Swati Bhalaik

An effective electrode selection strategy is crucial in motor imagery (MI)-based brain–computer interfaces (BCIs) to maintain competitive performance while reducing the number of electrodes and overall computational complexity. This study introduces a novel electrode selection method based on signal power and evaluates its impact on MI task classification using the BCI Competition IV-2a dataset. Electrodes were systematically reduced by selecting those with the highest signal power. The proposed method was evaluated using state-of-the-art deep learning models, EEGNet, ShallowConvNet, and DeepConvNet, with classification accuracy and F1-score as performance metrics. EEGNet, with all electrodes, achieved an average accuracy of 69.30% and an average F1-score of 0.6910. As the number of electrodes was progressively reduced, performance declined gradually, with a noticeable drop observed after a 50% reduction. Notably, even with 50% fewer electrodes, accuracy remained within 8.59% of the full-electrode configuration. Topographic analysis showed that electrodes near the motor cortex, exhibiting higher signal power, were most critical for classification. In contrast, peripheral electrodes with lower signal power were less informative and could be removed, demonstrating the effectiveness of the proposed method. Similar trends were observed for ShallowConvNet and DeepConvNet, further confirming the method’s generalizability. This approach provides a promising direction for developing more practical, faster, cost-effective, and resource-efficient BCI systems.

Motor imagery (MI)-based brain-computer interfaces (BCIs) require efficient electrode selection to reduce cost and computational complexity while preserving their performance. We propose a signal-to-noise ratio (SNR)-based electrode selection strategy and introduce a novel method to calculate SNR from EEG datasets. We further present the Pairwise Difference (PRD) technique for data augmentation to compensate for the accuracy loss resulting from electrode reduction. Using EEGNet, a state-of-the-art deep learning model, we achieved an accuracy of 68.99% and an F1-score of 0.6891 with all 22 electrodes as the baseline. As electrodes were progressively removed, performance declined, with a notable drop after a 50% reduction. However, applying the PRD technique maintained EEGNet's average accuracy within 3.96% of the baseline (66.26%), and F1-score at 0.6587, even with the above reduction. PRD effectively compensated for the loss of information due to electrode reduction. Simulations further validated the generalizability of our approach, demonstrating its effectiveness across other deep learning models, viz., ShallowConvNet and DeepConvNet. This work will help develop cheaper and resource-efficient BCI systems.

The major challenge in fully utilizing the motor imagery-based brain-computer interface (MI-BCI) capabilities is accurately classifying the MI electroencephalography (MI-EEG) signals. Despite numerous advancements in signal processing and deep learning techniques, there is significant scope for improvement in the accuracy currently available in the state-of-the-art. This can be achieved by leveraging spatial and temporal features of MI-EEG signal. We propose SWCNet, a convolutional neural network (CNN)-based model, and integrate it with the sliding window technique to increase the accuracy. In this work, a new CNN architecture has been proposed to extract more features from data, whereas the sliding window technique enhances temporal features by augmenting the input sensor data along the temporal dimension. We have thoroughly evaluated the performance of SWCNet using subject-dependent and subject-independent approaches for four different datasets. Our analysis includes general accuracy metrics, an ablation study, a parametric sensitivity study, and a detailed class-wise performance evaluation for the tongue, foot, left-hand, and right-hand movements. The proposed model achieves accuracies of 97.42%, 94.46%, 92.27%, and 90.82% for BCI Competition IV-2a, BCI Competition IV-2b, High Gamma, and OpenBMI datasets, respectively. SWCNet outperforms the state-of-the-art methods with higher accuracy for all the datasets, demonstrating its superior generalizability. SWCNet holds promise in enhancing the effectiveness of BCI applications, especially in medical rehabilitation.