Brain-Computer Interfaces (BCI)

 

Problem's Description

The aim of EEG-based Brain-Computer Interfaces (BCI) is to detect patterns in user EEG signals in order to control a computer or an external device. As an example, patterns corresponding to motor imagery (the user imagines that one of his body parts is moving), object rotation imagery or thinking of words, can be recognized in the EEG signal. The high variability of EEG patterns among different subjects makes machine learning classification techniques the tool of choice. Thus, classifiers can be trained from user-generated data by means of supervised machine learning techniques. The classifier can then be used to detect patterns on real-time EEG data.

However, it is very difficult to learn classifiers from the raw EEG data for two reasons. First, the number of attributes is too large and second, it is known that patterns are best detected on the frequency-domain rather than in the time-domain of the original raw data. Therefore, the raw EEG signal is usually preprocessed before the training stage.

An appropriate preprocessing of the signal is acknowledged to be very important in order to get high classification accuracy. For instance, figures show class distribution of a three-class problem with respect to the two attributes most correlated with the class. The image on the left corresponds to the application of just the FFT (Fast Fourier Transform) filter and the image on the right with the FFT and a spatial filter. It also shows the accuracy of the classifier (percentage). It can be noticed that the use of appropriate filters may affect the quality of the attributes used for building the classification.

Three kinds of transformations are commonly used: spatial filters, the Fourier Transform (to convert to the frequency domain), and band-pass filters. These filters are typically adjusted by hand, following a process of trial and error. This requires some experience on building and adjusting filters that at the end, may turn out not to be optimal.

Our interest here is related to computing filters automatically (both spatial and frequency-selection filters) by means of multi-objective evolutionary optimization algorithms. A mono-objective evolutionary algorithm has been already used to solve the problem of computing filters automatically. In that work, the fitness function is defined as:

F=Error+µ*|B|

where Error is the classification error and |B| is the number of frequency-bands selected.

The proposed work is to study the success in the classification task using a multi-objective approach versus the mono-objective perspective. Also, the efficiency of different multi-objective evolutionary algorithms will be studied.

The first multi-objective approach is formulated by means of the following two objectives (although other objectives could be also defined):

F1=Error

F2=|B|

Documents

  • A general description of the whole system (iSpanish) is here.
  • A general description and a detailed explanation to compute the fitness function can be found here.

Data sets

Three data sets acquired in the IDIAP Research Institute will be used (Millán, 2004). These data sets have been previously used in the 2005 BCI-III Competition: http://www.bbci.de/competition/iii/.

Each data set contains data from a different subject (subject1, subject2 and, subject3) during 4 non-feedback sessions (01, 02, 03 and, 04) . 32 electrodes were located on the subjects's scalp. There are 3 mental tasks, so this is a three-class classification problem:

  • Imagination of repetitive self-paced left hand movements
  • Imagination of repetitive self-paced right hand movements
  • Generation of words beginning with the same random letter

All 4 sessions of a given subject were acquired on the same day, each lasting 4 minutes with 5-10 minutes breaks in between them. The subject performed a given task for about 15 seconds and then switched randomly to another task at the operator's request. EEG data is not splitted in trials since the subjects are continuously performing any of the mental tasks.

For each subject, data from sessions 01+02+03 are used to compute both the filters and the classifier. Data from session 04 are used as test data.

  Training Testing
Subject1 train_subject1_raw01.asc+ train_subject1_raw02.asc+ train_subject1_raw03.asc train_subject1_raw04.asc
Subject2 train_subject2_raw01.asc+ train_subject2_raw02.asc+ train_subject2_raw03.asc train_subject2_raw04.asc
Subject3 train_subject3_raw01.asc+train_subject3_raw02.asc+ train_subject3_raw03.asc train_subject3_raw04.asc

 

References

(Millán, 2004) J. del R. Millán. On the need for on-line learning in brain-computer interfaces. In Proc. of the International Joint Conference pn Neural Networks. July, 2004. Hungary IDIAP-RR 03-30.

Guido Dornhege and José del R. Millán and Thilo Hinterberger and Dennis J. McFarland and Klaus Robert Müller (eds). Toward Brain-Computer Interfacing. MIT Press, 2007