During CopyTask, an adaptive classifier was applied, which was trained on data from preceding sessions.
Thus, while each trial was continuously classified, Fig. 4 and Fig. 5B show the online selection accuracy as a bar for each block. It should be noted that the type of feature used for classification was changed between (sometimes also within) sessions. This happened especially within the first three BCI-sessions, since the experimenters could not be sure which feature would suit best for each patient. Fig. S1 pictures theses modifications over time, which resembles the closed-loop design cycle following a user-centered design. One example for such a transition can be found in patient 4 between session 2 and session 3: in session 2 (the first feedback session), a classifier in the μ band was used, resulting in a poor BCI control. After reanalyzing the data of session 1 and 2, a new classifier was generated for session 3. The new classifier evaluated an ERD in the beta-band, leading to a considerably increased offline accuracy (estimated with cross validation) and also the online performance all following sessions was increased considerably.