Objective: This work proposes a machine-learning based system for a scalp EEG that flags an alarm in advance of a clinical seizure onset.
Methods: EEG recordings from 12 patients with drug resistant epilepsy were marked by an expert neurologist for clinical seizure onset. Scalp EEG recordings consisted of 56 seizures and 9.67 h of interictal periods. Data from six patients were reserved for testing, and the rest was split into training and testing sets. A global spatial average of a cross-frequency coupling (CFC) index, , was extracted in 2 s windows, and used as the feature for the machine learning. A multistage state classifier (MSC) based on random forest algorithms was trained and tested on these data. Training was conducted to classify three states: interictal baseline, and segments prior to and following EG onset. Classifier performance was assessed using a receiver-operating characteristic (ROC) analysis.
Results: The MSC produced an alarm 45 16 s in advance of a clinical seizure onset across seizures from the 12 patients. It performed with a sensitivity of 87.9%, a specificity of 82.4%, and an area-under-the-ROC of 93.4%. On patients for whom it received training, performance metrics increased. Performance metrics did not change when the MSC used reduced electrode ring configurations.
Conclusion: Using the scalp , the MSC produced an alarm in advance of a clinical seizure onset for all 12 patients. Patient-specific training improved the specificity of classification.
Significance: The MSC is noninvasive, and demonstrates that CFC features may be suitable for use in a home-based seizure monitoring system.