Abstracts

Rationale: The principal goal of this project is to develop and optimize a seizure detection algorithm for the reliable detection of epileptic seizures and its integration into a fully implanted and autonomously powered intervention device. One of the key parameters for the efficiency of intervention in seizure suppression is the early detection of the seizure onset. Here, a seizure detection algorithm is proposed which can detect seizure onset in the early stages of seizure formation when seizure suppression by a closed-loop system is still possible. Methods: A. Database Data in the European Epilepsy Database (EPILEPSIAE) cover diverse types of epileptic seizures. The iEEG recordings of 10 patients were selected from this database. The selected patients had a total number of 160 clinical seizures, annotated by experienced epileptologists. The EEG recordings were split into 1 hour segments, each with one seizure. B. Feature selection Because the ultimate goal of this project is the implementation of this algorithm in a fully implanted device, features requiring low computational power were selected. Ten features in the time and frequency domain were selected, including: mean, mean absolute deviation, variance, skewness, kurtosis, line length, autocorrelation, power in beta (13-30Hz) and gamma (30-128Hz) frequency bands and finally the power ratio of gamma band power, divided by the sum of power in the alpha and beta bands. These features were computed in non-overlapping 1 sec time windows. C. Classification A Support Vector Machine (SVM) with a radial basis function kernel was used for classification. Features were normalized between 0 and 1. Ten-fold cross validation for sigma and the box constraint parameters was done to achieve optimized values for each seizure detection channel of every patient. The sequential minimal optimization (SMO) algorithm was used as convergence method. For analysis, from every patient the channel with the minimum delay was selected. Results: Of the total number of160 seizures, 156 were detected correctly. The mean seizure detection delay was 4.1 seconds. The mean sensitivity was 98%, with a false detection rate (FDR) of 5.6/h. For seven out of ten patients all seizures were detected correctly. Conclusions: The advantage of the SVM classifier is its ability to realize short detection delays combined with a high sensitivity. According to the importance of the seizure detection at earlier stage of seizure formation for an effective seizure suppression, application of the SVM classifier seems to be a promising method for closed-loop intervention devices. The exploration of more representative feature sets and focusing on the seizure onset pattern for training should further improve the performance of the classifier. Funding: This study has been supported by the Excellence Cluster BrainLinks-BrainTools (DFG, Grant Number EXC 1086).