Subject-Independent Brain-Computer Interfaces: A Comparative Study of Attention Mechanism-Driven Deep Learning Models

This research examines the employment of attention mechanism driven deep learning models for building subject-independent Brain-Computer Interfaces (BCIs). The research evaluated three different attention models using the Leave-One-Subject-Out cross-validation method. The results showed that the Hybrid Temporal CNN and ViT model performed well on the BCI competition IV 2a dataset, achieving the highest average accuracy and outperforming other models for 5 out of 9 subjects. However, this model did not perform the best on the BCI competition IV 2b dataset when compared to other methods. One of the challenges faced was the limited size of the data, especially for transformer models that require large amounts of data, which affected the performance variability between datasets. This study highlights a beneficial approach to designing BCIs, combining attention mechanisms with deep learning to extract important inter-subject features from EEG data while filtering out irrelevant signals.