Background: Brain-computer interfaces (BCIs) can provide alternative, motor-independent means of communication for people who have lost motor function. A promising variant is the functional magnetic resonance imaging (fMRI)-based BCI, which exploits information on hemodynamic brain activity evoked by performing different mental tasks. However, due to the sluggish nature of the hemodynamic response, a current challenge is to make these BCIs as efficient and fast as possible to allow useful clinical application. Furthermore, there is yet no consensus on optimal mental-task selection for multi-voxel pattern analysis-based decoding, nor whether certain tasks generalize well across users, or if individualized task selection would yield a higher decoding accuracy. Methods: To increase BCI efficiency, we tested whether distributed patterns of 3T-fMRI brain activation evoked by two-second mental tasks could be reliably discriminated in 2- to 7-class classification. In addition, we identified optimal mental-task combinations for high-accuracy classification across all classes. Finally, we examined whether individualized task selection-based on subjects' previous decoding performance (accuracy-based tasks) or their subjective preference (preference-based tasks)-was superior to the other in a yes/no communication paradigm. Results: The 2-class decoding resulted in a mean accuracy of 78% and 3- to 7-class accuracies were above chance level. Mental calculation and spatial navigation were most frequently associated with the highest decoding accuracy. Furthermore, subjects could encode yes/no answers using their accuracy-based and preference-based tasks with mean accuracies of 83% and 81%, respectively. This implies that this paradigm, using short encoding durations, is well-suited to the diversity of patients and could greatly increase BCI efficiency.
Keywords: brain-based communication; brain–computer interface; mental imagery; multi-voxel pattern analysis; multiclass classification; syndrome; “locked-in”.