1,720,982 research outputs found
Exploring code families and event-types for cVEP BCIs
No full textA brain-computer interface (BCI) is a system that makes it possible for a
user to interact with a computer without muscular control. Code-modulated
visual evoked potential (cVEP) BCI can be used to create visual speller systems.
To make a step towards practical, real-world cVEP BCI by improving
performance and reducing calibration time, this thesis compares three methods
of decomposing codes, called event-types, when using the reconvolution
method by Thielen et al. (2015) [12] to generate templates. This is done on
two di erent data sets, one using random stimulation codes and one using
modulated Gold codes. It is found that event-types that take into account the nonlinearity of
the visual system perform better in terms of classi cation accuracy, auto
explained variance and cross explained variance than event-types that assume
linearity. It is concluded that using the edge responses of a
ash as an event-type is the most practical approach, as it models only two events
and therefore requires less training data than event-types with more events.
The contrast event-type is also more versatile, as it does not require events
in the validation set to have appeared in the training set
6-channel to stereo audio transformation optimizing stimuli distinguishability in AMUSE paradigm.
No full textBrain-computer interfaces provide novel means of communication through
induced brain signals, bypassing the need for motor function. A novel auditory
BCI paradigm has been shown to facilitate rehabilitation in native
German post-stroke aphasia patients. The paradigm employed an auditory
oddball task incorporating spatial information into the auditory stimuli.
Spatial information was introduced with the aid of a ring of six speakers
surrounding a patient. Long and intense training sessions were required for
patients to improve on clinical aphasia assessments. This study aims to find
novel ways of introducing supplementary information to auditory stimuli
such that the ring of speakers might be replaced by stereo headphones. Five
different transformations from 6-channel to stereo audio were developed and
assessed. Additionally, a web application was developed to facilitate future
researchers in the creation of novel auditory stimuli incorporating spatial
information
A Deep Learning approach to Noise Tagging
No full textBrain-computer interfaces (BCI) are systems that allow direct communication
between the brain and a computer. A speci c type of BCI is a codemodulated
visual evoked potential (cVEP) based spelling BCI. These cVEP
BCIs can use di erent approaches to decoding and classifying brain signals.
Reconvolution, linear discriminant analysis (LDA) and deep learning are
some of these approaches. In my research, I implement my own adaptation
of a deep learning model and compare it against a baseline model which is
LDA, and I compare both deep learning and LDA against a simple event
type reconvolution. The deep learning approach (96.0% accuracy) is shown
to be signi cantly better (p < :05) than the simple event type reconvolution
(56.8% accuracy). Furthermore, on trial classi cation the deep learning approach
(96.0% accuracy) does not perform signi cantly di erent (p > :05)
from LDA (95.7% accuracy). On epoch classi cation deep learning (65.3%
accuracy) performs signi cantly better (p < :05) than LDA (63.2% accuracy).
This proves deep learning to be a useful option for cVEP BCIs with
many ways to go further in research
Influence of Word Presentation via Stereo Loudspeakers in an Auditory ERP Paradigm for Brain-Computer Interfaces
No full textBrain-computer interfaces (BCI) allow users to interact or communicate with their environment
without depending on motoric output pathways. By classifying and translating
neurophysiological signals into command signals for an external device, BCI users can control
applications such as a wheelchair, spelling program, or prosthesis. Auditory BCI systems detect
attention to auditory stimuli based on the event-related potentials in electroencephalography
(EEG) signals. These auditory BCI systems can be used for cognitive rehabilitation. Musso et al.
(2022) developed a novel language training paradigm for aphasia rehabilitation in stroke
patients using an EEG-based BCI system. In their approach, patients are given individual and
immediate brain-state-dependent feedback during an auditory detection task, which serves as
an indicator of their task success. Hence, the approach exploits the notion that reinforcing an
appropriate language processing strategy may induce beneficial brain plasticity. During the
language training task, patients sit in a ring of six loudspeaker, which present the auditory
stimuli. However, his set-up is not feasible for at home- or practitioner-based training. Therefore,
the present study will explore the possibility of a simplified audio set-up with stereo headphones
by investigating the differences in offline classification accuracy and ERP components in the
described experimental set-up. In a within-subject study design, eight healthy participants tested
the proposed BCI paradigm in four audio conditions: six loudspeakers, stereo headphones,
stereo headphones incorporating pitch, and mono headphones. Results show that the stereo
headphones incorporating pitch are comparable to the six loudspeakers in terms of offline
classification accuracy
Exploring optimal codes and ensemble classifiers in a c-VEP BCI
No full textBrain-computer interfaces (BCIs) are used to interact with a computer using brain activity.
BCIs based on code-modulated visual evoked potentials (c-VEPs), whereby the background
of targets flickers with different codes, have recently shown the ability to outperform other
methods in terms of accuracy and speed. In a speller application, electroencephalogram
(EEG) data is recorded from the user while they focus on a target. Though c-VEP-based
BCIs outperform other methods, the accuracy and speed could still be improved. The present
research aims is to improve the accuracy and speed of classification in an offline analysis of two
datasets. First, to explore improving the accuracy, combining multiple canonical correlation
analysis (CCA, a feature extractor) components in an ensemble classifier is shown to not
significantly improve accuracy. Secondly, to improve speed, it is found that there is no
significant difference in the information transfer rate (ITR) of Gold codes within a set. With
a similar goal, amount of neighbours is found to not have a significant effect on the accuracy
of Gold codes. The combination of finding that there is no significant difference in ITR of
Gold codes within a set, and their low cross-correlation properties shows there is no bias in
either. Furthermore, it is found that the amount of neighbours does not significantly affect
Gold codes. Further research could focus on testing these findings on other popular code
families
Can we go faster than world's fastest brain-computer interface: An application of recurrency to EEG2Code Deep Learning
No full textIn this research the structure of recurrent neural networks was applied to
EEG2Code deep learning. A state of the art high-speed electroencephalogram
(EEG) brain-computer interface (BCI) speller, using a deep convolutional
neural network. The BCI predicts user intention in a noise-tagging
paradigm speller set-up by decoding the EEG data and predicting individual
ashes in the noise code. While EEG2Code prides itself as being the current
world fasts BCI with an information transfer rate (ITR) of 1237 bits/min
with a trial classi cation accuracy of 95:9%. There is still room for improvement
on the level of individual
ash prediction. Therefore I introduce an
extended variant of the EEG2Code deep learning model which makes use
of a recurrent neural network layer. The goal of this layer is to capture the
temporal dependencies that occur in the data when recording user intention
in the hopes of increasing the ability to predict individual
ashes with a
higher degree of certainty. The recurrent neural network used was a simple
recurrent neural network unit. While this addition did not seem to perform
as expected, by increasing the performance of individual
ash prediction
with the initial hyperparameters given, it could serve as a starting point
for future researchers to tweak the hyperparameters in an attempt to truly
capture the temporal dependencies on an individual
ash to
ash basis
Riemannian geometry for code-modulated visual evoked potential brain-computer interfaces: Towards calibration-less brain-computer interface
No full textA brain-computer interface (BCI) is a system that makes it possible to have
a direct communication between the brain and a computer. To make a step
towards practical BCI systems we want to improve the decoding and shorten
calibration time. A method that has been gaining more popularity in the
BCI community is the Riemannian geometry framework. To our knowledge,
so far no code-modulated visual evoked potential (cVEP) based BCI has
adapted a Riemannian approach. Therefore, it is yet unknown if the Riemannian
framework could possibly improve or have the same performance
as the current state of the art cVEP based methods. To answer the question
if this Riemannian framework could be used for cVEP BCIs I compare
the performance of the minimum distance to the Riemannian mean (MDM)
classi er to a baseline method, which is linear discriminant analysis (LDA).
Two dimensionality methods are introduced namely, canonical correlation
analysis (CCA) and cherry picking. Multiple scenarios are tested these include
a small and larger training dataset, the use of one or two (virtual)
channels and for the MDM an extra scenario where the use of both prototypes
or only the target prototype is investigated. The MDM classi er shows
comparable overall results with LDA. Additionally, the use of two (virtual)
channels signi cantly outperforms (p < 0:05) the use of one (virtual) channel.
Lastly, the use of both prototypes is not signi cantly different from
only using the target prototype. However, it could improve computational
effciency to only use the target prototype. Overall I can say that the use
of MDM and in turn the Riemannian framework can be recommended for
cVEP based BCIs
Towards a Simplified Audio Setup of a BCI-based Language Training Paradigm: a Behavioral Study
No full textBrain-computer interface (BCI) systems use brain signals to create a communication link
between the brain and an external device. Auditory BCIs detect focused attention on
acoustic stimuli based on event-related potentials (ERPs) in signals from
electroencephalography. A recent study by Musso et al. (2022) investigated the use of an
auditory BCI for aphasia rehabilitation in stroke patients. Musso et al. propose a new
BCI-based language training paradigm that reinforces effective language processing
strategies through brain-state-dependent feedback. The paradigm uses a ring of six
loudspeakers for the spatial presentation of stimuli which has been shown to improve
classification accuracy in auditory BCI paradigms. However, this setup is not feasible for
everyday BCI usage. Thus, the present study aims to investigate the possibility of a
simplified audio setup using stereo headphones. The proposed setup will be evaluated by
conducting a behavioral study and assessing the workload and ergonomic ratings. Eight
healthy participants were tested in a within-subject design using four audio conditions: six
loudspeakers, stereo headphones, stereo headphones incorporating pitch and mono
headphones. Results show the two stereo headphone conditions are comparable to the six
loudspeaker condition in terms of counting accuracy, workload ratings and ergonomic
ratings. This indicates that a spatial headphone paradigm might be a promising compromise
between the goal of high classification accuracy and easy applicability for everyday BCI use
Improving the Event-related Potential Classi cation Performance of the Riemannian Pipeline using Shrinkage Regularization
No full textThe aim in the eld of brain-computer interfaces is to control an electronic device using brain
responses that are recorded in an electroencephalogram. These brain signals can be decoded by
means of the Riemannian metrics that are associated with the Riemannian manifold. This way of
classifying allows for transfer learning possibilities and is more robust opposite to the Euclidean
approach. However, to get to these desired characteristics, the data points that represent a
single brain response need to be represented by their covariance matrices. This is a drawback of
the framework, as the sample covariance matrix is a sub-optimal estimator of the true covariance
matrix when the number of samples is low compared to the number of features. To improve the
estimator and enhance the classi cation performance, I apply shrinkage regularization on the
di erent submatrices of the epoch-based covariance matrix. To assess the e ect of this method,
I compare two decoding algorithms against a baseline using six datasets. The results show that
there is a signi cant increase in classi cation performance for one out of the six datasets
A comparison of non-linear methods for the decoding of motor performance.
No full textCurrently, work is underway to create a closed-loop deep brain stimulation
(DBS) system. This research attempts to contribute to the creation of a
closed-loop DBS system by improving motor labels used in the search for
informative brain signals. This was done by building on earlier research
that used a motor task called the copy-draw task, to evaluate ne-grained
hand motor performance under DBS-on and DBS-o conditions. The earlier
research used linear discriminant analysis (LDA) to reduce the large feature
dimensionality to a single scalar value representing the motor score, which
could then be used in conjunction with source power comodulation to nd
informative brain signals. However, LDA makes the assumption that the
underlying features distributions are Gaussian. We showed that this assumption
is not met by the data, and we therefore tested whether logistic
regression, random forest, and the support vector classi er, which do not
make the assumption of Gaussian features yield better motor decoding performance.
We evaluated the new methods using a nested cross-validation
procedure with hyperparameter optimization. We also tested whether removing
linear trends from the data, or splitting trials of the copy-draw task
to create more training data yielded better motor decoding performance.
We demonstrate that there is no signi cant increase in motor decoding performance
by logistic regression, random forest, or support vector classi er.
Moreover, we saw no increase in the motor decoding performance by using
the data where linear trends were removed, or where the trials were split to
create more training data. Our research therefore shows that in this problem
setting LDA is robust to non-Gaussian features
- …
