Clustering Sparse Swarm Decomposition for Automated Recognition of Upper Limb Movements From Nonhomogeneous Cross-Channel EEG Signals

Decoding motor imagery electroencephalogram (MI-EEG)-based upper limb movements become a prominent tool to people with neuromuscular diseases. In this letter, the clustering sparse swarm decomposition method (CSSDM) is proposed to extract homogeneous spectral characteristics across nonhomogeneous multichannel MI-EEG sensor data with significant channel selection for improving decomposition and enhancing the performance of automatic upper limb movement recognition. CSSDM, a novel approach proposed to address the limitation of processing nonhomogeneous signals, such as EEG, extends the capabilities of existing swarm decomposition. In CSSDM, first, the nonhomogeneous EEG signal is analyzed by a density-based spatial clustering algorithm based on canonical correlation analysis-mutual information measure into homogeneous EEG clusters. The CSSDM adopts modified swarm filtering and sparse spectrum to automatically deliver into optimal band-limited modes, which shows the mutual characteristics across channels. Further, the time-frequency graph spectral features are extracted from CSSDM modes. The experimental results on the 7-class BNCI EEG (001-2017) database reveal that CSSDM-based classification frameworks outperformed all baseline models and achieved the highest accuracy of 49.02 +/- 0.61% using tenfold cross-validation.