See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221712001

Application of Multiscale Amplitude Modulation Features and Fuzzy C-Means

to Brain-Computer Interface

Article  in  Clinical EEG and neuroscience: official journal of the EEG and Clinical Neuroscience Society (ENCS) · January 2012
DOI: 10.1177/1550059411429528 · Source: PubMed

CITATIONS READS

21 1,030

5 authors, including:

Wei-Yen Hsu Yu-Chuan Jack Li

National Chung Cheng University Taipei Medical University
21 PUBLICATIONS   879 CITATIONS    481 PUBLICATIONS   8,558 CITATIONS   

SEE PROFILE SEE PROFILE

Chien-Yeh Hsu Hung-Wen Chiu

Taipei Medical University Taipei Medical University
175 PUBLICATIONS   2,049 CITATIONS    116 PUBLICATIONS   1,565 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Head and Neck Cancer View project

Editorials View project

All content following this page was uploaded by Yu-Chuan Jack Li on 01 September 2014.

The user has requested enhancement of the downloaded file.

 Clinical EEG and Neuroscience
http://eeg.sagepub.com/

Application of Multiscale Amplitude Modulation Features and Fuzzy C-Means to Brain−Computer

Interface
Wei-Yen Hsu, Yu-Chuan Li, Chien-Yeh Hsu, Chien-Tsai Liu and Hung-Wen Chiu
Clin EEG Neurosci 2012 43: 32
DOI: 10.1177/1550059411429528

The online version of this article can be found at:

http://eeg.sagepub.com/content/43/1/32

Published by:

http://www.sagepublications.com

On behalf of:

EEG and Clinical Neuroscience Society

Additional services and information for Clinical EEG and Neuroscience can be found at:

Email Alerts: http://eeg.sagepub.com/cgi/alerts

Subscriptions: http://eeg.sagepub.com/subscriptions

Reprints: http://www.sagepub.com/journalsReprints.nav

Permissions: http://www.sagepub.com/journalsPermissions.nav

>> Version of Record - Jan 1, 2012

What is This?

Downloaded from eeg.sagepub.com by guest on October 11, 2013

 Clinical EEG and Neuroscience
43(1) 32-38
Application of Multiscale Amplitude ª EEG and Clinical Neuroscience
Society (ECNS) 2012
Reprints and permission:
Modulation Features and Fuzzy C-Means sagepub.com/journalsPermissions.nav
DOI: 10.1177/1550059411429528
to Brain–Computer Interface http://eeg.sagepub.com

Wei-Yen Hsu1, Yu-Chuan Li1, Chien-Yeh Hsu1,

Chien-Tsai Liu1, and Hung-Wen Chiu1

Abstract
This study proposed a recognized system for electroencephalogram (EEG) data classification. In addition to the wavelet-based
amplitude modulation (AM) features, the fuzzy c-means (FCM) clustering is used for the discriminant of left finger lifting and
resting. The features are extracted from discrete wavelet transform (DWT) data with the AM method. The FCM is then
applied to recognize extracted features. Compared with band power features, k-means clustering, and linear discriminant anal-
ysis (LDA) classifier, the results indicate that the proposed method is satisfactory in applications of brain–computer interface
(BCI).

Keywords
amplitude modulation, brain–computer interface, discrete wavelet transform, electroencephalography, fuzzy c-means (FCM)

Received January 31, 2011; accepted June 3, 2011.

Introduction use Daubechies DWT to decompose EEG signals into sub-

bands and then to obtain multiresolution envelope features
BCI provides an alternative communication channel by directly
by AM methods. In addition to multiscale characteristics, the
transmitting messages to computers from the human brain by
features also contain important envelope information.
analyzing the brain’s mental activities.1-5 BCI systems based
EEG data classified by supervised classifiers have been
on single trial EEG signal analysis associated with finger applied in many BCI systems. Parameters need to be trained
movement have grown rapidly in the last decade.3 It focuses
before the applications. FCM clustering, an unsupervised
on discriminating finger movements for EEG analysis using
approach, partitions a collection of feature vectors into a num-
event-related brain potentials (ERPs). It shows that there are spe-
ber of subgroups based on minimizing the trace of a within-
cial characteristics of event-related desynchronization (ERD)
cluster scatter matrix.14,15 The EEG data are nonstationary and
and synchronization (ERS) in mu and beta rhythms over the sen-
their characteristics vary with time. Unsupervised FCM clus-
sorimotor cortices during mental tasks.6,7
tering may be better than conventional supervised classifiers
Feature extraction is an important topic that greatly affects
in BCI applications since the former does not need training.
the classification rate. All kinds of feature extraction methods Several properties of FCM clustering are described. FCM can
have been proposed. Among them, the band power and adap-
discriminate the data without a priori knowledge, it is able to
tive autoregressive (AAR) parameter models are popular.8-11
make flexible partitions of a finite data set, and it is an adap-
Frequency components in alpha and beta bands are predomi-
tive approach able to recognize nonstationary biomedical sig-
nately involved in mental tasks,12 so feature extraction based
nals, such as EEG data.16 For these reasons, in this study,
on band power estimation is usually obtained by computing
EEG data were recognized by FCM clustering.
their powers at these bands. The estimated band powers are
then computed with their logarithm values as descriptive para-
meters for every channel,8 or estimated by averaging them.9
On the other hand, AAR models are also common in feature 1
Graduate Institute of Biomedical Informatics, Taipei Medical University,
extraction of mental tasks.10,11 EEG signals are modeled as fil- Taiwan
tered white noise with certain preferred energy bands by the
Corresponding Author:
all-pole AAR model. The AAR model can be intuitively
Wei-Yen Hsu, Graduate Institute of Biomedical Informatics, Taipei Medical
rephrased in the frequency domain as white noise source driv- University, 250 Wu-Xin Street, Taipei 110, Taiwan
ing an all-pole spectral shaping network.13 In this study, we Email: shenswy@stat.sinica.edu.tw; shen@csie.ncku.edu.tw
Hsu et al. 33

Figure 1. Flowchart of proposed BCI system. The system mainly consists of feature extraction and FCM clustering.

To assess the performance of the proposed system, band

power features, k-means clustering and linear discriminant
analysis (LDA) classifier were compared. Results show satis-
factory classification accuracy, and area under the receiver
operating characteristics curve (AUC) for the proposed
wavelet-based AM features and FCM clustering. It shows that
there is great potential for the proposed methods in BCI
applications.

Materials and Methods

A flowchart of the EEG analysis for the single-trial discriminant
is illustrated in Figure 1. The procedure consists of feature
extraction and FCM clustering. We extract features with the
AM method from the DWT data, and the FCM is then applied
for the discriminant.

Data Acquisition
The EEG signals were recorded from 7 untrained partici-
pants (6 males and 1 female, 3 left-handed and 4 right-
handed) in a shielded room using 7 silver/silver chloride
electrodes. As illustrated in Figure 2, they included 5 scalp
EEG channels (C3, C5, FC3, C1, and CP3. These five
channels were selected since they can be further processed
by surface Laplacian filtering as discussed in the next
section. One EMG channel for monitoring right muscle
activity, and 1 channel on the forehead to record possible
electro-oculogram (EOG) artifacts and eye blinks during
the experiment.17 All electrodes were referenced to the
A1 lead at the left earlobe. Before being sampled at the rate
of 256 Hz, the EEG data were filtered by an analog band-
pass filter with cutoff frequencies at 0.5 Hz and 100 Hz, and
amplified by a multiple of 10 000. During the experiments,
each participant was asked to perform 2 trials in each test,
which included left finger lifting and resting. Each trial was
10 seconds in length, so it took 20 seconds to complete a test.
In each lifting trial, the first 4 seconds were quiet and then an
acoustic stimulus was given as a cue to signify the beginning
of left finger lifting. At the same time, each participant was
asked to execute finger lifting. An example of a test is shown
in Figure 3A. We recorded 60 tests for each participant, and
thus there were 120 trials for each participant. No trials were
removed during the EEG data processing stage. Data segments
Figure 2. Electroencephalogram (EEG) electrode placement
according to International 10-20 system. All electrodes are referenced
to the position A1 at the left earlobe.
for finger lifting were acquired from 2 to 2 seconds, where
second 0 stands for the trigger of movement by detecting the
peak EMG signal after linear envelope processing. (Only the
data recorded between 2 and 2 seconds were considered to
be event related.) Data segments for finger resting were
randomly acquired from 4-second windows within the finger
resting sections in each trial.
Preprocessing
An appropriate filtering method can increase the signal-to-
noise ratio by reducing non-EEG noise. The surface
Laplacian filter is a simple but effective filtering method.18
It calculates the second derivative of the spatial voltage
distribution for a selected electrode. It is a high-pass spatial
filter that enhances localized activities and reduces back-
ground noise. This filter is achieved by subtracting the
average potential of a set of surrounding electrodes from the
electrode of interest,
34 Clinical EEG and Neuroscience 43(1)

Figure 3. Preprocessing with surface Laplacian filtering. A, Raw EEG signals. B, Electroencephalograph signals posterior to the preprocessing.
(The first channel in (B) is the result of surface Laplacian filtering performed from the first 5 channels in (a); the last 2 channels in (A) and (B) are
the same, and they are EMG and EOG signals, respectively.) EMG indicates electromyography; EOG, electro-oculogram.

surrounding the ith electrode and the number of surrounding

1X electrodes, respectively. The distance between the selected elec-
ViLap ¼ Vi  Vj ð1Þ trode and its surrounding electrodes demonstrates the
N j2S
i characteristic of surface Laplacian filtering (the greater the dis-
where Vi represents the potential between the ith electrode and tance, the greater the insensitivity to highly localized potentials).
the reference A1, and Si and N stand for the set of electrodes As illustrated in Figure 3B, the electrode of interest to each of its
Hsu et al.
surrounding electrodes performed the surface Laplacian filtering
is shown as follows:
Lap 1
VC3 ¼ VC3  ðVC5 þ VFC3 þ VC1 þ VCP3 Þ ð2Þ
4 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
The 4-second window for left finger lifting was acquired
from the finger-lifting trial with its center standing for the trig-
ger of movement, while the 4-second window for finger resting
was randomly acquired from the finger-resting trial.
Feature Extraction
Wavelet transform
Before classification, feature extraction is performed
on the event-related windows rather than directly classifying
the native EEG data without feature extraction. In this study,
we first band-pass filtered the windows to the wide frequency
range that contains all mu and beta rhythmic components
using a Butterworth band-pass filter. We then performed
the Daubechies discrete wavelet transform (DWT)19 on the fil-
tered windows. We chose the Daubechies wavelet due to the spe-
cial characteristic that Daubechies family wavelets are
compactly supported with extreme phase and highest number
of vanishing moments for a given support width.
Multiresolution analysis decomposed a signal into
numerous details at various resolutions, where each resolution
represented a class of distinct physical characteristics within
the signal. More specifically, a signal was characterized with
the formulation by decomposing it into sub-bands, each of
which could be treated individually based on its characteris-
tics. Multiresolution representation of filtered event-related
windows was achieved by Daubechies DWT. The event-
related window W for each trial is represented in terms of the
DWT as
X
1  
W ð xÞ ¼ SJ ðk Þ2J =2 j 2J x  k
k¼1
X
J X
1
ð3Þ
 
þ Dj ðkÞ2 c 2j x  k
j=2
j¼1 k¼1
where the expansion coefficients are determined by
  
SJ ðk Þ ¼ W ðxÞ; 2J =2 j2J x  k
Dj ðkÞ ¼ W ðxÞ; 2j=2 c 2j x  k ð4Þ
where SJ ðk Þ and Dj ðkÞ represent the approximation and detail
spaces of W, respectively, and 2J =2 jð2J x  k Þ and
2j=2 cð2j x  k Þ denote the dilated and translated versions of the
scaling function jðxÞ and wavelet function fðxÞ, respectively.
The event-related window W is then divided into individual sub-
bands SJ , DJ , . . . , and D1 , where the value of J is chosen as 3 in
the experiment.
Amplitude Modulation
The amplitude modulation (AM) method based on the Hil-
bert transform was applied to obtain the envelopes of all the
35
sub-bands and quantify event-related oscillatory activities.20
Each envelope or AM waveform was calculated as
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ASJ ðkÞ ¼ SJ ðkÞ2 þ HðSJ ðkÞÞ2
ADj ðkÞ ¼ Dj ðkÞ2 þ HðDj ðkÞÞ2 ð5Þ
where HðxÞ was in the Hilbert transform of single-trial EEG
signal x.
Fuzzy c-means Clustering
Clustering is a process for classifying training samples in
such a way that samples within a cluster are more similar
to one another than samples belonging to different clusters.
The fuzzy clustering method21,22 assigns the sample with a
number between 0 and 1 described as a membership function.
In this article, fuzzy clustering method classified feature vec-
tors extracted from the original EEG data and recognized
complicated brain mental tasks, such as left finger lifting and
resting.
The FCM approach minimizes the least squared error of an
objective function. For class number c (c  2), sample number
n, and fuzzification parameter mð1 P m < 1Þ, the algorithm
d
chooses ui : X ! ½0; 1 so that i ui ¼ 1 and wj 2 R for
j ¼ 1; 2; :::; c to minimize the objective function:
1X c X n  m  2
JFCM ¼ ui;j xi  wj  ; ð6Þ
2 j¼1 i¼1
where ui;j is the value of jth membership grade on ith sample xi .
The cluster centroids w1 ; :::; wj ; :::; wc can be regarded as
prototypes for the clusters represented by the membership
grades. For the purpose of minimizing the objective function,
the cluster centroids and membership grades are chosen so that
a high degree of membership occurs for samples close to the
corresponding cluster centroids.
Results and Discussion
Performance and Statistical Evaluation for Wavelet-
Based AM Features
We can expect higher classification accuracy if extracted
features are better. In this study, we combine DWT with
AM to extract multiscale envelope features. As the DWT
is often implemented by a pyramid-structured filter bank
that is very efficient in data representation, it is an efficient
and structured approach to ERP representation.23,25 More-
over, all kinds of feature extraction methods have been pro-
posed. The band power is common among them, so band
power features are compared with the wavelet-based
AM features.
The classification tests for EEG data are carried out using
5-fold cross validation in this study. More specifically, the data
set for each participant is divided into 5 subsets, and the follow-
ing procedure is repeated 5 times. Each time, 1 of the 5 subsets
is used as the test set and the other 4 are used as training set.
36 Clinical EEG and Neuroscience 43(1)

Table 1 Comparison of Classification Accuracy Between Band Power Table 2. Comparison of AUC Between Band Power and Wavelet-
and Wavelet-Based AM Features Under the FCM Clustering Based AM Features Under the FCM Clustering

Classification Accuracy Band Power Wavelet-Based AM AUC Band Power Wavelet-Based AM

S1 86.1% 84.7% S1 0.79 0.81

S2 88.3% 89.0% S2 0.84 0.82
S3 69.8% 83.3% S3 0.65 0.79
S4 80.5% 91.9% S4 0.80 0.89
S5 74.3% 75.4% S5 0.71 0.71
S6 53.2% 73.1% S6 0.55 0.70
S7 57.7% 74.2% S7 0.53 0.68
Average 72.8% 81.7% Average 0.70 0.77

Abbreviations: AM, amplitude modulation; FCM, fuzzy c-means. Abbreviations: AUC, area under the receiver operating characteristics curve;
AM, amplitude modulation; FCM, fuzzy c-means.

The average recognition rate is immediately evaluated across

all 5 folds. In addition to classification accuracy, the receiver Table 3. Comparison of Classification Accuracy Among LDA, k-
operating characteristics (ROC) curve is another popular Means, and FCM
approach to evaluate the performance of a method.26 An ROC
Classification Accuracy LDA k-Means FCM
curve is obtained by plotting all sensitivity values (true positive
fraction) on the y axis against their equivalent specificity val- S1 83.8% 84.1% 84.7%
ues (false positive fraction) on the x axis for all available S2 83.0% 85.2% 89.0%
thresholds of classified results of test data. The area under the S3 77.3% 80.6% 83.3%
ROC curve (AUC) is usually taken to be an important index S4 89.8% 86.3% 91.9%
S5 73.3% 72.8% 75.4%
because it provides a single measure of overall accuracy that
S6 62.7% 65.0% 73.1%
is not dependent upon a particular threshold. This study evalu- S7 67.4% 71.4% 74.2%
ates the performance of proposed system by means of both the Average 76.8% 77.9% 81.7%
classification accuracy and AUC.
An experiment is performed for evaluating the performance Abbreviations: LDA, linear discriminant analysis; FCM, fuzzy c-means.
of wavelet-based AM features. Tables 1 and 2 show the com-
parisons of classification accuracy and AUC between band Table 4. Comparison of AUC Among LDA, k-Means, and FCM
power and wavelet-based AM features under the FCM cluster-
ing. These 2 features are applied to the same clustering. In AUC LDA k-Means FCM
other words, the listed values demonstrate only the deviations S1 0.81 0.79 0.81
of performance between different features. Overall, the aver- S2 0.77 0.75 0.82
age recognition rate and AUC for band power features are S3 0.76 0.78 0.79
72.8% and 0.70, respectively, whereas using wavelet-based S4 0.87 0.82 0.89
AM as features yielded the better average rate (81.7%) and S5 0.69 0.69 0.71
AUC (0.77), the difference being 8.9% and 0.07, respectively. S6 0.63 0.63 0.70
In addition, we observe that the wavelet-based AM features S7 0.65 0.66 0.68
Average 0.74 0.73 0.77
give the best results for 6 participants, whereas band power
features lead to those for only 1 participant. It indicates that Abbreviations: AUC, area under the receiver operating characteristics curve;
the wavelet-based AM features can obtain the better perfor- LDA, linear discriminant analysis; FCM, fuzzy c-means.
mance in BCI applications.
Moreover, 2-way analysis of variance (ANOVA) is per- study. The FCM algorithm is implemented to recognize left
formed to verify whether the comparisons between band power finger lifting and resting EEG data without any training in
and wavelet-based AM features are significantly different. The advance. As FCM clustering is an unsupervised approach, it
results indicate that the differences between them are signifi- is adopted to automatically classify wavelet-based AM features
cant in both classification accuracy (P value .0345) and AUC into 2 clusters, left finger lifting and resting. Some attractive
(P value .0362). That is, wavelet-based AM features are signif- properties of FCM clustering are summarized as follows: first,
icantly better than band power features in BCI applications. FCM clustering is an unsupervised method, discriminating the
data without labeling them before training; second, FCM clus-
tering is capable of making flexible partitions of a finite data
Performance and Statistical Evaluation for FCM
set; and third, FCM clustering is a robust approach suitable for
Clustering the classification of nonstationary biomedical signals. There-
An unsupervised FCM method is proposed for left finger lifting fore, we use FCM clustering for the classification of EEG data
and resting classification from single-trial EEG data in this in this study.
Hsu et al.
We evaluate the performance of FCM clustering. Tables 3
and 4 list the comparisons of classification accuracy and
AUC under the same AM features among LDA, k-means, and
FCM. The results show that the average accuracy for LDA is
the worst, whereas that for FCM is the best. The average
AUC for FCM is the best, while that for k-means is the
worst. Classification accuracy and AUC with FCM obtains
the best results for all participants. Accordingly, the results
denote that the FCM is better than the LDA and k-means
in EEG classification.
In addition, 2-way ANOVA and multiple comparison tests
determined whether these 3 classifiers are significantly dif-
ferent. The results reveal that there are significant differences
in both classification accuracy and AUC: ‘‘between LDA and
FCM’’ (P values .0084 and .0105) and ‘‘between k-means
and FCM’’ (P values .0066 and .0096). Hence, the FCM is the
best of these 3 classifiers in BCI applications.
Conclusion
We have proposed an analysis system for single-trial EEG
classification. The system that contains wavelet-based AM
features and the FCM clustering is present for the discrimi-
nant of left finger lifting and resting. The features are
extracted by AM method from the DWT data. They are good
features that can improve the classification of mental tasks.
The FCM is then used for the recognition of EEG data.
Results indicate that the proposed methods are better in clas-
sification accuracy and AUC in comparison with band power
features, k-means, and LDA classifier. We will develop some
more powerful features and classifiers to further enhance the
performance.
Acknowledgments
The author would like to express his sincere appreciation for grant
under shared facilities supported by the Program of Top 100 Univer-
sities Advancement, Ministry of Education, Taiwan.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to
the research, authorship, and/or publication of this article.
Funding
The authors received no financial support for the research, authorship,
and/or publication of this article.
References
1. Nicolaou N, Georgiou J. The use of permutation entropy to
characterize sleep electroencephalograms. Clin EEG Neurosci.
2011;42(1):24–28.
2. Xu J, Zhao SZ, Zhang HS, Zheng CX. Decreased delta event-
related synchronization in patients with early vascular dementia.
Clin EEG Neurosci. 2011;42(1):53–58.
37
3. Parra L, Alvino C, Tang AC, et al. Linear spatial integration for
single trial detection in encephalography. Neuroimage. 2002;
7(1):223–230.
4. Ahmadlou M, Adeli H. Fuzzy synchronization likelihood with
application to attention-deficit/hyperactivity disorder. Clin EEG
Neurosci. 2011;42(1):6–13.
5. Hsu WY, Sun YN. EEG-based motor imagery analysis using
weighted wavelet transform features. J Neurosci Methods. 2009;
167(2):310–318.
6. Muller-Putz GR, Zimmermann D, Graimann B, Nestinger K, Kori-
sek G, Pfurtscheller G. Event-related beta EEG-changes during pas-
siveandattemptedfootmovementsinparaplegicpatients.BrainRes.
2007;1137(1):84–91.
7. Hsu WY. Continuous EEG signal analysis for asynchronous BCI
application. Int J Neural Sys. 2011;21(4):335–350.
8. Obermaier B, Neuper C, Guger C, Pfurtscheller G. Information
transfer rate in a five-classes brain-computer interface. IEEE
Trans Neural Sys Rehabil Eng. 2001;9(3):283–288.
9. Guger C, Edlinger G, Harkam W, Niedermayer I, Pfurtscheller G.
How many people are able to operate an EEG-based brain-
computer interface (BCI)?. IEEE Trans Neural Sys Rehabil Eng.
2003;11(2):145–147.
10. Burke DP, Kelly SP, Chazal deP, Reilly RB, Finucane C. A
parametric feature extraction and classification strategy for
brain-computer interfacing. IEEE Trans Neural Sys Rehabil
Eng. 2005;13(1):12–17.
11. Hsu WY. EEG-based motor imagery classification using neuro-
fuzzy prediction and wavelet fractal features. J Neurosci Methods.
2010;189(2):295–302.
12. Hsu WY. EEG-based motor imagery classification using
enhanced active segment selection and adaptive classifier. Com-
put Biol Med. 2011;41(8):633–639.
13. Kelly S, Burke D, Chazal deP, Reilly R. Parametric models and
spectral analysis for classification in brain-computer interfaces.
Proc 14th Int Digital Signal Processing. 2002;1:307–310.
14. Ismail MA, Selim SZ. Fuzzy c-mean: optimality of solutions and
effective termination of the algorithm. Pattern Recogn. 1986;
19(6):481–485.
15. Hsu WY. Analytic differential approach for robust registration of
rat brain histological images. Microscopy Res Tech. 2011;74(6):
523–530.
16. Hsu WY, Kuo WF. Unsupervised fuzzy c-means clustering for
motor imagery EEG recognition. Int J Innov Comput Inf Control.
2011;7(8):4965–4976.
17. Jasper H. The ten-twenty electrode system of the international
federation. Electroenceph Clin Neurophysiol. 1958;10(2):371–375.
18. McFarland DJ, McCane LM, David SV, Wolpaw JR. Spatial fil-
ter selection for EEG-based communication. Electroencepha-
logr Clin Neurophysiol. 1997;103(3):386–394.
19. Daubechies I. Ten Lectures on Wavelets. CBMS-NSF Lecture
Notes nr. 1992; 61 SIAM.
20. Clochon P, Fontbonne JM, Lebrun N, Etevenon P. A new method for
quantifyingEEGevent-relateddesynchronization:amplitudeenvel-
ope analysis. Electroencephalogr Clin Neurophysiol. 1996;98(2):
126–129.
38
21. Dunn JC. A fuzzy relative of the ISODATA process and its use in
detecting compact well-separated clusters. J Cybern. 1974;3(3):
32–57.
22. Bezdek JC. Fuzzy Mathematics in Pattern Classification. Ph.D.
Dissertation, Applied Mathematics. New York, NY: Cornell
University, Ithaca; 1973.
23. Demiralp T, Yordanova J, Kolev V, Ademoglu A, Devrim M,
Samar VJ. Time-frequency analysis of single-sweep event-
related potentials by means of fast wavelet transform. Brain Lang.
1999;66(1):129–145.
View publication stats
Clinical EEG and Neuroscience 43(1)
24. Hsu WY, Lin CC, Ju MS, Sun YN. Wavelet-based fractal fea-
tures with active segment selection: application to single-trial
EEG data. J Neurosci Methods. 2007;163(1):145–160.
25. Hsu WY, Poon WFP, Sun YN. Automatic seamless mosaicing of
microscopic images: enhancing appearance with color degrada-
tion compensation and wavelet-based blending. J Microscopy-
Oxford. 2008;231(3):408–418.
26. Hanley JA, McNeil BJ. The meaning and use of the area under
a receiver operating characteristic (ROC) curve. Radiology.
1982;143(1):29–36.