A Review of Classification Algorithms for EEG-based Brain-Computer Interfaces

Conclusions

This paper has surveyed classification algorithms used to design Brain-Computer Interfaces (BCI). These algorithms were divided into five categories: linear classifiers, neural networks, nonlinear Bayesian classifiers, nearest neighbor classifiers and combinations of classifiers. The results they obtained, in a BCI context, have been analysed and compared in order to provide the readers with guidelines to choose or design a classifier for a BCI system. In a nutshell, it seems that SVM are particularly efficient for synchronous BCI. This probably is due to their regularization property and their immunity to the curse-of-dimensionality. Furthermore, combinations of classifiers and dynamic classifiers also seem very efficient in synchronous experiments.

This paper focused on reviewing classifiers used in BCI research, i.e., related to published online or offline studies. However, other existing classification techniques, not currently used for BCI purposes, could be explored and may prove to be rewarding. Furthermore, it should be noted that once BCI will be more widely used in clinical practice, new properties will have to be taken into consideration, such as the availability of large data sets or long term variability of EEG signals.

One difficulty encountered in such a study concerns the lack of published objective comparisons between classifiers. Ideally, classifiers should be tested within the same context, i.e., with the same users, using the same feature extraction method and the same protocol. Currently, this is a crucial problem for BCI research. For this reason some researchers have proposed general purpose BCI systems such as the BCI2000 toolkit [80]. This toolkit is a modular framework which makes it possible to easily change the classification, preprocessing or feature extraction modules. With such a system it becomes possible to test several classifiers with the same features and preprocessings. With similar objectives of modularity, the Open-ViBE platform [81] proposes a framework to experiment BCI on various protocols using, for instance, neuro-feedback and Virtual Reality. An extensive use of such platforms could lead to interesting findings in the future.