Evaluation of power spectral and machine learning techniques for the development of subject-specific BCI

Evaluation and interpretation of massive amounts of brain data are a big challenge for the design of functional brain-computer interface (BCI) devices. In this chapter, three power spectral methods: Welch, Burg, and multiple signal classification (MUSIC) are investigated to improve the pattern mining of two-class motor imagery electroencephalography (EEG) signals. Specifically, freely available dataset IVa from BCI competition III was used for evaluation. This dataset comprises a total of 118 electrodes, while 18 electrodes around the motor cortex region were included in our experiments. The proposed study is described in threefold. First, the multiscale principal component analysis (MSPCA) method was used to obtain clean EEG signals. Second, power spectral density (PSD) values were determined using the Welch, Burg, and MUSIC methods, and these PSD vectors were used as feature vectors. At last, all feature vectors were provided to logistic regression (LR), multilayer neural perceptron, and support vector machine classifiers by varying different parameters for classification tests. The results showed that the Welch PSD process gives a cumulative sensitivity, specificity, and accuracy of 99.7%, 100%, and 99.8%, respectively, which is better than Burg and MUSIC. In comparison with other methods on dataset IVa, the proposed system obtains an increase of up to 26.3% in classification. The findings suggest that the combination of MSPCA, Welch method, and LR is very efficient that can be used to build a subject-specific BCI system for different applications such as games and movies, artificial intelligence, robotics, and rehabilitation.