Table 1.
Demographic and disease related data of all patients.
Figure 1.
The experimental design is shown in plot (A).
Plot (B) depicts the architecture of the flexible BCI system which simultaneously considers oscillatory features and slow potentials. Two classifiers are applied and the feedback application is receiving simultaneous output of both classifiers and their weighted combination. A screen shot of the “Connect-4” application in mode FR (foot vs. right hand) is plotted in (C). In the top-left corner, the cue is presented (an arrow pointing to the right) and based on the BCI output, the yellow bar is either extending rightwards or downwards. The rightmost column is currently selected and visually highlighted.
Figure 2.
Standard physiological screening of the four patients.
The top row shows the spectra at electrode ‘’ in the conditions eyes-open and eyes-closed. The spatial distribution of the channel-wise spectral power in the alpha-band [8–12 Hz] is depicted in the scalp maps of the lower row.
Figure 3.
Discriminative power of each feature across sessions, obtained with offine reanalysis of the CopyTask data.
Global parameters such as the frequency band and time interval were chosen individually for each patient after manually inspecting the data from all sessions. For each session, the same global parameters were taken – which might be suboptimal. The classification accuracy was then estimated with cross validation using the same parameters for each session. Note that the number of trails was varying across sessions with later sessions featuring less trials. Moreover, a β rebound was defined to as a discriminative feature in the β band, which was observed more than 500 ms after the end of a trial. As the β ERD of patient 4 was heavily delayed, it is also considered as β rebound in this analysis. Fig. S2 shows the corresponding spatial distribution of discriminative information as scalp maps.
Figure 4.
Binary online accuracies (left column) and estimated bit rates (middle column) in the CopyTask for patients 1–3.
Each bar represents one block of at least 20 trials. Session numbers are specified in blue color (left column). Session numbers with a * mark sessions with significant online BCI control across all trials ( test with p<0.05). For patient 2, results for session 3 had to be disregarded due to technical problems. The right column depicts the scalp patterns of the most discriminant spectral features, based on data from all sessions. Results for Patient 4 are shown in Fig. 5.
Figure 5.
BCI performance and scalp patterns of patient 4.
Online binary accuracies, estimated bit rates (left, middle) of the CopyTask, and CSP patterns (right) averaged across all sessions are depicted in the top row (A). Each bar represents one block of at least 20 consecutive trials. Middle row (B) relates the continuous online BCI output to the residual muscle control (button press) for a representative time segment. Colored areas mark trial periods where the patient was asked to initiate a motor action. The excerpt shown was extracted from session 6, revealing that the BCI can detect the users intention far before a muscle contraction can be initiated. The lower row (C) depicts the motor related patterns in the β band for each session individually.