Data set IVc
‹motor imagery, time-invariance problem›
Data set provided by Fraunhofer FIRST, Intelligent Data Analysis Group
(Klaus-Robert Müller, Benjamin Blankertz), and
Campus Benjamin Franklin of the Charité - University Medicine Berlin,
Department of Neurology, Neurophysics Group (Gabriel Curio)
Correspondence to Benjamin Blankertz
When taking a machine learning approach to Brain-Computer Interfacing,
the user usually has to perform a calibration measurement in the
beginning of a BCI experiment which provides the training data.
After that the user should be able to control BCI feedback applications
as long as s/he wants. With powerful algorithms the classification
performance on the training data is ofter better when using complex,
i.e., high-dimensional features. But these features may be affected
by signal characteristics that slowly change over time, and accordingly
the quality of the BCI feedback can degrade over time. When training
data is only available from a short time span at the beginning of a
measurement it is difficult to make the algorithm invariant to such
disturbances. This data set poses the challenge of finding a
classification that works on test data that was recorded several
hours after the training session.
This data set was recorded from one healthy subject. He sat in a
comfortable chair with arms resting on armrests. The training data set
consists of the first 3 (non-feedback) sessions. (It is the same as the
training data of data set IVb). Visual cues
(letter presentation) indicated for 3.5 seconds which of the following
3 motor imageries the subject should perform: (L) left hand,
(F) right foot, (Z) tongue (=Zunge in german). The
presentation of target cues were intermitted by periods of random
length, 1.75 to 2.25 seconds, in which the subject could
The test data was recorded more than 3 hours after the training
data. The experimental setup was similar to the training sessions,
but the motor imagery had to be performed for 1 second only,
compared to 3.5 seconds in the training sessions. The intermitting
periods ranged from 1.75 to 2.25 seconds as before. The other
difference was that the class tongue was replaced by the
class relax. The reason for including the relax
class into the test data without having training examples for it
is the same as for data set IVb, see there.
Format of the Data
Given are continuous signals of 118 EEG channels and, for the
training data, markers that indicate the time points of 210 cues and
the corresponding target classes. Only cues for the classes
left and foot are provided for the competition (since
tongue imagery was not performed in the test sessions).
Data are provided in Matlab format (*.mat) containing
info: structure providing additional information with fields
- cnt: the continuous EEG signals, size [time x channels].
The array is stored in datatype INT16. To convert it to
uV values, use cnt= 0.1*double(cnt); in Matlab.
- mrk: structure of target cue information with fields
(the file of test data contains only the first field)
- pos: vector of positions of the cue in the EEG signals given in
unit sample, length #cues
- y: vector of target classes (-1 for left or 1 for
foot), length #cues
As alternative, data is also provided in zipped ASC II format:
- name: name of the data set,
- fs: sampling rate,
- clab: cell array of channel labels,
- xpos: x-position of electrodes in a 2d-projection,
- ypos: y-position of electrodes in a 2d-projection.
- *_cnt.txt: the continuous EEG signals, where each
row holds the values for all channels at a specific time point
- *_mrk.txt: target cue information, each row represents one cue
where the first value defines the time point (given in unit sample),
and the second value the target class (-1 for left or 1 for
foot). The file of test data only contains time points.
- *_nfo.txt: contains other information as described for the
Requirements and Evaluation
Please provide an ASC II file (named 'result_IVc.txt') containing 420
lines of your classifier output (real number between -1 and 1) for
each trial of the test data set. The performance criterium is the mean
squared error with respect to the target vector that is -1 for class
left, 1 for foot, and 0 for relax, averaged
across all trials of the test set. Note that there are no training
samples for the class relax, see also the introductory
paragraph of the description of data set IVb.
This class must be defined by the absence of mental states
left and foot. The motivation for this performance
measure is that the idea is to have a system that is suitable for one
dimensional cursor control, see also the
description of data set IVb.
You also have to provide a description of the used algorithm (ASC II,
HTML or PDF format) for publication at the results web page.
The recording was made using BrainAmp amplifiers and a 128 channel
Ag/AgCl electrode cap from ECI. 118 EEG channels were measured at
positions of the extended international 10/20-system. Signals were
band-pass filtered between 0.05 and 200 Hz and then digitized at
1000 Hz with 16 bit (0.1 uV) accuracy. We provide also a version
of the data that is downsampled at 100 Hz (by picking each 10th
sample) that we typically use for analysis.
Note that the above reference describes an older experimental setup.
A new paper analyzing data sets similar to the one provided in this
competition and presenting feedback results will appear soon.
- Guido Dornhege, Benjamin
Blankertz, Gabriel Curio, and Klaus-Robert Müller.
Boosting bit rates in non-invasive EEG single-trial classifications by
feature combination and multi-class paradigms.
IEEE Trans. Biomed. Eng., 51(6):993-1002, June 2004.
[ BCI Competition III ]