Epileptic seizure prediction based on a bivariate spectral power methodology

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Epilepsy seizure prediction relies on analyzing electroencephalography (EEG) signals.
• Feature extraction and selection are critical for accurate seizure detection from complex EEG data.

Purpose of the Study:

• To develop and evaluate novel feature selection methods for EEG-based seizure prediction.
• To compare the performance of new methods against established techniques like mRMR.

Main Methods:

• Computed spectral power for 5 EEG frequency bands (Alpha, Beta, Gamma, Theta, Delta) across 6 channels.
• Created a high-dimensional feature space (435D) from pairwise combinations.
• Introduced two new feature selection methods and Support Vector Machines (SVMs) for classification.
• Applied a "Firing Power" regularization method to SVM outputs.

Main Results:

• The proposed feature selection methods demonstrated efficiency in reducing dimensionality and improving classification.
• Achieved a sensitivity of 76.09% (predicting 35 out of 46 seizures).
• Reported a false prediction rate of 0.15 per hour.

Conclusions:

• The novel bivariate feature selection approaches are effective for EEG-based seizure prediction.
• These methods offer a promising advancement in classifying preictal states for epilepsy management.