Limits...
Modern electrophysiological methods for brain-computer interfaces.

Menendez RG, Noirhomme Q, Cincotti F, Mattia D, Aloise F, González Andino S - Comput Intell Neurosci (2007)

Bottom Line: For this data, we show that although the patient received extensive training in mu-rhythm control, valuable information about movement imagination is present on the untrained high-frequency rhythms.This is the first demonstration of the importance of high-frequency rhythms in imagined limb movements. (2) Self-paced finger tapping task in three healthy subjects including the data set used in the BCI-2003 competition.We show that by selecting electrodes and frequency ranges based on their discriminative power, the classification rates can be systematically improved with respect to results published thus far.

View Article: PubMed Central - PubMed

Affiliation: Electrical Neuroimaging Group, Department of Clinical Neurosciences, Geneva University Hospital, 1211 Geneva, Switzerland. rolando.grave@hcuge.ch

ABSTRACT
Modern electrophysiological studies in animals show that the spectrum of neural oscillations encoding relevant information is broader than previously thought and that many diverse areas are engaged for very simple tasks. However, EEG-based brain-computer interfaces (BCI) still employ as control modality relatively slow brain rhythms or features derived from preselected frequencies and scalp locations. Here, we describe the strategy and the algorithms we have developed for the analysis of electrophysiological data and demonstrate their capacity to lead to faster accurate decisions based on linear classifiers. To illustrate this strategy, we analyzed two typical BCI tasks. (1) Mu-rhythm control of a cursor movement by a paraplegic patient. For this data, we show that although the patient received extensive training in mu-rhythm control, valuable information about movement imagination is present on the untrained high-frequency rhythms. This is the first demonstration of the importance of high-frequency rhythms in imagined limb movements. (2) Self-paced finger tapping task in three healthy subjects including the data set used in the BCI-2003 competition. We show that by selecting electrodes and frequency ranges based on their discriminative power, the classification rates can be systematically improved with respect to results published thus far.

No MeSH data available.


Selecting the number of features for the BCI data set. The picture depicts thepercentage (%) of correct classification (CC) on the training set (continuous trace)and the test set (discontinuous trace) as a function of the number of features.The number of features (180) is defined as the beginning of the first plateau, thatis, where increasing the number of features does not increase CC on thetraining set.
© Copyright Policy - open-access
Related In: Results  -  Collection


getmorefigures.php?uid=PMC2233873&req=5

fig5: Selecting the number of features for the BCI data set. The picture depicts thepercentage (%) of correct classification (CC) on the training set (continuous trace)and the test set (discontinuous trace) as a function of the number of features.The number of features (180) is defined as the beginning of the first plateau, thatis, where increasing the number of features does not increase CC on thetraining set.

Mentions: Figure 5 depicts the percentage (%) of correct classification (CC) on the training setand the test set as a function of the number of features. To be compatible withthe information available at the time of the competition, we selected thenumber of features based only on the training set. For the number of features[10, 20, 40, 60, 70, 80, 100, 120, 150, 180, 200], we obtained CC values of[64, 68, 72, 81, 80, 83, 84, 86, 88, 89, 89], respectively. The final number offeatures was selected as 180, corresponding to the value where the CC firststabilizes (reaches a plateau) at a value of 89%. Note that, as happens with linearinterpolation procedures, the CC might still increase with the number offeatures and attain a new plateau for a higher number of features. Nonetheless,for this number of features, the CC is 87% for the test set outperforming the bestresults obtained thus far for this data (i.e., best results are marked as a horizontaldotted line in Figure 5). The plot of the CC for the test set indicates thatthere are better solutions using only 60 or 70 features. At these points, performanceon the test set attains 88% and 89%, respectively.


Modern electrophysiological methods for brain-computer interfaces.

Menendez RG, Noirhomme Q, Cincotti F, Mattia D, Aloise F, González Andino S - Comput Intell Neurosci (2007)

Selecting the number of features for the BCI data set. The picture depicts thepercentage (%) of correct classification (CC) on the training set (continuous trace)and the test set (discontinuous trace) as a function of the number of features.The number of features (180) is defined as the beginning of the first plateau, thatis, where increasing the number of features does not increase CC on thetraining set.
© Copyright Policy - open-access
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC2233873&req=5

fig5: Selecting the number of features for the BCI data set. The picture depicts thepercentage (%) of correct classification (CC) on the training set (continuous trace)and the test set (discontinuous trace) as a function of the number of features.The number of features (180) is defined as the beginning of the first plateau, thatis, where increasing the number of features does not increase CC on thetraining set.
Mentions: Figure 5 depicts the percentage (%) of correct classification (CC) on the training setand the test set as a function of the number of features. To be compatible withthe information available at the time of the competition, we selected thenumber of features based only on the training set. For the number of features[10, 20, 40, 60, 70, 80, 100, 120, 150, 180, 200], we obtained CC values of[64, 68, 72, 81, 80, 83, 84, 86, 88, 89, 89], respectively. The final number offeatures was selected as 180, corresponding to the value where the CC firststabilizes (reaches a plateau) at a value of 89%. Note that, as happens with linearinterpolation procedures, the CC might still increase with the number offeatures and attain a new plateau for a higher number of features. Nonetheless,for this number of features, the CC is 87% for the test set outperforming the bestresults obtained thus far for this data (i.e., best results are marked as a horizontaldotted line in Figure 5). The plot of the CC for the test set indicates thatthere are better solutions using only 60 or 70 features. At these points, performanceon the test set attains 88% and 89%, respectively.

Bottom Line: For this data, we show that although the patient received extensive training in mu-rhythm control, valuable information about movement imagination is present on the untrained high-frequency rhythms.This is the first demonstration of the importance of high-frequency rhythms in imagined limb movements. (2) Self-paced finger tapping task in three healthy subjects including the data set used in the BCI-2003 competition.We show that by selecting electrodes and frequency ranges based on their discriminative power, the classification rates can be systematically improved with respect to results published thus far.

View Article: PubMed Central - PubMed

Affiliation: Electrical Neuroimaging Group, Department of Clinical Neurosciences, Geneva University Hospital, 1211 Geneva, Switzerland. rolando.grave@hcuge.ch

ABSTRACT
Modern electrophysiological studies in animals show that the spectrum of neural oscillations encoding relevant information is broader than previously thought and that many diverse areas are engaged for very simple tasks. However, EEG-based brain-computer interfaces (BCI) still employ as control modality relatively slow brain rhythms or features derived from preselected frequencies and scalp locations. Here, we describe the strategy and the algorithms we have developed for the analysis of electrophysiological data and demonstrate their capacity to lead to faster accurate decisions based on linear classifiers. To illustrate this strategy, we analyzed two typical BCI tasks. (1) Mu-rhythm control of a cursor movement by a paraplegic patient. For this data, we show that although the patient received extensive training in mu-rhythm control, valuable information about movement imagination is present on the untrained high-frequency rhythms. This is the first demonstration of the importance of high-frequency rhythms in imagined limb movements. (2) Self-paced finger tapping task in three healthy subjects including the data set used in the BCI-2003 competition. We show that by selecting electrodes and frequency ranges based on their discriminative power, the classification rates can be systematically improved with respect to results published thus far.

No MeSH data available.