(Abst. 1.181; ), 2007
Analysis of Ictal Dynamics Using Generative Neural Networks
Authors: P. Mirowski1, D. Madhavan2, Y. LeCun1, R. Kuzniecky2
Seizure onsets often appear widespread on intracranial EEG, leading to uncertainty in their localization. We propose a new dynamical model combined with Time-Delay Neural Networks (TDNN) that generates EEG time series prior and after simulated surgical resection. This is done for the purpose of predicting changes to seizure propagation upon alterations of certain channel parameters.
Ictal EEG recordings of 4 patients were obtained from the NYU Epilepsy Center database, and converted into numerical EEG data sampled at 400Hz. A dynamical model was constructed that generates EEG time-series from a time-series of hidden internal sources. Inference consists in finding the most likely sequence of sources given observed EEG and a trained dynamical model (a Time-Delay Neural Network). Learning consists of iteratively adjusting the parameters of the TDNN and inferring the most likely sequence of sources. Once the TDNN is trained, the dynamical model was applied to another EEG time series of the same patient with inference of the corresponding sources. The model can be used to predict in closed-loop a short-time horizon (1-3 sec) continuation of EEG. Simulation of surgical resections on the EEG time-series was performed by (1) setting some channels to zero, (2) inferring the time-series of sources, (3) predicting the continuation of the time-series of sources, (4) generating corresponding EEG. This method was used to compare (A) deactivating channels in the seizure onset zone vs. (B) deactivating remaining channels. Linear univariate and bivariate measures were used to assess influence of (A) and (B) on seizure propagation.
Numerical data from a partial seizure as recorded on a 32-channel EEG grid in a bipolar montage was used as target for the dynamical model. After training, 1-second EEG predictions were started at different times. The performance of the predictive dynamical model was assessed on unseen cross-validation data from another seizure (same patient). Similar signal-to-noise ratios were obtained for 1-step (14dB), 20-step (0.2sec, 5.2dB) and 100-step (1sec, 3dB) iterated predictions on both in-sample and out-of-sample data (see figure). This demonstrates that the generative model is able to learn a hidden dynamical model of the propagation of epileptic seizures, and generalize it to other seizures from the same patient. However, when trying to deactivate certain channels by setting them to zero, the inferred hidden states did not properly reconstruct EEG data. Further research to address that problem involves using more realistic generative models and more hidden states.
We have developed a dynamical generative model able to learn patterns in intracranial EEG at ictal onset, and are currently enhancing it so that it can predict dynamic changes in response to alteration of specific EEG channels. This system could potentially aid the epileptologist in the analysis and planning of surgery, and delineate diffuse epileptic networks in brain.
Back to search results