Abstracts

Rationale: Epilepsy is one of the most prevalent neurological diseases, affecting an estimated 50 million people worldwide. Seizures on one third of patients cannot be controlled with medications alone. For these individuals, the next best course of treatment is to surgically remove the parts of the brain responsible for the seizures. The current clinical gold standard for identifying the areas to resect relies on qualitative expert visual inspection of intracranial EEG (iEEG) data which is only successful in 40-80% of cases. Here, we report a method for analyzing and visualizing electrocortical source dynamics in epilepsy patients (Akalin Acar et al; Mullen et al, 2011, EMBC). By identifying sources of seizure activity and modeling dynamic relationships between seizure zones, this methodology can generate information and insight into seizure generation that can help experts make more accurate and objective clinical assessments and surgical plans. Methods: The main steps of our analysis methodology can be itemized as follows: 1. Adaptive mixture independent component analysis (AMICA) (Palmer et al, 2006) is applied to iEEG data to identify and separate seizure sources. When applied to epilepsy data, AMICA can adapt to the differences in spatio-temporal source structure during pre-seizure, seizure, post-seizure and non-seizure periods (Figure 1). 2. An electrical forward head model of the patient is generated from MR and CT images of the patient's head (Akalin Acar & Makeig, 2010, http://sccn.ucsd.edu/nft/). A sparse Bayesian method (SBL) including multi-scale patch-based source model is used to estimate the anatomical region of each iEEG independent component. 3. A pairwise mutual information (PMI) matrix is calculated to understand how the time courses differ across components. The component clusters showing high dependencies are identified and mapped on the patient's brain image. 4. Dynamics of ICA source clusters are analyzed by fitting a suitable time-varying vector autoregressive model (VAR) that captures distributed patterns of spectral dynamics in the data (Mullen et al, www.sccn.ucsd.edu/wiki/SIFT). Graph theoretic metrics are then used to summarize and visualize the complex network structure. These metrics are combined with the SBL component inverse maps to identify regions significantly implicated in seizure propagation. Results: These methods are tested on datasets collected from two patients at the Mayo Clinic (Rochester, MN) and at the Hoftsra North Shore LIJ (NY), respectively. Initial results suggest that spatiotemporal network dynamics, modeled in the absence of seizure, may provide a novel source of predictive information regarding the location and likelihood of later ictal dynamics. In both datasets, our localization of epileptogenic zone was confirmed by the (mainly) positive surgical outcomes. Conclusions: Our work to date suggests that this method a) may improve the spatial resolution of iEEG information, and b) may provide new insights into seizure genesis and propagation by revealing seizure network interactions too complex to be readily identified by visual inspection.