Figure 1.
Recorded scalp electrode positions.
The three bipolar derivations, indicated by the arrows were considered by the co-adaptive BCI. Feedback was provided from only one of these bipolar derivations. The bipolar derivation selected in the last re-calibration, was also used for the self-paced paradigm. The black circles mark electrodes, recorded for future analyses. The reference electrode was at the left ear-lobe (Ref.) and the ground electrode at AFz (Gnd.).
Table 1.
Information about the severely impaired participants.
Figure 2.
Overview of the measurement procedure.
The runs colored in blue were recorded with the co-adaptive paradigm (see Figure 3). The runs colored in green were recorded with the self-paced paradigm (see Figure 4). During the non-control runs, we recorded EEG while participants relaxed with eyes open looking at a black screen. These non-control runs are not analyzed in this paper.
Figure 3.
Schematic description of the co-adaptive BCI paradigm.
Panel (A) shows how the system initially collected trials for three classes non-control, left and right hand movement imagery (MI left/right hand). Panel (B) shows the trial structure for the “Initial calibration phase”. After nine “artifact-free” trials per class (TPC) were collected the system auto-calibrated, selected one of the hand MI classes and continued to provide visual, real-time feedback. Panel (C) shows the trial structure for the “Online phase”. The system re-calibrated whenever five new artifact-free TPC were available.
Figure 4.
The self-paced BCI paradigm in different states of operation.
The head of the arrow was generally rotating clockwise around the center. Panel (A) shows how the arrow is short and colored in blue, whenever class non-control is detected. The dialog above the window indicated the next target item. Panel (B) shows how the arrow changed its color to red, when movement imagery was above the activation threshold. The user scored one point for every second the arrow stayed above this threshold in a target segment. Panel (C) shows how the user received feedback if s/he scored at least one point. In this case the arrow stopped rotating and turned grey. After a refractory period of three seconds the paradigm returned back to the initial state depicted in Panel (A).
Figure 5.
Online performance for all 22 end users.
The blue dots show the overall peak accuracy, while the grey dots depict within session performance. The color coded maps show the Fisher criterion [48] over time (left to right) for the features
and
(bottom to top).
Figure 6.
Feature dominance after calibration.
Shows for what percentage of users, the different logarithmic band-power features were selected in the final classifier calibration step.
Figure 7.
The three panels show power spectra for the three participants for whom the BCI system worked most effectively. For participant P01 and P03, the system selected the β-feature and for participant P02 the α-feature. Here, all three users control the system by causing oscillatory power of the sensorimotor rhythms to decrease (event-related desynchronization, c.f. [2]).
Table 2.
Detailed results for both paradigms.