Instructions



The task in this benchmarking approach is to define a method extracting the information, whether the subject has pressed a button with the left or the right hand, from the recorded EEG data. This shall be done on a specific amount of data related to the event of the buttonpress and should respect causality. Hence, one has to take care that no information from before or after the specified time range is used for the estimation of the class information.

The basic idea is to train a classifier on labelled data (including class information for each trial, marker ‘S  4’ for left hand movements, marker ‘S  8’ for right hand movements) and test it on unlabelled data (only containing markers (‘response’) for events not naming the class of trials). The same type of preprocessing has to be applied on data from both blocks.

Classification results from three categories of data are relevant for our benchmarking approach. Each category is defined by a time window. It indicates the amount of data, which can be taken into account for training the classifier (on trials of block 1) and applying it to previously unseen data (block 2). Any subset of data points, including the set containing all data from each of the first 32 electrodes and within the specified time window, can be processed. This includes preprocessing (again, be careful with filter methods and check causality) and feature extraction.

Categories:

  1. Cat. 1: -500 ms to -200 ms
  2. Cat. 2: -500 ms to 0 ms
  3. Cat. 3: -500 ms to +200 ms


A result for one category consists of:

  • the estimated classification accuracy (in percent with two digits after the decimal point) of the defined method. This should be based on a 10 times repeated 10x(5x) (nested, if parameter search is included) crossvalidation on block 1. It can be checked for dependency on artifacts (muscle, eye) by comparison to the auxiliary channels (33 to 36).
  • a label vector consisting of an estimation of the class label (-1 for left hand movements, +1 for right hand movements) for each trial of block 2.
  • a detailed description of the used preprocessing
  • a detailed description of the used feature extraction method
  • a detailed description of used classification algorithm
  • a statement of the causality of used methods

 

If you want to publish your results on this website, please send them to mga(at)mms.tu-berlin.de