| Brain-Computer Interface (BCI) provides a pathway to communicate between human brain and machine environment and translates human thoughts into equivalent machine activity. The research on human BCI was initiated in 1978-79 and was developing very actively in later days. The original premise of BCI was to provide patients suffering from motor disabilities a platform to perform all activities that a regular person does via thoughts. Here, Electroencephalogram (EEG) might be viewed as the language of human brain that needs to be appropriately interpreted. EEG is a non-invasive technique recording the electrical activity of human brain as a function of time, frequently measured in terms of so called Event Related Potential (ERP) and acquired from the surface of human head. EEG contains the information related to different brain states and thought processes and, therefore, may be used as the basic input for all recent BCI prototypes. On the other hand, EEG varies notably from subject to subject even during the same thinking cycle. Different EEG patterns are observed during different thought processes.;In present work, the objective was, therefore, to classify the pattern for a particular brain activity. Motor imagery movements of right and left hand in different directions were taken into consideration rather than actual physical movement. A number of modern signal processing algorithms and mathematical methods, such as Principle Component Analysis (PCA), Linear Discriminant Analysis (LDA), Kozinec's Method, Perceptron, and Probabilistic Neural Networks (PNN), were thereafter used to classify Motor Imagery EEG. In the project, key features of EEG data were considered while utilizing linear projection techniques. Linear discriminant functions were initially applied to classify EEG data in two broad classes: right hand imagery movement and left hand imagery movement. However, the classification accuracy was observed as unsatisfactory (51.9%). Probabilistic Neural Network (PNN) was implemented next. PNNs exhibit fast training and lower complexity, and yielded the accuracy of 73.076% for classification of two patterns. Classification of three classes, i.e. right hand up, hold and down movement, was attempted next, yielding 71.795% accuracy. Similar classification for the left hand yielded 82.025% accuracy, although the accuracy of classification for 6 classes (three for right hand and three for left hand) was reduced to 69.24%.;Based on our observations, we conclude that PNN should be provided with the maximum amount of training data while attempting to discriminate among multiple classes. |
