Gaussian Connectivity-Driven EEG Imaging for Deep Learning-Based Motor Imagery Classification.

Area of Science:

• Neuroscience and Biomedical Engineering
• Brain-Computer Interfaces (BCIs)
• Machine Learning for Healthcare

Background:

• Electroencephalography (EEG)-based motor imagery (MI) BCIs offer potential for neuro-rehabilitation but face challenges like low spatial resolution and inter-subject variability.
• Conventional methods (CSP, CNNs) struggle with robustness, generalization, and interpretability in MI classification.
• Existing BCIs often fail to adequately address BCI illiteracy, limiting their practical application.

Purpose of the Study:

• To introduce EEG-GCIRNet, a novel network integrating Gaussian connectivity and a regularized LeNet for improved MI classification.
• To enhance the robustness, generalization, and interpretability of EEG-based BCIs.
• To overcome limitations of current methods and mitigate BCI illiteracy.

Main Methods:

• Developed EEG-GCIRNet, a variational autoencoder framework combining raw EEG signals with functional connectivity topographic maps.
• Employed a multi-objective loss function optimizing reconstruction, classification accuracy, and latent space regularization.
• Utilized interpretability techniques like latent space visualization and Grad-CAM++ for validation.

Main Results:

• EEG-GCIRNet achieved the highest average accuracy (81.82%) with lowest variability (±10.15%) in binary classification, outperforming state-of-the-art methods.
• The model completely eliminated the 'Bad' performance group in BCI illiteracy, improving accuracy by ~22% for these users.
• Demonstrated competitive accuracy (75.20% ± 4.63) in 5-class scenarios with statistical superiority (p=0.002).

Conclusions:

• EEG-GCIRNet offers a robust and interpretable end-to-end framework for EEG-based BCIs.
• The method effectively addresses BCI illiteracy and shows promise for reliable neurotechnology in rehabilitation and assistive applications.
• Interpretability analyses confirmed the model captures genuine neurophysiological mechanisms underlying MI classification.