Brain-computer interface speller system design from electroencephalogram signals with channel selection algorithms.

BACKGROUND AND OBJECTIVE

Brain-computer interfaces (BCI) have started to be used with the development of computer technology in order to enable individuals who are in this situation to communicate with their environment or move. This study focused on the spelling system that transforms the brain activities obtained with EEG signals into writing. In BCI systems working with P300 obtained from 64 electrodes, data recording and processing cause high cost and high processing load. By reducing the number of electrodes used, the physical dimensions, costs, and processing loads of the systems can be reduced. The main problem at this stage is to determine which electrodes are more effective. Randomness-based optimization methods perform their experiments within the framework of a specific fitness function, resulting in near-best results rather than the best result. The electrodes chosen as a result of the study are expected to contribute positively to the classifier performance. At the same time, an unbalanced data set is balanced, and an increase in system performance is expected.

METHOD

Electrode selection was performed in both the original dataset and ADASYN dataset using the Genetic Algorithm and Binary Particle Swarm Optimization methods. As a dataset, Wadsworth BCI Dataset (P300 Evoked Potentials) was used in the study. The channels chosen most frequently by optimization methods were determined and compared with the 64-channel classification results using LS-SVM and LDA.

RESULT

As a result of the optimization processes, the eight channels selected most frequently, the channels selected more than the average of all the selected channels and 64 channel results were compared. The highest accuracy was achieved with the LDA classifier for user A with 29 channels selected with BPSO with 97.250%.

CONCLUSIONS

The results obtained in the study showed that the number of channels decreased by optimization methods increases the classification performance. In addition, classifier training and test times have been greatly reduced. The application of the ADASYN method did not result in any significant difference.