BCI competition 2008-graz Data set B (Chinese translation)

R. Leeb 1, C. Brunner 1, G. R. M Uller-putz 1, A.SCHLOGL 2, and G.pfurtscheller 1

1 Institute for Knowledge Discovery, grazuniversity of Technology,austria

2 Institute for Human-computer Interfaces, grazuniversity of Technology, Austria

Experimental paradigm

As published in the paper [1], the data contained 9 subjects of EEG data. The subjects were right-handed, had normal or corrected normal vision, and received experimental remuneration. All the participants sat on a chair with armrests and looked at the screen 1m away from the eye. Each data acquisition process is divided into 5 segments, the first two include collecting data but (the screen) does not give feedback, the last three paragraphs collect data and record feedback.
Each section contains some groups (run), 1. At the beginning of each paragraph, approximately 5 minutes of data are recorded to assess the effects of ocular electrical EOG. This five minutes is divided into three parts, (1) Two minutes open eyes (staring at the fixed "+" on the screen), (2) a minute to close your eyes, (3) One minute eye movement. The Pseudo-Trace module (artifact block) is divided into four parts (15 seconds of Pseudo-trace time followed by 5 seconds of rest time), while the participants follow the onscreen instructions to perform blink, rotate the eyeball, up or down or turn the eyeball around. At the beginning there is a tone of the bass, at the end there is a tone of the treble. Note that eye power is not recorded for technical reasons in the B0102T and b0504e segments.

Data ingest

EEG of three electrode records (C3, CZ and C4) with a sampling frequency of 250Hz. Recorded EEG data includes dynamic range +-50UV of the dynamic range of the screen for +-100UV and feedback segments. The digital signal is provided with a notch filter at 50Hz with the 0.5-100hz bandpass filter. The electrodes are slightly different in each of the subjects ' heads (the size of the distance, or the front or rear, for more details see [1]). The electrodes are set to the EEG grounding in the FZ position.

In addition to the EEG channels, three electrodes were used to include the eye electrical data (2, the left electrode as a reference), and the same amplifier was used for the eye and EEG. The dynamic range is set at +-1MV. It is not necessary to classify the EEG after performing the operation of Pseudo-trace processing.
The screen display paradigm (see Figure 3a) consists of two categories, a left-handed (Category 1) motion picture, a right-handed (category 2) motion picture. Each participant in the two weeks of different days, in the case of not recording feedback, participated in the two session (session) Only the screen display of the test. Each session consisted of 6 groups (run), a group containing 10 (trial), two types of motion imaging. A total of 20 (trail) data are available for each group, with 120 data in a single issue. In general, you can get 120 duplicate data for each action image category for each person. Before the first motion imaging exercise, the participants were asked to perform and visualize the movements of different parts of the body, choosing the best possible action (such as pressing a ball or a pull off).
Each experiment starts with a fixed "+" and adds a short-term sound reminder to the pronunciation (1khz,70ms). After a few seconds, a visual cue (an arrow, left or right, according to the desired category) renders for 1.25 seconds. It was then imagined that the corresponding arm movement lasted for 4 seconds. Rest for at least 1.5 seconds after each experiment is completed. Add at any time within 1 seconds to interrupt the adaptability of the participants.

3 Online feedback periods (sessions) have 4 groups, smiley feedback is recorded (3b), and each group (run) contains 20 experiments in each type of motion picture. At the beginning of each experiment, (0 seconds out) The gray smiley face is in the center of the screen. At 2 seconds, a short reminder appears (1KHZ,70MS). Cues (cue) appear in 3 to 7.5 seconds. According to the clues provided, the subjects were asked to move the smiley face to the left or right by imagining the left or right hand movements. In the feedback period, when the imagination of the direction of movement in line with the smiling face turns green, if done wrong, the smiling face becomes red. The moving distance from the original position of the smiley face is based on the integrated classification output setting (details in [1]) of the last two second motion picture. In addition, the classifier output also draws the smiley corners of the mouth also has a upward downward bend change. At 7.5 seconds, the screen turns white and a random interval of 1-2 seconds is added to the test. The participants were asked to keep the smiley face on the right side for as long as possible, so that the action could be imagined as long as possible.

Data File Description

All data is stored in general data format, as a biomedical signal form (GDF), one file at a time (session). However, only the first three sessions have a classification label for all tests. The remaining two-phase seesion are used to test the classifier and then evaluate the performance of the classifier. All the files are listed in table 1. GDF files can be loaded using the Open source Toolbox Biosig, available for free at http://biosig.sourceforge.net/. Includes library files for Octave/matlab and C + +.

In Octave/matlab, use the Biosig Toolbox to load a GDF file with the following action (for C + +, the corresponding function hdrtype* sopen and size T sread must be CA lled)
[s, h] = sload (' B0101T.GDF ');
Note that the group (run) is split by 100 default values, which are encoded by default to have no number (NaN). Alternatively, the behavior will be stopped and the default data will be encoded into a negative maximum value stored in the file.
[s, h] = sload (' A01T.GDF ', 0, ' Overflowdetection:off ');
The workspace contains two variables, S is the signal, and H is the head structure. The signal variable consists of 6 channels (the first 3 channels are the EEG and the last 3 channels are EOG signals). The header structure contains event information that describes the structure of the data in the time stream. The following describes important information for evaluating this data.

The position of the event in the sample is contained in H. EVENT. Pos. The corresponding type is in H. EVENT. TYP, a specific event interval is stored in H. EVENT. DUR. The type of the data set is recorded in table 2 (16 binary digits, 10 notation in parentheses). Note that category tags (such as the corresponding event types 769 and 770) are only labeled in the training data and are not labeled in the test data.
The test (trials) contains artifacts that will be labeled by the expert as event type 1023. In addition, the h.artifactselection contains a list of all trials (trials). 0 indicates an experiment with no artifacts, and 1 indicates an experiment containing artifacts.
In order to observe the GDF file, the observation and scoring application terminal Sigviewer v0.2 or higher biosig can be used.


Evaluation
Participants must provide sequential output classification results for each sample in the form of a category number, including the marked test (trial) and the test (trial), which is marked as a pseudo-trace. For every point in time, all tests that do not have artifacts (trial) can be used to create a confusion matrix (confusion matrix). From these confusion matrices, we can get the accuracy rate of time flow and kappa coefficient. The Biosig will provide an evaluation algorithm. The winner will be the most kappa-valued algorithm.
The evaluation data set will not be sent until the end of the game. The submitted program must be compatible with the EEG data (the data structure must be the same as that used in all training sets), and the EEG data as input produces the class callout vectors mentioned above.
Because 3 EOG channels are provided, the pseudo-trace processing techniques, such as high-pass filtering or linear regression [4], are used to reject EOG artifacts before processing the data. In order to ensure the use of other remediation methods, we choose the maximum transparency measure to provide the EOG channel; At the same time we require artifacts not to affect the classification effect.
All algorithms must conform to causality, which means that the classification results at K-time are only x_k with current and past sampling; X_ k-1,..., x_0. In order to detect whether the submission algorithm conforms to the causality and the pseudo-trace processing requirements. All entries must be open source, including libraries used, compilers, programming languages, and so on. (such as Octave/freemat, C + +, Python, ...). Using the MATLAB submission algorithm can be developed in a closed-source environment as long as the code can be run on octave. Similarly, C + + programs can be compiled in a Microsoft or Intel environment, but the code must be compiled under g++.


Reference
[1] R. Leeb, F. Lee, C. Keinrath, R. Scherer, H. Bischof, G. Pfurtscheller. Brain-computer Communication:motivation, aim, and impact of exploring a virtual apartment. IEEE transactions on neural Systems and rehabilitation Engineering 15, 473{482, 2007.
[2] M. Fatourechi, A. Bashashati, R. K. Ward, G. E. Birch. EMG and EOG artifacts in Brain computer Interface Systems:a survey. Clinical neurophysiology 118, 480{494, 2007.
[3] A. Schlogl, J. Kronegg, J. E. Huggins, S. G. Mason. Evaluation criteria in BCI. In:g. Dornhege, J. D. R. Millan, T. Hinterberger, D. J. McFarland, K.-r. Muller (Eds.). Toward Brain-computer Interfacing,mit Press, 327{342, 2007.
[4] A. Schlogl, C. Keinrath, D. Zimmermann, R. Scherer, R. Leeb, G.pfurtscheller. A fully automated correction method of EOG artifacts in EEG recordings. Clin.neurophys. jan;118 (1): 98-104.

BCI competition 2008-graz Data set B (Chinese translation)