Chapter 15: Classification of Time

Report
Chapter 15: Classification of TimeEmbedded EEG Using Short-Time
Principal Component Analysis
by Nguyen Duc Thang
5/2009
Outline

Part one






Introduction
Principal Component Analysis (PCA)
Signal Fraction Analysis (SFA)
EEG signal representation
Short time PCA
Part two


Classifier
Experimental setups, results, and analysis
Outline

Overview the previous presentations




Common Spatial Patterns (CSP)
Classifiers
Experimental setups and results
Analysis, discussion, and conclusions
An architecture of EEG-based BCI
system
PCA, SFA,
Short time PCA
LDA, SVM
Feature extraction
Classification
The shortcomings of conventional PCA
projection line
Not good for large number
of samples
Short time PCA approach
Apply PCA on
short durations
Extract short time PCA features
PCA
stack
D
h
Time-embedded
features
D
window
h
n basic
vectors
D
1 X Dn
n
Short time PCA
features
The role of Singular Value Decomposition
(SVD) in PCA
w1
w2
Using SVD, we can
compute the eigenvector
w of covariance matrix Cx
(maximize variance)
T
maxw w Cx w
Generalized SVD
Maximize
variance
w
A
Using GSVD, we can find
generalized eigenvector w that
maximizes the variance when
projecting data A into w and
minimizes the variance when
projecting data B into w
maxw wT Ax w
Minimize
variance
B
minw wT B y w
Common Spatial Pattern

For 2-classes:




Choose m eigenvectors, that maximize the
variance of class A and minimize the variance of
class B
Choose m eigenvectors, that maximize the
variance of class B and minimize the variance of
class A
The basic vectors W = total 2m eigenvectors
Examples: Distinguish left-hand movement
and right-hand movement
Common Spatial Pattern

For n-classes (Combine with classifier)

n-classes are converted to n(n-1)/2 2-classes
CSP
A
B
(AB), (BC), (CD), (DA), (AC), (BD)
D
C

New trials are assigned to the class for which
most classifiers (2-classes) are voting
Outline

Overview the previous presentations




Common Spatial Patterns
Classifiers
Experimental setups and results
Analysis, discussion, and conclusions
Linear Discriminant Analysis (LDA)

LDA is a simple classification approach in which
the samples from each class are modeled by
Gaussian distribution
1
k 
Nk
x
xC k
1

N K
K
T
(
x


)(
x


)

k
k
k 1 xC k
P( x | Class  k )  e
1
 ( x   k )T  1 ( x   k )
2
Linear discriminant boundary
P ( x | Class  i )  P( x | Class  j )
e
1
 ( x   i )T  1 ( x   i )
2
e
1
 ( x   j )T  1 ( x   j )
2
1 T 1
1 T 1
T 1
x  i  i  i  x   j   j   j
2
2
1 T 1
T
T 1
 ij ( x)  x  ( i   j )  ( i  i   j  1 j )
2
T
1
Boundary
Linear discriminant boundary
13
12
12
 23
13
Outline

Overview the previous presentations




Common Spatial Patterns
Classifiers
Experimental setups and results
Analysis, discussion, and conclusions
The parameters of EEG representations
l+1
r EEG
channels
 x1 (t ) 


...


 x (t  l ) 
 1

x(t )   ...  r  (l  1 ) dimensions
 xr (t ) 
 ...



 xr (t  l ) 


l: the numbers of lags
W=[w1,w2,…wf…wf+m…] → choose m basic vectors
f is first chosen basic vector
Time-embedded
features
window
s is window size
Cross-Validation Training Procedure

The training process




The training trials are randomly partitioned into
80% for constructing classifier and 20% for
evaluating
This partition and evaluation process is repeated
five times
The set of parameters getting best validation
performances are chosen
The testing process

Use the learned parameters to apply to test trials
Experiment 1: Five-task dataset



The subjects perform five mental tasks: (1)
resting task, (2) mental letter writing, (3)
mental multiplication of two multi-digit
number, (4) visual counting, and (5) visual
rotation
Each task is repeated five trials
6 electrodes are used: C3, C4, P3, P4, O1, O2,
record each trial 10s/250 Hz
Learning parameters
Confusion matrix for short-time PCA
representation averaged over test trials
Visualize the classification results



Given a set of samples X={x1,x2,…,xn} that belong to k class and
have dimension D>3. How to visualize X ?
For each class, apply K-means clustering to find N cluster points
(center of cluster). We have a total K x N points
Using Multidimensional Scale (MDS) to map points in D dimension
to d ≤3 (Preserve distance between points)
Visualize the classification results
Three-task dataset




Three subjects perform 3 tasks: imagine left
hand movement, imagine right hand
movement, and generate words
The subjects perform given task 15s, then
switch to another task at the operator’s
request
There are three training dataset and one test
set of data
EEG signal are recorded at 512 Hz using 32
electrodes
Short-time PCA procedure for three-task
dataset



Bandpass-filtered data 8-30 Hz
Down-sampled to 128 Hz
The best parameters from the learning process are
given
Subject
Number
of lags
First
vector
Number
of
vectors
1
2
1
5
2
2
1
4
3
3
1
5
The other methods


S. Sun et al. remove 7 electrodes, bandpass filtered
8-13 Hz (subjects 1-2),11-15 Hz (subject 3) . Multiclass CSP for extract features and SVM for
classification
Schlögl et al. downsampled to 128 Hz, extract all
bipolar channels (496) +32 monopolar channels.
Each channels extract the features: AR (order=3),
Bandpower in α and β bands. LDA are used as
classifier
The other methods (cont.)


Arbabi et al. downsampled to 128 Hz, filtered to 0.545 Hz. Using some statistical features and Bayesian
classifier.
Salehi use all raw data, features: PSD and some
statistical time domain features (not mentioned).
Bayesian classifier
Comparison results
Visualize the classification results
Outline

Overview the previous presentations




Common Spatial Patterns
Classifiers
Experimental setups and results
Analysis, discussion, and conclusions
Improve the classifier performance by
smoothing

Many incorrect classification appear as single
samples

If n continuous samples have same class, we
can decide the majority class
With smoothing, the accuracy is improved from
78.7% to 82.7% (five-task dataset)

Analyze the parameters of EEG
representation

Number of lags = 2-3, window size 125, the first basic vector should
be early in order, number of basic = 20, subtract mean has minor
effects
Analyze the importance of electrodes



The weights of the discriminant functions are summarized
corresponding to each electrode
The variances of the weights grouped by this way are plotted
The parietal electrodes are most important for mental task
discrimination
Conclusion




This chapter describes a new approach of extracting
features from EEG signals by using short-time PCA
For five-mental dataset, combining short-time PCA
with simple classifier LDA, this approach achieves
80% accuracy
On three-task dataset, this approach places second
among five compared methods
Some analysis about the parameters of the system
and the roles of electrodes are also given

similar documents