Seizure prediction with bipolar spectral power features using Adaboost and SVM classifiers

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computational Medicine

Background:

• Epilepsy seizure prediction is crucial for patient care.
• Existing methods often face challenges with complexity and robustness.
• Intracranial electroencephalogram (iEEG) offers high-resolution data for seizure detection.

Purpose of the Study:

• To develop and evaluate a low-complexity, robust seizure prediction method using iEEG.
• To compare the performance of Adaboost and Support Vector Machine (SVM) classifiers for seizure prediction.
• To identify and utilize effective spectral power features for improved prediction accuracy.

Main Methods:

• Extracted bipolar and time-differential spectral power features from iEEG recordings.
• Employed Adaboost classifier for simultaneous feature classification and ranking.
• Compared Adaboost with a nonlinear Gaussian kernel SVM using selected top features.
• Utilized a moving-average window and threshold for alarm generation.
• Applied double-cross validation on the EPILEPSIAE database.

Main Results:

• Adaboost classifier achieved a sensitivity of 77.1% (27/35 seizures) with a false alarm rate of 0.18/hour.
• Adaboost demonstrated significantly lower computational complexity compared to SVM.
• Reduced feature sets selected by Adaboost improved classification performance on average.
• The method was validated on 8 invasive iEEG recordings from the EPILEPSIAE database.

Conclusions:

• Adaboost offers a robust and computationally efficient alternative for iEEG-based seizure prediction.
• Feature selection by Adaboost enhances prediction accuracy while reducing complexity.
• The proposed method shows significant potential for clinical application in epilepsy management.