Overview
Motor imagery decoding based on electroencephalogram (EEG) signals presents a fundamental challenge due to significant variations between subjects. These inter-subject variations complicate the development of robust brain-computer interfaces (BCIs). The existing approaches often utilize Riemannian geometry, representing EEG signals as covariance matrices on the symmetric positive definite (SPD) manifold. However, these methods frequently overlook subject-specific differences in covariance dispersion and orientation. This work introduces geometry-aware congruence transformations to mitigate these specific challenges.
Research Context
EEG-based brain-computer interfaces (BCIs) rely on decoding motor imagery. A core issue in this field is the substantial inter-subject variability, which hinders the generalization of decoding models across different individuals. While Riemannian geometry has been applied to represent EEG signals as covariance matrices on the SPD manifold, current methods have primarily focused on these manifold-based representations without adequately addressing subject-specific covariance dispersion and orientation variations. This gap in addressing specific geometric properties of inter-subject variability forms the basis for the current investigation.
Approach
The research addresses the challenges of inter-subject variability through geometry-aware congruence transformations. This approach led to the development of three complementary models:
- Discriminative Congruence Transform (DCT): A foundational model leveraging congruence transformations.
- Deep Linear DCT (DLDCT): An extension building upon the DCT with a deep learning architecture.
- Deep DCT-UNet (DDCT-UNet): A more complex deep learning model incorporating a U-Net-like structure within the DCT framework.
These proposed models were evaluated in two primary capacities:
- As manifold alignment modules for integration with downstream classifiers.
- As end-to-end discriminative architectures.
The end-to-end architectures were optimized using cross-entropy, incorporating a custom logistic regression head. The evaluation was conducted on challenging cross-subject motor imagery benchmarks.
Findings
Experiments on cross-subject motor imagery benchmarks indicated consistent improvements in transductive decoding performance. The proposed geometry-aware congruence learning methods achieved 2-3% higher accuracy compared to strong baselines. These results suggest the effectiveness of geometry-aware congruence learning in mitigating inter-subject variability in EEG decoding.
Why This Matters
The consistent improvements in transductive decoding performance, achieving 2-3% higher accuracy than strong baselines, suggest that geometry-aware congruence learning can effectively mitigate inter-subject variability in EEG decoding. This addresses a fundamental challenge in EEG-based brain-computer interfaces.