Abstracts

Rationale: Predicting seizures will significantly improve the quality of life for patients with pharmacoresistant epilepsy. So far, many methods have been developed with varying degrees of success to identify preictal periods and consequently seizure precursors by investigating various linear and nonlinear properties of scalp and intracranial electroencephalography (iEEG). In this study, we developed and evaluated a patient-specific seizure prediction method based on integrated univariate and bivariate nonlinear measures. Methods: The iEEG recordings of 11 patients with focal neocortical epilepsy were obtained from the Freiburg seizure prediction EEG (FSPEEG) database with authorization. Informed consent had been also obtained from each patient. Intracranial contacts were surgically placed on the cortex and/or inserted in the brain to record iEEG data from each patient. To analyze iEEG data, we used six intracranial contacts, three near the epileptic focus and three in remote locations. In total for each patient, 49 hours of iEEG data containing 49 seizures and 267 hours of interictal data were individually analyzed, including at least 24 hours of iEEG recordings with no seizure activity. The seizure prediction system comprised preprocessing, feature extraction, and decision making stages. After band-pass filtering, iEEG data were divided into quasi-stationary segments. Following this, features including five nonlinear univariate measures, correlation dimension, correlation entropy, noise level, Lempel-Ziv complexity, and largest Lyapunov exponent, as well as one bivariate nonlinear measure, and nonlinear interdependence were extracted from iEEG segments. Finally, the decision making stage including spatial and feature combiners, utilized several guidelines to integrate feature values extracted from the iEEG segments of the channels to raise preseizure flags. For each patient, a sample seizure and 4-h seizure-free interictal data considered as a baseline were used to optimize the system. The ultimate performance of the system was then evaluated in terms of sensitivity and false prediction rate on the remaining seizures and seizure-free interictal data. Results: An average sensitivity of 79.9% and 90.3% with an average false prediction rate of 0.17/h and 0.12/h were achieved, respectively, within a prediction horizon of 30 and 60 minutes. The preictal changes in feature values were equally observed on the electrodes located in the epileptogenic zone and remote areas. Table 1 compares the performance of the seizure prediction methods applied to the FSPEEG database.Conclusions: In this study, the spatio-temporal patterns observed prior to the seizures showed patient-specific dynamic signatures that could be used for seizure prediction and preictal identification. These preictal observations will undoubtedly help elucidate the mechanisms underlying seizure generation. Furthermore, the system showed a clinically acceptable performance across different patients and seizure types, thus giving credence to it potential as an effective and accurate seizure prediction tool.