Integrating Language Model and Dynamic Stopping to Improve a Code-Modulated Visual Evoked Potentials Brain-Computer Interface
Keywords
Loading...
Authors
Issue Date
2021-06-24
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Objective: The code-modulated visual evoked potentials (cVEP) speller is
a brain-computer interface (BCI) system that can facilitate communication
for people with severe motor and speech impairments. However, its implementation
is not yet fast and accurate enough in order to be a practical
solution for these patients. The goal of this project is to explore a method
that has the potential of improving classi cation speed. The dynamic stopping
technique is proven to be a way of achieving faster classi cation by
dynamically calculating a signal parameter to stop a trial when a certain
condition is met. In this paper, a new dynamic stopping method is proposed.
It uses language-speci c information extracted from a language model and
integrates them with probabilities calculated from another dynamic stopping
method to improve the speed of signal classi cation in a cVEP BCI
speller. Approach: The proposed method implements a Long-Short Term
Memory (LSTM) neural network to build a language model that outputs
a probability distribution over all possible characters in a speci c vocabulary.
The dynamic stopping method uses the features of a Beta distribution
to obtain a probability distribution over all possible classes in each classi
cation. Finally, a posterior distribution is updated by combining those
dynamically derived probabilities with the language-speci c priors using a
Bayesian approach. A con dence level parameter emits a class label when
a prede ned threshold is reached. Main results: This method was evaluated
by comparing its performance to the original dynamic stopping which
does not incorporate language information. For that, an o ine simulation
is created which aims to spell a di erent targeted sentence in each session.
The results demonstrate a signi cant increase of 6.6% in speed as measured
by symbols per minute. Signi cance: This study shows the potential for
classi cation improvement by integrating language information in a cVEP
BCI speller.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen