Abstract— Developing efficient and usable brain-computer
interfaces (BCIs) requires well-designed trade-off between
accuracy and computational time. This paper presents a very
fast and accurate method to classify asynchronous brain signals
from a multi-class mental tasks dataset using time-domain
features. Five different statistical time-domain features were
extracted to characterize various properties of three mental
tasks electroencephalograms (EEGs). Versatile Elliptic Basis
Function Neural Network (VEBFNN) was employed to classify
single EEG features as well as multi-feature set. Discriminating
power of single features was evaluated and compared by
considering the classification accuracy and computational cost
consumed during the training stage. Finally, the performance
of the best single EEG feature was compared to the multi-
feature set. The results indicated the usefulness of Willison
Amplitude EEG feature in classifying the different motor tasks
as it provided the highest discrimination ratio. Classification
results showed the high potential of VEBFNN by the average
89.78% accuracy and 0.21 seconds computation time obtained
for its offline training. Moreover, VEBFNN outperformed the
conventional support vector machine classifier in both terms of
accuracy and speed.
I. INTRODUCTION
In the past two decades, there has been growing interest
for the biomedical and neural engineers as well as the
neuroscientists to connect brain signals with
computers/machines in order to provide a reliable and
comfortable communication pathway for disabled people
with external environment. Brain computer interface (BCI)
systems aim to detect users’ thoughts and transform them
into input commands that control devices like wheelchair and
prosthesis hand [1-2]. Various BCI systems using EEG
signals have been proposed among which imagination of
Motor Imagery (MI)-based, imagination of different limb
movements, is one of the most promising [3].
M. H. is with Center for Biomedical Engineering, Transportation
Research Alliance, Universiti Teknologi Malaysia, 81310, Johor Bahru,
Malaysia (phone: +6014-7730-290; e-mail: hamedi.mahyar@ieee.org).
Sh-H. S. is with Center for Biomedical Engineering, Transportation
Research Alliance, Universiti Teknologi Malaysia, 81310, Johor Bahru,
Malaysia (e-mail: hussain@fke.utm.my).
I. M-R. is with David Geffen School of Medicine, Semel Institute for
Neuroscience and Human Behavior, University of California, Los Angeles
& Center for Mind and Brain, University of California, Davis, USA (e-mail:
irezazadeh@ucdavis.edu).
M. A. is with department of Biomedical Engineering, Science and
Research Branch, Islamic Azad University Tehran, Iran (e-mail:
astarakee@yahoo.ca ).
A. M. N. is with Center for Biomedical Engineering, Transportation
Research Alliance, Universiti Teknologi Malaysia, 81310, Johor Bahru,
Malaysia (e-mail: alias@mail.fkm.utm.my).
Imagination of different motor tasks like right/left hand
movements generate different EEG features in the brain
hemispheres corresponding to specific sensorimotor areas.
Such electrical brain activities produced by underlying
groups of neurons can be detected non-invasively by surface
electrodes. To classify commands from these
electrophysiological time-varying signals several stages of
processing are necessary. Feature extraction and
classification play the most dominant roles for designing
BCIs as they directly influence the performance indices of
the designed system. Since BCIs have been mainly proposed
to be used for online applications, some considerations need
to be taken into account like a reliable trade-off between
classification accuracy and computational complexity.
Although many robust techniques have been introduced
for feature extraction and classification of MI signals, they
have not considered the trade-off criteria and could not
efficiently tackle the observed problems in real-time BCI
applications. In terms of feature extraction, more studies have
focused on the extraction of spectral, time-frequency, and
spatial features like Quantification of Event-Related
Synchronization/ Desynchronization (ERS/ERD)
phenomenon [4]; Morlet Wavelet transform [5]; Empirical
Mode Decomposition (EMD) [6]; Hilbert-Huang Transform
(HHT) [7]; Short-Time Fourier Transform (STFT) [8];
Principal Component Analysis [9]; Independent Component
Analysis (ICA) [7]; Common Spatial Pattern (CSP) [8].
These features are computationally expensive as they need
different levels of transformation which lead to longer
processing time. On the other hand, classification problem
has been addressed by numbers of linear and non-linear
algorithms with different complexity and efficiency such as
SVM [9], [10], K-Nearest Neighbor (KNN) [11], and Linear
Discriminant Analysis (LDA) [12]. Even though high
accuracy might have been achieved by the mentioned
methods, computational cost has neither been investigated
nor reported vividly. Recently, the effectiveness of time-
domain (TD) features for characterizing the MI movements
for BCI systems has been investigated [9], [13-14] and it is
reported that not only these features are very simple and easy
to compute but also they could perfectly represent the
discriminative information of different MI EEGs and lead to
high degree of classification accuracy. Although very few
studies extracted this type of features for MI-based BCIs, it
has been shown that they deliver different classification
performance when classified by various techniques.
In this paper, by considering the TD feature extraction
approach we offered a new methodology to provide a reliable
trade-off between computational cost and accuracy when
EEG signals were recorded asynchronously. Five different
Asynchronous Multiclass Mental Tasks Classification through Very
Fast Versatile Elliptic Basis Function Neural Network
Mahyar Hamedi, EMBS Member, Sh-Hussain Salleh, IEEE Member, Iman Mohammad-Rezazadeh,
Mehdi Astaraki, Alias Mohd Noor
2014 IEEE Conference on Biomedical Engineering and Sciences, 8 - 10 December 2014, Miri, Sarawak, Malaysia
978-1-4799-4084-4/14/$31.00 ©2014 IEEE 295