EEG feature extraction for multiclass motor imagery task classification /|cby Aida Khorshid Talab

Abstract

The importance of Brain Machine Interface (BMI) calls for its continued improvement. BMI is a direct communication link between the brain and an external electronic device. BMIs aim to translate brain activities into control commands. Designing a system that translates brain waves into desired commands requires motor imagery task classification. Improvement of this translation not only depends on how capable the classifier is but also depends on the data fed to the classifier. Feature extraction highlights the properties of a signal that make it distinct from the signals of other mental tasks. The performance of BMIs directly depends on the effectiveness of the applied feature extraction and classification algorithms. If a feature provides a significant interclass difference for different classes, the applied classifier exhibits a better performance. In this work, to realize an interface between motor imagery task and an external device, a feature extraction method for Electroencephalogram (EEG) signal is introduced. This method uses signal dependent orthogonal transform based on the decomposition of linear prediction filter of impulse response matrix of the signal of interest. Two decomposition methods, known as singular value decomposition and QR decomposition, are applied for the proposed feature extraction method. Additionally, a new EEG channel selection method based on wrapper type method is proposed. This study is conducted on BCI completion III, dataset IIIa, which is a multiclass cued motor imagery EEG dataset. The obtained results are benchmarked against discrete cosine transform (DCT), adaptive autoregressive (AAR) based method, and different extensions of the state-of-the-art, Common Spatial Pattern (CSP) method, as well as the obtained results of the winners of competitions on this dataset. The best obtained results of the proposed methods reached an average accuracy of 81.38% with the capacity of generalization, which is the second best in terms of accuracy among all the available methods. A point to consider is that the best result belongs to a subject specific method with fine tuning of all parameters. The future direction of this research is to investigate more about the best choice of the relevant parameters to further improve performance.