EEG-TriNet++: A Transformer-Guided Meta-Learning Framework for Robust and Generalizable Motor Imagery Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI) classification using electroencephalography (EEG) is crucial for brain-computer interfaces (BCIs).
• Challenges include low signal-to-noise ratio, non-stationarity, and significant inter-subject variability.
• Existing methods struggle with robust and generalizable MI classification.

Purpose of the Study:

• To introduce EEG-TriNet++, a multi-branch deep learning architecture for enhanced MI classification.
• To improve both classification accuracy and cross-subject generalization in BCIs.
• To enable rapid adaptation to new users with minimal data.

Main Methods:

• Developed EEG-TriNet++, integrating convolutional spatial-spectral encoders, bidirectional LSTMs, and a Transformer head.
• Employed patchwise tokenization and neural architecture search for optimized efficiency and capacity.
• Incorporated model-agnostic meta-learning (MAML) for fast user adaptation.

Main Results:

• Achieved 79.1% and 78.6% accuracy in within-subject MI classification tasks.
• Reached 72.4% and 71.3% accuracy in leave-one-subject-out (LOSO) cross-subject generalization.
• Ablation studies confirmed the efficacy of individual model components.

Conclusions:

• EEG-TriNet++ significantly advances MI classification performance and cross-subject generalization.
• The architecture effectively addresses key challenges in EEG-based BCIs.
• Demonstrated superior performance compared to state-of-the-art methods under identical conditions.