Existing deep-learning-based motor imagery (MI) electroencephalogram (EEG) decoding methods face challenges in generalizing across sessions and providing channel-level physiological interpretability. These limitations hinder the practical application of MI-EEG systems. Accordingly, DO-PI-EATCNet (Dream-Optimization-Enhanced, Physics-Inspired, Efficient-Attention Temporal Channel Network) is proposed to improve generalization and interpretability in MI-EEG classification. Unlike models that simply combine multiple components, DO-PI-EATCNet assigns distinct roles to feature representation, temporal channel modeling, temporal regularization, and channel compactness. Latent-Projected Attention (LPA) enhances spatiotemporal discriminability by aligning attention in a low-dimensional latent space, and Temporal Channel Cascaded Collaborative Attention (TCCA) refines dependencies between time and channels. Fractional-Order Difference Temporal Consistency Loss (FD-TCL) is introduced as a neurodynamics-inspired temporal regularizer to reduce high-frequency fluctuations in prediction sequences and improve within-subject cross-session prediction stability. The Multi-Population Dream Optimization Algorithm (MPDOA) is used for channel selection to obtain a compact EEG channel subset and reduce computational load, although it introduces a slight accuracy decrease compared with the uncompressed full model. Under a within-subject cross-session protocol on the BCI Competition IV-2a four-class MI dataset, the final compact model achieves an average accuracy of 84.4% and Cohen’s κ of 0.790, outperforming the reimplemented baselines. Compared with the uncompressed LPA-TCCA-FD-TCL variant, MPDOA slightly decreases accuracy from 84.9% to 84.4%, but reduces EEG channels from 22 to about 15 and decreases MACs by 27%. Scalp topographies and selected-channel visualizations provide qualitative support for channel-level anatomical plausibility, as the selected electrodes are mainly located over expected sensorimotor-related regions, while t-SNE offers a descriptive visualization of the learned feature distributions.
Shen et al. (Sun,) studied this question.