A Computational Model of MI-EEG Association Prediction Based on SMR-DCT and LS-SVM

The theory of feature extraction and pattern classification of Motion Imaging Electroencephalogram (MI-EEG) plays an important role in Brain-computer Interface (BCI) system, which is widely used in rehabilitation medicine, intelligent control and other fields. Although advanced acquisition technologies have generated considerable EEG data for various brain areas, it has inevitable drawbacks such as high cost, time consumption, and inherently high false positive rate of existing methods. In this paper, considering the contradiction between data quantity and accuracy, we used a two-step feature extraction method based on Discrete Cosine Transform (DCT) together with Least Squares Support Vector Machine (LS-SVM) to perform feature extraction and classification for single-trial MI EEG. Based on the dataset constructed from the Project BCI-EEG motor activity, the average ACC, PE, SN and MCC of seven channels in the central and temporal lobes of the brain are 90.27%, 91.28%, 89.35%, and 82.34% were obtained with 5-fold cross validation, respectively. Furthermore, the same method was applied to the BCI Competition IV dataset, and the above results were also confirmed. The performance comparison with previous prediction models show that our method used fewer channels, but the accuracy was higher and more reliable. It is anticipated that the proposed method can be used as an effective computational tool for future BCI researches.