**Figure 4**.

*Brain-Computer Interface: Use of Electroencephalogram in Neuro-Rehabilitation DOI: http://dx.doi.org/10.5772/intechopen.110162*

**Figure 4.**

*The architecture of a brain computer system consists of signal acquisition from the brain, pre-processing, feature extraction, feature selection, classification and eventually application to external devices that provides feedback to users.*

#### *2.2.1 Data acquisition and pre-processing*

After the raw EEG data has been detected via the scalp electrodes, the data needs to remove any signals originated in areas other than the brain such as using the 60 Hz notch to clean the intervening frequency and the EMG activity before further analysis [1, 3]. After this pre-processing step, signal processing is performed using many different feature extraction techniques to identify specific brain signals that would later be translated to system commands [5].

#### *2.2.2 Feature extraction*

There are numerous techniques that enable proper signals to be retrieved during feature extraction. We will not go into detail into each of the techniques, but some common ones are briefly discussed in this section. Time-domain and frequencydomain are two basic techniques often applied in studies. Using quaternions to represent objects within a three-dimensional space offers a better method to aid in extracting signals in time-domain analysis especially from motor imagery EEG. Fast Fourier transform theory and local characteristic-scale decomposition are approaches that are often utilized in frequency-domain analysis. In order to relate the frequency content to the temporal domain and vice versa, time-frequency domain analysis helps compensate each other's deficit in decomposing signals in a more dynamic fashion. Common spatial pattern (CSP) is advantageous in motor imagery EEG processing as it can extract particular information from a particular frequency band. Different modifications of CSP are available, and sub-band common spatial pattern offers a much better classification accuracy by initially filtering EEG at different sub-bands and then tabulating CSP features for each of the bands [4, 5].

#### *2.2.3 Feature selection and classification*

The most common feature selection techniques include principal component analysis (PCA), filter bank selection and evolutionary algorithms. PCA helps to reduce dimensionality, while filter bank selection is specific for CSP extraction technique. Due to the high computational demands and large size feature set, evolutionary algorithm can further select a more appropriate feature by hybrid approach so to improve accuracy at the cost of time [4, 5]. For classification and modeling of the control system, linear discriminant analysis (LDA), support vector machines (SVM) and artificial neural networks (ANN) are the frequently used classifiers [6]. LDA is a linear classifier that is simple to use but it may not be good enough to process non-linear EEG data. SVM is a non-linear classifier that handles well with high dimensionality data; however, it takes more time for processing. ANN is another non-linear classifier that requires long handling time to process large computational data. It is known to be highly adaptive but also over-fitting; therefore, it may fail to predict future observations reliably [6]. Eventually, neurofeedback system relays back to the users so they can make modification in their brain patterns and improve the system.
