Abstracts

Convolutional Neural Networks (CNNs) are specifically designed to extract and classify high-dimensional patterns such as time-series. Currently, CNNs have not been applied to the analysis of EEG data. We are currently developing a CNN that would have the capability to learn the complex interrelationships between intracranial EEG channels, providing a powerful method of delineating electrically interconnected areas of brain in the generation of an electrographic seizure discharge. This network will also have the ability to predict changes to the seizure upon alterations of certain channel parameters, thereby potentially simulating the effects of epilepsy surgery., Ictal EEG samples were obtained from the NYU Epilepsy Center database, and converted into numerical matrices representing EEG channels values every 25ms (400Hz sampling). We implemented CNN software that would take as input a slice of 400-800 time samples (1-2 sec) of EEG channels and predict values 1 time step into the future. The training of the CNN consisted of iteratively optimizing the thousands of parameters of the CNN in order to minimize the error between the predicted values and the real measurements. The CNN was trained to correctly predict outputs 1 time step into the future, and to re-use its outputs as successive inputs at the following time steps, predicting 2.5 second EEG waveforms. After adequate training, channels determined to be the seizure onset zone via visual analysis were [quot]deactivated[quot] by replacing their signals with random low frequency noise, and the output EEG was examined. A similar operation was performed for channels not included in the seizure onset zone., Numerical data from a partial seizure as recorded on a 32 channel EEG grid was obtained, and used as input for the CNN (Fig. 1A, dotted line). After training, 3-second EEG predictions were started at different times (Fig. 1B). There was a marked reduction in ictal activity with deactivation the visual seizure onset zone (Fig. 1D) vs. channels 0-15 (Fig. 1C)., We have developed a CNN that has the ability to robustly learn patterns in intracranial EEG at ictal onset, and predict dynamic changes in response to alteration of specific EEG channels. This system can potentially aid the epileptologist in the analysis and planning of surgery, and delineate diffuse epileptic networks in brain. Further investigation includes integrating other signal attributes to enhance the fidelity of network predictions.[figure1], (Supported by the Epilepsy Foundation Finding A Cure for Epilepsy and Seizures (FACES).)