Hello,
I am using openvibe with a g.tec. mobilab+ device (8 channels) for motor imagery experiments. Yet without any success until now,
although I get 65 - 75% accuracy by the classifier. I am using the motor imagery scenarios delivered with openvibe with laplacian and common spatial filters. The electrodes are placed on C3,C4,C5,C6,CP3,CP4,FC3,FC4.
For the laplacian variant I choosed the following coefficients: 4*C3-C5-CP3-FC3 and 4*C4-C6-CP4-FC4 and named the electrodes in the acquisition server respectively. For the CSP-method the naming of the electrodes is not necessary - as far as I understood that.
We did imagination of the movement of the left and rigth hand as well as performing the real movement. The blue feedback bars in the online scenarios were nevertheless not controlled at all.
I only rented the EEG-device and have to send it back soon. If anyone has an idea how to improve the performance of the motor imagery BCI system, I would be very happy.
Thanks
motor imagery - How to improve performance?
-
- Posts: 775
- Joined: Tue Dec 04, 2012 3:53 pm
- Location: INRIA Rennes, FRANCE
Re: motor imagery - How to improve performance?
Hi,
about the non-CSP scenario, there seems to be a bug in the scenario
The classifier trainer box needs to have a classifier path starting with
The $ has apparently been forgotten in the last few versions. You can just add this missing character to the beginning of the path. This bug has a result that the trained and used classifier won't be the same.
Why the CSP version doesn't work any better, one possibility is the people. Generally motor imagery is hard to do. The last time we did it with 6 'new' subjects, only 1 could get some level of control. Literature often mentions that motor imagery participants(!) were trained to do the imagery many hours over several weeks before reaching accuracies like 80%.
ps. make sure that none of the openvibe scenarios you run print any errors to the log window. If they do, these need to be addressed first.
Best,
Jussi
about the non-CSP scenario, there seems to be a bug in the scenario
Code: Select all
applications/demos/motor-imagery/bci-examples/motor-imagery/motor-imagery-bci-2-classifier-trainer.xml
Code: Select all
${Player_ScenarioDirectory}
Why the CSP version doesn't work any better, one possibility is the people. Generally motor imagery is hard to do. The last time we did it with 6 'new' subjects, only 1 could get some level of control. Literature often mentions that motor imagery participants(!) were trained to do the imagery many hours over several weeks before reaching accuracies like 80%.
ps. make sure that none of the openvibe scenarios you run print any errors to the log window. If they do, these need to be addressed first.
Best,
Jussi
Re: motor imagery - How to improve performance?
Thanks a lot for your fast reply.
We will adjust the files by removing the bug you mentioned and have an eye on error messages.
Thanks also for your your experience with MI, that explains our low success. But shouldn't there be a reliable signal in all cases and immediately when really moving your hands? By moving your hands in reality the motor-cortex-signal should change significantly without learning or training in my opinion.
Thanks for help and regards,
Sebastian
We will adjust the files by removing the bug you mentioned and have an eye on error messages.
Thanks also for your your experience with MI, that explains our low success. But shouldn't there be a reliable signal in all cases and immediately when really moving your hands? By moving your hands in reality the motor-cortex-signal should change significantly without learning or training in my opinion.
Thanks for help and regards,
Sebastian
-
- Posts: 775
- Joined: Tue Dec 04, 2012 3:53 pm
- Location: INRIA Rennes, FRANCE
Re: motor imagery - How to improve performance?
In my experience this also depends somewhat on the person, with some its difficult to get reliable classification with this either. I'd suggest performing a repeated motion, perhaps a hand-squeeze, continuously during the training trials and then test the classifier with that same kind of motion. If you're suspicious about the signal processing chain, you can try to test with "artifact-based bci": Blink your eye repeatedly during left trial, and remain still during the right trial. This should give you 100% online accuracy when you do the same, as the effect of blinking on the recording should be quite noticeable -- it works as a test that the components are at least properly connected.skirsch82 wrote: Thanks also for your your experience with MI, that explains our low success. But shouldn't there be a reliable signal in all cases and immediately when really moving your hands? By moving your hands in reality the motor-cortex-signal should change significantly without learning or training in my opinion.
There are more experienced motor imagers on the forum that have larger experience though. Any additional thoughts, anyone?
Cheers,
Jussi
Re: motor imagery - How to improve performance?
Thanks for help. I will try the artefact-based-trial with eye blinking as a control on saturday.
I come back with results then
I come back with results then