Graph-based feature learning methods for subject-dependent and subject-independent motor imagery EEG decoding

The significant intra-individual variability and inter-individual differences in scalp electroencephalogram (EEG) make it difficult to learn task-distinguishable features, posing a challenge for motor imagery brain-computer interfaces. Current feature learning methods often produce an incomplete feature space, struggling to accommodate these variations and differences. Additionally, the weak discriminative nature of this feature space results in diminished EEG classification performance. This paper introduces novel graph-based feature learning methods to improve motor imagery decoding performance in both subject-dependent and subject-independent contexts. Firstly, construct a complete time-frequency-spatial-graph (TFSG) feature space. The original EEG signals are segmented into multiple time-frequency units using filter banks and sliding time windows. Spatial and brain network-based graph features are then extracted from each time-frequency unit and fused to create the TFSG features. This fused feature space is larger and more inclusive, effectively accommodating both intra- and inter-individual EEG variations. Secondly, learn a discriminative TFSG feature space. Two advanced methods are proposed. The first method employs a nonconvex sparse optimization model with log function regularization, which reduces bias in model estimation, thereby enabling more accurate learning of EEG patterns. The second method incorporates Fisher's criterion regularization into a sparse optimization framework to improve feature separability. A unified algorithmic framework is developed to solve the two new models. Our methods are validated on two motor imagery EEG datasets, achieving the highest average classification accuracies of 82.93, 68.52, and 71.69% for subject-dependent, subject-independent, and subject-adaptive evaluation methods, respectively. Experimental results demonstrate that the developed TFSG features significantly enhance both subject-dependent and subject-independent decoding performance, while the proposed regularization models improve the discriminability of the feature space, leading to further advancements in motor imagery decoding performance.

The online version contains supplementary material available at 10.1007/s11571-025-10291-5.