Design and implementation of a writing-stroke motor imagery paradigm for multi-character EEG classification.

Motor imagery (MI) based brain-computer interfaces (BCI) decode neural activity to generate command outputs. However, the limited number of distinguishable commands in traditional MI-BCI systems restricts practical applications. To overcome this limitation, we propose a multi-character classification framework based on Electroencephalography (EEG) signals. A structurally simplified MI paradigm for stroke writing is designed, and maximize Euclidean distance trajectory optimization enhances neural separability among five stroke categories. The EEG data cover 11 motor imagery tasks, including five stroke-writing tasks and six related movement tasks such as hand, foot, tongue movements and eye blinks, collected from ten participants. Ensemble Empirical Mode Decomposition (EEMD) eliminates artifact-related Intrinsic Mode Functions (IMFs) and reconstructs the signals. Kernel Principal Component Analysis (KPCA) then conducts nonlinear dimensionality reduction to extract discriminative features. Finally, a recurrent neural network based on Gated Recurrent Units (GRU) performs classification, effectively modeling the temporal dynamics of EEG signals. Experimental results indicate that the optimized stroke paradigm achieves an average classification accuracy of 84.77%, outperforming the unoptimized version at 76.83%. Compared to existing MI-BCI methods, the proposed framework improves classification accuracy and expands the set of distinguishable commands, demonstrating enhanced practicality and effectiveness.