This repository contains an algorithm that is able to classify the four class motor imagery EEG data from the BCI Competition IV, data set 2a (http://www.bbci.de/competition/iv/#dataset2a).
The algorithm reaches 0.46 kappa on the test set, so it would be able to reach the third place in the competition (http://www.bbci.de/competition/iv/results/index.html#dataset2a).
It was developed as a part of my master's thesis:
Piotr Szachewicz, "CLASSIFICATION OF MOTOR IMAGERY FOR BRAIN-COMPUTER INTERFACES",
Poznan University of Technology, Poznan, 2013
How to install and use this code:
- INSTALL PREREQUISITES
sudo apt-get install octave octave-signal liboctave-dev octave-tsa octave-struct octave-specfun octave-optim octave-miscellaneous
- CLONE REPOSITORIES
Clone this repository and a repository containing BioSig's code (I forked this code in order to fix some errors in it).
cd ~
git clone https://github.com/piotr-szachewicz/event-related-desynchronization.git
git clone https://github.com/piotr-szachewicz/biosig.git
- CONFIGURE
In the event-related-desynchronization directory there is a settings_example.m file which contains paths used by the code. If you download the code to your home directory (~) using the git clone commands written above, you only need to copy the settings_example.m file to settings.m:
cd event-related-desynchronization/code
cp settings_example.m settings.m
- DOWNLOAD THE EEG DATA
Download the BCI Competition data from http://www.bbci.de/competition/iv/#datasets and put the *.gdf files inside of the ~/event-related-desynchronization/data/2a directory.
Download the true labels for the 2a data set from http://www.bbci.de/competition/iv/results/index.html#labels. Unpack it and put the *.mat files inside of the ~/event-related-desynchronization/data/2a/true_labels/ directory.
- RUN EXPERIMENTS
You are ready to run your first experiment.
cd ~/event-related-desynchronization/code/experiments
octave lda_bp_experiment.m
This code runs the final evaluation experiment using an band power features (BP) and an LDA classifier. The mean test kappa obtained should be around 0.46. This is the best algorithm I obtained in my master's thesis.
Other files contain:
- lda_tdp_experiment - LDA classifier + TDP features
- svm_bp_experiment - SVM classifier + BP features
- svm_tdp_experiment - SVM classifier + TDP features
In order to run the experiments that use the SVM classifier you need to install some more things:
sudo apt-get install swig
Download and unpack http://shogun-toolbox.org/archives/shogun/releases/2.1/sources/shogun-2.1.0.tar.bz2, then go to the src/ directory and run the following commands:
./configure --interfaces=octave_modular
make
sudo make install