Pattern
recognition and classification under MATLAB
 What is g.CLASSIFYtoolbox for g.BSanalyze?
g.CLASSIFYtoolbox is a software package for linear
and non-linear data set classification. — Classification enables to
categorize patterns and signal features of biosignals into different
classes.
In the learning phase of supervised learning, classifiers
are trained with signal patterns and signal features with known class
information. After the training phase the classifier is ready to categorize
new signal patterns into the specific classes.
One application of classification is to discriminate EEG
patterns in brain-computer-interface experiments (e.g. into LEFT and RIGHT
hand motor imageries). Optimized feature selection can be performed via
Distinction Sensitive Learning Vector Quantization (DSLVQ). This method
yields the most relevant features for optimal data set classification.
The graphical user interface for g.CLASSIFYtoolbox
enables you to classify signal patterns of whole trials, within specified
time segments or at certain time points. Signal features are combined in a
so-called "feature matrix". This matrix is the basis for training of all
linear and non-linear classification methods, among them a multi-class
linear discriminant analysis, radial basis function neural network or
multi-layer perceptron neural network. You can load and save your preferred
processing steps as a script program and automatically process your data in
batch mode. Toolbox highlights
- Generate feature matrix
- Multi-class linear discriminant analysis
- Minimum distance classifier
- Backpropagation neural network
- Reveiver operator curve
- Radial basis function neural network
- DSLVQ (distinction sensitive learning vector quantization)
- DSLVQ for feature weighting
- KMEANS clustering
Example: Classification of EEG alpha and beta band
patterns
 Feature extraction
EEG during right hand and
foot movement imagery was recorded over electrode positions C3 and Cz.
The type of movement imagery was given via experimental instructions on the
computer screen. A total of 160 imaginations of movements were performed (80
hand and 80 foot movements).
The first step in data set classification is to define features
best describing the signal patterns. In this case the band power features in
the ALPHA and BETA bands are computed for the 2 EEG channels. The
computation yields 2 feature channels per EEG channel.
 Feature matrix generation
The second step is to generate learning examples at
certain time points for the training phase of the used classifier. The
classifier also needs - in supervised learning - the class information, i.e.
to which type of movement a certain EEG pattern is related. The following
steps explain how-to select the time points, the class information and to
select an appropriate classifier.
- The classification interval is specified between 1000 and 6000 ms
- EEG patterns for RIGHT and FOOT movement imagery classes are
selected for classification
- The ALPHA and BETA band power are selected as feature channels
- The classification method Linear discriminant analysis is selected
 Classification results
The third step is to classify the
features by the chosen classification method. As a result of the
classification an error curve is computed yielding information about the
successful or unsuccessful classification as function of time. The result
below is based on a 10 times 10 fold cross validation of the input data.
This means that 90% of the input examples are chosen randomly for training
of the classifier. The classification of the data is done on the remaining
10% of the examples. This yields the first classification results. In a
second run another 90% of the data are randomly selected and a second
classification on the remaining 10% of examples is performed.
This procedure is repeated 10 times yielding an average error curve.The selected features and class information
define now the feature matrix which is the basis for classification via a
selected classification method. The classification error (figure to the
right) for the movement imagery experiment displays that at the beginning of
the imagination the classification error is around 50 %. At second 7 the
error drops down to about 13 %. Therefore RIGHT and FOOT movement imagery
patterns can be discriminated successfully.
 DSLVQ Feature Weighting for most relevant feature
identification
g.CLASSIFYtoolbox can perfom a fourth step in data set
classification!
In order to identify the most relevant feature contributing to the
discrimination task, the method of Distinction Sensitive Learning Vector
Quantization (DSLVQ) can be applied. The linear vector quantizer classifier
is initially trained on all features. Through implicit feature relevance
analysis the system finds non-relevant features and adapts the weighting of
the features accordingly. For the RIGHT and FOOT
movement imagery experiment a total of 4 feature channel were created. For
channels C3 and Cz the band power in the ALPHA and BETA bands were computed.
The graphs to the right yield an overview about the contribution of the
individual features to the classification result.
At second 1 feature 1 is the most important one, followed by 4, 2 and 3.
But the classification error is 51,75 % and therefore the result is random.
At second 7 the error is 11,75 % and therefore the feature weighting can
be considered as reliable. Therefore, feature 4 (bandpower in the beta range
of channel 2) is the most important one. The bar of feature 3 is much
smaller but the feature can still be considered as important for the
classification task. Features 1 and 2 are not important and should not be
considered for the discrimination Package includes
- Software modules
- help
manual
- hardlock
Technical Requirements
MATLAB, g.BSanalyze base versionCopyright © g.tec
|
