Designing class discrepancy-guided sub-band filter using common frequency pattern for preprocessing EEG signals in MI-BCIs

Motor imagery-based brain-computer interfaces (MI-BCI) using electroencephalogram (EEG) signals translate brain activity into control commands for output devices. Nowadays, BCI systems with fewer electrodes have gained attention; however, designing efficient feature extraction algorithms remains challenging. The common spatial pattern (CSP) is a well-known feature extraction method in MI-BCI; however, non-stationarity, noise, small training sets, and limited channels could affect its performance. In this regard, selecting informative and subject-specific frequency bands is crucial, especially when recording channels are limited. This work introduces a novel approach that applies the CSP objective function to the time-frequency representations of each EEG channel to design subject-specific, class-discrepancy-guided sub-band filters as one stage before conventional CSP feature extraction, thereby forming a two-stage CSP framework. These filters maximize the power difference between classes and generate enhanced signals treated as more informative virtual channels, with the number exceeding the initial ones. The second CSP then operates on these augmented channels to extract higher-level spectral-spatial patterns, enhancing the discriminative power of the extracted features. We evaluated our method on three public MI-BCI datasets, comparing it with conventional CSP and related state-of-the-art methods (CDFCSP and tCSP). Results demonstrate the significant enhancement of our method in classification outcomes (p < 0.001), by augmenting the effect of informative frequency components, with up to 10 % increase among subjects over benchmarks. Moreover, using the method's first stage as a spectral preprocessing technique significantly improved the outcomes of an advanced deep learning model (LMDA-Net) compared to its original and CDFCSP-preprocessed configurations. Outcomes indicate the proposed method could lead to more robust BCI applications, especially for limited-channel scenarios.