Seizure onset zone classification based on imbalanced iEEG with data augmentation.

Area of Science:

• Neuroscience and Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Accurate identification of the seizure onset zone (SOZ) is crucial for epilepsy surgery.
• Clinical intracranial electroencephalogram (iEEG) data often exhibit class imbalance, posing challenges for machine learning models.
• Existing machine learning approaches struggle with the time-consuming and subjective nature of manual iEEG analysis.

Purpose of the Study:

• To address the challenge of imbalanced iEEG data in focal epilepsy.
• To develop a novel data augmentation method, EEGAug, for generating synthetic minority class data.
• To improve the applicability of clinical iEEG data in machine learning models for SOZ identification.

Main Methods:

• Introduced EEGAug, a method that transforms minority class iEEG samples into the frequency domain.
• Composed new synthetic data by combining different frequency bands from selected minority samples.
• Converted the synthesized frequency-domain data back to the time domain to create new samples.

Main Results:

• The EEGAug method successfully balanced imbalanced clinical iEEG datasets for machine learning applications.
• A one-dimensional convolutional neural network classifier demonstrated improved performance using EEGAug-generated data.
• EEGAug outperformed other data augmentation techniques and the focal loss function in balancing iEEG data.

Conclusions:

• EEGAug offers a flexible solution for generating high-fidelity synthetic data for minority classes in imbalanced clinical datasets.
• The method achieves high distribution similarity between synthetic and raw data.
• EEGAug enhances the utility of clinical iEEG data for machine learning-driven epilepsy research and diagnostics.