Hi Josh,
Just thought I'd separate questions for later users reference.
For my project I want to conduct Mu feedback sessions.
So from my understanding I first need to perform an initial session for each participant. Then do an offline analysis of this EEG recording to determine what are the best channels for training them specifically. Then I need to configure subject specific parameters to load for the feedback sessions. So each subject will have separate feedback parameters. These parameters established in the initial session will be used each time they come to a feedback session. Is this correct?
Also...
Does the initial session have a min or max number of electrodes required?
Does the offline analysis then determine the best electrodes to use for each subject or do you use the same montage for the feedback as you used for the initial session?
Or can you only use one feedback channel for the feedback training like it suggests on the classifier configuration page...http://www.bci2000.org/wiki/index.php/U ... e_Feedback (the use of the phrase "enter the location of *the* feedback electrode" suggests it only uses one electrode)?
thanks
channels for Mu
Hi Emily,
That is the process in a nutshell. Of course there will be similarities between subjects for related movements (e.g., right hand movements will typically be correlated to responses in the left motor cortex), but each subject is unique. Thus, we have to determine the optimal feature set for each subject before performing a feedback session. Unfortunately, there are often changes in a given subjects response across sessions as well. So, you'll typically want to analyze an initial dataset at the beginning of each session. It's been our experience that the amount of change exhibited by the response across sessions typically decreases over time. Still, for best results, you'll probably want to go through the offline analysis.
As far as the number of electrodes, I would take two things into consideration. First, you want to use a montage that corresponds to the movements. Again, the right hand corresponds to a specific area of the left motor cortex. This page from the wiki provides a nice overview of the physical properties of Mu rhythms: http://www.bci2000.org/wiki/index.php/U ... _Mu_Rhythm. So, for right hand movement, we'll probably want to use a montage that places electrodes over C3 and Cz (and probably others). The second consideration is that different caps are likely to have different noise and impedance characteristics. Of course, in performing experiments, you want to keep the conditions as similar as possible across sessions. So, it's best to use the exact same cap if possible. As far as how many electrodes that cap should have, that depends on how much data you want to collect and where you want to collect it from. Many experiments can be done using an 8 or 16 electrode cap. My guess is that you'll probably be fine with either of those. You can also obtain caps that are specifically configured for particular types of experiments (e.g., Mu, P300, etc...).
The wording on the wiki page you referenced does appear to be a bit confusing...I'll make sure to change that. You can use as many features as you would like in the feedback session. Using too few or too many, however, can degrade the results. Also, the more you use, the greater the computational cost. In my experience, 2 features has been enough to achieve decent results. This is something you should probably experiment with. Now, Offline Analysis (OA) will not, unfortunately, pick the best subset of features for you. Instead, it will help you to choose the best features by presenting you with various visualizations of the data. For instance, if your subject blinks every time he/she moves his/her right hand, the topography might show a response over the frontal lobe. Having some knowledge of the mapping of the brain, however, we would know to omit that response from our feature set.
Well, hopefully this all helps. Let us know if you have any more questions.
- Josh
That is the process in a nutshell. Of course there will be similarities between subjects for related movements (e.g., right hand movements will typically be correlated to responses in the left motor cortex), but each subject is unique. Thus, we have to determine the optimal feature set for each subject before performing a feedback session. Unfortunately, there are often changes in a given subjects response across sessions as well. So, you'll typically want to analyze an initial dataset at the beginning of each session. It's been our experience that the amount of change exhibited by the response across sessions typically decreases over time. Still, for best results, you'll probably want to go through the offline analysis.
As far as the number of electrodes, I would take two things into consideration. First, you want to use a montage that corresponds to the movements. Again, the right hand corresponds to a specific area of the left motor cortex. This page from the wiki provides a nice overview of the physical properties of Mu rhythms: http://www.bci2000.org/wiki/index.php/U ... _Mu_Rhythm. So, for right hand movement, we'll probably want to use a montage that places electrodes over C3 and Cz (and probably others). The second consideration is that different caps are likely to have different noise and impedance characteristics. Of course, in performing experiments, you want to keep the conditions as similar as possible across sessions. So, it's best to use the exact same cap if possible. As far as how many electrodes that cap should have, that depends on how much data you want to collect and where you want to collect it from. Many experiments can be done using an 8 or 16 electrode cap. My guess is that you'll probably be fine with either of those. You can also obtain caps that are specifically configured for particular types of experiments (e.g., Mu, P300, etc...).
The wording on the wiki page you referenced does appear to be a bit confusing...I'll make sure to change that. You can use as many features as you would like in the feedback session. Using too few or too many, however, can degrade the results. Also, the more you use, the greater the computational cost. In my experience, 2 features has been enough to achieve decent results. This is something you should probably experiment with. Now, Offline Analysis (OA) will not, unfortunately, pick the best subset of features for you. Instead, it will help you to choose the best features by presenting you with various visualizations of the data. For instance, if your subject blinks every time he/she moves his/her right hand, the topography might show a response over the frontal lobe. Having some knowledge of the mapping of the brain, however, we would know to omit that response from our feature set.
Well, hopefully this all helps. Let us know if you have any more questions.
- Josh
Hi Josh,
thanks for all of this information it's really helping. Apologies but I still have questions!
After reading the Wolpaw and Mcfarland papers I was beginning to think that I would have to record from all 64 electrodes for the initial session and all subsequent feedback sessions. Was the reason they used 64 for both initial and training sessions for accuracy?
So, I can get adequate information for offline analysis and CAR from 8 - 16 sites. I'm guessing the more you use though the more accurate it becomes.
And, I could also get adequate cursor control from only 2-3 sites eg. C3, C4 and CZ. I am only focusing on training Mu rhythm though so would I be better off just using C3 and C4 (CZ being more likely to be beta related?) and screening for participants that had strong 8-12Hz responses over these areas so would most likely respond to the training.
The Mcfarland and Wolpaw papers all seem to use both Mu and Beta in the training even for the 1D cursor experiments. Is there a reason for this?
Also if you have been able to replicate the error I was getting with the matlab runtime component it would be great because I will need to get it working on my test computer if I'm going to be doing offline analysis at each session.
thanks very much
thanks for all of this information it's really helping. Apologies but I still have questions!
After reading the Wolpaw and Mcfarland papers I was beginning to think that I would have to record from all 64 electrodes for the initial session and all subsequent feedback sessions. Was the reason they used 64 for both initial and training sessions for accuracy?
So, I can get adequate information for offline analysis and CAR from 8 - 16 sites. I'm guessing the more you use though the more accurate it becomes.
And, I could also get adequate cursor control from only 2-3 sites eg. C3, C4 and CZ. I am only focusing on training Mu rhythm though so would I be better off just using C3 and C4 (CZ being more likely to be beta related?) and screening for participants that had strong 8-12Hz responses over these areas so would most likely respond to the training.
The Mcfarland and Wolpaw papers all seem to use both Mu and Beta in the training even for the 1D cursor experiments. Is there a reason for this?
Also if you have been able to replicate the error I was getting with the matlab runtime component it would be great because I will need to get it working on my test computer if I'm going to be doing offline analysis at each session.
thanks very much
channels
Emily,
For successful experiments, you only need a few sites, e.g, C3 or C4. You typically want to do spatial filtering, e.g., a Laplacian Filter. Thus, will need at least the surrounding channels also. However, in order to get a comprehensive picture of what's going on, and to rule out EMG artifacts, you should be recording from as many channels as possible.
Gerv
For successful experiments, you only need a few sites, e.g, C3 or C4. You typically want to do spatial filtering, e.g., a Laplacian Filter. Thus, will need at least the surrounding channels also. However, in order to get a comprehensive picture of what's going on, and to rule out EMG artifacts, you should be recording from as many channels as possible.
Gerv
Thanks Gerv,
Could you please confirm my understanding is correct. If I'm using C3 and C4 as the feedback channels then for the subject specific parameters the only thing that will change is what the feedback frequency its centred on. Which is manually confirmed by offline analysis and may change over sessions.
cheers
Could you please confirm my understanding is correct. If I'm using C3 and C4 as the feedback channels then for the subject specific parameters the only thing that will change is what the feedback frequency its centred on. Which is manually confirmed by offline analysis and may change over sessions.
cheers
channels
Emily,
Pretty much. In general, you would slowly make changes to both frequencies and locations, and the changes would not be dramatic. For example, if in two sessions in a row, in offline analyses CP3 is better than C3 you used online, or 11 Hz is better than the 12 Hz you used online, I would switch over.
Gerv
Pretty much. In general, you would slowly make changes to both frequencies and locations, and the changes would not be dramatic. For example, if in two sessions in a row, in offline analyses CP3 is better than C3 you used online, or 11 Hz is better than the 12 Hz you used online, I would switch over.
Gerv
Who is online
Users browsing this forum: No registered users and 3 guests