Early Detection of Epileptic Seizures in Sparse Domains

This work presents a method for early detection of epileptic seizures from EEG data, taking into account information about both the temporal and the spatial evolution of the seizures. The system was designed using over 8 hours of EEG, including 10 seizures in 5 patients. Seizure detection was accomplished in three main stages: multiresolution overcomplete decomposition by the à-trous algorithm, feature extraction by computing power spectral density and sample entropy values of subbands and detection by using z-test and support vector machines (SVM). Results highlight large differences between the sub-band sample entropy values during ictal and normal EEG epochs, respectively, reveling a substantial increase of such parameter during the seizure. This enables high detection accuracy and specificity especially in beta and gamma bands (16-125 Hz). The detection performance of the proposed method was evaluated based on the ground truth provided by the expert neurophysiologist, and the results show that our technique is capable to obtain a high accuracy (above the 95% on average), with a high temporal resolution. This enables reaching very low detection latency and early detection of the seizures onset. Furthermore, spatial information, within the limits of the acquisition, on the evolution of the seizure is maintained since all the channels are separately processed.