Epilepsy EEG classification method based on supervised locality preserving canonical correlation analysis.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• High-dimensional electroencephalogram (EEG) features pose challenges for automatic epileptic seizure detection systems, leading to noise and reduced response speed.
• Existing methods struggle to effectively utilize both sample category information and nonlinear relationships within EEG features.

Purpose of the Study:

• To propose and validate a novel epileptic signal classification method based on supervised locality preserving canonical correlation analysis (SLPCCA).
• To address the limitations of high-dimensional EEG features in automatic seizure detection systems.

Main Methods:

• Feature extraction from EEG fragments using power spectral density and fluctuation index of frequency slice wavelet transform.
• Application of SLPCCA to obtain optimal projection directions, maximizing correlation between samples and their neighbors.
• Fusion of features in the optimal direction to create a reduced-dimensionality representation.
• Classification using Least Squares Support Vector Machine (LS-SVM) with the generated fusion features.

Main Results:

• Achieved an average classification accuracy of 99.16% on the Bonn dataset for various classification tasks.
• Attained an average accuracy of 97.18% on the CHB-MIT dataset for binary classification of inter-seizure and seizure epileptic EEG.
• Demonstrated superior performance compared to several state-of-the-art methods.

Conclusions:

• The proposed SLPCCA-based method effectively reduces high-dimensional EEG features, enhancing classification accuracy and system efficiency.
• The method shows stability and effectiveness, validated by parameter sensitivity analysis and performance on benchmark datasets.
• This approach offers a promising solution for accurate and rapid automatic epileptic seizure detection.