A Novel Motor Imagery EEG Decoding Approach via Sparse Latent-Space Regression With Multi-Norm Regularization

Motor imagery (MI) electroencephalography (EEG) has shown significant potential in consumer electronics (CE) applications enabled by brain–computer interfaces (BCIs); however, current limitations in decoding MI-EEG signals still restrict the broader development and adoption of this technology in the CE industry. Although deep learning methodologies have recently gained prominence in EEG decoding, existing approaches often encounter challenges due to high computational demands, long training times, and limited sample sizes. To address these limitations, we introduce a fast and robust MI EEG decoding algorithm that captures spectral power features from the latent-space. Our approach simultaneously optimizes spatial filters and classifiers within a convex optimization framework, employing L21 and L1 norm regularization to ensure sparsity. We rigorously tested our algorithm against several EEG decoding methods, including CSP-based algorithms and state-of-the-art deep learning approaches, across four MI BCI datasets and various small datasets. Our results demonstrate that our algorithm surpasses existing methods in both accuracy and decoding speed. Furthermore, visualization of the optimized spatial filters reveals neurophysiologically meaningful patterns, affirming the validity of our algorithm.