A Domain Adaptation Framework By Aligning the Inverse Gram Matrices for Cross-Subject Motor Imagery Classification

Motor imagery (MI)-based brain-computer interfaces (BCIs) using electroencephalography (EEG) have been widely adopted due to their portability and safety. However, the non-stationary nature of EEG signals introduces significant variability between individuals, making it difficult to directly use data from other subjects (source domain) to decode the intentions of a new subject (target domain). To tackle this issue, this paper proposes a domain adaptation framework by aligning the inverse Gram matrices for cross-subject motor imagery classification (DAIG), which consists of a feature extraction module and the aligns the inverse Gram matrices module. Specifically, in the feature extraction module, EEGBiMamba is proposed to capture the temporal dependencies in one-dimensional EEG time-series representations with a bidirectional (forward and backward) modeling. For the extracted EEG features from both domains, the model aligns their inverse Gram matrices from the perspectives of angle and scale, effectively minimizing the differences in feature distributions. Furthermore, the multilayer perceptron (MLP) layer is replaced with a Kolmogorov-Arnold Network (KAN) to generate an M-dimensional vector for classification, improving the model’s expressiveness and adaptability. The proposed method is evaluated on two public datasets, BCI IV IIa and BCI IV IIb, along with the AMI_MD dataset collected in our laboratory. The results show that the proposed method achieved average cross-subject decoding accuracies of 68.58%, 86.72%, and 81.65% across different datasets, representing remarkable performance improvements of 17.06%, 13.6%, and 15.76% over baseline methods, respectively.