Classification of epilepsy using high-order spectra features and principle component analysis

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Epileptic electroencephalogram (EEG) signal classification is complex due to high nonlinearity, dimensionality, and hidden states.
• Distinguishing the preictal state from the ictal state in EEG recordings presents a significant challenge.

Purpose of the Study:

• To develop a framework for improving the detection rate of epileptic EEG signal classes.
• To address the nonlinearity and high dimensionality inherent in epileptic EEG data.

Main Methods:

• Extracted 15 high-order spectra (HOS) features from EEG data.
• Applied Principal Component Analysis (PCA) to extract principal component features from HOS features.
• Evaluated eight classifiers using true positive (TP) rate and area under the curve (AUC) of receiver operating characteristics (ROC).

Main Results:

• A logistic regression model achieved a 97.5% TP rate for the 'preictal' class and an average TP rate of 96.8%.
• The logistic regression model, utilizing PCA with 100% variance, achieved the highest AUC of 99.5%.
• PCA effectively reduced dimensionality and uncovered nonlinearity in epileptic EEG signals.

Conclusions:

• The proposed framework using PCA and logistic regression significantly improves epileptic EEG signal classification.
• This approach enhances the detection of the preictal state, offering potential for improved seizure prediction systems.