On the suitability of random forests for detecting mental imagery for non-invasive brain-computer interfacing

Brain-Computer Interfaces (BCIs) are devices that directly convert a user's brain activity into actions. One class of BCI is based on the detection of changes in oscillatory activity of noninvasive electroencephalographic (EEG) signals. Motor imagery (MI) is typically used to induce such changes, and machine learning and pattern recognition methods for translating corresponding EEG activity pattern into messages for devices.
The aim of this thesis is to explore the usefulness of the Random Forests classifier (RF) for
the classification of MI tasks. The RF classifier is an ensemble classifier, which consist of
many uncorrelated decision trees. The output of the RF classifier is chosen by a vote. To
ensure more diverse votes, each decision tree is built up by randomized parameters.
The RF method was applied to EEG data recorded from ten able-bodied subjects while
performing left hand (L), right hand (R) and feet (F) MI. The results of extensive offline
cross-validation tests and offline BCI simulations suggest that RF are suitable for the
classification of oscillatory EEG activity pattern. Peak (mean ± std computed by averaging
the peak accuracies for each subject) accuracies of 82% (59 ± 14%) for the 3-class problem,
and 93% (67 ± 15%) for L vs R, 91% (77 ± 12%) for L vs F and 94% (77 ± 10%) for R vs F,
respectively, were computed. The calculated results are comparable with state-of-the-art
methods used in BCI research. Furthermore, online feedback experiments were performed
with three able bodied subjects. Two were able to successfully operate the BCI. Subjects
achieved peak accuracies of 92% and 88%, respectively.