bioRxiv (Cold Spring Harbor Laboratory),
Journal Year:
2024,
Volume and Issue:
unknown
Published: Dec. 23, 2024
Abstract
The
increasing
digitalisation
of
multi-modal
data
in
medicine
and
novel
artificial
intelligence
(AI)
algorithms
opens
up
a
large
number
opportunities
for
predictive
models.
In
particular,
deep
learning
models
show
great
performance
the
medical
field.
A
major
limitation
such
powerful
but
complex
originates
from
their
’black-box’
nature.
Recently,
variety
explainable
AI
(XAI)
methods
have
been
introduced
to
address
this
lack
transparency
trust
AI.
However,
majority
solely
evaluated
on
single
modalities.
Meanwhile,
with
XAI
methods,
integrative
frameworks
benchmarks
are
essential
compare
different
tasks.
For
that
reason,
we
developed
BenchXAI,
benchmarking
package
supporting
comprehensive
evaluation
fifteen
investigating
robustness,
suitability,
limitations
biomedical
data.
We
employed
BenchXAI
validate
these
three
common
tasks,
namely
clinical
data,
image
signal
biomolecular
Our
newly
designed
sample-wise
normalisation
approach
post-hoc
enables
statistical
visualisation
robustness.
found
Integrated
Gradients,
DeepLift,
DeepLiftShap,
GradientShap
performed
well
over
all
while
like
Deconvolution,
Guided
Backpropagation,
LRP-
α
1-
β
0
struggled
some
With
acts
as
EU
Act
application
domain
becomes
more
essential.
study
represents
first
step
toward
verifying
suitability
various
domains.
Wiley Interdisciplinary Reviews Data Mining and Knowledge Discovery,
Journal Year:
2025,
Volume and Issue:
15(2)
Published: May 19, 2025
ABSTRACT
In
machine
learning,
exploring
data
correlations
to
predict
outcomes
is
a
fundamental
task.
Recognizing
causal
relationships
embedded
within
pivotal
for
comprehensive
understanding
of
system
dynamics,
the
significance
which
paramount
in
data‐driven
decision‐making
processes.
Beyond
traditional
methods,
there
has
been
shift
toward
using
graph
neural
networks
(GNNs)
given
their
capabilities
as
universal
approximators.
Thus,
thorough
review
advancements
learning
GNNs
both
relevant
and
timely.
To
structure
this
review,
we
introduce
novel
taxonomy
that
encompasses
various
state‐of‐the‐art
GNN
methods
used
studying
causality.
are
further
categorized
based
on
applications
causality
domain.
We
provide
an
exhaustive
compilation
datasets
integral
with
serve
resource
practical
study.
This
also
touches
upon
application
across
diverse
sectors.
conclude
insights
into
potential
challenges
promising
avenues
future
exploration
rapidly
evolving
field
learning.
Research Square (Research Square),
Journal Year:
2025,
Volume and Issue:
unknown
Published: April 8, 2025
Abstract
Large-scale
neural
networks
have
revolutionized
many
general
knowledge
areas
(e.g.,
computer
vision
and
language
processing),
but
are
still
rarely
applied
in
expert
healthcare),
due
to
data
sparsity
high
annotation
expenses.
Human-in-the-loop
machine
learning
(HIL-ML)
incorporates
domain
into
the
modeling
process,
effectively
addressing
these
challenges.Recently,
some
researchers
started
using
large
models
substitute
for
certain
tasks
typically
performed
by
humans.
Although
limitations
areas,
after
being
trained
on
trillions
of
examples,
they
demonstrated
advanced
capabilities
reasoning,
semantic
understanding,
grounding,
planning.
These
can
serve
as
proxies
human,
which
introduces
new
opportunities
challenges
HIL-ML
area.Based
above,
we
summarize
a
more
comprehensive
framework,
Agent-in-the-Loop
Machine
Learning
(AIL-ML),
where
agent
represents
both
humans
models.
AIL-ML
efficiently
collaborate
human
model
construct
vertical
AI
with
lower
costs.This
paper
presents
first
review
recent
advancements
this
area.
First,
provide
formal
definition
discuss
its
related
fields.
Then,
categorize
methods
based
processing
development,
providing
definitions
each,
present
representative
works
detail
each
category.
Third,
highlight
relative
applications
AIL-ML.
Finally,
current
literature
future
research
directions.
Computers in Biology and Medicine,
Journal Year:
2025,
Volume and Issue:
191, P. 110124 - 110124
Published: April 15, 2025
The
increasing
digitalization
of
multi-modal
data
in
medicine
and
novel
artificial
intelligence
(AI)
algorithms
opens
up
a
large
number
opportunities
for
predictive
models.
In
particular,
deep
learning
models
show
great
performance
the
medical
field.
A
major
limitation
such
powerful
but
complex
originates
from
their
'black-box'
nature.
Recently,
variety
explainable
AI
(XAI)
methods
have
been
introduced
to
address
this
lack
transparency
trust
AI.
However,
majority
solely
evaluated
on
single
modalities.
Meanwhile,
with
XAI
methods,
integrative
frameworks
benchmarks
are
essential
compare
different
tasks.
For
that
reason,
we
developed
BenchXAI,
benchmarking
package
supporting
comprehensive
evaluation
fifteen
investigating
robustness,
suitability,
limitations
biomedical
data.
We
employed
BenchXAI
validate
these
three
common
tasks,
namely
clinical
data,
image
signal
biomolecular
Our
newly
designed
sample-wise
normalization
approach
post-hoc
enables
statistical
visualization
robustness.
found
Integrated
Gradients,
DeepLift,
DeepLiftShap,
GradientShap
performed
well
over
all
while
like
Deconvolution,
Guided
Backpropagation,
LRP-α1-β0
struggled
some
With
acts
as
EU
Act
application
domain
becomes
more
essential.
study
represents
first
step
towards
verifying
suitability
various
domains.
