A brain-computer interface (BCI) for the locked-in: comparison of different EEG classifications for the thought translation device.

OBJECTIVE

The Thought Translation Device (TTD) for brain-computer interaction was developed to enable totally paralyzed patients to communicate. Patients learn to regulate slow cortical potentials (SCPs) voluntarily with feedback training to select letters. This study reports the comparison of different methods of electroencephalographic (EEG) analysis to improve spelling accuracy with the TTD on a data set of 6,650 trials of a severely paralyzed patient.

METHODS

Selections of letters occurred by exceeding a certain SCP amplitude threshold. To enhance the patient's control of an additional event-related cortical potential, a filter with two filter characteristics ('mixed filter') was developed and applied on-line. To improve performance off-line the criterion for threshold-related decisions was varied. Different types of discriminant analysis were applied to the EEG data set as well as on wavelet transformed EEG data.

RESULTS

The mixed filter condition increased the patients' performance on-line compared to the SCP filter alone. A threshold, based on the ratio between required selections and rejections, resulted in a further improvement off-line. Discriminant analysis of both time-series SCP data and wavelet transformed data increased the patient's correct response rate off-line.

CONCLUSIONS

It is possible to communicate with event-related potentials using the mixed filter feedback method. As wavelet transformed data cannot be fed back on-line before the end of a trial, they are applicable only if immediate feedback is not necessary for a brain-computer interface (BCI). For future BCIs, wavelet transformed data should serve for BCIs without immediate feedback. A stepwise wavelet transformation would even allow immediate feedback.