The Impact of Non-Task-Related Neural Activity in EEG-Based Motor Imagery Classification

In the context of Brain-Computer Interfaces, detecting motor imagery events from electroencephalography signals plays a crucial role. Traditionally, this task has been formulated as a classification problem among different imagined movements. However, despite remarkable accuracy in distinguishing among different events, real-world biomedical applications introduce the additional challenge of differentiating motor imagery events from non-task-related background neural activity. To address this challenge, in this work, we conduct a comparative analysis between two deep learning based approaches (end-to-end and hierarchical) for also taking into account background class alongside motor imagery classes (specifically, left and right hand motor imagery). We demonstrate that the end-to-end approach achieves performance comparable to the hierarchical one, with only a negligible drop in accuracy, while offering significant advantages in terms of model complexity and suitability for portable applications.