Time Window Optimization for Riemannian Geometry-based Motor Imagery EEG Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Current brain-computer interface (BCI) methods often use fixed time windows, which are suboptimal due to inter-subject variability.
• Existing Riemannian classifiers have limitations in adapting to dynamic changes in neural signals.

Purpose of the Study:

• To develop an unsupervised time window optimization method for motor imagery classification within the Riemannian geometric framework.
• To introduce a time window selection confidence metric (TWSCM) for enhanced BCI performance.

Main Methods:

• Proposed a time window selection confidence metric (TWSCM) operating on the manifold of symmetric positive definite (SPD) matrices.
• Employed an unsupervised optimization process suitable for online BCI scenarios without requiring training labels.
• Utilized Riemannian geometry for theoretically grounded and computationally efficient time window selection.

Main Results:

• Significant improvements in classification performance were observed for most subjects on the BCI competition IV dataset IIa.
• The average classification performance across six subjects improved by 7.52%.
• Simulated online experiments demonstrated enhanced performance compared to baseline methods without time window optimization.

Conclusions:

• The TWSCM offers the first Riemannian geometry-based approach for unsupervised time window optimization in motor imagery BCI.
• This method provides an effective, interpretable, and promising perspective for EEG signal analysis and BCI development.
• The approach addresses the limitations of fixed time windows and enhances BCI adaptability and accuracy.