Intrinsic mode function synchronization measures for the anticipation of seizures in epilpsy

Epileptic seizures affect as many as 50 million people and often occur without warning or apparent provocation. We explore the applicability of noise-assisted Ensemble Empirical Mode Decomposition (EEMD) for patient-specific seizure anticipation synchronization measures as applied to the EEMD intrinsic mode function (IMF) output. Intracranial EEG data were obtained from pre-surgical monitoring at the Epilepsy Center of the University Hospital of Freiburg. Data from twenty patients were analyzed. For each recorded channel, non-overlapping time windows were submitted to the EEMD algorithm, producing twelve levels of IMFs. IMF synchronization measures (mean and maximum coherence, mean and maximum cross-correlation, correlation coefficient and synchronized phase-locking value) for channel pairs were computed and smoothed with a 20-point moving average, producing IMF-x data. Statistical distributions of IMF-x synchronization data were determined for three hours of interictal training data. Three hours of interictal validation data were used to determine the smallest zero-false-positive threshold (multiples of 0.5 standard deviations of IMF-x data) for each channel pair and IMF level. These patient-, IMF level-, and channel pair-specific IMF-x thresholds were compared against periictal (60 minutes preictal with 15 minutes ictal/postictal) IMF-x data for each seizure. Our study shows that while not all channel pairs are able to detect every ictal event, low IMF levels containing frequency components greater than –,1 Hz can discriminate between interictal and periictal activities. The anticipation window for channel pairs detecting all ictal events frequently ranged from 30 to 53 minutes prior to clinical manifestation. We propose an anticipation optimality index for a joint indicator of sensitivity and earliest anticipation times useful for selection of relevant channel pairs and IMF levels. Generalization of the analyzed synchronization measures may be appropriate for some patients, while other patients may require preferential selection of these measures. For the majority of patients, the electrode pairing type holds some relevance to performance assessment values. A strong indication of IMF-level dependence of anticipation performance data was shown, suggesting seizure dynamics in the patient-specific scenario manifest within certain frequency bandwidths. The patients with a hippocampal seizure origin show better sensitivity with our algorithm than patients with neocortical seizure origin