Automated Seizure Detection Based on State-Space Model Identification.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Automated seizure detection from electroencephalogram (EEG) recordings is crucial for epilepsy management.
• Traditional methods often struggle with accuracy and efficiency.
• System identification offers a novel approach to extract meaningful features from EEG signals.

Purpose of the Study:

• To develop and evaluate a machine learning model for automated seizure detection.
• To utilize system identification techniques for feature extraction from EEG data.
• To compare the performance of different epoch lengths and classifiers for seizure detection.

Main Methods:

• Developed a machine learning model using system identification techniques on EEG recordings.
• Extracted features using fifth-order state-space dynamic systems across various epoch lengths (1s, 2s, 5s, 10s).
• Tested multiple classifiers, including decision trees, on seizure and non-seizure EEG datasets from two institutions.

Main Results:

• Decision tree classifier with 1s epochs achieved 96.0% accuracy, 92.7% sensitivity, and 97.6% specificity on the Jefferson dataset.
• Performance decreased with increased epoch length.
• High accuracy (94.1%) and specificity (97.5%) were observed on the CHB-MIT dataset, with subject-specific models reaching 98.3% accuracy.

Conclusions:

• System identification, particularly state-space modeling, combined with decision tree classifiers, is an effective and efficient method for automated seizure detection.
• Shorter epoch lengths (1s) yield superior performance in automated seizure detection.
• The developed model shows significant potential for clinical application in epilepsy monitoring.