Region-Specific Topographic Representations for Deep Learning-based Brain-Computer Interfaces

Electroencephalography (EEG) signals are widely used across disciplines and are promising for future medical and technological applications. However, current EEG analysis methods often fall short in classification accuracy, limiting progress in brain-computer interfaces (BCIs), early diagnosis of neurological disorders, and AI-driven health technologies. This paper proposes a region-specific spatial-spectral topographic mapping approach to enhance EEG-based classification using deep learning. The core objective is to improve discriminative feature learning by focusing only on brain regions relevant to motor-related activity and utilizing spectral representations (alpha, beta, and alpha/beta ratio). Pre-trained deep learning architectures-ResNeSt-50d, HRNet-W18, and ConvNeXt-Base were fine-tuned on topographic images generated from the BCI Competition IV dataset 2a for classifying left- and right-hand motor imagery tasks. The proposed method achieved classification accuracies exceeding 75% when data from multiple sessions were aggregated. These results demonstrate the effectiveness of spatially and spectrally informed topographic representations for robust EEG-based BCI systems and lay a foundation for integrating EEG signals into advanced artificial intelligence applications.