What BCI research needs

The state of the art of Brain Computer Interfacing, though promising, is still far from what one would need for fast and reliable control of games and interfaces. A mature non-invasive systems is e.g. the discrete P300 visual speller, in which alternating rows and columns of a letter matrix flash and gaze and attention is directed by the user to a specific letter (Sellers & Donchin, 2006). The EEG signal is analysed, and in a handful of flashes the selection can be determined with high accuracy. But even in optimizations of the flash regime (Hill, Farquhar, Martens, Biessmann & Schölkopf, 2007) the maximum data rate is still very low. And this BCI may not be independent, i.e. it needs gaze direction and cannot be controlled by mental activity alone. A second example of a mature BCI, for the continuous case, is the neuro-feedback trained slow cortical potential or e.g. theta-power for cursor control. (Birbaumer, 1977, Birbaumer et al., 1999). This BCI is also only able to achieve a very modest data rate of control in one dimension. Why is it that, in all these years of development, not more progress has been achieved? What makes the problem so complex and the signals that hard to decipher? Should we continue to build upon these examples and try to improve them? Or would we need new approaches? What are the fundamental issues that need addressing? Below we identify a series of topics that need consideration and may lead to fundamentally different, and better, brain computer interfaces.