
2,
pp.
614–624,
2025,
doi:
10.11591/ijict.v14i2.pp614-624.
[15]
A.
Este
v
a
et
al.,
”Dermatologist-le
v
el
classication
of
skin
cancer
with
deep
neural
netw
orks,
”
Natur
e
,
v
ol.
542,
pp.
115–118,
2017,
doi:
10.1038/nature21056.
[16]
Y
.
Zhang,
Y
.
Y
an,
R.
L.
K
umar
,
and
S.
Juneja,
”Impro
ving
colle
ge
ideological
and
political
education
based
on
deep
learning,
”
Int.
J
.
Information
and
Communication
T
ec
hnolo
gy
,
v
ol.
24,
no.
4,
pp.
431–447,
2024,
doi:
10.1504/IJICT
.2024.138778.
[17]
F
.
Ponti,
F
.
Frezza,
P
.
Simeoni,
and
R.
P
arisi,
”A
generalized
learning
approach
to
deep
neural
netw
orks,
”
J
ournal
of
T
elecommuni-
cations
and
Information
T
ec
hnolo
gy
,
no.
3,
2024,
doi:
10.26636/jtit.2024.3.1454.
[18]
R.
Agg
arw
al
et
al.,
”Diagnostic
accurac
y
of
deep
learning
in
medical
imaging:
a
systematic
re
vie
w
and
meta-analysis,
”
NPJ
Digital
Medicine
,
v
ol.
4,
p.
65,
2021,
doi:
10.1038/s41746-021-00438-z.
[19]
B.
Azdine
and
M.
El
Kabbouri,
”L
’analyse
sentimentale
et
l’apprentissage
automatique:
M
´
ethodes
h
ybrides
et
perspecti
v
es
fu-
tures,
”
International
J
ournal
Accounting
,
F
inance
,
A
uditing
,
Mana
g
ement
and
Economics
,
v
ol.
6,
no.
1,
pp.
166–190,
2025,
doi:
10.1515/9782760525702-007.
[20]
M.
Bakator
and
D.
Radosa
v
,
”Deep
learning
and
medical
diagnosis:
A
re
vie
w
of
literature,
”
Multimodal
T
ec
hnolo
gies
and
Inter
ac-
tion
,
v
ol.
2,
no.
3,
p.
47,
2018,
doi:
10.3390/mti2030047.
[21]
H.
T
aud
and
J.
F
.
Mas,
”Multilayer
perceptron
(MLP),
”
in
Geomatic
Appr
oac
hes
for
Modeling
Land
Chang
e
Scenar
ios
,
Springer
,
Cham,
2017,
pp.
451–455,
doi:
10.1007/978-3-319-60801-3
27.
[22]
J.
Naskath,
G.
Si
v
akamasundari,
and
A.
A.
S.
Be
gum,
”A
study
on
dif
ferent
deep
learning
algorithms
used
in
deep
neural
nets:
MLP
,
SOM
and
DBN,
”
W
ir
eless
P
er
sonal
Communications
,
v
ol.
128,
no.
4,
pp.
2913–2936,
2023,
doi:
10.1007/s11277-022-10079-4.
[23]
E.
Bisong,
”The
multilayer
perceptron
(MLP),
”
in
Building
Mac
hine
Learning
and
Deep
Learning
Models
on
Goo
gle
Cloud
Plat-
form
,
Apress,
Berk
ele
y
,
2019,
pp.
401–405,
doi:
10.1007/978-1-4842-4470-8
31.
[24]
Y
.
N.
Ifriza
et
al.,
”Optimization
articial
neural
netw
ork
(ANN)
models
with
adam
optimizer
to
impro
v
e
customer
satisf
action
b
usi-
ness
banking
prediction,
”
J
urnal
T
eknik
Informatika
(J
utif)
,
v
ol.
6,
no.
3,
pp.
1419–1430,
2025,
doi:
10.52436/1.jutif.2025.6.3.4776.
[25]
S.
A.
Salihu
et
al.,
”Detection
and
classication
of
potato
lea
v
es
diseases
using
con
v
olutional
neural
netw
ork
and
adam
optimizer
,
”
Pr
ocedia
Computer
Science
,
v
ol.
258,
pp.
2–17,
2025,
doi:
10.1016/j.procs.2025.04.159.
[26]
G.
D.
Chen
et
al.,
”Associations
between
th
yroid
function
and
gestational
diabetes
mellitus
in
Chinese
pre
gnant
w
omen:
a
retro-
specti
v
e
cohort
study
,
”
BMC
Endocrine
Disor
der
s
,
v
ol.
22,
no.
1,
p.
44,
2022,
doi:
10.1186/s12902-022-00959-y
.
[27]
C.
Zou
et
al.,
”Association
of
maternal
th
yroid
function
and
gestational
diabetes
with
pre
gnanc
y
outcomes:
a
retrospecti
v
e
cohort
study
,
”
F
r
ontier
s
in
Endocrinolo
gy
,
v
ol.
16,
p.
1555409,
2025,
doi:
10.3389/fendo.2025.1555409.
[28]
Z.
Zhang
et
al.,
”Machine
learning
prediction
models
for
gestational
diabetes
mell
itus:
meta-analysis,
”
J
ournal
of
Medical
Internet
Resear
c
h
,
v
ol.
24,
no.
3,
p.
e26634,
2022,
doi:
10.2196/26634.
[29]
H.
Zak
y
et
al.,
”Machine
learning-based
model
for
the
early
detection
of
gestational
diabetes
mellitus,
”
BMC
Medical
Informatics
and
Decision
Making
,
v
ol.
25,
no.
1,
p.
130,
2025,
doi:
10.1186/s12911-025-02947-3.
[30]
H.
Zhao
et
al.,
”Diagnosis
of
th
yroid
disease
using
deep
con
v
olutional
neural
netw
ork
models
applied
to
th
yroid
scintigraph
y
images:
a
multicenter
study
,
”
F
r
ontier
s
in
Endocrinolo
gy
,
v
ol.
14,
p.
1224191,
2023,
doi:
10.3389/fendo.2023.1224191.
MLP-DT
:
a
deep
learning
model
for
early
pr
ediction
of
diabetes
and
thyr
oid
disor
der
s
.
.
.
(Aouatef
Chaib)
Evaluation Warning : The document was created with Spire.PDF for Python.