Epileptic electroencephalography classification using Embedded Dynamic Mode Decomposition.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Accurate seizure prediction is crucial for managing drug-resistant epilepsy.
• Current electroencephalography (EEG) analysis for seizure prediction often involves complex features and classifiers.
• Identifying the pre-seizure state is challenging but vital for improving patient quality of life.

Purpose of the Study:

• To develop a novel spatio-temporal EEG feature set for improved classification of interictal and preictal states.
• To simplify the process of early seizure identification for epilepsy patients.
• To enhance the effectiveness of seizure prediction algorithms.

Main Methods:

• Derived spectral features from Embedded Dynamic Mode Decomposition (EmDMD) of brain state dynamics.
• Utilized EmDMD to linearize complex spatio-temporal EEG dynamics into a spectral basis.
• Identified relative subband spectral power and mean phase locking values of EmDMD modes as key indicators.

Main Results:

• Analyzed linear separability and classification of preictal vs. interictal states using EmDMD features.
• Achieved high classification accuracy (up to 92% sensitivity, 89% specificity) using lightweight classifiers (SVM, Random Forest).
• Validated features on CHB-MIT scalp EEG and Kaggle AES Seizure Prediction Challenge intracranial EEG databases.

Conclusions:

• EmDMD-derived features effectively separate preictal and interictal states in EEG data.
• The proposed features significantly improve classification accuracy for seizure prediction.
• These findings motivate the integration of EmDMD features into advanced seizure prediction algorithms.