Subject Dependent Feature Extraction Method for Motor Imagery based BCI using Multivariate Empirical Mode Decomposition

Extracting efficient features is an important stage in Brain Computer Interface (BCI). In motor imagery (MI) based BCI, features that characterize ERS/ERD phenomenon could favor the discriminant of different kinds of MI tasks. Common Spatial Pattern (CSP), a spatial filtering method which is often used in MI-EEG feature extraction, shows outstanding performance due to its effectiveness in maximizing the difference of variance between each class. But one of the shortcomings of CSP is that performance is greatly affected by the frequency of the signal, because the variances can be used to discriminate each class only by processing those task related frequency bands. However, individual differences existed in EEG is an obstacle for CSP. Although MI related frequency bands focus on mu and beta rhythms, for every subject, the distinguishing hands could be different. In this paper, a novel feature extraction method (Subject Dependent Multivariate Empirical Mode Decomposition, SD-MEMD) for MI based BCI is proposed, it utilizes MEMD algorithm to decompose the multi-channel EEG into a set of Intrinsic Mode Functions (IMFs), each IMF represents an inherent oscillation mode of the raw signal. Then an enhanced EEG is re-constructed after a careful selection of the task related IMF subset. Classification accuracy for right hand and left hand MI tasks is improved by 5.76% on our dataset.