Optimized efficient channel attention-based ShuffleNet framework for EEG graph-based brain topology modeling in motor imagery tasks.

BACKGROUND: Brain topology modeling in motor imagery tasks represents Electroencephalography (EEG) channels as graph nodes and their interactions as edges, which captures neural activity patterns essential for distinguishing motor imagery tasks. NEW METHOD: An Optimized Deep Learning-based Brain Topology in Motor Imagery Tasks (ODL-BTMIT) has been proposed to increase the classification accuracy of motor imagery signals. The ODL-BTMIT starts with the acquisition of EEG signals, which are used to construct a graph that captures the intricate interactions between brain regions. From this graph, topological features, including Node Degree and Hybrid Weighted Node Centrality (HW-NC), are extracted. These features are then fed into the Efficient Channel Attention-based ShuffleNet (ECA-ShN), a lightweight convolutional neural network optimized for efficient computation. To further improve performance, the hyperparameters of Efficient Channel Attention-based ShuffleNet are fine-tuned using the Self-Improved Red Panda Optimization (SI-RPO) algorithm. RESULTS: At 90% training data, the proposed ODL-BTMIT approach achieved an accuracy of 96.5%. COMPARISON WITH EXISTING METHODS: The proposed ECA-ShN outperforms all existing methods, including Multi-Scale Spatial-Temporal Convolutional Neural Network (MSCNet), EEG Graph Lottery Ticket (EEG-GLT), GoogLeNet, EfficientNet, LinkNet, Recurrent Neural Network (RNN), and ShuffleNet, with the highest mean accuracy of 0.936, median accuracy of 0.939, and maximum accuracy of 0.966. CONCLUSIONS: The proposed ODL-BTMIT framework, leveraging the ECA-ShN model and fine-tuned with the SI-RPO algorithm, effectively enhances motor imagery classification accuracy by efficiently modeling brain connectivity and achieving stable, low-error performance.