A dissertation submitted to the Faculty of Science,
University of the Witwatersrand, in ful llment of the
requirements for the degree of Master of Science. October 31, 2016. / Brain Computer Interface (BCI) technology is a promising new channel for communication
between humans and computers, and consequently other humans. This technology has the
potential to form the basis for a paradigm shift in communication for people with disabilities or
neuro-degenerative ailments. The objective of this work is to create an asynchronous BCI that
is based on a commercial-grade electroencephalography (EEG) sensor. The BCI is intended
to allow a user of possibly low income means to issue control signals to a computer by using
modulated cortical activation patterns as a control signal. The user achieves this modulation
by performing a mental task such as imagining waving the left arm until the computer performs
the action intended by the user. In our work, we make use of the Emotiv EPOC headset to
perform the EEG measurements. We validate our models by assessing their performance when
the experimental data is collected using clinical-grade EEG technology. We make use of a
publicly available data-set in the validation phase.
We apply signal processing concepts to extract the power spectrum of each electrode from
the EEG time-series data. In particular, we make use of the fast Fourier transform (FFT).
Specific bands in the power spectra are used to construct a vector that represents an abstract
state the brain is in at that particular moment. The selected bands are motivated by insights
from neuroscience. The state vector is used in conjunction with a model that performs classification. The exact purpose of the model is to associate the input data with an abstract
classification result which can then used to select the appropriate set of instructions to be executed
by the computer. In our work, we make use of probabilistic graphical models to perform
this association.
The performance of two probabilistic graphical models is evaluated in this work. As a
preliminary step, we perform classification on pre-segmented data and we assess the performance
of the hidden conditional random fields (HCRF) model. The pre-segmented data has a trial
structure such that each data le contains the power spectra measurements associated with only
one mental task. The objective of the assessment is to determine how well the HCRF models the
spatio-spectral and temporal relationships in the EEG data when mental tasks are performed
in the aforementioned manner. In other words, the HCRF is to model the internal dynamics
of the data corresponding to the mental task. The performance of the HCRF is assessed over
three and four classes. We find that the HCRF can model the internal structure of the data
corresponding to different mental tasks.
As the final step, we perform classification on continuous data that is not segmented and
assess the performance of the latent dynamic conditional random fields (LDCRF). The LDCRF
is used to perform sequence segmentation and labeling at each time-step so as to allow the
program to determine which action should be taken at that moment. The sequence segmentation
and labeling is the primary capability that we require in order to facilitate an asynchronous
BCI protocol. The continuous data has a trial structure such that each data le contains the
power spectra measurements associated with three different mental tasks. The mental tasks
are randomly selected at 15 second intervals. The objective of the assessment is to determine
how well the LDCRF models the spatio-spectral and temporal relationships in the EEG data,
both within each mental task and in the transitions between mental tasks. The performance of
the LDCRF is assessed over three classes for both the publicly available data and the data we
obtained using the Emotiv EPOC headset. We find that the LDCRF produces a true positive
classification rate of 82.31% averaged over three subjects, on the validation data which is in the
publicly available data. On the data collected using the Emotiv EPOC, we find that the LDCRF
produces a true positive classification rate of 42.55% averaged over two subjects.
In the two assessments involving the LDCRF, the random classification strategy would
produce a true positive classification rate of 33.34%. It is thus clear that our classification
strategy provides above random performance on the two groups of data-sets. We conclude that
our results indicate that creating low-cost EEG based BCI technology holds potential for future
development. However, as discussed in the final chapter, further work on both the software and
low-cost hardware aspects is required in order to improve the performance of the technology as
it relates to the low-cost context. / LG2017
| Identifer | oai:union.ndltd.org:netd.ac.za/oai:union.ndltd.org:wits/oai:wiredspace.wits.ac.za:10539/21648 |
| Date | January 2016 |
| Creators | Khoza, Phumlani Rueben Nhlanganiso |
| Source Sets | South African National ETD Portal |
| Language | English |
| Detected Language | English |
| Type | Thesis |
| Format | Online resource (116 leaves), application/pdf |
Page generated in 0.0016 seconds