Article Figures & Data
Figures
- Figure 1.
Conceptual illustration of neural adaptation process induced by brain-computer interfacing. A, Setup of a brain-computer interface. Online acquired scalp electroencephalograms were fed into a classifier to detect the presence/absence of attempted movement. Predicted brain state was shown to participants as movement of visual object on display. B, Conceptual visualization of cortical adaptation. Scaling adaptation reflects improvement in voluntary regulation of a specific component. If the centers of gravity determined from datapoints in two conditions are separated after brain-computer interfacing, it suggests the separability of two conditions is enhanced by adaptation. Deforming adaptation suggests that activity patterns are allocated to a specific brain state to adapt to the classifier. If the geometric relationships between two conditions are deformed with respect to a specific axis, it suggests the adaptation process progressed such that the two conditions are separated along the axis.
- Figure 2.
Experiment setup and protocol. A, Electrode locations. The three classifiers used in the study had different channels of interest. The model-based classifier used only channel C3 indicated in blue around the left sensorimotor cortex. The adaptive classifier used whole-head EEG channels (purple) to construct a common spatial pattern. The de novo classifier used only the Cz channel, shown here in green. B, Experimental protocol and time course of a trial. C, Visual feedback object. For the model-based or adaptive classifiers, an illustration of a hand was shown that matched the attempted movements of the users while an illustration of a tail was used in the de novo task to encourage users to acquire novel mental actions that enhanced controllability of the BCI.
- Figure 3.
Low dimensional visualization of EEG data by t-SNE. A, Changes in geometric relationships between dataset and classifier plane. As training progressed, the geometric relationship of points from two brain states changed with respect to the classifier plane (black plane). The large points indicate the centers of gravity of points from each brain state. The black line orthogonal to the classifier plane is the classifier normal vector (see also Figure 3C). B, An example of t-SNE-based data visualization in embedded space (Model-based classifier user). Each datapoint is colored with its SMR-ERD value derived from the C3 electrode around the left sensorimotor cortex. The black plane represents the classifier plane (see also Classifier plane and geometric assessment of EEG data for mathematical details). The large points indicate the centers of gravity of points from each brain state. The black line orthogonal to the classifier plane is the classifier normal vector (see also Figure 3C). C, The t-SNE-based quantification of the adaptation process with respect to the classifier plane. tNormp is defined as a component of tVec with respect to the classifier vector, while θp is defined as a subtended angle between tVec and the classifier vector.
- Figure 4.
Changes in BCI operation performance and time-frequency representations of scalp electroencephalogram signals. A, Group results of performance scores from users of model-based, de novo and adaptive classifiers. Solid lines indicate mean values while shaded areas represent 1 SE across participants. B, Changes in the acquired scores during BCI operation. Asterisks indicate statistical significance (p < 0.05). C, Changes in time-frequency representations of scalp electroencephalogram signals from representative channels.
- Figure 5.
Spatial activity patterns during brain-computer interfacing. Results of source estimation analysis from representative participants. The colored regions indicate voxels where activities were significantly different during Rest and Imagine periods (p < 0.05 unc.). Areas colored with blue and green indicate those for model-based and de novo classifiers, respectively. While significant voxels were localized around the contralateral hemisphere of the imagined hand for the model-based classifier, those for the de novo classifier were located bilaterally, including in the pre/postcentral gyrus and supplementary motor area (peak voxel was in the postcentral gyrus; MNI coordinates: −40, −25, 45). Note that a representative source estimation for the adaptive classifier is not shown because of variable activity patterns among participants. sLoreta analyses of statistical nonparametric mapping for estimated cortical sources of band power in the α band (8–13 Hz). Areas colored with blue and green indicate those from model-based and de novo classifiers, respectively. Masks superimposed on a standard brain template were visualized by MRIcroGL (https://www.mccauslandcenter.sc.edu/mricrogl/home).
- Figure 6.
Overall changes in distance between brain states. Changes over time in the norm of tVec for participants operating under the model-based classifier (A), the de novo classifier (B), and the adaptive classifier (C). Asterisks indicate statistical significance (p < 0.05).
- Figure 7.
Quantitative comparison of cortical adaptation processes in embedded. Changes over time in
In this issue
