Features importance in seizure classification using scalp EEG reduced to single timeseries

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Seizure detection and classification typically rely on high-density electroencephalogram (EEG) recordings.
• Wearable systems face limitations with few electrodes, reducing signal spatial resolution.
• This study addresses seizure classification challenges in low-electrode-count wearable devices.

Purpose of the Study:

• To evaluate machine learning classifiers for seizure-type classification using limited EEG data.
• To compare the performance of Random Forest (RF) and Support Vector Machine (SVM) classifiers.
• To identify optimal feature extraction strategies for reduced-dimensionality EEG signals.

Main Methods:

• Tested RF and SVM classifiers on a subset of EEG recordings from wearable systems.
• Employed single-trace selection and hemispherical dimensionality reduction techniques.
• Utilized permutation tests to analyze feature importance across different frequency bands (delta, theta, alpha, gamma).
• Applied advanced sampling for imbalanced datasets and leave-patients-out cross-validation.

Main Results:

• Random Forest classifiers demonstrated superior efficiency and stability compared to SVMs.
• Feature importance varied by classifier: SVMs prioritized low frequencies (delta, theta), while RF favored higher frequencies (alpha, gamma).
• Achieved up to 94.3% accuracy in classifying absence versus tonic-clonic seizures.

Conclusions:

• Random Forest classifiers are highly effective for seizure-type classification with limited EEG data.
• Frequency band relevance differs between RF and SVM classifiers, offering insights into signal processing.
• The study validates a robust method for wearable epilepsy monitoring systems.