Med Bio Eng Comput
DOI 10.1007/s11517-006-0107-4
TECHNICAL NOTE
Adaptive feature extraction for EEG signal classification
Shiliang Sun Æ Changshui Zhang
Received: 14 March 2006 / Accepted: 21 August 2006

International Federation for Medical and Biological Engineering 2006
Abstract One challenge in the current research of
brain–computer interfaces (BCIs) is how to classify
time-varying electroencephalographic (EEG) signals
as accurately as possible. In this paper, we address this
problem from the aspect of updating feature extractors
and propose an adaptive feature extractor, namely
adaptive common spatial patterns (ACSP). Through
the weighed update of signal covariances, the most
discriminative features related to the current brain
states are extracted by the method of multi-class
common spatial patterns (CSP). Pseudo-online simu-
lations of EEG signal classification with a support
vector machine (SVM) classifier for multi-class mental
imagery tasks show the effectiveness of the proposed
adaptive feature extractor.
Keywords Brain–computer interface (BCI) Æ
Common spatial patterns (CSP) Æ EEG signal
classification Æ Feature extraction
1 Introduction
The research of brain–computer interfaces (BCIs),
which aim to provide their users communication and
control capabilities that do not depend on the brain’s
S. Sun (&) Æ C. Zhang
State Key Laboratory of Intelligent Technology
and Systems, Department of Automation,
Tsinghua University, Beijing 100084, China
e-mails: shiliangsun@gmail.com; sunsl02@mails.tsinghua.edu.cn
C. Zhang
e-mail: zcs@mail.tsinghua.edu.cn
normal output channels of peripheral nerves and
muscles, arouses more and more interests of late
years [9, 13–15]. Up to now, study of BCI systems has
mainly involved recording of electroencephalographic
(EEG) signals using surface electrodes, as this kind
of recording is relatively convenient, harmless and
inexpensive compared with other methods [2]. In this
paper, we focus on the classification problem of
EEG signals, a crucial component embodied in gen-
eral EEG-based BCIs. For an EEG-based BCI,
adaptive learning algorithms are necessary in princi-
ple, because the recorded EEG signals usually change
over time due to both biological and technical causes,
such as subject attention, subject fatigue, disease
progression, electrode impedances, amplifier noise,
and environmental noise [13]. The high variability of
EEG recordings makes it a difficult task to classify
different EEG signals accurately and necessitates
adaptive learning to boost up the performance of
existing BCIs.
With respect to adaptive learning for EEG signal
classification, one can choose to update classifiers or
alternatively the feature extractors. However, up to
now, there is not much work addressing this problem.
The adaptive update of Bayesian statistical classifier
with Gaussian mixture model (GMM) is recently
studied in several papers [6, 7, 10, 11], whereas the
performance is still very moderate. Wolpaw and
McFarland [16] used the least-mean-square (LMS)
algorithm to adaptively adjust weights for a two-
dimensional movement control and found out that
people with severe motor disabilities could use scalp
EEG signals to operate a robotic arm or a neuro-
prosthesis. In this paper, we propose to address the
adaptive learning problem in EEG signal classification
123

via updating feature extractors. The basic feature
extractor is named common spatial patterns (CSP),
whose essence is to project EEG signals to the most
discriminative directions found after the simultaneous
diagonalization of covariance matrices from different
signal categories [8]. Because of the inherent variabil-
ity of EEG patterns, the discriminative directions for
classification tend to shift over time. In order to resolve
this problem, the method of adaptive common spatial
patterns (ACSP) is thus presented to improve the CSP
method.
2 Adaptive feature extraction
2.1 Common spatial patterns and its extension
to the multi-class paradigm
The original feature extractor of CSP can be seen as
linear spatial filters that lead to signals which discrim-
inate optimally between two conditions. It is based on
a decomposition of raw multi-channel signals into
spatial patterns that are extracted from the data of two
populations of EEGs in a manner that maximizes their
differences. These spatial patterns provide a weighting
of the electrodes, which is derived directly from the
data (for detailed description and computational is-
sues, please refer to [8]). Recently, some approaches
are presented to extend CSP to the multi-class para-
digm, such as using CSP within the classifier, one ver-
sus the rest CSP (OVR), approximate simultaneous
diagonalization [3, 8]. In this paper, the idea of OVR is
adopted to carry out feature extraction as suggested in
[3]. To be exact, if three sets of EEG trials or segments
corresponding to three different conditions A, B, and C
are given (e.g., three class of mental imaginary tasks),
we use each set to obtain covariance matrices CA, CB,
and CC, respectively. Then we can combine the sam-
ples belonging to conditions B and C to obtain a
covariance matrix CA_Rest. Likewise, covariance
matrices CB_Rest and CC_Rest could be obtained. Se-
quently, we use each pair of covariance matrices (e.g.,
CA and CA_Rest) to carry out the standard CSP proce-
dures. Thus the optimal spatial filters related to the
corresponding conditions A, B, and C are reserved.
The final feature extractor is the combination of the
three sets of spatial filters.
2.2 Adaptive common spatial patterns
Given K EEG trials or segments x(k) (k = 1,...,K), the
covariance matrix can be usually estimated as
123
Med Bio Eng Comput
1X K
CðkÞ ¼ xðkÞx> ðkÞ; ð1Þ
K k¼1
where x(k) is a N · T EEG recording matrix, and N
and T are, respectively, the number of recording elec-
trodes and recording points. As we know, the CSP
feature extractor adopts fixed covariances obtained
from training sessions. But in the ACSP feature
extractor when encountering a new trial or EEG
segment from test sessions, e.g., x(k), we update the
corresponding covariance matrix as follows:
CðkÞ ¼ lCðk
1Þ þ ð1
lÞxðkÞx> ðkÞ; ð2Þ
where l2[0,1] is defined as variability coefficient in our
paper. This kind of adaptive strategy embodies the idea
of weighted average, i.e., the current covariance matrix
is described as the weighted sum of historical covari-
ance and the covariance of the current recording seg-
ment. In general, those two covariance components
given in the right side of Eq. 2 are both necessary.
C(k – 1) contains historical information and can also
benefit the robust computation of C(k). xðkÞx> ðkÞ
contains the newly added information and reflects the
time-varying characteristic of EEG signals. Concretely,
in the multi-class paradigm if the new EEG segment
x(k) belongs to condition A, we would update the re-
lated variables as follows:
CA ðkÞ ¼ lCA ðk
1Þ þ ð1
lÞxðkÞx> ðkÞ;
CB Rest ðkÞ ¼ lCB Rest ðk
1Þ þ ð1
lÞxðkÞx> ðkÞ; ð3Þ
>
CC Rest ðkÞ ¼ lCC Rest ðk
1Þ þ ð1
lÞxðkÞx ðkÞ:
Then, we follow the standard procedures of CSP to
derive projection directions.
The ACSP method is a superset of the basic CSP
method because when variability coefficient l = 1, it
degenerates to the CSP method. The selection of l is
very flexible, which can reflect the discrepancies of
different subjects. For subjects whose signal vari-
ability is very slow, l should take large values to
retain more historical information and vice versa.
When the ACSP feature extractor is followed by a
classification task in BCI applications, we employ the
last EEG segment to update the feature extractor
and then use the updated feature extractor to extract
the features of the current segment. Based on the
extracted features, the classifier trained on the pre-
vious training sessions could give the estimated label
of the current segment. This process runs iteratively
as new EEG segments are continually recorded and
sent for classification.
Med Bio Eng Comput
3 Experiments
The data set used in this paper contains EEG record-
ings from three normal subjects (denoted by S1, S2, S3,
respectively) during mental imagery tasks, which are
imagination of repetitive self-paced left-hand move-
ments (class C1), imagination of repetitive self-paced
right-hand movements (class C2) and generation of
different words beginning with the same random letter
(class C3). For every subject, there were three record-
ing sessions acquired on the same day, each lasting
about 4 min with breaks of 5–10 min in between [1].
Galán et al. [5] show that subjects S1, S2, and S3 rep-
resent three different levels of mental consistency,
which are, respectively, consistent, scarcely consistent,
and inconsistent. Hence the data set is representative
and in this paper it is employed to assess algorithms.
3.1 Signal preprocessing
As mental imagery tasks are mainly related to the
activities of brain’s sensorimotor cortices, from the
entire electrode cap we retain the central 15 electrodes
covering this cortical region for signal analysis, which
are F3, Fz, F4, Fc1, Fc2, C3, Cz, C4, Cp1, Cp2, Pz, P3,
P4, Po3, Po4. Because the usual frequency of sponta-
neous EEG recordings is normally below 50 Hz, we
convert the original sampling rate to 128 Hz without
fear of information loss. The signals are then referred
by the Common Average Reference method [15].
Further, the continuous recording sessions are,
respectively, partitioned to segments of 1 s with 0.5 s
overlapped to provide an output every 0.5 s using the
last second of data, as in [6]. Resultantly, an EEG
segment would be 128 points, with a time length of 1 s.
To emphasize the l rhythm which is highly discrimi-
native in distinguishing different mental tasks, these
segments are temporally filtered (forward and reverse
filtering) with pass bands 8–13 Hz (for subjects S1 and
S2) and 11–15 Hz (for subject S3). The filtering bands
are determined by observing the first sessions of these
three subjects, and the slight variation of filtering bands
among subjects reflects the individual specialty. In
addition, each EEG segment is normalized to have unit
energy, and to avoid the imperfection of temporal fil-
tering to the start and end parts of a segment, 81 points
(with index from 20 to 100) are retained for analysis
from the original 128 points.
3.2 Training the classifier
Throughout this paper, the support vector machine
(SVM) classifier with a radial basis function (RBF)
kernel is adopted to classify EEG segments [12]. The
experimental paradigm for training a classifier is as
follows. First we choose two recording sessions from
the same subject, one serving as training session and
the other test session. Then we use the stationary CSP
method to extract features from the training session,
and use these features further to train a SVM classifier.
The number of sources (number of spatial patterns
related to one condition, such as ‘left-hand movement’)
varies from 2 to 6 in our experiments, for this is
empirically enough to describe the sources of mental
imagery tasks. Therefore, the maximum dimensions of
one EEG segment after feature extraction would be 18
(each 6 dimensions belong to one kind of brain state).
The optimal parameters, i.e., the penalty parameter for
error terms in the SVM classifier, the variance
parameter in RBF kernel, and the number of sources
related to each mental task, are selected through 20-
fold cross validation [4] on each training set. Finally,
SVM is retrained using these optimal parameters and
the whole training set.
3.3 Experimental results
For the sake of objectively evaluating the ACSP feature
extractor, two other feature extractors are employed
for performance comparisons. One is stationary CSP
(SCSP), the standard CSP method which does not up-
date spatial filters at all on new test sessions. The spatial
filters computed from the training session are used to
implement feature extraction of the test session. The
other feature extractor, which we name as windowed
CSP (WCSP), updates the signal covariance by adding a
new EEG segment and removing the first segment from
the original segment entries for calculating covariance
matrix. Like ACSP, WCSP uses updated covariances to
construct a new feature extractor and then extracts
features of the current EEG segment. The variability
coefficient l in ACSP feature extractor takes 0.95 in
our experiments empirically.
The experimental results of classification rates on
all the available recording sessions using the above
three feature extractors are given in Table 1. Through
a paired t test, no significant differences are found
between feature extractors SCSP and WCSP (P
value = 0.25), although the average accuracy of WCSP
is slightly better than that of SCSP. However, with
regard to SCSP and ACSP, 16 out of 18 classification
results of using ACSP are superior to those using SCSP.
Through a paired t test, significant differences are
found between feature extractors SCSP and ACSP
(P value < 1 · 10–3). From these results, we can draw
the conclusion that the ACSP feature extractor is the
123
 Med Bio Eng Comput

Table 1 The classification accuracies (%) using different feature and the present study does not take this into account.
extractors A complete adaptive algorithm should also address the
Subject Training session Test session Accuracy adaptive behavior of the brain itself.
SCSP WCSP ACSP The computational complexity of ACSP is much less.
Because the number of electrodes (data dimensions) in
S1 1 2 69.03 68.60 70.32 BCI utilities is usually small, such as 64, 128, and 256 (15
1 3 68.74 68.31 76.66
electrodes in our experiments), the simultaneous diag-
2 1 61.59 65.24 63.73
2 3 69.38 69.16 64.24 onalization of covariance matrices could almost be
3 1 51.93 59.23 63.30 implemented in real time. This could completely meet
3 2 58.71 57.42 67.96 the needs of online learning. Besides, as the subjects
S2 1 2 51.29 57.33 65.09 used in this article represent three different levels of
1 3 56.71 60.82 74.46
2 1 54.31 53.88 63.36 mental consistency, which are respectively consistent,
2 3 58.01 57.36 69.70 scarcely consistent, and inconsistent [5], we are confi-
3 1 53.02 56.03 70.47 dent that ACSP would function well in a wide range of
3 2 59.91 56.03 65.52 users. We believe that ACSP would show more merits
S3 1 2 56.49 58.87 65.80
1 3 52.60 48.48 57.58 in future developments of BCI technology.
2 1 54.63 54.85 53.96
2 3 55.19 54.33 56.06 Acknowledgments The authors would like to thank IDIAP
3 1 51.98 54.85 57.27 Research Institute of Switzerland for providing the analyzed
3 2 64.94 63.64 66.67 data. The authors are also grateful to the anonymous editor and
Average 58.25 59.14 65.12 reviewers for giving valuable comments. This work was sup-
ported by the Chinese Natural Science Foundation (60475001)
and Postdoctoral Science Foundation (2005038075).
best among SCSP, WCSP, and ACSP, at least on the
used data set. At the same time, this also manifests the
necessity and applicability of our adaptive feature References
extractor. 1. Chiappa S, Millán JR (2005) Data set V <metal imagery,
multi-class> [http://www.ida.first.fraunhofer.de/projects/bci/
competition_iii/ desc_V.html]. IDIAP Research Institute,
4 Discussions and conclusions Switzerland
2. Curran EA, Stokes MJ (2003) Learning to control brain
activity: a review of the production and control of EEG
In this paper, we propose the ACSP method for the
components for driving brain–computer interface (BCI)
feature extraction of EEG signals. Its efficacy and systems. Brain Cogn 51:326–336
superiority over the SCSP and WCSP methods are 3. Dornhege G, Blankertz B, Curio G, Müller KR (2004)
validated through classification experiments on multi- Boosting bit rates in noninvasive EEG single-trial classifi-
cations by feature combination and multiclass paradigms.
ple recording sessions of three subjects. Theoretically,
IEEE Trans Biomed Eng 51:993–1002
because EEG signals are time-varying, if we use SCSP 4. Duda RO, Hart PE, Stork DG (2000) Pattern classification.
to extract features on new EEG segments, the distri- Wiley, New York
bution of signal features would be transferred. In this 5. Galán F, Oliva F, Guàrdia J (2005) BCI competition III, data
set V: algorithm description [http://www.ida.first.fraunhofer.
sense, adaptive feature extraction has a sound basis.
de/projects/bci/competition_iii/results/ martigny/FerranGalan_
However, because WCSP treats all the EEG entries in desc.pdf]. Faculty of Psychology, University of Barcelona
computing covariance matrix equally, it cannot catch 6. Millán JR (2004) On the need for on-line learning in brain–
hold of the variability of EEG signals effectively. This computer interfaces. In: Proceedings of the international
joint conference on neural networks, Budapest, Hungary
is the reason why WCSP is inferior to ACSP which
7. Millán JR, Renkens F, Mouriño J, Gerstner W (2004) Brain-
considers the weighted problem of EEG entries actuated interaction. Artif Intell 159:241–259
advisably. 8. Müller-Gerking J, Pfurtscheller G, Flyvbjerg H (1999)
It should be noted that although ACSP has obtained Designing optimal spatial filters for single-trial EEG classi-
fication in a movement task. Clin Neurophysiol 110:787–798
the best results generally, pitfalls also exist. From
9. Nicolelis MAL (2001) Actions from thoughts. Nature
Table 1, we can find that there are several training and 409:403–407
test trails where ACSP does not show improvements in 10. Sun S, Zhang C (2005) Learning on-line classification via
the classification. We provide an explanation for this decorrelated LMS algorithm: application to brain–computer
interfaces. Lect Notes Comput Sci 3735:215–226
phenomenon. Though the subjects are doing mental
11. Sun S, Zhang C, Lu N (2005) On the on-line learning algo-
imagery tasks, there exist some irrelevant activities rithms for EEG signal classification in brain computer
from the brain. These activities also change over time, interfaces. Lect Notes Comput Sci 3614:638–647

123
Med Bio Eng Comput
12. Vapnik V (2000) The nature of statistical learning theory.
Springer, New York
13. Vaughan TM (2003) Guest editorial brain–computer inter-
face technology: a review of the second international meet-
ing. IEEE Trans Neural Syst Rehabil 11:94–109
14. Wolpaw JR, McFarland DJ, Neat GW, Forneris C (1991) An
EEG-based brain–computer interface for cursor control.
Electroencephalogr Clin Neurophysiol 78:252–259
15. Wolpaw JR, Birbaumer N, McFarland DJ, Pfurtscheller G,
Vaughan TM (2002) Brain–computer interfaces for commu-
nication and control. Clin Neurophysiol 113:767–791
16. Wolpaw JR, McFarland DJ (2004) Control of a two-dimen-
sional movement signal by a non-invasive brain–computer
interface in humans. Proc Natl Acad Sci 101:17849–17854
123