Abstracts

Immediate detection of epileptic seizures in the EEG represents a significant problem in epileptology. The underlying morphology of a seizure usually exhibits a high intra- and interindividual variability and thus, a precise detection usually requires detailed analysis for every particular case. Previous studies have shown that neural networks represent an appropriate paradigm for the recognition of hidden patterns in noisy and non-stationary environments. Here, we propose a novel approach to the problem of automated real-time seizure recognition in EEG using Cellular Neural Networks (CNN). Such networks have a massive computing power, allow parallel computation, and are already available as integrated circuits. The proposed method exploits the phenomenon of induced pattern formation within a locally perturbed nonlinear medium, simulated by a CNN. This system along with an appropriately chosen set of internal parameters exhibits spatial-temporal disorder. The process of induced pattern formation can be regarded as detection of certain transient rhythms within applied local perturbation, i.e., the EEG. In this retrospective study, we applied our method to automatically detect seizures in intracranial multi-channel, multi-day EEG recordings from patients undergoing presurgical evaluation. First observations show that our proposed technique allows to detect seizures with a high sensitivity and specificity. By construction, our method is able to immediately detect these events. However, we observed a trade-off between the number of false detections and the latency of detection relative to the electrical seizure onset as defined by an expert reader. Our preliminary findings indicate that, in principle, the proposed method can be used as a real-time seizure detector. However, further improvements are necessary in order to achieve a high degree of generalization. Nevertheless, we expect a future implementation as a miniaturized real-time detection device to allow a variety of clinical applications. (Supported by The Deutsche Forschungsgemeinschaft)