The advancements in Brain-Computer Interface (BCI) have substantially evolved people’s lives by enabling direct communication between the human brain and external peripheral devices. In recent years, the integration of machine larning (ML) and deep learning (DL) models have considerably imrpoved the performances of BCIs for decoding the motor imagery (MI) tasks. However, there still exist several limitations, e.g., extensive training time and high sensitivity to noises or outliers with those existing models, which largely hinder the rapid developments of BCIs. To address such issues, this paper proposes a novel extreme learning machine (ELM) based self-attention (E-SAT) mechanism to enhance subject-specific classification performances. Specifically, for E-SAT, ELM is employed both to imrpove self-attention module generalization ability for feature extraction and to optimize the model’s parameter initialization process. Meanwhile, the extracted features are also classified using ELM, and the end-to-end ELM based setup is used to evaluate E-SAT performances on different MI EEG signals. Extensive experiments with different datasets, such as BCI Competition III Dataset IV-a, IV-b and BCI Competition IV Datasets 1,2a,2b,3, are conducted to verify the effectiveness of proposed E-SAT strategy. Results show that E-SAT outperforms several state-of-the-art (SOTA) existing methods in subject-specific classification on all the datasets, with an average classification accuracy of 99.8%,99.1%,98.9%,75.8%, 90.8%, and 95.4%, being achieved for each datasets, respectively. The experimental results not only show outstanding performance of E-SAT in feature extractions, but also demonstrate that it helps achieves the best results among nine other robust ones. In addition, results in this study also demonstrate that E-SAT achieves exceptional performance in both binary and multi-class classification tasks, as well as for noisy and non-noisy datatsets.
