Optimizing motion imagery classification with limited channels using the common spatial pattern-based integrated algorithm

The extraction of effective classification features from electroencephalogram (EEG) signals in motor imagery is a popular research topic. The Common Spatial Pattern (CSP) algorithm is widely employed in this field. However, the performance of the traditional CSP method depends significantly on the choice of a specific frequency band and channel number of EEG data. Furthermore, inter-class variance among these frequency bands and the limited number of available EEG channels can adversely affect the CSP algorithm’s ability to extract meaningful features from the relevant signal frequency bands. We hypothesize that multiple Intrinsic Mode Functions (IMFS), into which the raw EEG signal is decomposed, can better capture the non-Gaussian characteristics of the signal, thus compensating for the limitations of the CSP algorithm when dealing with nonlinear and non-Gaussian distributed data with few channels. Therefore, this paper proposes a novel method that integrates Variational Mode Decomposition (VMD), Phase Space Reconstruction (PSR), and the CSP algorithm to address these issues. VMD is used to filter and enhance the quality of the collected data, PSR is employed to increase the effective data channels (data augmentation), and the subsequent CSP filtering can obtain signals with spatial features, which are decoded by Convolutional Neural Networks (CNN) for action decoding. This study utilizes self-collected EEG data to demonstrate that the new method can achieve a good classification accuracy of 82.30% on average, confirming the improved algorithm’s effectiveness and feasibility. Furthermore, this study conducted validation on the publicly available BCI Competition IV dataset 2b, demonstrating an average classification accuracy of 87.49%. Graphical Abstract