Abstract
This study explores the use of attention mechanism-based deep learning models to construct subject-independent motor-imagery based brain-computer interfaces (MI-BCIs), which present unique and intricate challenges from a machine learning perspective. By comparing four attention mechanism-based models and employing nested LOSO methods for robust model selection, the study enhances the reliability of performance estimates and offers unique insights into the application of attention mechanisms in building subject-independent BCIs. The results indicate the potential of the Spatio-Temporal CNN + ViT model for practical BCI applications, as it outperforms other models on several datasets. Additionally, the study presents a realistic approach to building subject-independent BCIs by combining attention mechanisms and deep learning models to identify informative features common across subjects while filtering out noise and irrelevant data. While there are limitations and areas for future work to enhance the potential of these models, transformer-based models could become even more valuable in the BCI research field, leading to more robust and accurate subject-independent BCIs for various applications. The need for subject-independent MI-BCIs is amplified due to their potential in assisting individuals with severe neurological conditions, such as ALS and locked-in syndrome, which severely limit mobility and communication.