Detection of movement-related cortical potentials based on subject-independent training

Abstract

To allow a routinely use of brain-computer interfaces (BCI), there is a need to reduce or completely eliminate the time-consuming part of the individualized training of the user. In this study, we investigate the possibility of avoiding the individual training phase in the detection of movement intention in asynchronous BCIs based on movement-related cortical potential (MRCP). EEG signals were recorded during ballistic ankle dorsiflexions executed (ME) or imagined (MI) by 20 healthy subjects, and attempted by five stroke subjects. These recordings were used to identify a template (as average over all subjects) for the initial negative phase of the MRCPs, after the application of an optimized spatial filtering used for pre-processing. Using this template, the detection accuracy (mean ± SD) calculated as true positive rate (estimated with leave-one-out procedure) for ME was 69 ± 21 and 58 ± 11 % on single trial basis for healthy and stroke subjects, respectively. This performance was similar to that obtained using an individual template for each subject, which led to accuracies of 71 ± 6 and 55 ± 12 % for healthy and stroke subjects, respectively. The detection accuracy for the MI data was 65 ± 22 % with the average template and 60 ± 13 % with the individual template. These results indicate the possibility of detecting movement intention without an individual training phase and without a significant loss in performance.

Fig. 1

Normalized individual and global MRCP. MRCP obtained with OSF for healthy and stroke subjects. The solid line shows the average across all subjects. The dashed lines show the individual MRCPs. The initial negative phase of the MRCP (part before the vertical dashed line) was used as template for the matched filter

Fig. 2

Performance for healthy and stroke subjects. Results from healthy motor execution (ME) and motor imagination (MI) (N = 20) and stroke patients (N = 5) testing data set with the TPR (%) (mean ± SD) and FPs per minute (mean ± SD). TST1 represents all the runs of the subject being tested with global detector (GD), TST2 are the identical testing run(s) as used for testing of the ID in motor execution/attempted task in healthy and stroke subjects, respectively