Fusion Algorithm for Imbalanced EEG Data Processing in Seizure Detection.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Seizure detection algorithms (SDAs) using electroencephalography (EEG) face challenges due to imbalanced ictal and interictal data.
• Existing methods struggle with the technical difficulties posed by imbalanced data distribution.

Purpose of the Study:

• To propose a novel algorithm for improved seizure detection performance.
• To address the technical challenge of imbalanced data distribution in EEG seizure detection.

Main Methods:

• A hybrid sampling technique combining synthetic minority oversampling and undersampling TomekLink was employed at the data level.
• A cost-sensitive (CS) support vector machine (SVM) classifier was utilized at the algorithm level.
• The CS classifier assigns differential costs to misclassifications to mitigate performance degradation from imbalanced data.

Main Results:

• The proposed algorithm demonstrated significant improvements, increasing average sensitivity by 46.67% and AUC by 0.0482 compared to baseline.
• Achieved an average sensitivity of 86.34% and an AUC of 0.9837 across all cases.
• Outperformed existing methods in seizure detection performance evaluations.

Conclusions:

• A fusion of data- and algorithm-level methods effectively enhances seizure detection performance.
• The proposed algorithm achieves high sensitivity and AUC, surpassing current imbalanced EEG data processing techniques.
• Improved SDA performance facilitates their potential clinical application in EEG-based seizure management systems (SMSs).