An efficient seizure prediction method using KNN-based undersampling and linear frequency measures.

Area of Science:

• * Neuroscience
• * Biomedical Engineering
• * Signal Processing

Background:

• * Seizure prediction using electroencephalogram (EEG) signals is a significant research area.
• * Existing algorithms often have high computational demands, limiting clinical application.
• * There is a need for reliable seizure prediction with minimal computational requirements for medical facilities.

Purpose of the Study:

• * To propose an optimized novel method for predicting epileptic seizures with reduced computational complexity.
• * To develop a clinically applicable seizure prediction algorithm.
• * To improve the trade-off between sensitivity and false prediction rate.

Main Methods:

• * Utilized univariate linear features from eight frequency sub-bands of EEG signals.
• * Employed Principal Component Analysis (PCA) for dimension reduction and feature selection.
• * Addressed class imbalance using K-nearest neighbor (KNN)-based undersampling combined with a Support Vector Machine (SVM) classifier.
• * Evaluated two postprocessing methods.

Main Results:

• * Achieved 100% seizure prediction accuracy.
• * Reported an average false alarm rate of 0.13 per hour (range: 0-0.39).
• * Validated performance using G-Mean (0.97) and F-measure (0.90).

Conclusions:

• * The proposed algorithm demonstrates high discriminative ability for seizure prediction.
• * It offers improved sensitivity and a reduced false prediction rate compared to other studies.
• * The method's low computational requirements and high performance make it suitable for potential use in implantable devices.