Wednesday, September 17, 2008



In this work six EEG-based brain computer interface systems were reviewed and compared. Experiments lasting five days with three subjects were done with the new Adaptive Brain Interface system.
The comparison of the BCI systems, especially their training duration and performance, proved to be difficult. This was because the results were reported inadequately and differently in most of the papers. Reporting the experiments and results should be standardized.
The results from each recording or training day should be presented in order to see the evolution of the performance. Instead of hit rates and correct classification rates, the results should be presented using confusion matrices and channel capacities. This would make the comparison of the performances possible.
In this work the BCI systems were divided into the pattern recognition and the operant conditioning approaches. From the two approaches, the pattern recognition approach seems more plausible. Compared to the operant conditioning approach, the training duration is much shorter. However, the high variability in the EEG between the days and and changing EEG patterns during the actual use cause problems with this approach. This means that the classifier needs to re-trained often. In the future, online learning could be used, in which the classifier is updated after every recorded EEG sample.
Accuracy, speed, usability and feedback methods should be improved in the current BCI systems. Accuracy is the most important and affects greatly on the performance of the BCI. Many of the BCI systems are operated in a synchronous way, using trials lasting many seconds each. This means that time required for making one selection is long. This time should be kept short (below one second). Feedback methods could be improved, maybe using games like in the EEG biofeedback. Some of the mental tasks used in the ABI and the experiments in this work are not good. The relax task is the easiest to classify,
but it includes eye opening and closing, which is not permitted in a BCI by the definition presented in the beginning of the second chapter. It can be argued if people suffering from locked-in-syndrome can use the relax task. In addition, it is not good in applications, because eyes are closed. Subtraction, word association and cube rotation tasks are not very natural and practical in applications. The left and the right hand movement are the most natural of the current tasks.
In the future, an exhaustive research about the mental tasks should be done. A study of the left and right hand movements using high-resolution EEG and MEG is planned. Research topics would include the localization of the brain activity during the mental tasks and how the EEG changes in process of time. Other research areas would be feedback methods and online learning.
There are many challenges in the future of the BCI field. Currently none of the BCIs are capable of proper cursor control, which could be used to control ordinary computer applications. In the near future it is not possible and special applications must be developed for BCIs. Today, special writing applications or Internet browser can provide communication tools for severely disabled people.
These applications could be improved. In the future, BCIs could be used to control a hand prosthesis. How well that can be achieved with EEG-based BCIs is not yet known. Non-invasive BCIs recording activity directly from the motor cortex may be used for this kind of purpose in the future.