Evaluation of the continuous detection of mental calculation episodes as a BCI control input

This paper presents an evaluation of the continuous detection of mental calculation episodes, which may be useful for users who strive to operate current BCI paradigms or even for augmenting degrees of freedom. The experimentation consisted in the alternated realization of basic arithmetic mental calculations and resting periods. EEG data were analyzed using sliding windows of 2s length. The experimental population was comprised of fifteen healthy subjects who participated in three sessions on different days. The features used for the classification process were the power spectral density over the beta band ([14-35] Hz) and the scaling exponent obtained via detrended fluctuation analysis. Both indices were estimated over four channels, specifically selected for each subject. The performance was evaluated using the Area Under the ROC Curve (AUC) by measuring the overall classification performance of each experimental session with a cross-validation procedure, and by transferring the model obtained from one session to the others called inter Session Validation (iSV). The best AUC values computed in each cross-validation session were: 0.87±0.067, 0.89±0.056 and 0.88±0.040 respectively; and the iSV provided a value of 0.67±0.122. These high values indicate that a mental calculation paradigm and a combination of features can efficiently control a BCI system. Notwithstanding that several days passed between sessions, the AUC mean value estimated for the iSV is similar to the performance of a motor imagery-BCI calibrated on the same day.