Misclassification with Handball scenario

[bxxworld]

Hi everyone,

I'm new to Openvibe, and has just figured out how to use it a few weeks ago. My classmates and I are using Emotiv 1.0.0.3 with Openvibe 0.9.0 and we've been familiarizing ourselves with the scenarios so as to build a BCI game out of it. So far, the tie-fighter seemed to run perfectly well, yet when it comes to the handball scenario, the classification accuracy is merely 50%. However we tried, the final online test would always classify the right/left hand movement as ONE movement, i.e. for actual right hand movement, it would classify correctly; yet for actual left hand movement, it would still classify as right...

We've switched the classifier from LDA to SVM, hoping that the classification might get better, yet nothing seemed to improve.. Hence I wonder if there is anyway I can get down to the original source code of this scenario, so that I might be able to know how to deal with this problem? Or is there anyone else who have encountered similar situations?

Any suggestion is appreciated, thanks in advance!!

-bxxworld

[ddvlamin]

Hi,

Left and right hand motor imagery experiments can indeed be hard. First of all there are two types of imagined movement: the first one is simply seeing yourself (or someone else) make the movement, most of the time this is the bad way of doing it. What I do and how I instruct people to complete the task, is to really try to move your hand (having the intent to move), but not to really move. After recording a few calibration (one or two) sessions of around 10 minutes (i thinks this is about 40-50 trials), some people will then be able to achieve good (above 70%) performance, for others a few sessions won't be enough and a minority won't be able to do it at all. At least this is my experience with the few people I recorded data from. So it could be that you just need to practice a bit more.

The classifier should also be trained each time on each individual and in case a supervised spatial filter is used, that one should be trained too.

For motor imagery the difference between LDA and SVM is not that big (in case you only use very few features based on the power in some pass-band and some electrodes), unless maybe there are some outliers due to eye-artifacts, try not to have any eye-artifacts during the trials.

Apart from that, I see that you use Emotiv headcap, but the emotiv headcap does not have any electrodes above the motor cortex? How do you record C3, C4 and Cz, by replacing the electrodes manually or reorientating the cap? The electrodes closest to that are FC6 and FC5, although not optimal for motor-imagery. I wonder how you did this as you say that the foot-imagery worked fine which mostly uses Cz?

Normally it should not be necessary to dive into the source code as the provided boxes and scenarios are able to accomplish this. Nevertheless, with svn you can download the sources http://openvibe.inria.fr/documentation/ ... stall.html

I hope these hints can be of use to you.

Best regards,
Dieter Devlaminck

[bxxworld]

Hi Dieter,

First of all thank you so much for the comprehensive explanation! It is really helpful!

I think we might have been conducting the experiment in the 'bad' way you've specified.. I would try to do it again as soon as possible with a real intention to move as you talked about, and might as well try not to produce any eye artifacts. Though I'm curious why would 'seeing someone/myself moving the hand' be a bad way compared with the intention of really trying to move? Is it because different regions of the brain is activated under this two different scenarios?

As for Emotiv's electrode location, that's one of the biggest problems we have.. We did try to replace the electrodes manually, to move it a little bit backward, so that we can look at F3 and F4 instead for signals over the motor cortex, though we haven't come up with other good ideas of tackling this problem.. But I think it indeed worked, for we did manage to fly the tie-fighter, and the power spectrum increased dramatically over the two channels. We didn't record the signal so I cannot remember clearly if the increase was only for the two channels or over all the channels.. We might need to experiment again, but if it were over all the channels, then would it indicate a strong EMG artifact or other related factors?

So far, our Emotiv cap can only capture mu rhythms effectively, for the power spectrum does soar up significantly when spacing out. But when building a BCI game, mu rhythm cannot really do anything, for most of time it would require the subject to close his/her eyes so as to space out, and this would make all visual display impossible.

Regarding the source code, I was actually trying to measure the level of the spectrum over our trails, ideally to record the magnitude of certain frequency over time. But I don't think this is applicable with built-in Openvibe components?

Thanks again for the help! Though I've come up with even more questions..

Best regards,
bxxworld

[ddvlamin]

I think we might have been conducting the experiment in the 'bad' way you've specified.. I would try to do it again as soon as possible with a real intention to move as you talked about, and might as well try not to produce any eye artifacts. Though I'm curious why would 'seeing someone/myself moving the hand' be a bad way compared with the intention of really trying to move? Is it because different regions of the brain is activated under this two different scenarios?

That's a very good question, however, I'm no neurologist so take my answer with the necessary portion of criticism. My guess is that the motor area is mainly involved in planning and executing movement. When you just imagine someone or yourself doing some movement then other processes play a part in the brain for example memory recollection. In the motor area however there are also so-called mirror neurons that activate when you observe someone else performing some kind of movement. Thus one could argue that maybe these mirror neurons are also active when you see yourself making a movement (through internal respresentation), but maybe these activations are too weak to be seen on the surface?

As for Emotiv's electrode location, that's one of the biggest problems we have.. We did try to replace the electrodes manually, to move it a little bit backward, so that we can look at F3 and F4 instead for signals over the motor cortex, though we haven't come up with other good ideas of tackling this problem.. But I think it indeed worked, for we did manage to fly the tie-fighter, and the power spectrum increased dramatically over the two channels. We didn't record the signal so I cannot remember clearly if the increase was only for the two channels or over all the channels.. We might need to experiment again, but if it were over all the channels, then would it indicate a strong EMG artifact or other related factors?

Is it not easier to reorientate the FC electrodes as they are closer to C3 and C4?
If there are some strong activations on all channels at each time frame (during the whole course of the experiment) I would say it is an external artifact (outside the body). EMG artifacts can occur, but will most likely have very specific location, such as prefrontal electrodes (eye-blinks) or temporal electrodes (chewing) and occur only at some time frames. Off course all these artifacts can also be visible at other electrodes.

So far, our Emotiv cap can only capture mu rhythms effectively, for the power spectrum does soar up significantly when spacing out. But when building a BCI game, mu rhythm cannot really do anything, for most of time it would require the subject to close his/her eyes so as to space out, and this would make all visual display impossible.

Alpha rhythms can also be modulated by focusing versus relaxing and wakeful staring/dreaming, but indeed the difference is most noticeable when closing/opening eyes. mu rhythms are more associated with motor tasks, but display discriminative information in the same frequency band as the alpha rhythms. It's subject to learning (operant conditioning)

Regarding the source code, I was actually trying to measure the level of the spectrum over our trails, ideally to record the magnitude of certain frequency over time. But I don't think this is applicable with built-in Openvibe components?

I think I do not exactly understand what you mean. I know there are boxes to compute the spectrum and to perform stimulus based or time based epoching. There is also a box for time-frequency analysis that you could use to see how power at certain frequencies change over time?

Best regards,
Dieter Devlaminck

[bxxworld]

Hi Dieter,

Thanks again for the thoughtful explanation! We've tried again during the weekend, but the classification is still quite poor. We checked for the classification accuracy, and the average is merely 50%, despite of the fact that we were indeed trying to have the intention of real moving, instead of visualizing other's hand movement.

So it seems that the one and only explanation we can draw from such performance is that the Emotiv cap is not suitable for measuring SMR or brain signals from the motor cortex. I wonder if you've had experience with Emotiv before? How was the cap's performance considering motor cortex?

The only consolation was that we were indeed able to make the tie-fighter to work, through shifting the cap backward, so that F3 and F4 can lie right on the motor cortex (we're still not familiar with the orientation process of the cap..). With a channel selector from Openvibe, we've monitored only these two electrodes, and was able to fly the tie-fighter, though only once....After some time, the brain signal just wouldn't stop oscillating to a quite considerable extent, no matter how we tried. I guess there must have been some really strong noise which prevented the signal from keeping steady. Wonder if you've encountered similar situations?

Thanks again for the reply!
bxxworld

[karthiks]

Hi Dieter,

Thanks again for the thoughtful explanation! We've tried again during the weekend, but the classification is still quite poor. We checked for the classification accuracy, and the average is merely 50%, despite of the fact that we were indeed trying to have the intention of real moving, instead of visualizing other's hand movement.

So it seems that the one and only explanation we can draw from such performance is that the Emotiv cap is not suitable for measuring SMR or brain signals from the motor cortex. I wonder if you've had experience with Emotiv before? How was the cap's performance considering motor cortex?

As the Emotiv doesn't work heavily on motor areas (primarily due to the difficulty of maintaining proper pressure, I think), the visualization of each hand will not work. Why it does work for their own control panel application is because the headset concentrates on occipital areas which are also involved in such imagination/visualisation.

[ddvlamin]

Hi,

I do not have a Emotiv cap, but I tested it briefly with an SSVEP paradigm (thanks to my colleagues at the University of Leuven of the Computational Neuroscience Research Group) and this worked great (it uses occipital regions). I do not know if they used it for motor-imagery experiments too.

Why it does work for their own control panel application is because the headset concentrates on occipital areas which are also involved in such imagination/visualisation.

@karthiks: I can imagine that the visual cortex plays a role in guidance of precision motor tasks, but never heard of it being used during motor imagery. Could you give me a reference about that?

Best regards,
Dieter Devlaminck

[karthiks]

Why it does work for their own control panel application is because the headset concentrates on occipital areas which are also involved in such imagination/visualisation.

@karthiks: I can imagine that the visual cortex plays a role in guidance of precision motor tasks, but never heard of it being used during motor imagery. Could you give me a reference about that?

Best regards,
Dieter Devlaminck

Dieter, I do not think they use the visual area for motor imagery of the subject. Think they use it about the change in the box. I am not sure. This is just a conjecture. I could be very wrong

What do you think about the headset's algorithms?

Karthik

[Tiernan]

Hi Guys

Very interesting topic. I am exploring a similar scenario..... I also would like to explore differences between trials and I believe EEGlab is really good at ERP analysis and visualisation for scientific publication, but I'm having problems with the OpenVibe GDF. Also a reversal of the EPOch headset may be a solution to the sensorimotor problem.. If you get the chance can you look at my topic that I posted under the following heading Openvibe GDF will not open in EEGlab and Emotic EPOC reverse

Kind regards

Tiernan