Surrogate-Driven Multi-Objective SMEPO Framework for Optimal EEG Channel and Feature Selection in Motor Imagery BCI

Abstract Brain-computer interfaces (BCIs) hold significant potential in communication, mechatronic control, and rehabilitation. However, existing BCI systems are often expensive, and require laborious preparation. A major challenge in motor imagery (MI)-based BCIs is the effective selection of EEG channels due to the non-stationary, high-dimensional, and subject-specific nature of EEG data. Consequently, most MI-BCI pipelines fail to simultaneously optimise classification accuracy and computational efficiency across subjects. Although nature-inspired algorithms have been explored, they often suffer from premature convergence and poor balance between exploration and exploitation, while conventional methods typically rely on single-objective optimisation that ignores the trade-off between accuracy and model complexity. To address these limitations, this paper introduces the novel Surrogate-Assisted Multi-Objective SMEPO (Emperor Penguin Optimiser) algorithm for efficient EEG channel selection in MI-BCIs. SMEPO integrates temperature-gradient dynamics and adaptive phase-gate scheduling strategy within a multi-objective Pareto archive framework, enabling effective navigation of the complex binary search space while maintaining computational efficiency via a kNN-5 surrogate model. To address inter-subject variability and prevent data leakage, spatial covariance features are extracted using Multivariate Empirical Mode Decomposition (MEMD). Subject-specific normalisation is performed using statistics derived exclusively from the training set. Extensive experiments on three BCI Competition datasets (BCI-IV Dataset 2a, BCI-IV Dataset 1, and BCI-III Dataset IVa) show that SMEPO achieves superior mean SVM accuracies of 85.79%, 83.28%, and 84.75%, respectively, consistently outperforming five state-of-the-art surrogate-assisted multi-objective baselines across SVM, Naïve Bayes, and Decision Tree classifiers. Further validation on eight UCI benchmark datasets confirms that SMEPO wins best accuracy on six out of eight datasets with a mean accuracy of 92.56% and mean feature reduction of 56.2%. These results establish SMEPO as a robust, efficient, and generalisable channel and feature selection solution.