Classifying EEG Signals

(Original picture here )


Description:

      Human-Computer interaction is a very important issue that has been addressed in many ways, from graphical user interfaces, to joystick control for games. The Brain Computer Interface (BCI) is an attempt to communicate directly a brain with a computer. This is of the utmost interest for people with severe motor disabilities, who cannot use the standard communication devices like keyboards or mouses. Most usually, the BCI relies on non-invasive EEG (electroencephalogram) electrodes, which are attached to the scalp. The electrodes detect the EEG signals related to motor intentions, like the preparation to move the left hand, or just imagining making such movement. Once the EEG signal has been decoded, it can be used to move a cursor on the screen, or to execute commands in a computer. For insntance, the intention to move the right hand can be used to move the cursor to the right, and so on.

 

      Decoding the EEG signal is not a straightforward task. The signal is very weak and many artifacts can be present (just blinking an eye may add noise to the signal). But most importantly, there is no simple function to map EEG signals to intentions. In addition, the mapping function can change from person to person, or even for the same person on different days. A common approach is to inductively learn the mapping function from hundreds of labelled data. For this task, inductive algorithms like Neural Networks or Support Vector Machines, can be used.

Click here to get this description in tex format and here to get the figure in eps format.

Instances and best known solutions for those instances:

      Several BCI competitions have already taken place. The problem described in this section was posed in the NIPS 2001 Brain Computer Interface Workshop . The problem was named "EEG self-paced key typing" and the dataset was courtesy of Blankertz and Mueller (Fraunhofer FIRST), and Curio, (FU-Berlin) [BCM01].

EEG signals were recorded from one subject in 3 sessions with some minutes breaks inbetween. The subject was sitting in a normal chair, relaxed arms resting on the table, fingers in the standard typing position at the computer keyboard (index fingers at 'f','j' and little fingers at 'a',';'). The task was to press the forementioned keys with the corresponding fingers in a self-chosen order and timing ('self-paced key typing'). The average speed was 1 key every 2.1 seconds. Brain activity was measured with 27 Ag/AgCl electrodes referenced to nasion at 1000 Hz using a band-pass filter from 0.05 to 200 Hz. Data has also been downsampled to 100Hz simply by picking each 10th sample.

The supplied data consists of 27 EEG channels. Windows 1500 ms long were cutted out of the continuous raw signals each ending at 120 ms before the respective keystroke. The reason for choosing the endpoint at -120 ms is that from that point on there is EMG activity in a significant number of trials [BCM01].

The classification task is to create a classifier to predict which of the two hands (left or right) the user intends to use (before actually pressing the key).

Results from the competition can be found here . The best result was obtained by P. Sottas using a recurrent neural network optimized by dynamic noise annealing [SG02] . 96% accuracy was obtained on the test set. 95\% accuracy was obtained by other approaches [SG02] . Estébanez et al. obtained 94.2% accuracy by means of a Genetic Programming approach [EVA05] . The following table summarizes current results on this problem:

Algorithm

Author

Classification Rate

Recurrent Neural Network Optimized by Dynamic Noise Annealing

Sottas [SG02]

96%

Common Spatial Subspace Decomposition (CSSID) and Multiple Electrode Activity Subtaction (MEAS) Linear Discriminant Analysis

Shangkai Gao

95%

Slow Potential Shift nu-Support Vector Classifier with CV Criterion Parameter Tuning

Rosipal, Trejo, and Wheeler

95%

Genetic Programming Based Data Projection and Linear Discrimination

Estebanez [EVA05]

94%

Decision Tree and Neural Network Classifier

Dam, Tosevski, Belista, and El-Ali

87%

AutoRegressive Model with eXogenous (ARX) Linear Discriminant Analisys

Burke, Kelly, Chazal, and Reilly

78%

Two files are provided:
  • selfpaced2s_aa01.zip (27 channel, 151 instants, 100Hz (517 records)) : It contains two files:
    • selfpaced2s_aa01.txt: It includes the whole 517-dataset, downsampled at 100Hz. For every of the 27 channels, 151 instants of the signal are provided. The format of the data is a matrix with 4077 rows x 517 columns. The first 151 rows correspond to the first channel, the next 151 to the second one, etc. ( 27 channels * 151 instants/channel = 4077) Bear in mind that there is a column for every data record. In order to use this dataset with tools like Weka , the matrix should have to be transposed first. .
    • selfpaced2s_aa01_labels.txt: Labels are provided in a separate file. It contains two lines with 517 columns. The first one gives the class (0 or 1) for each one of the 517 records of the previous file.
  • selfpaced_description.txt: Full description of the data.

selfpaced2s_aa01.zip selfpaced_description.txt

Related Papers:

[BCM01] Benjamin Blankertz, Gabriel Curio, and Klaus-Robert Müller. Classifying single trial EEG: Towards brain computer interfacing. In T. G. Diettrich, S. Becker, and Z. Ghahramani, editors, Advances in Neural Inf. Proc. Systems (NIPS 01), volume 14, pages 157-164, 2002

[SG02] Sottas, P, Gerstner W. Dynamic noise annealing for learning temporal sequences with recurrent neural networks. International Conference on Artificial Neural Networks (ICANN2002). LNCS 2415. 1144-1149. 2002.

[SG02] Paul Sajda, Adam Gerson, Klaus-Robert Muller, Benjamin Blankertz, Lucas Parra. A Data Analysis Competition to Evaluate Machine Learning Algorithms for use in Brain-Computer Interfaces. IEEE Transactions on Neural Systems and Rehabilitation Engineering, Vol 11(2).184-185. 2003.

[EVA05] César Estébanez, José M. Valls, Ricardo Aler. A first attempt at constructing genetic programming expressions for EEG classification. International Congress on Artificial Neural Networks (ICANN'05). Special Session on Information-theoretic Concepts in Biomedical Data Analysis. Warsaw (Poland). 2005

Click here to get the bibliography in bibtex format.

Last Updated: 1/7/05                                                                               For any question or suggestion, click here to contact with us.