DO-PI-EATCNet: Efficient-Attention- and Dream-Optimization-Based Channel Selection for EEG Motor Imagery...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep learning models for motor imagery (MI) electroencephalogram (EEG) decoding struggle with cross-session generalization and channel-level interpretability.
• These limitations impede the real-world use of MI-EEG brain-computer interfaces.

Purpose of the Study:

• To introduce DO-PI-EATCNet, a novel deep learning framework designed to enhance generalization and interpretability in MI-EEG classification.
• To improve the practical applicability of MI-EEG systems.

Main Methods:

• DO-PI-EATCNet employs distinct modules for feature representation, temporal channel modeling, temporal regularization, and channel compactness.
• Key components include Latent-Projected Attention (LPA) for spatiotemporal discriminability, Temporal Channel Cascaded Collaborative Attention (TCCA) for dependency refinement, and Fractional-Order Difference Temporal Consistency Loss (FD-TCL) for temporal stability.
• The Multi-Population Dream Optimization Algorithm (MPDOA) is utilized for efficient channel selection.

Main Results:

• The compact DO-PI-EATCNet model achieved 84.4% accuracy and a Cohen's κ of 0.790 on the BCI Competition IV-2a dataset under a within-subject cross-session protocol.
• MPDOA reduced the number of EEG channels from 22 to approximately 15 and decreased computational load (MACs) by 27%, with a minimal accuracy drop from 84.9% to 84.4%.
• Selected channels demonstrated anatomical plausibility, primarily located over sensorimotor regions, supported by scalp topography visualizations.

Conclusions:

• DO-PI-EATCNet effectively addresses generalization and interpretability challenges in MI-EEG decoding.
• The proposed model offers a balance between classification performance, computational efficiency, and physiological interpretability.
• The findings suggest DO-PI-EATCNet is a promising advancement for practical brain-computer interface applications.