Neural network analysis of seizure activity in subdural EEG

Detection of the onset of seizure-activity in the EEG record of an epileptic patient is an important step in the localization of seizure foci. This research examines the use of three recurrent neural-network architectures to detect seizure-activity onset in multi-channel subdural EEG data. The neural networks´ output represents the degree of seizure-like activity for each of the 94 channels and four closest neighboring electrodes. A decrease in the standard-deviation of the neural network output at each seizure-like event was noted. Also, by spectrogram analysis, a large increase in the low-frequency components of the variation at the time of the seizure event in each sample was seen. This suggests that, at the onset of seizure-like activity: 1) the magnitude of the signal standard-deviation across the subdural space is small; 2) the time-dependent change of the standard-deviation is small; 3) the signal power is concentrated in a small range of low-frequency components. These results suggest the entrainment of epileptogenic neural components.