Source-driven enhanced semantic style transfer for addressing BCI illiteracy in subject-independent EEG-based motor imagery classification

초록

Motor-imagery brain-computer interface (MI-BCI) provides a direct connection between the brain and external devices by decoding user intentions. Despite its potential, the integration of MI-BCI systems into real-world ap plications faces challenges due to the prevalence of BCI illiteracy, where users encounter difficulties in device operation and maintaining focus during training sessions. Although one-to-one style transfer methods have been explored to mitigate BCI illiteracy, they have not been successfully extended to a subject-independent setting that involves unseen users. However, in practical BCI applications, this subject-independent setting is essential because subject-specific information is typically unavailable before BCI system implementation. To overcome these limitations, we propose a novel style transfer framework adept at tackling the BCI illiteracy problem within a subject-independent setting. Specifically, to improve style transfer performance, we enhance the framework through a source-driven selection mechanism that identifies and utilizes the most class-representative sample. Furthermore, we leverage the domain generalization technique to mix styles from multiple BCI illiterate users in the target domain, enabling robust performance on previously unseen users. This approach ensures that the model captures a diverse range of class-relevant features across users, enhancing its ability to generalize to new, unseen users. We demonstrate the effectiveness of our proposed approach through a leave-one-subject-out evaluation, which shows clear performance improvements across unseen subjects.

키워드