Subject-Independent Deep Learning Framework for Motor Imagery Electroencephalogram Decoding in Neurorehabilitation.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) is crucial for brain-computer interfaces (BCIs) in neurorehabilitation.
• Traditional MI-EEG models struggle with inter-subject variability and require extensive subject-specific data.
• This necessitates the development of subject-independent BCI systems.

Purpose of the Study:

• To introduce a deep learning framework, DSGNet, for subject-independent MI-EEG classification.
• To enable generalization to unseen subjects without requiring labeled data for each individual.
• To enhance the reliability of BCIs for neurorehabilitation.

Main Methods:

• Developed a Dual-branch Subject-aligned Generalization Network (DSGNet) utilizing deep learning.
• DSGNet extracts temporal and spectral EEG features via dual convolutional branches.
• Implemented a class alignment loss for domain-invariant representation across subjects.

Main Results:

• DSGNet demonstrated superior accuracy on three-class and four-class MI-EEG datasets compared to baseline models.
• Achieved performance improvements of 0.22% and 2.15% on these datasets, respectively.
• Maintained comparable performance on binary-class datasets, validating its generalization capability.

Conclusions:

• Class-structure alignment is effective for developing reliable subject-independent BCI systems.
• DSGNet offers a promising solution for neurorehabilitation by overcoming data limitations.
• The framework advances non-invasive BCIs for motor function restoration.