Multi-class classification of EEG in a brain-computer interface

The brain-computer interface (BCI) has drawn much interest for its broad potential in clinical applications, to restore motor control and communication ability to the disabled. Using electroencephalography (EEG) to record brain activity, data collected can be used to train classifiers for predicting an output. The objective of this project was to investigate the performance of multi-class classification on an EEG-based BCI, by developing a user interface for conducting experiments and data acquisition, building linear-discriminant analysis classifiers trained on the data, and evaluating the classifiers’ performance with k-fold cross validation. Average validation accuracy of 31.4% and 45% were obtained for four-class classification, using feature selection by the top 10 individual features and sequentially searching for the best combination of features respectively. Binary classification of different combinations of two classes achieved 54% and 71.4% average validation accuracy using the same two methods of feature selection. EEG frequency bands delta and alpha were found to be more commonly selected as the best features for four-class classification. For binary classification of up v.s. left, the delta and theta bands comprised a larger proportion of best features selected, and similarly with the alpha band for classification of up v.s. right.Bachelor of Engineering (Computer Science