Enhancing brain–computer interface performance through source-level attention mechanism: An EEG motor imagery study

Abstract

Background:

Brain–computer interfaces (BCIs) enable direct communication between humans and machines by translating brain signals into control commands. Electroencephalography (EEG) is a commonly used modality in BCI systems due to its non-invasiveness and high temporal resolution. However, EEG-based BCIs often suffer from low signal-to-noise ratios and limited spatial resolution, primarily due to the small number of recording electrodes. Although source estimation techniques can improve spatial specificity, they typically require subject-specific information such as individual brain anatomy or electrode positions, which may not always be available. This study aims to address these challenges by proposing a framework that enhances task-relevant EEG signals using an attention-guided source estimation approach based on coarse predefined brain regions.

New method:

We developed an attention-guided neural network that estimates source-level activity most relevant to the BCI task, without requiring subject-specific structural data. The model uses predefined regions of interest to guide attention mechanisms toward informative spatial features.

Results:

The framework was validated using publicly available motor imagery EEG datasets, achieving strong performance. Comparison with existing methods: Comparative analyses were conducted against baseline models using traditional EEG signals and standard feature extraction methods. This study presents an effective approach for improving EEG-based BCI performance by integrating an attention-guided source estimation network into the decoding pipeline. The method does not rely on subject-specific anatomical information, making it broadly applicable.

Conclusion:

By emphasizing task-relevant source activity, the framework enhances signal quality and classification accuracy, thereby advancing the potential of BCIs for precise and practical applications.