Neural Manifolds in Brain-Machine Interfaces: Evaluating Their Potential for Decoding Motor Imagery

Brain-computer interfaces (BCIs) constitute an emerging technology with significant potential to enhance communication and control for individuals with severe motor impairments. Motor imagery (MI), which involves decoding cortical activity linked to imagined movements, represents a promising approach within BCI research. Nonetheless, effective MI decoding remains challenging due to the complexity of neural dynamics and the need for accurate, real-time performance. Traditional techniques, such as Power Spectral Density (PSD) methods, have demonstrated substantial accuracy, yet may inadequately capture the intricate structure of neural data. Alternatively, neural manifold methods may offer a solution by reducing the dimensionality of the neural data while preserving essential relations. This study systematically compares traditional and manifold-based approaches at two data segmentation levels: whole-trials and shorter temporal intervals, namely windows, resembling real-time data streaming. While PSD-based methods outperformed others in whole-trial decoding, Principal Component Analysis (PCA)-based manifolds improved classification stability and maintained competitive performance in window-based scenarios, suggesting potential suitability for real-time applications. Despite the limited dataset size hindering the generalizability of the results, the findings offer useful insights into dimensionality-reduction strategies and suggest that linear manifold techniques may have the potential to advance effective, real-time MI decoding in brain-computer interfaces.