Seizure prediction using cost-sensitive support vector machine.

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Epilepsy affects 300,000 Americans, with no current cure.
• Seizure prediction devices could significantly enhance patient quality of life.
• Distinguishing preictal (seizure-imminent) from interictal (between seizures) EEG states is crucial for prediction.

Purpose of the Study:

• To develop and evaluate a patient-specific algorithm for distinguishing preictal from interictal EEG features.
• To assess the efficacy of a Cost-Sensitive Support Vector Machine (CSVM) classifier for seizure prediction.

Main Methods:

• Extraction of linear features from the power spectrum across 9 frequency bands in EEG recordings.
• Application of a patient-specific CSVM classifier to differentiate preictal and interictal states.
• Validation using EEG data from 9 patients (45 seizures, 219 hours interictal) from the Freiburg EEG database via double-cross validation.

Main Results:

• The CSVM classifier achieved high sensitivity and specificity in distinguishing preictal from interictal EEG features.
• A sensitivity of 77.8% (35/45 seizures) was obtained.
• A zero false positive rate was achieved using a 5-minute preictal window.

Conclusions:

• The proposed CSVM-based algorithm effectively distinguishes preictal from interictal EEG features, demonstrating potential for seizure prediction.
• This method enables power-efficient real-time analysis for implantable seizure prediction devices.
• Offline optimization of computationally intensive feature extraction combined with real-time analysis offers a practical approach for device implementation.