Small Sample Motor Imagery Decoding Algorithm Based on Clustered Centered Spatial Pattern

Decoding motor imagery (MI) signals under small sample conditions remains a key challenge in brain-computer interface (BCI) research. This paper proposes a few-shot MI decoding algorithm based on Clustered Centered Spatial Pattern (CCSP) to improve cross-subject decoding performance. The method first applies Euclidean alignment through covariance matrix whitening to eliminate cross-subject feature differences. Then, K-means clustering is applied to CSP filters from multiple training subjects to construct a universal filter dictionary shared across subjects. Features are extracted from limited target subject samples (5 per class), and classification templates are built using cosine similarity-weighted strategies. Experimental results demonstrate that the proposed method achieves an average accuracy of 6 8. 6 7 pm 8. 9%, outperforming the traditional filter bank common spatial pattern (FBCSP) method by 11.52 percentage points. Ablation experiments further verify the effectiveness of Euclidean alignment in enhancing cross-subject feature consistency. The proposed algorithm shows significant potential for MI-BCI applications under small sample conditions.