Data selection-based method to improve gait-related multi-session motor imagery brain-computer interface

Abstract

Prominent performance in lower limb motor imagery (MI)-based brain-computer interfaces (BCIs) is crucial for their successful application in real-life scenarios. Nonetheless, achieving stable performance across multiple sessions remains a challenge due to the nonstationarity of electroencephalogram (EEG) data. To address this issue, we propose a data selection-based method to mitigate the observed instability in gait-related multi-session MI. In order to evaluate our method, we collected three gait-related classes of MIs (gait, sitdown, rest) from two healthy male participants and two participants with spinal cord injury (SCI). Each session included data acquisition of 30 trials for each class, and a total of 4 sessions were recorded for each participant. A convolutional neural network (CNN)-based model was used to extract features from the individual sessions, which were evaluated using the cosine similarity function with respect to the target session. Based on an empirically defined criterion of a cosine similarity above 0.1, we selected relevant feature data from each session. The selected target-relevant features corresponding to each class from previous sessions were then utilized in the training phase. The proposed method showed an increase in accuracy compared to self-training, which employs only the training data from the target session (Data selection: 70.83% vs Selftraining: 65.72%). These findings suggest that incorporating relevant data from past sessions enhances the stability of performance across multiple sessions in MI EEG data.