Single-Trial EEG Classification via Orthogonal Wavelet Decomposition-Based Feature Extraction.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Electroencephalography (EEG) signal classification faces challenges due to non-stationarity and low signal-to-noise ratio (SNR).
• Existing spatial filtering methods often overlook individual differences in temporal or frequency information.
• Improving EEG signal discriminability is crucial for accurate brain-computer interfaces.

Purpose of the Study:

• To develop an advanced method (SEOWADE) for enhancing EEG signal classification performance.
• To effectively integrate spatial and spectral filtering techniques for motor imagery signals.
• To address overfitting issues and improve the neurophysiological interpretability of spatial filters.

Main Methods:

• Utilized orthogonal wavelet decomposition to separate EEG signals into sub-band components.
• Implemented channel-wise spectral filtering combined with spatial filtering, incorporating L2-norm regularization.
• Applied sparse Bayesian learning with Gaussian prior to extracted power features, resulting in a Recursive Vector Machine (RVM) classifier.

Main Results:

• SEOWADE demonstrated significantly superior classification performance compared to established algorithms like CSP, FBCSP, CSSP, CSSSP, and shallow ConvNet.
• Optimized spatial filters generated by SEOWADE exhibited more neurophysiologically meaningful scalp weight maps.
• The method effectively extracted relevant spatio-temporal information for single-trial EEG classification.

Conclusions:

• SEOWADE offers a robust and effective approach for single-trial EEG classification, particularly for motor imagery tasks.
• The integration of wavelet decomposition, spectral-spatial filtering, and sparse Bayesian learning enhances classification accuracy.
• The improved interpretability of spatial filters provides valuable insights into EEG signal processing.