bci2000.org BBS

Hello,

I have a question about the linear equations that are used to determine cursor movement in the standard mu/beta rhythm task. Two strategies for training a subject to control a cursor in 1-dimension can be:

1 - Think about some type of imagery to move the cursor right, and relax to move left.
2 - Think about separate types of imagery for moving the cursor right and left (so use 2 channels, each which responds to different types of imagery, for example, C3 and C4 can be used for right and left hand imagery.)

In case 1, I'm pretty sure I understand how the linear equation and adaptation works - it finds the running average of the signal (Ymean) and above that the cursor moves right and below that the cursor moves left (or vice versa depending on the weights).

In case 2, it does something similar, but both signals are added together (after being multipled by their weights) and the average of this combined signal is used for Ymean.

If I turn off adaptation (for case 2), both signals still contribute to each direction because they are being added together in the linear equation.

Finally my questions:

1 – Is it correct that in case 2, Ymean will be the average signal after all contributing signals (for two different directions of movement imagery) have been added together?

2 – If so, couldn’t this lead to problems with non-independence of the signals? For example, right hand imagery may lead to a large response on C3 at a certain freq band, but may also have a response in C4. However, if you have C4 setup for left hand imagery (and cursor movement in the opposite direction), it seems there is potential for a problem due to correlation of the signals.

2 – Is there some way to make each direction independent (so each type of imagery is solely responsible for ‘right’ or ‘left’, and not both rolled into the overall linear equation)? If not, is there a reason why the current method is better?

Thanks for all of your help.

Elizabeth Felton
UW-Madison

Elisabeth,

As always, you are asking excellent questions. Your questions contain two components. First, one can train people to use different mental strategies (such as using one or more different mental strategies). Second, one can use either one or more brain signal features.

In regards to the first component, it is not clear which method is better. With EEG, and also ECoG, we have typically used only one mental strategy for each dimension.

Regarding (2), BCI2000 always adjusts the mean and gain of the resulting control signal (X/YMean/Gain). As you indicate, it is possible that individual brain signal features exhibit correlation or otherwise unevenly contribute to the linear movement prediction. As you point out, this problem is not solved by calculating statistics on the output control signal. This is why BCI2000 also implements the LMS algorithm to adjust (on a much slower time scale than is done for the mean/gain adaptation) the weights that are given to each feature (in MUD/MLR). To do this, you need to turn on weight adaptation (WeightUse to 2) and choose an appropriate WtLrnRt. If this is too high, the system becomes unstable. If it is too low, the system takes a long time to adapt. If you turn this on, you also need to set the predicted positions for the targets in TargetPos correctly (i.e., columns 5-7). If any of these settings are not correct, this will not work correctly and thus you need to be quite careful. I suggest that you first check out the 2004 Wolpaw and McFarland article on 2D control that talks about this conceptually. Then, look at the User manual again for the TargetPos definitions. Before you actually try this on a patient, you should use a function generator to make sure the system does what you think it should. If you have more questions about this, please email Dennis McFarland (mcfarlan@wadsworth.org). He wrote the adaptation code.

I hope this helps,
Gerv