P300 Online Classification Configuration--negative response?

[lynn]

Hi,

In the Time-Domain Offline Analysis of EEG Data, we get a Waveform Plot range from -0.006 to 0.004, and the peak of waveform is not really a "peak", it's a "valley". Does this mean the response is negative?

Also, according to "User Tutorial:Configuring P300 Online Classification", when Configuring the Classifier, "in the fourth column, enter the weights. If the response is positive at the specified temporal offset, enter 1. If it is negative, enter -1". So.....if it is negative, we need to set it -1, right?

Thank you.

Smile,
lynn

[gschalk]

Lynn,

It is very possible that you are getting a negative peak. Typically, you may see a negative peak at posterior electrodes, which is then followed by a positive peak at central-parietal locations. As described in the tutorial, negative responses would need a negative weight in the classifier.

Gerv

[lynn]

Lynn,

It is very possible that you are getting a negative peak. Typically, you may see a negative peak at posterior electrodes, which is then followed by a positive peak at central-parietal locations. As described in the tutorial, negative responses would need a negative weight in the classifier.

Gerv

3x~~~ We try the initial session again today with another subject, the response are all positive now:)

The problem is, the P300 Spelling session does not work well even the result of initial session looks pretty good ( all of the clusters appear to be aligned at about 300.50 ms, and 3 out of 4 channels are good enough to serve as an effective recognition feature).

I modified the source code of Neuroscan to meet our need.... We have 4 input channels in total, the frequency is 1000HZ.... So.....Do we need more input channels? Is 1000HZ too high for BCI2000?

Thank you~

Smile,
Lin

[lynn]

When P300 starts to run, a window named "Timing" will show up(it contains "Block","Roundtrip" and "Stimulus" on the right side, the left side of this window show two waveforms).

We don't find any description about this window in the P300 Tutorial. Just for curious, what's this window show us?

Thank you.

Smile,
Lin

[gschalk]

Lynn,

When you say that the results are not too great, what does this mean? In other words, what percentage of the letters is spelled correctly? With how many flashing sequences per character? Just because the responses in the offline analyses look nice (on the average), this does not mean that they are large enough to be classified correctly in single trials.

Also, for the timing, please see the documentation on the wiki:

http://www.bci2000.org/wiki/index.php/U ... nce:Timing

Gerv

[lynn]

Hi,

Thank you for the url of timing window.

Well, the system could not get any correct spelling when we tried it last week.

About the timing window, the Block curve we saw during the test kept jumping up and downing like crazy , while in the documentation on wiki says " the block duration curve should be a straight line". Is this a serious problem?

Beside, the Roundtrip curve went up above block curve sometimes....Ｉ guess this is the main reason we didn't have any correct spelling.


And... when you say " large enough to be classified correctly in single trials", could you suggest us how large incoming data is usually large enough ?

Thank you so much.

Smile,
Lynn

Lynn,

When you say that the results are not too great, what does this mean? In other words, what percentage of the letters is spelled correctly? With how many flashing sequences per character? Just because the responses in the offline analyses look nice (on the average), this does not mean that they are large enough to be classified correctly in single trials.

Also, for the timing, please see the documentation on the wiki:

http://www.bci2000.org/wiki/index.php/U ... nce:Timing

Gerv

[gschalk]

Lynn,

This definitely indicates that you have a performance problem. I would try the system on a fast computer with a dedicated graphics card (e.g., a recent desktop, or a laptop with a dedicated video card -- not an integrated graphics chip), and then checking out the timing again before you do any experiments.

It is difficult to state how large the r^2 should be, but in good subjects, you should get about 0.01 to 0.05. For inferring classification performance, you should be looking at the r^2 plots, not the waveform amplitude averages.

Gerv

[lynn]

Thanks~We'll try the system again with a Mac Pro tomorrow.

Smile,
Lynn

Lynn,

This definitely indicates that you have a performance problem. I would try the system on a fast computer with a dedicated graphics card (e.g., a recent desktop, or a laptop with a dedicated video card -- not an integrated graphics chip), and then checking out the timing again before you do any experiments.

It is difficult to state how large the r^2 should be, but in good subjects, you should get about 0.01 to 0.05. For inferring classification performance, you should be looking at the r^2 plots, not the waveform amplitude averages.

Gerv