978-1-4673-7678-5 ©2015 IEEE 1056
2015 11th International Conference on Natural Computation (ICNC'15)
A Feature Extraction Method Based on Dictionary
Learning for EEG
Lingyue Xie; Han Zhang; Feng Duan
College of Computer and Control Engineering
Nankai University
Tianjin, China 300071
Abstract—For decades, it has been widely used to extract EEG
features on every single trial, while in this article, features are
extracted based on one fixed dictionary basis. Here, by designing
a feature extraction method applying dictionary learning on EEG
signals and by using the BCI competition EEG data of two classes,
we show that the degree of every used dictionary component
related to task state and relaxed state are different and could be
used as the feature of EEG. What’s more, we use Bayesian
classifier to classify our features compared with wavelet features
and find that our accuracy is a lot higher than wavelet.
Keywords- dictionary learning; sparse coding; EEG; Bayesian
classifier; relaxed state; task state
I. INTRODUCTION
Recently sparse coding and dictionary learning are widely
used in machine learning, neuroscience, signal processing and
image classification, for example, the iterative least squares
dictionary learning algorithm provided by Engan [1] and the K-
SVD method for dictionary learning mentioned by Aharon [2].
A lot of dictionary learning methods were provided for
different conditions, such as online dictionary learning by
Mairal [3] to obtain the overlap dictionary components and
fisher discrimination dictionary learning by Yang [4] which
could get dictionary components belong to different classes.
The dictionary learning methods have gotten great results on
image denoising and classification. The dictionary learning
method models data vectors as sparse linear combinations of
basis elements which make up the dictionary. Different from
the methods before, dictionary learning does not need the
predefined dictionary components just like the ones based on
wavelet. The online dictionary learning method model also
does not impose that the basis vectors be orthogonal, allowing
more flexibility to adapt the representation to the data. Though
dictionary learning method is great to represent the natural
signals, but only several people used these methods on brain
study. The University of Georgia used online dictionary
learning [5] and fisher discrimination dictionary learning [6] to
analysis the Functional Magnetic Resonance Imaging. While
almost nobody take this method on the EEG analysis.
The EEG has developed a lot since Hans Berger recorded
the first human EEG in 1924. EEG is the recording of brain
electrical activity along the scalp whose amplitude is very
small. When the brain is activated, the neurons of the brain will
discharge and produce ionic current flow. Then the current
flow spread to the scalp. The EEG measures voltage
fluctuations to estimate the state of the brain. There are several
different types of EEG, for example, steady-state visual evoked
potentials (SSVEP), P300 and motor imagery. The steady-state
visual evoked potentials are evoked by external stimulus of
flicking checkerboard at specific frequencies [7].When the
retina is excited by a visual stimulus ranging from 3.5 Hz to 75
Hz, the brain generates electrical activity at the same frequency
of the visual stimulus. SSVEP is useful in research because of
the excellent signal-to-noise and relative immunity to artifacts.
While the motor imagery EEG is produced by imaging
movement but without real action [8]. Though there is not any
real movement, there are still EEG voltage fluctuations. Just
image you are writing words with your hand, a few times later
you will feel sore. P300 is an event related potential component
elicited in the process of decision making. It is considered to be
an endogenous potential, as its occurrence links not to the
physical attributes of a stimulus, but to a person’s reaction to it.
The P300 signal contains a large positivity that peaks around
300 milliseconds after the stimulus appeared. Besides these
different types of EEG, there are a lot of methods to do feature
extraction and classifying, the extracted features could belong
to time domain, frequency domain or the time-frequency
domain, for example the adaptive autoregressive (AAR) model
mentioned by Pfurtscheller [9], the wavelet feature used in the
article by Xu [10]. The AAR is the feature from the time
domain while the wavelet is the time-frequency feature and the
Fourier Transform belongs to the frequency domain. After
feature extraction, the raw signals could be expressed by some
typical features. So feature extraction is important, if the
extracted features are not different enough of different classes,
the classification results will be very bad. There are also a lot
of classifiers to be used to distinguish the EEG signals. These
classifiers could be divided into two types, linear classifier and
nonlinear classifier. Linear classifiers for example linear
discriminant analysis [11], the least square estimation and
nonlinear classifiers k-nearest neighbor [12], neural network
[13] and Bayesian.
In this article, we use online dictionary learning method to
extract features of EEG signals and Bayesian classifier to do
classification. The datasets and preprocessing are described in
section II and the method will be introduced in section III. The
results are in the section IV and the section V is the conclusion.
This work is supported by the National Natural Science Foundation of
China (No. 61203339), the Key Technologies R & D Program of Tianjin
(No.14ZCZDSY00008), the Tianjin International Science and technology
cooperation project (14RCGFGX00848), and the Research Fellowship for
International Young Scientists (No. 61450110444) Tianjin Research Program
of Application Foundation and Advanced Technology (No.15JCYBJC18900)