Abstracts

Rationale: The large inter-patient variability regarding seizure characteristics poses a great challenge for developers of an automatic seizure detector. Patient-specific systems yield better performance, due to the rather consistent intra-patient seizure patterns. The disadvantage of such a solution is the fact that the detector has to be trained for each patient individually, involving both medical professionals and technicians. We propose a seizure detector capable of exploiting crucial structural information underlying the EEG data, therefore, providing better performance and requiring lower amount of seizure data for training than other approaches.Methods: A total of 16 features were extracted from each channel of each 2s long epoch, and were stored in the form of a matrix. The feature matrices extracted from the seizures of a given patient are expected to follow a similar pattern. Thus, a suitable classifier is a matrix with well-defined structure, where the features and channels characteristic for the patient's seizures are highlighted. Such classifier is determined by a convex optimization algorithm using nuclear norm penalty, ensuring a low-rank structure underlying the solution. This method will be referred to as nuclear norm learning (NNL). NNL was compared to two commonly applied solutions using the same feature set, but a training algorithm not designed for exploiting such structural information. The late-integration method (LI-LSSVM) uses the features extracted from each channel to train an independent least squares support vector machine (LS-SVM) using linear kernel. During the detection task the outputs of the channel classifiers are integrated with an OR function to obtain the final decision. On the contrary, following the early-integration method (EI-LSSVM) the features of all channels are stacked in a long feature vector, which is used to train the LS-SVM.Results: The proposed methods were tested on 20 adult patients with refractory partial epilepsy. Given each patient, different classifiers were trained including an increasing number of seizures in the training set. Performances are evaluated as the area under the ROC curve. All three methods successfully differentiate between seizure and non-seizure epochs. NNL performs better than EI-LSSVM given arbitrary number of seizures in the training set (p=0.0001). Moreover, NNL outperforms EI-LSSVM even if the latter is trained with one additional seizure (p=0.0084). NNL performs better than LI-LSSVM as well, given at least three training seizures (p=0.0056). Finally, LI-LSSVM performs better than EI-LSSVM up to two training seizures (p=0.0001). However, when training with three or more seizures is performed, this difference is no longer significant. Table 1 shows the average performance over all patients obtained by each method, in terms of sensitivity given 90% specificity.Conclusions: Nuclear norm learning is a suitable method for patient-specific seizure detection. It successfully exploits structural information from the EEG, and can achieve better performance than traditional methods, even when a limited amount of seizure data is available for training.