Methodology

CSP

CSP aims to find spatial filters that minimize the variance of one class while maximizing the variance of another. To achieve this, it is often useful to bandpass filter the multichannel EEG signals ^[1]. The band power in any given frequency band gives the variance of the filtered EEG signals in the selected band. The CSP method obtains optimal discrimination for MI-based BCI tasks based on band-power features ^[1:1].

Fisher's LDA

The LDA-based classifier tries to reduce the dimensionality, while preserving most of the class discriminatory information.

Fisher's LDA:

• enhances the LDA classifier by determining a decision boundary or a hyperplane in the feature space to classify the features into distinct classes.
• does not make some of the assumptions of LDA such as normally distributed classes or equal class covariances.
• determines the separation boundary between two given distributions in terms of the ratio of two group variances.
• When the assumptions of LDA are satisfied, it is equivalent to LDA.

Sliding Window Techniques

SW-LCR

LCR is an element in a sequence, which has the longest subsequence containing just itself. If there is more than one candidate for LCR, then the earliest one is chosen.

Example:
In sequence “112221121”, "11" is the longest subsequence containing only 1, while “222” is the longest subsequence containing only 2, thus SW-LCR returns 2.

SW-Mode

In this approach, the number of occurrence of classes 1 and 2 is counted. The class containing the maximum count is attributed to that particular trial.

Example:
In “121111222”, the count of class 1 is 5, while the count of class 2 is 4, therefore SW-Mode returns 1.

This mode-based approach is especially important when the person is unable to concentrate on the task for a long time such as in the case of stroke patients although the overall focus is on the cued task despite fluctuations in concentration.