R Squared Anomaly in Mu Tutorial

Forum for discussion on different signal processing algorithms
Locked
jwlbci
Posts: 6
Joined: 13 Aug 2009, 13:56

R Squared Anomaly in Mu Tutorial

Post by jwlbci » 14 Aug 2009, 16:26

Hello,
I'd like to ask a question about the R values on the mu tutorial. When I plot the features, certain subjects have significantly lower R values than than I expected.

The subjects have gone through the non-feedback protocol several times, but only reaching smaller R values (.02 or .06 at most). We're getting some features in the 8-14 hz range (as expected), but the R values between classes are consistently low. In other words, the features are there, but expected R values are not.

I understand every subject is unique, but this seems rather low. Out of four subjects (all normal young male subjects), only one achieved R^2 squared values higher than .1. This one subject only achieved an R^2 value of .2.

We are using a Neuroscan EEG system with Scan 4.4 software, and the system works and records fine.

A few questions:

1) What can cause R squared values to be so low, across several sessions and subjects?

2) Can some form of training or subject selection help increase these R values?

3) How can you train a subject on BCI? Is the feedback protocol more effective than the calibration one?

4) What if the subject is unable to perform the feedback protocol task effectively for several runs?

Any assistance for this would be welcome. Regards, John

gschalk
Posts: 615
Joined: 28 Jan 2003, 12:37

r^2 values ...

Post by gschalk » 18 Aug 2009, 13:11

John,

These are good questions. As you point out, every subject is different. This is the reason we need to do these initial sessions in the first place.

In general, it is not clear to what extent an initially low r^2 predicts success with BCI performance. What is clear is that the r^2 values should be distributed spatially and spectrally at least somewhat similarly to what is shown in http://www.bci2000.org/wiki/index.php/U ... hm_Session

In response to your question 2, r^2 values will definitely increase with BCI training. This training is described in the tutorial (which we are close to finish revising). If the subject is unable to perform the feedback task, I would first double check the BCI2000 configuration. It may also be helpful to simulate activity using the SignalGenerator and the mouse, then to configure BCI2000 based on what you find in the offline analyses, and then run a feedback session using the mouse. If this does not work, you are not configuring BCI2000 correctly. If you are confident that you configure BCI2000 correctly and the subject still cannot control the cursor, that subject will probably not be able to learn the task. However, about 75% of the subjects should be able to get some or even good control.

I hope this helps.

Gerv

Locked

Who is online

Users browsing this forum: No registered users and 3 guests