Mixture of Experts for EEG-Based Seizure Subtype Classification.

Area of Science:

• Neurology
• Machine Learning
• Signal Processing

Background:

• Epilepsy affects 50 million globally, necessitating accurate seizure classification for treatment.
• Automatic electroencephalogram (EEG) based seizure subtype classification faces challenges like class imbalance and the need for extensive labeled data.
• Current deep learning models often require large datasets due to a lack of a priori knowledge integration.

Purpose of the Study:

• To propose novel Mixture of Experts (MoE) models for improved EEG-based seizure subtype classification.
• To address the challenges of class imbalance and limited labeled data in EEG analysis.
• To enhance deep learning model performance by incorporating prior knowledge of EEG features.

Main Methods:

• Development of two novel MoE models: Seizure-MoE and Mix-MoE.
• Implementation of a novel imbalanced sampler within Mix-MoE to handle class imbalance.
• Integration of a priori knowledge of manual EEG features into deep neural networks.

Main Results:

• The proposed Seizure-MoE and Mix-MoE models demonstrated superior performance in cross-subject EEG-based seizure subtype classification.
• Mix-MoE effectively addressed significant class imbalance using its novel sampling technique.
• Both models outperformed existing approaches on two public EEG datasets.

Conclusions:

• The developed MoE models offer a promising solution for accurate EEG-based seizure subtype classification.
• These models effectively tackle class imbalance and leverage prior knowledge for better performance.
• The proposed MoE framework is adaptable for other EEG classification tasks with severe class imbalance, such as sleep stage classification.