A Voting Optimized Strategy Based on ELM for Improving
Classification of Motor Imagery BCI Data
Lijuan Duan
•
Hongyan Zhong
•
Jun Miao
•
Zhen Yang
•
Wei Ma
•
Xuan Zhang
Received: 22 August 2013 / Accepted: 8 April 2014 / Published online: 22 April 2014
Springer Science+Business Media New York 2014
Abstract This paper presents an approach to classifying
electroencephalogram (EEG) signals for brain–computer
interfaces (BCI). To eliminate redundancy in high-dimen-
sional EEG signals and reduce the coupling among dif-
ferent classes of EEG signals, we use principle component
analysis and linear discriminant analysis to extract features
that represent the raw signals. Next, we introduce the
voting-based extreme learning machine to classify the
features. Experiments performed on real-world data from
the 2003 BCI competition indicate that our classification
method outperforms state-of-the-art methods in speed and
accuracy.
Keywords Brain–computer interface Principle
component analysis Linear discriminate analysis
Voting-based extreme learning machine
Introduction
Human brains are natural cognitive systems that have
powerful abilities to communicate with each other through
external sensing and internal computation. Discovering and
simulating the cognitive mechanisms of human brains is
always the goal of cognitive computation societies.
Although human brains are skull-enclosed, electrical
signals at the skull can be detected and analyzed to build
artificial cognitive systems that can communicate with the
natural cognitive systems. Such artificial cognitive systems
are capable of interacting with humans by human–com-
puter interaction (HCI) technologies, such as detecting
keywords in human speech [1], predicting human gaze
location in dynamic scenes [2], handwritten text recogni-
tion [3] and controlling external devices with brain activity
in brain–computer interfaces (BCI) [4]. Among these
technologies, BCI is noninvasive and allows a brain to
directly control a device, bypassing the use of muscular
activity. Motor imagery is considered to be mental
rehearsal of muscular activity, and various acquisition
techniques are available for capturing motor imagery brain
activity. Among these techniques, the electroencephalo-
gram (EEG) is most commonly employed for monitoring
brain activity in BCI systems. It requires relatively simple
and inexpensive equipment and it is more convenient to use
than other methods [5].
There are two fundamental stages in the processing of
EEG signals: feature extraction and classification [6]. A
great variety of features represented on EEG signals have
been explored, such as amplitude values [7], band powers
(BP) [8], power spectral density (PSD) values [9] and
autoregressive (AR) and adaptive autoregressive (AAR)
parameters [10]. Classification methods widely used
include k-nearest neighbor [7], support vector machines
(SVM) [11], neural networks [12], naive Bayes [13], and so
on. In this paper, we use LDA-after-PCA to obtain low-
dimensional feature representations of original EEG sig-
nals. PCA is commonly used in dimension reduction;
however, it does not take inter-class differences into
account. Thus, features obtained by PCA are not discrim-
inative enough. LDA is capable of obtaining the best
projection directions and generates features with maximum
L. Duan H. Zhong Z. Yang W. Ma X. Zhang
Department of Computer Science and Technology, Beijing
University of Technology, Beijing 100124, China
J. Miao (&)
Key Laboratory of Intelligent Information Processing of Chinese
Academy of Sciences (CAS), Institute of Computing
Technology, CAS, Beijing 100190, China
e-mail: jmiao@ict.ac.cn
123
Cogn Comput (2014) 6:477–483
DOI 10.1007/s12559-014-9264-1