In recent years, electroencephalogram (EEG) signals have attracted significant interest for their potential in interpreting human intentions within Brain-Computer Interface (BCI) systems. However, identifying human intentions from these signals with accuracy is difficult since they are inherently non-linear and non-stationary. In order to improve the classification accuracy of MI-based EEG signals, there is a great need for efficient and reliable algorithms that can help advance the capability of BCI technology. The present work proposes a new signal decomposition technique called swarm decomposition, which is proposed for the first time in the literature. This work aims at MI-based EEG classification accuracy enhancement, which involves preliminary feature extraction procedures based on common spatial patterns of the signals considered. These, in turn, are reduced in features using Cuckoo Search during optimal feature identification, with crucial information retained across the reduced data space. The proposed approach is evaluated using the BCI Competition IV Dataset 2b, demonstrating its effectiveness in classifying binary EEG data with an average classification accuracy of 92.76%. These results highlight the potential of the proposed methodology for enabling real-time applications in Brain-Computer Interface systems.