Spectral characterization and classification of motor-imagery EEG using PSD and RNN-LCSO

Motor imagery-based brain-computer interface systems require accurate classification of Electroencephalography signals despite their non-stationary characteristics and temporal variability. Existing Electroencephalography-based classification approaches often process spectral information and temporal dependencies separately, while optimization-based neural networks may suffer from slow convergence and premature convergence. To address these limitations, this work proposes a Recurrent Neural Network with Lévy Flight-based Crow Swarm Optimization framework for Electroencephalography-based motor imagery classification. The proposed framework integrates Electroencephalography preprocessing, spectral feature extraction, temporal sequence modeling, and population-based parameter optimization within a unified architecture. The Common Spatial Pattern and Power Spectral Density analyses are employed to extract discriminative spatial and spectral characteristics from the physiologically relevant μ (8–13 Hz) and β (13–30 Hz) frequency bands. The Recurrent Neural Network is designed to model the temporal dependencies among sequential Electroencephalography observations from the extracted spectral features. Additionally, the Lévy Flight-enhanced Crow Swarm Optimization is established to optimize the parameters of the Recurrent Neural Network, where the Lévy Flight mechanism improves the global exploration capability of the optimization and reduces the likelihood of premature convergence during parameter optimization. Experimental results demonstrate that the Recurrent Neural Network with Lévy Flight-based Crow Swarm Optimization achieves an average accuracy of 98.51%, precision of 98.33%, recall of 98.07%, specificity of 97.72%, F1-score of 98.19%, and Matthews Correlation Coefficient of 0.980 when compared to existing methods. The framework also achieves an Area Under the Receiver Operating Characteristic Curve of 0.985 and 0.981 on the two datasets, indicating consistent discriminative performance across the evaluated datasets. The proposed approach is generally applicable to Electroencephalography-based brain-computer interface systems in which spectral characteristics and temporal neural dynamics need to be jointly represented for motor imagery decoding. These outcomes unveiled that the Recurrent Neural Network with Lévy Flight-based Crow Swarm Optimization provides an accurate and computationally efficient classification model for practical motor imagery brain-computer interface applications.