(IJCSIS) International Journal of Computer Science and Information Security,Vol. 11, No. 5, May 2013
desired components are needed to be extracted for furthercontrolling of devices.
In autoregressive (AR) techniques, a model iscreated where a current voltage can be predicted from N pastvoltages where the model order is N . Thus the model canbe represented as:
it xat x
is the ith order AR coefficient for electrode e.These AR coefficients can be used as features. To obtain thesecoefficients, EEG data is generally windowed into blocks of data with more than N samples. Then, as the value of t isshifted through the window of data, we obtain numerous modelequations which allow us to compute optimum AR coefficients.Thus, these AR coefficients can be used to represent the mentalstate during that window of time.
To date, little has been published usingwavelets as a feature extraction method for a BCI system.However, they have been used in a variety of other EEGpattern recognition work [50, 51] including neural networks[52,53]. Wavelets are essentially a compromise between time-domain and frequency-domain since they allow the user toview change in frequency bands over time (with less resolutionthan just time-domain or frequency-domain). The DiscreteWavelet Transform (DWT) can be computed as a series of filters. To date, little has been published using wavelets as afeature extraction method for a BCI system. However, theyhave been used in a variety of other EEG pattern recognitionwork, including neural networks. Wavelets are essentially acompromise between time-domain and frequency-domain sincethey allow the user to view change in frequency bands overtime (with less resolution than just time-domain or frequency-domain).
Common Spatial Filter:
Common spatial patterns (CSP)method was firstly suggested for classification of multi-channelEEG during imagery hand movements by Ramoser et al..The main idea is to use a linear transform to project the multi-channel EEG data into a low-dimensional spatial subspace witha projection matrix, of which each row consists of weights forchannels. This transformation can maximize the variance of two-class signal matrices. CSP method is based on thesimultaneous diagonalization of the covariance matrices of both classes.
Experimental results suggested that ICA is a usefuland feasible method for spatial filtering and feature extractionin motor imagery based multi-class BCIs. When using EEGrecordings as the input signals of a BCI system, the researchermay face a problem of extracting features used forclassification in the presence of artifacts such aselectrooculogram (EOG) or electromyogram (EMG). Theamplitude of the disturbances may be higher than that of brainsignals. This requires an efficient method to separate brainsignals from artifacts. ICA happens to be a suitable approach tocarry out the separation. This approach is based on theassumption that the brain activity and the artifacts areanatomically and physiologically separate processes, and thisseparation is reflected in the statistical independence betweenthe electrical signals generated by those processes .
The original EEG signals potentials recorded from the scalpare very complex so they are needed to be processed anddesired components are needed to be extracted for furthercontrolling of devices.
: LD classifier is one of the linear classificationmethods that require fewer examples in order to obtain areliable classifier output  It is also a simpler andcomputationally attractive as compared to other classifiers
LDwas used to classify different combinations of mental.
: An SVM also uses a discriminant hyper plane toidentify classes. However, concerning SVM, the selectedhyper plane is the one that maximizes the margins, i.e., thedistance from the nearest training points. Maximizing themargins is known to increase the generalizationcapabilities. As RFLDA, an SVM uses a regularizationparameter C that enables accommodation to outliers andallows errors on the training set. Such an SVM enablesclassification using linear decision boundaries and is known aslinear SVM. This classifier has been applied, always withsuccess, to a relatively large number of synchronous BCIproblems[57,58]. However, it is possible to create nonlineardecision boundaries, with only a low increase of theclassifier’s complexity, by using the ‘kernel trick’. It consistsin implicitly mapping the data to another space, generally of much higher dimensionality, using a kernel function K(x, y).The kernel generally used in BCI research is the Gaussian orradial basis function (RBF).
Neural networks (NN) are, togetherwith linear classifiers, the category of classifiers mostlyused in BCI research. Let us recall that an NN is an assemblyof several artificial neurons which enables us to producenonlinear decision boundaries.
The k-nearest neighbor (k-NN)  is a classifierthat assigns the class label of a new data based on the classwith the most occurrences in a set of k nearest training datapoints usually computed using a distance measure such as theEuclidean distance.
An MLP is composed of severallayers of neurons: an input layer, possibly one or severalhidden layers and an output layer. Each neuron’s input isconnected with the output of the previous layer’s neuronswhereas the neurons of the output layer determine the class of the input feature factor.