pbrunner wrote: ↑
25 Apr 2016, 10:23
please note that the 109-subject data on Physionet is screening data in which the subjects had to perform one imagery task at a time. The reference you cite below states that to achieve good 2-dimensional control you need training. Specifically, the training is necessary to control the 2-dimensions simultaneously and independently. In other words, most naive subjects would not be able to control the x- and y- dimension of a 2-dimensional task independently. I hope this explains your question.
I think I do not agree with the premise that the term "2-choice" BCI means 2-dimesnional control. It is only 2 choice: choosing either moving to left or moving to right.
This is more clear by this paragraph from the same book above by Pfurtscheller:
Let’s consider an example of a navigation/selection application, in which we want to move a cursor to items on a computer screen and then
we want to select them. A BCI based on selective attention could rely on five stimuli. Four stimuli are associated with the commands for cursor movement: left, right, up, and down. The fifth stimulus is for the select command. This system would allow two dimensional navigation and selection on a computer screen. A 5-choice BCI like this could be based on visual stimuli ...etc.
Of course when the book says 4 choice for the x and y navigation , it does not mean the user will control the 4 choices simultaneously.
Another paragraph stating the same meaning:
If a 4-choice BCI allows a user to select one of four wheelchair directions, ...etc.
So your description of "control the x- and y- dimension" is a 4-choice BCI and not 2 choice BCI.
The book when says:
BCIs based on motor imagery usually do not work very well during the first session. Instead, unlike BCIs on selective attention, some training is necessary. While performance and training time vary across subjects, most subjects can attain good control in a 2-choice task with 1–4 h of training (see chapters “The Graz Brain–Computer Interface”, “BCIs in the Laboratory and at Home: The Wadsworth Research Program”, and “Detecting Mental States by Machine Learning Techniques: The Berlin Brain–Computer Interface” in this book). However, longer training is often necessary to achieve sufficient control. Therefore, training is an important component of many BCIs.
So this Physionet EEG Motor Movement/Imagery Dataset is a 2 choice task. Either move left or move right on the x-axis. And of course nobody will do them both simultaneously. Just like the 4-choice BCI, nobody will do all the 4 directions simultaneously. So a 2-choice BCI (like what is mentioned in the book) is like what the volunteers of this dataset are doing.
And a more important statement in the same book:
Work with a large group of subjects that were not pre-screened stated that most untrained people can use an ERD BCI, even if their ITR is only a few bits per minute . However, that article assumed that subjects who attained greater than 60% accuracy in a 2 choice task could use a BCI effectively. If the more common threshold of 70% accuracy were used instead, more subjects would have been illiterate. Similar work validated P300 and SSVEP BCIs across many users [2, 32].
It is true that, in some cases, people who cannot initially use a BCI can attain control after training (see chapters “Neurofeedback Training for BCI Control” and “BCIs in the Laboratory and at Home: The Wadsworth Research Program” in this book). Improved signal processing or different instructions can also help. However, some subjects are unable to produce the brain activity patterns needed to control a specific BCI approach (such as ERD) no matter what. ....
This paper, and other recent work [2, 8, 32, 54] also suggested that illiteracy may be worse in ERD BCIs than SSVEP or P300 BCIs.
For the 109 volunteers of this dataset, if they were not trained, there would not be much value of the imagery part of the dataset, since "BCIs based on motor imagery usually do not work very well during the first session. Instead, unlike BCIs on selective attention, some training is necessary.
" - as the book by Pfurtscheller says. And as you know, the imagery part of the dataset is the important part in BCI and not the actual movement.
I think this is an important insight that should be considered.. The inability of pattern classifiers and machine learning models to get high scores on this dataset, perhaps from lacking of training of the volunteers on top of the BCI illeteracy of some others that even training is not useful..
Please correct me if I miss something or my conclusion is flawed..
And as I said, I like your work, and appreciate the huge efforts that you have been doing for the scientific community. The only thing I am trying is to think openly, and trying to understand this dataset more, for a better scientific outcome..