Multiclass Somatosensory Selective Attention BCI Classification Based on Riemannian Geometry Approach Combing Time-Frequency Sub-Band Selection and Feature Reduction

Somatosensory selective attention (SSA)-based BCI has received a great deal of attention due to its no training and non-visual characteristics. However, there are still two issues that constrain its application. One is that the control commands of SSA-BCI is limited. The other is that the high-accuracy classification for multiclass SSA modality remains a challenge. Methods To address these issues, this work constructed a multiclass SSA-based BCI system and presented a novel Riemannian geometry (RG) framework combing subject-specific time-frequency sub-band selection and feature reduction. In this study, two linear resonant tactors were separately adhered to both index fingers of each subject, and three SSA tasks (left-SSA, right-SSA, and idle-SSA) were implemented according to the random displayed cues. To validate its performance, 10 subjects took part in the experiment and were collected EEG signals. In addition, a RG framework was provided to identify this three-class SSA modality. Firstly, EEG signals were partitioned into multiple subsegments and bandpass filtered by a filter bank. Subsequently, the optimal time-frequency sub-bands were chosen according to Laplacian score (LS), and then the RG features achieved from each chosen sub-band were reduced using neighborhood component analysis (NCA) algorithm. Finally, a support vector machine (SVM) was trained for SSA-BCI identification. This approach yields an average accuracy of 79.48% in three-class SSA modality, which is considerably better than other competing methods. This study demonstrates that our developed RG framework is an attractive option for enhancing the multiclass SSA-BCI modality's performance.