Human expert supervised selection of time-frequency intervals in EEG signals for brain-Computer interfacing

Abstract

In the context of brain-computer interfacing based on motor imagery, we propose a method allowing a human expert to supervise the selection of user-specific time-frequency features computed from EEG signals. Indeed, in the current state of BCI research, there is always at least one expert involved in the first stages of any experimentation. On one hand, such experts really appreciate keeping a certain level of control on the tuning of user-specific parameters. On the other hand, we will show that their knowledge is extremely valuable for selecting a sparse set of significant time-frequency features. The expert selects these features through a visual analysis of curves highlighting differences between electroencephalographic activities recorded during the execution of various motor imagery tasks. We compare our method to the basic common spatial patterns approach and to two fully-automatic feature extraction methods, using dataset 2A of BCI competition IV. Our method (mean accuracy m = 83.71 ± 14.6 std) outperforms the best competing method (m = 79.48 ± 12.41 std) for 6 of the 9 subjects.