MAGCANet: A multiscale adaptive graph-convolutional attention network for MI-EEG decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery electroencephalography (MI-EEG) decoding faces challenges from low signal-to-noise ratios and inter-subject variability.
• Existing deep learning models may suffer from temporal leakage and fixed spatial topologies, limiting their adaptability.

Purpose of the Study:

• To develop MAGCANet, a novel deep learning architecture for robust and interpretable MI-EEG decoding.
• To address limitations of existing models by enforcing causality and adapting spatial interactions.

Main Methods:

• MAGCANet integrates Multiscale Causal Convolution, Temporal Convolution, Adaptive Graph Convolution, and Multi-Head Self-Attention modules.
• The architecture enforces temporal causality and learns subject- and trial-specific spatial connectivity patterns.

Main Results:

• Achieved high single-subject accuracies (88.58% on IV-2a, 91.13% on IV-2b) and competitive cross-subject generalization (70.49% on IV-2a, 79.49% on IV-2b).
• Demonstrated a lightweight design (0.0194M parameters) with low inference latency (2.23 ms).
• Qualitative analyses confirmed model interpretability and ability to capture relevant EEG patterns.

Conclusions:

• MAGCANet offers a computationally efficient and highly accurate solution for MI-EEG decoding.
• The model's interpretability and robustness make it suitable for real-time brain-computer interface applications.