feedback settings in motor imagery BCI CSP

[Joanllo]

Hello,

I’m currently starting my PhD project based on EEG signal processing and neurofeedback. I’ve been following the Motor Imagery BCI with Common Spatial Pattern filter scenarios as an example since I’ll perform a motor imagery paradigm (right and left hand) to enhance SMR rhythm on sensorimotor cortex.
I have a question about real-time feedback and the mathematical processes carried out: I understand the training on CSP filter and LDA classifier but I have problems on how the program actually adds the real-time feedback to the visualization. In this scenario the feedback is presented to the subject through a blue bar, so: the bar length and direction reflect the more or less “left hand imagery movement” comparing with LDA classifier? How is the mathematical process to match the LDA classifier info and the current online cortical activity?

Sorry for the inconvenience I’m really new on feedback software.

Thank you.

[fabien.lotte]

The output of the LDA classifier is actually the distance of the feature vector (i.e., the signal band power for each spatially filtered CSP channel) to the LDA separating hyperplane. This output is 0 if the feature vector is on the hyperplane, positive if it is on one side of it (i.e., if it belongs to one class) and negative if it is on the other side of it (i.e., if it belongs to the other class). The larger the outcome (in absolute value), the further away from the separating hyperplane the feature vector. This output, i.e., the distance to the separating LDA hyperplane, is what is mapped to the blue bar : the blue bar goes left or right is the LDA output is positive or negative, and the largest the absolute LDA output, the longer the longer the bar.

I hope this helps,

Best regards,
Fabien

[Joanllo]

Thank you so much Fabien, it has been helpful!

Regards,
Joanllo

[Joanllo]

Sorry for double post

Edit:

I have few more questions about MotorImagery scenario:

There’s a way to save a file with the classifier accuracy measure displayed online? This accuracy means: [feature vectors located on the correct side of hyperplane (proper class side) / total of feature vectors]?

I’m interested on select the best eeg feature extraction based on R2 coefficient, i.e. perform an screening session (left and right imagery movement without feedback) to pick the best frequency (within SMR band) to train online, based on each subject features. Something like compute a plot [Channel vs Frequency] with correlation coefficients of class 1 vs class 2. Can I do that with OpenViBE?

Than you very much again
Joanllo