Robust Feature Engineering Techniques for Designing Efficient Motor Imagery-Based Systems Using Multidomain Electroencephalography

Motor imagery-based brain-computer interfaces (BCIs) have potential applications in neurorehabilitation; however, the high dimensionality and complexity of electroencephalography (EEG) signals present challenges for accurate classification. This study aimed to improve the classification of motor and motor imagery activities using multidomain feature engineering and feature selection on the MILimbEEG dataset. Time-domain, frequency-domain, and wavelet-based features were extracted from 16-channel EEG signals, yielding 416 features. Maximum relevance minimum redundancy (MRMR) was then applied to select the four most informative and minimally redundant features. These features were used to train support vector machine (SVM), K-nearest neighbors (KNN), Naive Bayes, and decision tree classifiers, with performance evaluated using 10-fold cross-validation based on testing accuracy, precision, recall, and F1-score. The Gaussian-kernel SVM achieved the highest testing accuracies of 92.50% for motor activities and 95.48% for motor imagery activities. For the highest-performing motor imagery classification, precision, recall, and F1-score were 86.97%, 88.33%, and 87.04%, respectively. These findings demonstrate that reducing 416 multidomain EEG features to four discriminative features using MRMR can maintain strong classification performance. The proposed feature-engineering framework may support the development of computationally efficient EEG-based BCI systems for neurorehabilitation applications.