Deep
learning-based
predictive
models,
leveraging
Electronic
Health
Records
(EHR),
are
receiving
increasing
attention
in
healthcare.
An
effective
representation
of
a
patient's
EHR
should
hierarchically
encompass
both
the
temporal
relationships
between
historical
visits
and
medical
events,
inherent
structural
information
within
these
elements.
Existing
patient
methods
can
be
roughly
categorized
into
sequential
graphical
representation.
The
focus
only
on
among
longitudinal
visits.
On
other
hand,
approaches,
while
adept
at
extracting
graph-structured
various
fall
short
effectively
integrate
information.
To
capture
types
information,
we
model
as
novel
heterogeneous
graph.
This
graph
includes
nodes
events
nodes.
It
propagates
structured
from
event
to
visit
utilizes
time-aware
changes
health
status.
Furthermore,
introduce
transformer
(TRANS)
that
integrates
edge
features,
global
positional
encoding,
local
encoding
convolution,
capturing
We
validate
effectiveness
TRANS
through
extensive
experiments
three
real-world
datasets.
results
show
our
proposed
approach
achieves
state-of-the-art
performance.
JMIR Medical Informatics,
Год журнала:
2025,
Номер
13, С. e67748 - e67748
Опубликована: Июнь 2, 2025
Abstract
Background
Diabetes
affects
millions
worldwide.
Primary
care
physicians
provide
a
significant
portion
of
care,
and
they
often
struggle
with
selecting
appropriate
medications.
Objective
This
study
aimed
to
develop
model
that
accurately
predicts
what
drug
an
endocrinologist
would
prescribe
based
on
the
current
measurements.
The
goal
was
create
system
assist
nonspecialists
in
choosing
medications,
thereby
potentially
improving
diabetes
treatment
outcomes.
Based
performance
previous
studies,
we
set
target
achieving
receiver
operating
characteristic
area
under
curve
(ROC-AUC)
above
0.95.
Methods
A
transformer-based
encoder-decoder
whether
44
types
drugs
will
be
prescribed.
uses
sequences
age,
sex,
history
for
12
laboratory
tests,
prescribed
as
inputs.
We
assessed
using
electronic
health
records
from
7034
patients
seeing
endocrinologists
between
2012
2022
at
University
Tokyo
Hospital.
trained
data
subsets
spanning
different
time
periods
(2,
5,
10
years)
micro-
macro-averaged
ROC-AUC
hold-out
test
comprising
solely
2022.
model’s
compared
against
LightGBM.
Results
past
5
years
(2017‐2021)
yielded
best
predictive
performance,
microaverage
(95%
CI)
0.993
(0.992-0.994)
macroaverage
0.988
(0.980-0.993).
achieved
0.95
43
out
drugs.
These
results
surpassed
predefined
outperformed
both
studies
LightGBM
(0.985-0.990)
terms
prediction
accuracy.
Furthermore,
training
short-term
high
accuracy
years,
suggesting
learning
more
recent
prescribing
patterns
might
advantageous.
Conclusions
proposed
demonstrates
feasibility
predicting
next
model,
prescriptions
endocrinologists,
has
potential
information
can
making
diabetes-treatment
decisions.
Future
focus
incorporating
important
factors
such
prescription
contraindications
constraints
enhance
safety,
well
leveraging
large-scale
clinical
across
multiple
hospitals
improve
generalizability
model.
Many
diagnostic
errors
occur
because
clinicians
cannot
easily
access
relevant
information
in
patient
Electronic
Health
Records
(EHRs).
In
this
work
we
propose
a
method
to
use
LLMs
identify
pieces
of
evidence
EHR
data
that
indicate
increased
or
decreased
risk
specific
diagnoses;
our
ultimate
aim
is
increase
and
reduce
errors.
particular,
Neural
Additive
Model
make
predictions
backed
by
with
individualized
estimates
at
time-points
where
are
still
uncertain,
aiming
specifically
mitigate
delays
diagnosis
stemming
from
an
incomplete
differential.
To
train
such
model,
it
necessary
infer
temporally
fine-grained
retrospective
labels
eventual
"true"
diagnoses.
We
do
so
LLMs,
ensure
the
input
text
Deep
learning-based
predictive
models,
leveraging
Electronic
Health
Records
(EHR),
are
receiving
increasing
attention
in
healthcare.
An
effective
representation
of
a
patient's
EHR
should
hierarchically
encompass
both
the
temporal
relationships
between
historical
visits
and
medical
events,
inherent
structural
information
within
these
elements.
Existing
patient
methods
can
be
roughly
categorized
into
sequential
graphical
representation.
The
focus
only
on
among
longitudinal
visits.
On
other
hand,
approaches,
while
adept
at
extracting
graph-structured
various
fall
short
effectively
integrate
information.
To
capture
types
information,
we
model
as
novel
heterogeneous
graph.
This
graph
includes
nodes
events
nodes.
It
propagates
structured
from
event
to
visit
utilizes
time-aware
changes
health
status.
Furthermore,
introduce
transformer
(TRANS)
that
integrates
edge
features,
global
positional
encoding,
local
encoding
convolution,
capturing
We
validate
effectiveness
TRANS
through
extensive
experiments
three
real-world
datasets.
results
show
our
proposed
approach
achieves
state-of-the-art
performance.
