Abstracts

Rationale: A reliable identification of seizure precursors from the EEG of epilepsy patients may enable both investigations of basic mechanisms leading to seizure initiation in humans and the development of adequate seizure warning and prevention strategies. Previous studies have shown that particularly analysis techniques characterizing synchronization phenomena between different brain regions are capable of identifying seizure precursors with a performance that exceeds chance level. Synchronization phenomena can reliably and efficiently be detected on the EEG with Cellular Neural Networks (CNN) that provide a computational power for real-time, multi-channel analyses while minimizing energy and space consumption. Given the high intra- and interindividual variability of synchronization phenomena, a CNN with sufficient generalization properties would allow application across different patients without individual adaption. Methods: We investigated multi-channel, multi-day intracranial EEG recordings from up to now 10 patients (1658 hours; 60 seizures) and calculated for each channel combination the temporal evolution of a measure for generalized synchronization (nonlinear interdependency N). In order to find optimal network settings we iteratively trained the network on only 5 min EEG that was recorded at a preselected channel combination from a single patient. We then evaluated generalization properties of the network by performing an out-of-sample validation study that included the EEG data from all channel combinations from all patients. In order to quantify within- and across-subject results we compared the CNN-based estimates of N with the analytically obtained values using a statistical null hypotheses test and, moreover, benchmarked seizure prediction performance. Results: For both within- and across-subject approximations the out-of sample validation indicates that our CNN is able to estimate N at a high accuracy. For 9 patients the absolute deviation ranged between 0.08 and 0.17 without applying any patient-specific (re-)optimization. For 7 patients the null hypotheses of a random time series exhibiting a smaller deviation than the CNN-based approximation could be rejected (p< 0.05) for at least half of the channel combinations. Interestingly, in 6 patients the CNN-based estimate of N allowed us to identify statistically significant seizure precursors at more widespread brain regions within the epileptic network than the analytically derived N. Conclusions: Our CNN-based characterization of synchronization phenomena exhibits promising generalization properties, which may even enable across-subject applications without the need of re-optimizing the network. Considering the fact that seizure prediction performance even increased in some cases, it is likely that our CNN is able to identify common spatial-temporal preseizure synchronization processes on the EEG. This could pave the way for an improved prediction of epileptic seizures. (Supported by Deutsche Forschungsgemeinschaft)