Computer Methods in Biomechanics & Biomedical Engineering,
Journal Year:
2023,
Volume and Issue:
28(1), P. 51 - 60
Published: Nov. 20, 2023
To
enhance
the
accuracy
of
motor
imagery(MI)EEG
signal
recognition,
two
methods,
namely
power
spectral
density
and
wavelet
packet
decomposition
combined
with
a
common
spatial
pattern,
were
employed
to
explore
feature
information
in
depth
MI
EEG
signals.
The
extracted
features
subjected
series
fusion,
F-test
method
was
used
select
higher
content.
Here
regarding
classification,
we
further
proposed
Platt
Scaling
probability
calibration
calibrate
results
obtained
from
six
basic
classifiers,
random
forest
(RF),
support
vector
machines
(SVM),
Logistic
Regression
(LR),
Gaussian
naïve
bayes
(GNB),
eXtreme
Gradient
Boosting
(XGBoost),
Light
Machine
(LightGBM).
From
these
12
three
four
selected
for
model
fusion.
validated
on
Datasets
2a
4th
International
BCI
Competition,
achieving
an
average
data
nine
subjects
reached
91.46%,
which
indicates
that
fusion
effective
improve
classification
accuracy,
provides
some
reference
value
research
brain-machine
interface.
IEEE Transactions on Neural Systems and Rehabilitation Engineering,
Journal Year:
2024,
Volume and Issue:
32, P. 1535 - 1545
Published: Jan. 1, 2024
The
motor
imagery
brain-computer
interface
(MI-BCI)
based
on
electroencephalography
(EEG)
is
a
widely
used
human-machine
paradigm.However,
due
to
the
non-stationarity
and
individual
differences
among
subjects
in
EEG
signals,
decoding
accuracy
limited,
affecting
application
of
MI-BCI.In
this
paper,
we
propose
EISATC-Fusion
model
for
MI
decoding,
consisting
inception
block,
multi-head
selfattention
(MSA),
temporal
convolutional
network
(TCN),
layer
fusion.Specifically,
design
DS
Inception
block
extract
multi-scale
frequency
band
information.And
new
cnnCosMSA
module
CNN
cos
attention
solve
collapse
improve
interpretability
model.The
TCN
improved
by
depthwise
separable
convolution
reduces
parameters
fusion
consists
feature
decision
fusion,
fully
utilizing
features
output
enhances
robustness
model.We
two-stage
training
strategy
training.Early
stopping
prevent
overfitting,
loss
validation
set
are
as
indicators
early
stopping.The
proposed
achieves
within-subject
classification
accuracies
84.57%
87.58%
BCI
Competition
IV
Datasets
2a
2b,
respectively.And
cross-subject
67.42%
71.23%
(by
transfer
learning)
when
with
two
sessions
one
session
Dataset
2a,
respectively.The
demonstrated
through
weight
visualization
method.Index
Terms-Brain-computer
(BCI)
Electronics,
Journal Year:
2023,
Volume and Issue:
12(12), P. 2743 - 2743
Published: June 20, 2023
A
preponderance
of
brain–computer
interface
(BCI)
publications
proposing
artificial
neural
networks
for
motor
imagery
(MI)
electroencephalography
(EEG)
signal
classification
utilize
one
the
BCI
Competition
datasets.
However,
these
databases
encompass
MI
EEG
data
from
a
limited
number
subjects,
typically
less
than
or
equal
to
10.
Furthermore,
algorithms
usually
include
only
bandpass
filtering
as
means
reducing
noise
and
increasing
quality.
In
this
study,
we
conducted
comparative
analysis
five
renowned
(Shallow
ConvNet,
Deep
EEGNet,
EEGNet
Fusion,
MI-EEGNet)
utilizing
open-access
with
larger
subject
pool
in
conjunction
IV
2a
dataset
obtain
statistically
significant
results.
We
employed
FASTER
algorithm
eliminate
artifacts
processing
step
explored
potential
transfer
learning
enhance
results
on
artifact-filtered
data.
Our
objective
was
rank
networks;
hence,
addition
accuracy,
introduced
two
supplementary
metrics:
accuracy
improvement
chance
level
effect
learning.
The
former
is
applicable
varying
numbers
classes,
while
latter
can
underscore
robust
generalization
capabilities.
metrics
indicated
that
researchers
should
not
disregard
Shallow
ConvNet
they
outperform
later
published
members
family.
Frontiers in Neuroscience,
Journal Year:
2025,
Volume and Issue:
19
Published: Feb. 25, 2025
Easing
the
behavioral
restrictions
of
those
in
need
care
not
only
improves
their
own
quality
life
(QoL)
but
also
reduces
burden
on
workers
and
may
help
reduce
number
countries
with
declining
birthrates.
The
brain-machine
interface
(BMI),
which
appliances
machines
are
controlled
by
brain
activity,
can
be
used
nursing
settings
to
alleviate
stress
for
care.
It
is
expected
workload
workers.
In
this
study,
we
focused
motor
imagery
(MI)
classification
deep-learning
construct
a
system
that
identify
MI
obtained
electroencephalography
(EEG)
measurements
high
accuracy
low
latency
response.
By
completing
edge,
privacy
personal
data
ensured,
ubiquitous,
user
convenience.
On
other
hand,
however,
edge
limited
hardware
resources,
implementation
models
huge
parameters
computational
cost,
such
as
deep-learning,
challenging.
Therefore,
optimizing
measurement
conditions
various
model,
attempted
power
consumption
improve
response
minimizing
cost
while
maintaining
accuracy.
addition,
investigated
use
3-dimension
convolutional
neural
network
(3D
CNN),
retain
spatial
locality
feature
further
We
propose
method
maintain
enabling
processing
size
kernels
layer
structure.
Furthermore,
develop
practical
BMI
system,
introduced
dry
electrodes,
more
comfortable
daily
use,
optimized
memory
proposed
model
even
fewer
less
recall
time,
lower
sampling
rate.
Compared
EEGNet,
3D
CNN
parameters,
multiply-accumulates,
footprint
approximately
75.9%,
16.3%,
12.5%,
respectively,
same
level
eight
3.5
seconds
sample
window
size,
125
Hz
rate
4-class
dry-EEG
MI.
