Abstracts

Rationale: The occurrence and frequency of secondary generalized tonic-clonic seizures (SGTCS) is one of the most important risk factors of sudden unexpected death in epilepsy and seizure related serious injuries. In this study, we propose a system based on the characteristics of the first five seconds of an intracerebrally recorded focal seizure, to predict the evolution of the seizure into a SGTCS, a seizure with propagation but without secondary generalization, or a seizure without propagation beyond the seizure onset zone (SOZ).  Methods: The system was designed using 39 seizures of 38 consecutive patients with drug-resistant focal epilepsy who underwent intracerebral EEG recordings for the presurgical investigations at the Montreal Neurological Institute and Hospital. The seizures had a focal onset and were grouped as follows: seizures with SGTC evolution (type A), seizure with propagation beyond the SOZ but without SGTC (type B) and seizures without propagation beyond the SOZ (type C). Only the channels within the SOZ were considered for the prediction system. Five seconds of signals from the seizure onset were utilized. Seventeen features were used for the support vector machine based classification of seizure types, including spectral and entropy measures (Table 1). Further, important features were identified using a genetic algorithm to improve the performance of the classification model. A permutation test was conducted to assess the probability that results occurred by chance. Results: We used 21 seizures for training (type A: 5, type B: 11, and type C: 5). All features except relative power of delta, beta and power ratio of delta to beta, had significantly different distributions in the three seizure types (p < 0.01) (Table 1). The magnitude of theta and alpha activity was higher in type A seizures compared to other types.  For type C seizures, the magnitude of gamma (30-150 Hz) activity was higher. The higher value of entropy features in type C seizures indicates that these signals are more random than in the other seizure types. Eighteen seizures were used for testing (type A: 5, type B: 8, and type C: 5). Our system was able to predict the evolution of 10 seizures (type A: 2, type B: 6, and type C: 2) using all features. When selecting the most discriminating features, our system could predict 11 seizures, and with higher confidence (type A: 2, type B: 7, and type C: 2) (Table 2). The probability that results occurred by chance was less than 5%. Conclusions: Our system was able, using the first 5 seconds of a seizure, to predict the evolution of most seizures into three types. The proposed system aiming to predict SGTCS could alert the health care team when a patient is hospitalized for intracerebral EEG, decrease the morbidity and mortality, and improve the safety of the hospitalization of patients investigated using invasive electrodes. Eventually, an implanted system could alert patients if a seizure likely to become generalized is starting. Funding: This work was supported by grant FDN 143208 of the Canadian Institutes of Health Research.