A mixture of experts network structure for EEG signals classification

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Signal Processing

Background:

• Electroencephalogram (EEG) signal classification is crucial for diagnosing neurological disorders.
• Accurate and efficient EEG analysis requires sophisticated modeling techniques.
• Existing methods may lack the adaptability needed for diverse EEG signal types.

Purpose of the Study:

• To develop and evaluate a mixture of experts (ME) network for improved EEG signal classification.
• To leverage the Expectation-Maximization (EM) algorithm for training the ME network.
• To compare the performance of the ME network against stand-alone neural network models.

Main Methods:

• EEG signals were transformed into time-frequency representations using discrete wavelet transform.
• Statistical features were extracted to characterize EEG signal distributions.
• A mixture of experts (ME) network was trained using the Expectation-Maximization (EM) algorithm with extracted features.

Main Results:

• The ME network successfully classified three types of EEG signals (healthy, epilepsy seizure-free, epilepsy seizure) with 93.17% accuracy.
• The proposed ME network structure demonstrated superior performance compared to individual neural network models.
• The decoupled learning process facilitated by the EM algorithm enhanced model modularity.

Conclusions:

• The mixture of experts (ME) network provides a robust and accurate framework for EEG signal classification.
• This approach offers a significant advancement over conventional neural network models for EEG analysis.
• The ME network's modularity and efficient training are key to its high performance in complex signal classification tasks.