Generative diffusion meets domain adaptation: a framework for EEG cross-subject motor imagery classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Cross-subject motor imagery classification using electroencephalogram (EEG) data is limited by data scarcity and significant inter-subject variability.
• Existing methods struggle to generalize effectively across different individuals, hindering practical applications of brain-computer interfaces (BCIs).

Purpose of the Study:

• To develop a novel framework that improves cross-subject motor imagery classification by integrating generative data augmentation and domain adaptation.
• To enhance the generalization capability of EEG-based BCIs for diverse user populations.

Main Methods:

• Utilized a diffusion probabilistic model to generate high-fidelity synthetic EEG samples, thereby augmenting the limited training data.
• Proposed the Adaptive Multi-Scale Convolution-Domain Adversarial Neural Network (AMSC-DANN) architecture.
• Integrated an Adaptive Multi-Scale Convolution (AMSC) module for multi-granular feature extraction and a Domain Adversarial Neural Network (DANN) for feature distribution alignment across subjects.

Main Results:

• The proposed framework significantly outperformed state-of-the-art baselines on the BCI Competition IV datasets 2a and 2b.
• Demonstrated superior cross-subject generalization capabilities compared to existing methods.
• Validated the effectiveness of generative augmentation and domain adaptation in addressing EEG data challenges.

Conclusions:

• The novel framework effectively enhances cross-subject motor imagery classification by leveraging generative data augmentation and domain adaptation.
• The AMSC-DANN architecture successfully learns discriminative temporal-spectral representations and aligns feature distributions across subjects.
• This approach offers a promising solution for developing more robust and generalizable BCIs.