EEGCA-Net: Channel-Attention Framework with Subject-Wise Fine-Tuning for Motor Imagery Classification

Unnati Chaurasia, Himanshu Kumar Pathak, Koushlendra Kumar Singh, Marios Antonakakis, Michalis E. Zervakis

Abstract: Disabilities related to motor function significantly impact the daily lives of people worldwide. Advancements in Brain-Computer Interfaces (BCIs) paved the way for Motor Imagery (MI) based systems for assistive applications. The use of EEG signals in these systems to interpret a user's thoughts and further analysis are performed on these signals. This work proposes EEG Channel Attention Network (EEGCA-Net), a novel framework that enhances MI classification tasks by integrating channel attention mechanisms and Bayesian optimization. The channel attention mechanism prioritizes EEG channels carrying the most task-relevant information, while Bayesian optimization automates fine-tuning of hyperparameters for optimal performance. Additionally, subject-wise fine-tuning is applied to the BCI Competition IV-2a, which includes EEG recordings for four MI tasks. This improves accuracy and generalization across individuals. Two evaluation approaches are used for this dataset: global evaluation and subject-specific evaluation. The global evaluation consists of data from all the subjects, and an overall accuracy of 74.58% is achieved with Cohen's Kappa coefficient of 0.71. For subject-specific evaluation, transfer learning is applied for the recorded EEG signal of each individual. In this case, the average subject-wise accuracy is raised to 78.99 %, and the average subject-wise Kappa coefficient is 0.78. These results highlight the EEGCA-Net's ability to generalize across users while adapting to individual neural patterns, improving the practicality of MI-BCI systems for real-world applications.

External IDs:dblp:conf/memea/ChaurasiaPSAZ25