Indonesian
J
our
nal
of
Electrical
Engineering
and
Computer
Science
V
ol.
36,
No.
2,
No
v
ember
2024,
pp.
846
∼
853
ISSN:
2502-4752,
DOI:
10.11591/ijeecs.v36.i2.pp846-853
❒
846
De
v
elopment
of
explainable
machine
intelligence
models
f
or
heart
sound
abnormality
detection
J
ee
v
akala
Si
v
a
Ramakrishna
1
,
Sonagiri
China
V
enkateswarlu
2
,
K
ommu
Na
v
een
K
umar
3
,
P
arikipandla
Shr
eya
2
1
Department
of
Computer
Science
and
Engineering
(Articial
Intelligence
and
Machine
Learning),
Institute
of
Aeronautical
Engineering,
Hyderabad,
India
2
Department
of
Electronics
and
Communication
Engineering,
Institute
of
Aeronautical
Engineering,
Hyderabad,
India
3
Department
of
Electrical
and
Computer
Engineering,
Marri
Laxman
Reddy
Institute
of
T
echnology
and
Management,
Hyderabad,
India
Article
Inf
o
Article
history:
Recei
v
ed
Mar
8,
2024
Re
vised
Jul
10,
2024
Accepted
Jul
14,
2024
K
eyw
ords:
Explainable
learning
models
Heart
abnormalities
Heart
sound
detection
LIME
SHAP
ABSTRA
CT
De
v
eloping
e
xplainable
machine
intelligence
(XAI)
model
s
for
heart
sound
ab-
normality
detection
is
a
crucial
area
of
research
aimed
at
impro
ving
the
inter
-
pretability
and
transparenc
y
of
machine
learning
algorithms
in
medical
diagnos-
tics.
In
this
study
,
we
propose
a
frame
w
ork
for
b
uilding
XAI
models
that
can
ef
fecti
v
ely
detect
abnormalities
in
heart
sounds
while
pro
viding
interpretable
e
xplanations
for
their
predi
ctions.
W
e
le
v
erage
techniques
suc
h
as
SHaple
y
ad-
diti
v
e
e
xPlanations
(SHAP)
and
local
interpreta
ble
model-agnostic
e
xplanations
(LIME)
to
generate
e
xplanations
for
model
predictions,
enabling
clinicians
to
understand
the
rationale
behind
the
algorithm’
s
decis
ions.
Our
approach
in-
v
olv
es
preprocessing
heart
sound
data,
training
machine
learning
models,
and
inte
grating
XAI
techniques
to
enhance
the
interpretability
of
the
models.
W
e
e
v
aluate
the
performance
of
our
XAI
models
using
standard
metrics
and
demon-
strate
their
ef
fecti
v
eness
in
accurately
detecting
heart
sound
abnormalities
while
pro
viding
insightful
e
xplanations
for
their
predictions.
This
research
contrib
utes
to
the
adv
ancement
of
XAI
in
medical
applications,
particularly
in
the
domain
of
cardiac
diagnostics,
where
interpretability
is
crucial
for
clinical
decision-
making.
This
is
an
open
access
article
under
the
CC
BY
-SA
license
.
Corresponding
A
uthor:
Jee
v
akala
Si
v
a
Ramakrishna
Department
of
Computer
Science
and
Engineering
(Articial
Intelligence
and
Machine
Learning)
Institute
of
Aeronautical
Engineering
Hyderabad,
India
Email:
j.si
v
aramakrishna@iare.ac.in
1.
INTR
ODUCTION
The
human
heart
produces
characteristic
sounds,
kno
wn
as
heart
sounds,
that
result
from
the
closure
of
heart
v
alv
es
and
the
mo
v
ement
of
blood
within
the
heart
chambers.
These
sounds
are
crucial
indicators
of
cardiac
health
and
can
pro
vide
v
aluable
diagnostic
inform
ation
about
the
condition
of
the
heart.
Heart
sounds
are
typically
cate
gorized
into
tw
o
main
t
y
pe
s:
S1
(the
rst
heart
sound)
and
S2
(the
second
hear
t
sound).
S1
is
produced
by
the
closure
of
the
mit
ral
and
tricuspid
v
alv
es
at
the
be
ginning
of
systole,
while
S2
is
produced
by
the
closure
of
the
aortic
and
pul
monary
v
alv
es
at
the
be
gi
nning
of
diastole.
The
purpose
of
detecting
heart
sound
abnormalities
lies
in
the
early
identication
of
cardiac
pathologies,
such
as
v
alvular
disorders,
murmurs,
and
other
structural
abnormalities.
The
heart
sound
detetion
mechanism
requires
an
accurate
and
reliable
detection
algorithms,
the
interpretat
ion
of
comple
x
heart
sound
patterns,
and
the
inte
gration
of
these
algorithms
into
J
ournal
homepage:
http://ijeecs.iaescor
e
.com
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
❒
847
clinical
practice.
The
use
of
machine
learning
(ML)
algorithms
in
medical
diagnostics
has
sho
wn
promising
results
in
detecting
v
arious
health
conditions
[1]-[3].
The
comple
xity
of
ML
algorithms
often
mak
es
it
dif
cult
for
clinicians
to
trust
and
understand
the
decisions
made
by
these
models,
which
can
hinder
their
adoption
in
real
w
orld
healthcare
settings
[4].
In
the
conte
xt
of
cardiac
diagnostics,
the
interpretation
of
heart
sounds
plays
a
crucial
role
in
identifying
abnormalities
and
making
informed
clinical
decisions.
T
raditional
methods
of
heart
sound
analysis
are
often
subjecti
v
e
and
rely
hea
vily
on
the
e
xpertise
of
clinicians.
ML-based
approaches
ha
v
e
the
potential
to
automate
and
impro
v
e
the
accurac
y
of
the
de
v
eloped
models
[5]
b
ut
their
lack
of
interpretability
limits
their
practical
utility
in
clinical
settings.
Se
v
eral
studies
[6],
[7]
ha
v
e
e
xplored
the
application
of
ML
algorithms,
such
as
neural
netw
orks
and
decision
trees,
in
heart
sound
analysis
for
detecting
abnormalities.
The
studies
ha
v
e
demonstrated
promising
re-
sults
in
terms
of
accurac
y
,
the
y
often
f
all
short
in
pro
viding
transparent
e
xplanations
for
their
predictions.
This
limitation
has
led
to
gro
wing
interest
in
the
de
v
elopment
of
e
xplainable
machine
intelligence
(XAI)
models
that
can
not
only
achie
v
e
high
accurac
y
b
ut
also
pro
vide
interpretable
e
xplanations
for
their
decisions.
In
this
w
ork,
we
propose
a
no
v
el
approach
to
address
the
interpretability
challenge
in
ML-based
heart
sound
abnor
-
mality
detection.
Our
approach
in
v
olv
es
le
v
eraging
state-of-the-art
XAI
techniques,
such
as
SHaple
y
additi
v
e
e
xPlanations
(SHAP)
and
local
interpretable
model
agnostic
e
xplanations
(LIME)
[8]
to
generate
e
xplanations
for
the
predictions
of
our
ML
models.
The
con
v
er
gence
of
engineering
and
medical
e
xpertise
has
rapidly
adv
anced
the
health
care
domain.
In
t
his
conte
xt,
we
re
vie
w
the
progression
of
heart
sound
classication
and
disease
diagnosis.
The
combination
of
deep
con
v
olutional
neural
netw
orks
and
mel-frequenc
y
cepstral
coef
cients
(MFCC)
emplo
yed
to
classify
normal
and
abnormal
phonocardiograph
y
signals.
The
model
has
achie
v
ed
a
competiti
v
e
score
b
ut
did
not
address
the
generalizability
of
their
approach
to
dif
ferent
subjects
[9].
The
automatic
recognition
of
heart
rate
v
ariations
from
phonocardiograms
(PCG)
using
transfer
learning
with
MFCC
features
has
achie
v
ed
high
ac-
curac
y
.
Ho
we
v
er
the
limited
a
v
ailability
of
PCG
recordings
raises
concerns
about
the
generalizability
of
their
method
[10].
The
heart
sound
classication
using
impro
v
ed
MFCC
features
and
CRNN,
pro
vides
an
impro
v
e-
ment
in
classication
accurac
y
.
Ho
we
v
er
,
the
model
suf
fers
with
computational
comple
xity
,
need
for
further
in
v
estig
ation
into
generalizability
and
other
e
xperiment
conditions
[
1
1]
.
The
reliance
on
traditional
classiers
may
limit
their
ability
to
capture
comple
x
patterns,
and
the
e
v
aluation
based
on
the
datasets
[12].
The
recurrent
neural
netw
ork
(RNN)
based
frame
w
o
r
k
pro
vides
promising
results.
Ho
we
v
er
,
the
computational
comple
xity
of
RNNs
and
their
performance
dependence
on
dataset
used
for
the
analysis
[13],
[14].
The
se
gmented
and
unse
gmented
PCG
signals
ha
v
e
been
e
xplored
for
heart
s
oun
d
classication
[15].
The
heartbeat
sound
clas-
sication
using
normal,
murmur
,
and
e
xtra-systole
heartbeat
sounds,
achie
v
es
competiti
v
e
performance
[16].
The
deep
learning
and
ensembl
e
learning
techniques
for
phonocardiogram
classication,
achie
v
es
competiti
v
e
performance
[17].
The
classication
of
normal
and
abnormal
heart
sounds
using
an
ensemble
approach
along
with
handcrafted
features
may
limit
its
ability
to
capture
comple
x
patterns
[18].
Heart
sound
classication
using
deep
learning
based
CNNs
produce
an
impro
v
ed
classication
accu-
rac
y
[19]-[21].
The
learnable
lter
banks
is
another
approach
for
heart
sound
detection
[22].
The
interpretable
machine
learning
techniques
applied
on
medical
image
pro
vides
better
e
xplanati
ons
for
the
analysis
[23],
[24].
In
this
we
ha
v
e
inte
grated
the
XAI
techniques
into
the
de
v
elopment
of
ML
model
s
for
heart
sound
analysis.
The
proposed
model
has
the
potential
to
enhance
the
trust
and
adoption
of
ML
technologies
in
cardiac
diagnostics.
2.
PR
OPOSED
METHOD
The
proposed
methodology
for
heart
sound
abnormality
detection
using
e
xplainable
machine
learning
(XML)
models
is
sho
wn
in
Figure
1.
The
method
comprises
a
series
of
well-dened
steps
aimed
at
de
v
eloping
a
rob
ust
and
interpretable
model.
The
methodology
can
be
summarized
as
follo
wing
steps:
-
Step
1:
Data
collection:
the
data
collection
process
in
v
olv
es
g
athering
heart
sound
recordings
for
normal
and
abnormal
conditions.
These
recordings
are
typically
obtained
from
medical
databases
[25].
-
Step
2:
Data
preprocessing:
preprocessing
be
gins
with
a
thorough
inspection
of
the
collected
data
to
identify
and
remo
v
e
an
y
outliers,
artif
acts,
or
c
orrupted
recordings
that
could
adv
ersely
af
fect
model
training.
Noise
reduction
techniques,
such
as
ltering
or
denoising
algorithms,
are
applied
to
impro
v
e
the
quality
of
the
recordings.
-
Step
3:
Feature
e
xtraction:
feature
e
xtraction
in
v
olv
es
transforming
ra
w
heart
sound
recordings
int
o
a
set
of
rele
v
ant
features
that
can
be
used
as
a
fetures
while
b
uilding
the
model.
The
features
include
spectral
De
velopment
of
e
xplainable
mac
hine
intellig
ence
models
...
(J
ee
vakala
Siva
Ramakrishna)
Evaluation Warning : The document was created with Spire.PDF for Python.
848
❒
ISSN:
2502-4752
characteristics
and
statistical
measures
that
capture
the
underlying
patterns
in
the
heart
sounds.
-
Step
4:
Dataset
split
ting:
the
dataset
is
split
into
training,
v
alidation,
and
testing
sets
using
a
stratied
approach
to
ensure
that
each
set
contains
a
proportional
representation
of
normal
and
abnormal
heart
sounds.
-
Step
5:
Feature
scaling:
feat
ure
scaling
is
applied
to
normalize
the
e
xtracted
features
to
a
consistent
range,
pre
v
enting
certain
features
from
dominating
the
model
training
process
due
to
their
lar
ger
magnitudes.
-
Step
6:
Model
training:
multiple
XML
algorithms
such
as
random
forest,
gradient
boosting
(GBoost),
AdaBoost,
K-nearest
neighbors
(KNN),
and
logistic
re
gression
(LR)
are
trained
using
the
preprocessed
and
scaled
features.
The
training
process
in
v
olv
es
optimizing
each
algorithm’
s
parameters
through
techniques
lik
e
grid
search
or
randomized
search
to
nd
the
best-performing
conguration.
-
Step
7:
LIME
e
xplanation:
LIME
are
applied
to
pro
vide
local
e
xplanations
for
indi
vidual
predictions
made
by
the
XML
model.
LIME
generates
interpretable
e
xplanations
by
perturbing
the
input
features
around
a
specic
instance
and
observing
the
changes
in
the
model’
s
predictions.
-
Step
8:
SHAP
e
xplanation:
SHAP
is
used
to
pro
vide
global
e
xplanations
for
the
XML
model’
s
predictions,
highlighting
the
contrib
utions
of
each
feature
to
the
o
v
erall
predictions.
SHAP
v
alues
quantify
the
impact
of
each
feature
on
the
model’
s
output
and
help
understand
the
model’
s
decision-making
process
at
a
global
le
v
el.
-
Step
9:
V
isualization:
the
results
of
LIME
and
SHAP
e
xplanations
are
visualized
using
v
arious
techniques
such
as
bar
plots,
and
summary
plots.
V
isualization
aids
in
presenting
the
e
xplanations
in
an
intuiti
v
e
and
understandable
manner
,
f
acilitating
their
interpretation
by
domain
e
xperts
and
stak
eholders.
Figure
1.
Proposed
block
diagram
for
heart
sound
classication
3.
RESUL
TS
AND
DISCUSSION
The
obtained
results
demonstrate
that
emplo
ying
LIME
and
SHAP
signicantly
enhances
the
inter
-
pretability
of
machine
learning
models
used
for
detecting
heart
sound
abnormalities.
These
techniques
of
fer
v
aluable
insights
into
the
decision-making
processes
of
the
models
by
pinpointing
the
k
e
y
f
actors
inuencing
their
classications.
Through
visualizing
the
impact
of
dif
ferent
v
ariables,
we
ha
v
e
achie
v
ed
a
deeper
under
-
standing
of
the
models’
beha
vior
,
thereby
enhancing
transparenc
y
and
interpretability
3.1.
LIME
explanations
Across
the
man
y
classiers
used
in
our
analysis,
LIME
consistently
identied
the
most
important
v
ariables
for
heart
sound
cate
gorization.
LIME
is
crucial
in
our
analysis
because
it
consistently
identies
the
mos
t
important
v
ariables
for
heart
sound
cate
gorization
across
v
arious
classiers.
F
or
e
xample,
in
the
random
forest
classier
,
LIME
e
xplanations
demonstrated
ho
w
k
e
y
spectral
and
statistical
properties,
such
as
spectral
centroid,
spectral
bandwidth,
and
zero-crossing
rate,
were
critical
in
discriminating
between
normal
and
pathological
heart
sounds.
3.1.1.
Random
f
or
est
classier
Figure
2
illustrat
es
LIME
visualization
using
a
random
forest
classier
,
emphasizing
features
such
as
consistent
spectral
patterns
and
statistical
properties
that
play
a
crucial
role
in
the
classication
process.
These
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
36,
No.
2,
No
v
ember
2024:
846–853
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
❒
849
include
the
re
gularity
in
spectral
centroid
and
the
lo
w
v
ariability
in
zero-crossing
rate.
Figure
3
illustrates
ho
w
LIME
visual
ization
re
v
eals
that
irre
gularities
in
spectral
patterns
and
statistical
features,
such
as
high
v
ariation
in
spectral
centroid
and
pronounced
changes
in
zero-crossing
rate,
are
crucial
f
actors
inuencing
the
classication.
Figure
2.
LIME
visualization
where
the
person
has
normal
heartbeat
using
random
forest
Figure
3.
LIME
visualization
where
the
person
has
abnormal
heartbeat
using
random
forest
3.1.2.
GBoost
classier
Figure
4
illustrates
the
LIME
visualization
with
a
GBoost
classier
,
elucidating
ho
w
distinct
spectr
al
and
statistical
features
contrib
ute
to
a
consistent
and
predictable
pattern.
This
includes
stable
v
alues
in
spectral
centroid
and
smooth
transitions
in
zero-crossing
rate.
Figure
5
illustrates
ho
w
LIME
visualization
with
a
GBoost
classier
can
unco
v
er
irre
gularities
in
spectral
and
statistical
features,
such
as
abrupt
shifts
in
spectral
centroid
and
signicant
v
ariability
in
zero-crossing
rate,
which
play
a
crucial
role
in
distinguishing
abnormal
heart
sounds.
Figure
4.
LIME
visualization
where
the
person
has
normal
heartbeat
using
GBoost
classier
Figure
5.
LIME
visualization
where
the
person
has
normal
heartbeat
using
GBoost
classier
De
velopment
of
e
xplainable
mac
hine
intellig
ence
models
...
(J
ee
vakala
Siva
Ramakrishna)
Evaluation Warning : The document was created with Spire.PDF for Python.
850
❒
ISSN:
2502-4752
3.1.3.
XGB
classier
Figure
6
illustrates
ho
w
LIME
visualization
with
an
e
xtreme
gradient
boosting
(XGBoost)
cl
assier
might
sho
w
ho
w
certain
consistent
spectral
patterns
and
statistical
features,
lik
e
stable
spectral
centroid
and
minimal
zero-crossing
rate,
contrib
ute
to
the
classication.
Figure
7
illustrates
ho
w
LIME
visualization
for
abnormal
heart
sounds
with
an
XGBoost
classier
could
re
v
eal
ho
w
irre
gularities
in
spectral
patterns
and
statistical
features,
such
as
uctuations
in
spectral
centroid
and
high
zero-crossing
rate,
play
a
crucial
role
in
the
classication.
T
able
1
sho
ws
the
accuracies
of
the
classiers
using
LIME,
with
the
XGB
and
GBoost
achie
ving
the
highest
accurac
y
for
heart
sound
classication,
follo
wed
by
the
random
forest
and
decision
tree.
KNN
also
sho
wed
respectable
accuracies
b
ut
were
slightly
lo
wer
compared
to
the
XGB
and
GBoost
classiers.
Figure
6.
LIME
visualization
where
the
person
has
normal
heartbeat
using
XGB
classier
Figure
7.
LIME
visualization
where
the
person
has
abnormal
heartbeat
using
XGB
classier
T
able
1.
Classication
accurac
y
of
v
arious
classiers
for
normal
and
abnormal
heart
sound
detection
Classier
Accurac
y
(%)
Random
forest
88.2
GBoost
89.2
XGB
91.8
KNN
70
Decision
tree
88.2
3.2.
SHAP
explanations
In
our
analysis,
SHAP
consistently
identied
k
e
y
v
ariables
for
heart
sound
cate
gorization
across
v
ar
-
ious
classiers.
F
or
e
xample,
in
the
GBoost
classier
,
SHAP
highlighted
the
importance
of
features
such
as
spectral
rol
lof
f,
RMS
ener
gy
,
and
tempo
in
distinguishing
normal
from
pathological
heart
sounds.
These
in-
sights
pro
vided
a
clear
understanding
of
feature
contrib
utions,
enhancing
model
interpretability
.
In
addition,
SHAP
analysis
re
v
ealed
the
consistent
impact
of
t
hese
features
across
dif
ferent
classiers,
reinforcing
their
signicance
in
heart
sound
cate
gorization
and
pro
viding
v
aluable
insights
for
model
renement
and
clinical
decision-making.
3.2.1.
Random
f
or
est
In
SHAP
the
features
are
rank
ed
based
on
their
importance
in
the
model’
s
predictions.
F
or
normal
heart
sounds,
features
with
higher
SHAP
v
alues,
such
as
spect
ral
centroid
and
spectral
bandwidth,
indicate
a
strong
inuence
on
the
classication,
highlighting
their
importance
in
distinguishing
normal
heart
sounds.
In
a
SHAP
force
plot
for
a
random
forest
classier
,
the
plot
sho
ws
ho
w
indi
vidual
feature
v
alues
contrib
ute
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
36,
No.
2,
No
v
ember
2024:
846–853
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
❒
851
to
the
model’
s
prediction
for
a
specic
instance.
F
or
normal
heart
sounds,
features
lik
e
spectr
al
centroid
and
spectral
bandwidth
with
positi
v
e
SHAP
v
alues
indicate
their
contrib
ution
to
classifying
the
sound
as
normal,
pro
viding
a
detai
led
understanding
of
the
model’
s
decision
process.
An
e
xplanation
of
SHAP
for
the
random
forest
classier
can
be
seen
in
Figure
8.
Figure
8.
SHAP
e
xplanations
for
random
forest
classier
3.2.2.
Gradient
boosting
In
the
summary
plot
as
sho
wn
in
Figure
9,
SHAP
v
alues
are
displayed
horizontally
,
representing
the
impact
of
each
feature
on
the
model’
s
output.
Features
are
sorted
by
importance,
with
color
indicating
the
v
alue
of
the
feature
(red
for
high,
blue
for
lo
w).
This
plot
pro
vides
a
clear
o
v
ervie
w
of
which
features
are
most
inu-
ential
in
the
model’
s
decision-making
process.
In
a
force
plot,
SHAP
e
xplanations
for
indi
vidual
predictions
are
visualized.
F
or
e
xample,
for
normal
heart
sounds,
the
force
plot
might
sho
w
ho
w
certain
spectral
and
statistical
features
lead
to
a
classication
of
normal,
pro
viding
a
detailed
breakdo
wn
of
the
contrib
ution
of
each
feature
to
the
nal
predicti
on.
In
comparison
with
[15]
the
proposed
technique
pro
vides
an
alternate
perspec
ti
v
e
on
heart
sound
abnormality
detection.
While
the
preceding
publicati
o
n
focused
on
time-frequenc
y
approaches
and
deep
learning
for
representation
learning,
our
resear
ch
focuses
on
the
construction
of
XAI
models.
W
e
combine
di
v
erse
machine
learning
algorithms
with
e
xplainable
AI
approaches
such
as
LIME
and
SHAP
,
resulting
in
impro
v
ed
accurac
y
and
clear
e
xplanations
for
model
predictions.
The
potential
applications
of
ECG
(Electro-
cardiograph
y)
data
analys
is
in
healthcare,
wellness
monitoring,
and
be
yond
are
v
ast
and
promising.
Lik
e
wise,
e
xplainable
machine
learning
models
are
poised
to
be
pi
v
otal
in
guaranteeing
trans
p
a
renc
y
,
trustw
orthiness,
and
accountability
in
AI-dri
v
en
ECG
analysis
systems.
Figure
9.
Summary
plot
of
SHAP
using
GBoost
De
velopment
of
e
xplainable
mac
hine
intellig
ence
models
...
(J
ee
vakala
Siva
Ramakrishna)
Evaluation Warning : The document was created with Spire.PDF for Python.
852
❒
ISSN:
2502-4752
4.
CONCLUSION
Our
study
successfully
created
machine
intelligence
models
for
heart
sound
anomaly
identi
cation
utilising
methods
such
as
random
forest,
GBoost,
XGBoost,
and
decision
trees.
The
use
of
e
xplainable
AI
approaches
such
as
LIME
and
SHAP
impro
v
ed
model
interpretability
,
of
fering
insights
into
predictions.
These
models
present
a
viable
method
for
automated
and
transparent
heart
sound
analysis,
with
potential
applications
in
early
identication
and
m
on
i
toring
of
cardiac
diseases.
Future
research
m
ight
concentrate
on
increasing
the
dataset
and
conducting
real
w
orld
trials
to
conrm
their
benecial
ef
fects.
Ov
erall,
our
ef
fort
highlights
the
utility
of
e
xplainable
machine
intelligence
models
in
cardiac
diagnosis.
REFERENCES
[1]
M
.
A.
H.
Islam,
S.
M.
Shahriyar
,
M.
J.
Alam,
M.
Rahman,
and
M.
R.
K.
R.
Sark
er
,
“Skin
disease
detection
emplo
ying
transfer
learning
approach-
a
ne-tune
visual
geometry
group-19,
”
Indonesian
Journal
of
Electrical
Engineering
and
Computer
Science
(IJEECS)
,
v
ol.
31,
no.
1,
pp.
321–328,
Jul.
2023,
doi:
10.11591/ijeecs.v31.i1.pp321-328.
[2]
R
.
Battur
and
J.
Narayana,
“Classication
of
medical
X-ray
images
using
supervised
and
unsupervised
learning
ap-
proaches,
”
Indonesian
Journal
of
Electrical
Engineering
and
Computer
Science
(IJEECS)
,
v
ol.
30,
no.
3,
p.
1713,
Jun.
2023,
doi:
10.11591/ijeecs.v30.i3.pp1713-1721.
[3]
B.
J.
Khadhim,
Q.
K.
Kadhim,
W
.
K.
Shams,
S.
T
.
Ahmed,
and
W
.
A.
W
ahab
Alsiadi,
“Diagnose
CO
VID-19
by
using
h
ybrid
CNN-
RNN
for
chest
X-ray
,
”
Indonesian
Journal
of
Electrical
Engineering
and
Computer
Science
(IJEECS)
,
v
ol.
29,
no.
2,
pp.
852–860,
Feb
.
2023,
doi:
10.11591/ijeecs.v29.i2.pp852-860.
[4]
N.
Chinnamadha,
R.
Z.
Ahmed,
and
K.
Kale
go
wda,
“De
v
elopment
of
health
monitori
ng
system
using
smart
intelligent
de-
vice,
”
Indonesian
Journal
of
Electrical
Engineering
and
Computer
Science
(I
JEECS)
,
v
ol.
28,
no.
3,
p.
1381,
Dec.
2022,
doi:
10.11591/ijeecs.v28.i3.pp1381-1387.
[5]
S.
C.
V
enkatesw
arlu,
S.
R.
Jee
v
akala,
N.
U.
K
umar
,
P
.
Munas
w
amy
,
and
D.
Pendyala,
“Emotion
recognition
from
speech
and
te
xt
using
long
short-term
memory
,
”
Engineering,
T
echnology
and
Applied
Science
Research
,
v
ol.
13,
no.
4,
pp.
11166–11169,
Aug.
2023,
doi:
10.48084/etasr
.6004.
[6]
F
.
Renna,
J.
Oli
v
eira,
and
M.
T
.
C
oimbra,
“Deep
con
v
olutional
neural
netw
orks
for
heart
sound
se
gmentation,
”
IEEE
Journal
of
Biomedical
and
Health
Informatics
,
v
ol.
23,
no.
6,
pp.
2435–2445,
No
v
.
2019,
doi:
10.1109/JBHI.2019.2894222.
[7]
A.
Rath,
D.
Mishra,
G.
P
anda,
and
M.
P
al,
“De
v
elopment
and
assessment
of
machine
learning
based
heart
disease
de-
tection
using
imbalanced
heart
sound
signal,
”
Biomedical
Signal
Processing
and
Control
,
v
ol.
76,
p.
103730,
Jul.
2022,
doi:
10.1016/j.bspc.2022.103730.
[8]
A.
Holzinger
,
A.
Sarant,
C.
Molnar
,
P
.
Biecek,
and
W
.
Samek,
“Explainable
AI
methods-a
brief
o
v
ervie
w
,
”
in
International
w
orkshop
on
e
xtending
e
xplainable
AI
be
yond
deep
models
and
classiers
,
2022,
pp.
13–38,
doi:
10.1007/978-3-031-04083-2
2.
[9]
V
.
Maknickas
and
A.
Maknickas,
“Recognition
of
normal-abnormal
phonocardiographic
signals
using
deep
con
v
olutional
neu-
ral
netw
orks
and
mel-frequenc
y
spectral
coef
cients,
”
Ph
ysiological
Measurement
,
v
ol.
38,
no.
8,
pp.
1671–1684,
Jul.
2017,
doi:
10.1088/1361-6579/aa7841.
[10]
T
.
Al
af,
M.
Boulares,
A.
Barna
wi,
T
.
Alaf,
H.
Althobaiti,
and
A.
Alferaidi,
“Normal
and
abnormal
heart
rates
recognition
using
transfer
learning,
”
in
Proceedings
-
2020
12th
International
Conference
on
Kno
wledge
and
Systems
Engineering,
KSE
2020
,
No
v
.
2020,
pp.
275–280,
doi:
10.1109/KSE50997.2020.9287514.
[11]
M.
Deng,
T
.
Meng,
J.
Cao,
S.
W
ang,
J.
Zhang,
and
H.
F
an,
“Heart
sound
classication
based
on
impro
v
ed
MFCC
features
and
con
v
olutional
recurrent
neural
netw
orks,
”
Neural
Netw
orks
,
v
ol.
130,
pp.
22–32,
Oct.
2020,
doi:
10.1016/j.neunet.2020.06.015.
[12]
Z.
Abduh,
E.
A.
Nehary
,
M.
A.
W
ahed,
and
Y
.
M.
Kadah,
“Classication
of
heart
sounds
using
fractional
fourier
transform
based
mel-frequenc
y
spectral
coef
cients
and
traditional
classiers,
”
Biomedical
Signal
Processing
and
Control
,
v
ol.
57,
p.
101788,
Mar
.
2020,
doi:
10.1016/j.bspc.2019.101788.
[13]
S.
Latif,
M.
Usman,
R.
Rana,
and
J.
Qadir
,
“Phonocardiographic
sensing
using
deep
learning
for
abnormal
heartbeat
detection,
”
IEEE
Sensors
Journal
,
v
ol.
18,
no.
22,
pp.
9393–9400,
No
v
.
2018,
doi:
10.1109/JSEN.2018.2870759.
[14]
T
.
C.
I.
Y
ang
and
H.
Hsieh,
“Classication
of
acoustic
ph
ysiologica
l
signals
based
on
deep
learning
neural
netw
orks
with
augmented
features,
”
Computing
in
Cardiology
,
v
ol.
43,
pp.
569–572,
2016,
doi:
10.22489/cinc.2016.163-228.
[15]
F
.
A.
Khan,
A.
Abid,
A.
Abid,
M.
S.
Khan,
and
M.
S.
Khan,
“
Automatic
heart
sound
classication
from
se
gmented/unse
gmented
phonocardiogram
signals
using
time
and
frequenc
y
features,
”
Ph
ysiological
Measurement
,
v
ol.
41,
no.
5,
p.
055006,
Jun.
2020,
doi:
10.1088/1361-6579/ab8770.
[16]
A.
Raza,
A.
Mehmood,
S.
Ullah,
M.
Ahmad,
G.
S.
Choi,
and
B.
W
.
On,
“Heartbeat
sound
signal
classication
us
ing
deep
learning,
”
Sensors
(Switzerland)
,
v
ol.
19,
no.
21,
p.
4819,
No
v
.
2019,
doi:
10.3390/s19214819.
[17]
J.
M.
T
.
W
u
et
al.
,
“
Applying
an
ensemble
con
v
olutional
neural
netw
ork
with
Sa
vitzk
y–Golay
lter
to
construct
a
phonocardiogram
prediction
model,
”
Applied
Soft
Computing
Journal
,
v
ol.
78,
pp.
29–40,
May
2019,
doi:
10.1016/j.asoc.2019.01.019.
[18]
C.
Potes,
S.
P
arv
aneh,
A.
Rahman,
and
B.
Conro
y
,
“Ensemble
of
feature-based
and
deep
learning-based
classiers
for
detection
of
abnormal
heart
sounds,
”
in
Computing
in
Cardiology
,
Sep.
2016,
v
ol.
43,
pp.
621–624,
doi:
10.22489/cinc.2016.182-399.
[19]
F
.
Li
et
al.,
“Feature
e
xtraction
and
classication
of
heart
sound
using
1D
con
v
olutional
neural
netw
orks,
”
Eurasip
Journal
on
Adv
ances
in
Signal
Processing
,
v
ol.
2019,
no.
1,
p.
59,
Dec.
2019,
doi:
10.1186/s13634-019-0651-3.
[20]
F
.
Li,
H.
T
ang,
S.
Shang,
K.
Mathiak,
and
F
.
Cong,
“Classica
tion
of
heart
sounds
using
con
v
olutional
neural
netw
ork,
”
Applied
Sciences
(Switzerland)
,
v
ol.
10,
no.
11,
p.
3956,
Jun.
2020,
doi:
10.3390/app10113956.
[21]
B.
Xiao,
Y
.
Xu,
X.
Bi,
J.
Zhang,
and
X.
Ma,
“Heart
sounds
classication
using
a
no
v
el
1-D
con
v
olutional
neural
netw
ork
with
e
xtremely
lo
w
parameter
consumption,
”
Neurocomputing
,
v
ol.
392,
pp.
153–159,
Jun.
2020,
doi:
10.1016/j.neucom.2018.09.101.
[22]
A.
I.
Humayun,
S.
Ghaf
f
arzade
g
an,
M.
I.
Ansari,
Z.
Feng,
and
T
.
Hasan,
“T
o
w
ards
domain
in
v
ariant
heart
sound
abnormal-
ity
detection
using
learnable
lterbanks,
”
IEEE
Journal
of
Biomedical
and
Health
Informatics
,
v
ol.
24,
no.
8,
pp.
2189–2198,
Aug.
2020,
doi:
10.1109/JBHI.2020.2970252.
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
36,
No.
2,
No
v
ember
2024:
846–853
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
❒
853
[23]
S.
R.
Jee
v
akala
and
H.
Ramasangu,
“Identication
of
signicant
instants
of
v
ox
els
for
cogniti
v
e
state
classication
using
in-
terpretable
machine
learning
models,
”
Journal
of
Medicinal
and
Chemical
Sciences
,
v
ol.
6,
no.
6,
pp.
1291–1301,
2023,
doi:
10.26655/JMCHEMSCI.2023.6.9.
[24]
Z.
W
ang,
K.
Qian,
H.
Liu,
B.
Hu,
B.
W
.
Schuller
,
and
Y
.
Y
amamoto,
“Exploring
interpretable
representations
for
heart
sound
abnormality
detection,
”
Biomedical
Signal
Processing
and
Control
,
v
ol.
82,
p.
104569,
Apr
.
2023,
doi:
10.1016/j.bspc.2023.104569.
[25]
“Ph
ysioNet
Challenge2016:Heart
sound
classication,
”
Kaggle
,
2016.
https://www
.kaggle.com/datasets/bjoernjostein/ph
ysionet-
challenge
(accessed
Aug.
05,
2022).
BIOGRAPHIES
OF
A
UTHORS
J
ee
v
akala
Si
v
a
Ramakrishna
is
associate
professor
in
the
Department
of
CSE
(A
rticial
Intelligence
and
Machine
Learning),
Institute
of
Aeronautical
Engineeri
ng,
Hyderabad,
India.
He
Holds
a
Ph.D.
de
gree
in
Electronics
and
Communication
Engineering
with
specialization
in
biomedi-
cal
signal,
image
processing
and
Machine
Learning.
His
research
areas
include
image
and
signal
pro-
cessing,
medic
al
image
analysis,
and
FMRI
data
analysis.
Dr
.
J.
Si
v
a
Ramakrishna
has
o
v
er
14
years
of
e
xperience
in
teaching,
including
4
years
as
a
researcher
.
He
published
se
v
eral
research
papers
in
international
journals
and
conferences.
He
can
be
contacted
at
email:
j.si
v
aramakrishna@iare.ac.in.
Sonagiri
China
V
enkateswarlu
professor
in
the
Department
of
Electronics
and
Commu-
nication
Engineering
at
the
Institute
of
Aeronautical
Engineering
(IARE).
He
holds
a
Ph.D.
de
gree
in
Electronics
and
Communication
Engineering
with
a
specialization
in
Digital
Speech
Processing.
Dr
.
S.
China
V
enkatesw
arlu
has
more
than
40
citations
and
paper
publications
ac
ross
v
arious
pub-
lishing
platforms,
and
e
xpertise
in
teaching
subjects
such
as
microprocessors
and
microcontrollers,
digital
signal
processing,
digital
image
processing,
and
speech
process
ing.
W
ith
20
years
of
teaching
e
xperience,
he
can
be
contacted
at
email:
c.v
enkatesw
arlu@iare.ac.in.
K
ommu
Na
v
een
K
umar
is
a
professor
in
the
Department
of
Electronics
and
Communi-
cation
Engineering,
MLR
ITM,
Hyderabad,
India.
He
hol
ds
a
Ph.D.
de
gree
in
Electronics
and
Com-
munication
Enginee
ring
with
specialization
in
Digital
Image
Processing.
His
research
areas
include
digital
image
processing,
and
embedded
systems.
Dr
.
K.
Na
v
een
K
umar
has
o
v
er
15
years
of
e
xperi-
ence
in
teaching.
He
published
paper
in
se
v
eral
national
and
international
journals
and
conferences.
He
can
be
contacted
at
email:
na
v
eenkarun
ya@gmail.com.
P
arikipandla
Sh
r
eya
is
a
student
pursuing
bachelor’
s
de
gree
in
Electronics
and
Com-
munication
Engineering
in
Institute
of
Aeronautical
engineering.
She
has
strong
interest
in
Articial
Intelligence
and
Machine
Learning
as
well
as
cloud
computing.
She
is
v
ery
k
een
in
research
and
al
w
ays
eager
to
learn
and
apply
ne
w
concepts.
She
is
determined
to
contrib
ute
to
the
adv
ancements
of
technology
and
e
xcited
about
the
opportunities
that
li
e
ahead.
She
can
be
contacted
at
email:
shre
yaparikipandlla@gmail.com.
De
velopment
of
e
xplainable
mac
hine
intellig
ence
models
...
(J
ee
vakala
Siva
Ramakrishna)
Evaluation Warning : The document was created with Spire.PDF for Python.