Real-Time EEG Decoding of Motor Imagery via Nonlinear Dimensionality Reduction (Manifold Learning) and Shallow...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Decoding electroencephalography (EEG) signals for motor imagery is crucial for brain-computer interfaces (BCIs).
• High dimensionality and nonlinearity of EEG data present significant challenges for accurate and real-time decoding.
• Previous work explored manifold learning for EEG data from individuals with spinal cord injury (SCI).

Purpose of the Study:

• To introduce and evaluate a real-time processing framework for decoding motor imagery EEG signals.
• To compare the effectiveness of various nonlinear dimensionality reduction techniques integrated with shallow classifiers.
• To assess the framework's performance across different classification complexities and in comparison to previous findings in SCI populations.

Main Methods:

• EEG data from six healthy participants performing five motor imagery tasks were collected.
• Five nonlinear dimensionality reduction methods (t-SNE, ISOMAP, LLE, Spectral Embedding, MDS) were evaluated.
• Each method was combined with three shallow classifiers (k-NN, Naive Bayes, SVM) for binary, ternary, and five-class settings.

Main Results:

• The t-SNE + k-NN pairing achieved the highest accuracies: 99.7% (2-class), 99.3% (3-class), and 89.0% (5-class).
• ISOMAP and MDS demonstrated competitive performance, especially in multi-class scenarios.
• The framework achieved low-latency processing (approx. 150 ms per trial), meeting real-time BCI requirements.

Conclusions:

• Nonlinear dimensionality reduction, particularly t-SNE, significantly enhances real-time EEG decoding accuracy for motor imagery.
• The proposed framework offers a low-complexity, high-accuracy solution for BCIs in both healthy individuals and those with neurological impairments.
• Consistent advantages of t-SNE and ISOMAP in preserving class separability were observed across healthy and SCI groups.