Improved patient specific seizure detection during pre-surgical evaluation.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Automated seizure detection from EEG is crucial for reducing workload in monitoring units.
• Subject-specific methods outperform subject-independent ones but require pre-surgical data, which is impractical.
• An alternative approach is needed for practical, subject-specific seizure detection.

Purpose of the Study:

• To develop and evaluate an automated seizure detection method that adapts a subject-independent classifier to individual subjects using user feedback.
• To improve the efficiency and accuracy of offline electroencephalogram (EEG) analysis for seizure detection.

Main Methods:

• A subject-independent quadratic discriminant classifier was developed using modified features.
• Subject-specific classifiers were derived by adapting posterior probability thresholds through user interaction.
• The method was tested on intracranial EEG data from 15 subjects (529 hours, 63 seizures) and compared to the standard Gotman algorithm.

Main Results:

• The subject-independent scheme reduced false positive rates by 51% (0.23 to 0.11 h⁻¹) and increased sensitivity from 53% to 62%.
• Subject adaptation further improved sensitivity to 78% with a minor increase in false positive rate (0.18 h⁻¹).
• The adapted classifier performance approximated that of the optimized subject-specific Gotman algorithm.

Conclusions:

• Subject-independent classifiers with modified features effectively reduce false positives in EEG seizure detection.
• Subject adaptation significantly enhances sensitivity, approaching the performance of traditional subject-specific methods.
• The proposed method offers a practical solution to improve EEG analysis productivity and can be generalized to other algorithms.