Abstracts

Rationale: Seizure propagation within brain networks typically occurs on a very brief time scale. To analyze the patterns of propagation in such transient systems, short-window multivariate autoregressive modeling (MVAR, Ding et al., 2000) has been applied to ECoG signals from multiple seizures (multiple realizations of a brain process) to obtain one estimate of the directions, intensities, and spectral contents of propagating ictal signals. 2D penalized thin-plate splines (Ruppert et al., 2003) have then been used to test for the significance of changes in the patterns of seizure propagation, relative to pre-ictal intervals. However, two problems typically have arisen: 1) the required spline interpolation fails to converge when applied to epileptic signals, and 2) the patterns of propagation may differ between seizures. We designed a new approach to statistical testing that overcomes these limitations. Methods: We constructed a joint 95% confidence interval using bivariate smoothing with a rectangular kernel (BSRK), which is free of the problem of convergence of spline interpolation, and allows for analyzing long time series e.g. epileptic seizures, because it does not need to hold multiple thin-plate splines in computer memory. The method is also many times faster than the 2D penalized thin-plate splines method. We used BSRK to test the significance of changes in propagation of epileptic activity revealed by short-time direct directed transfer function (SdDTF, Korzeniewska et al., 2008, 2014), a method based on the concept of Granger causality. Results: Bivariate smoothing with a rectangular kernel (BSRK), applied to multiple seizures recorded in a single patient, allowed us to identify the sources of high-frequency (70-175 Hz) activity propagation, which overlapped with the seizure onset zone (SOZ) identified by clinicians. The time-frequency dynamics of the observed patterns of ictal propagation revealed a spatiotemporally complex seizure evolution. Moreover, bivariate smoothing with a rectangular kernel over a broad frequency range (e.g. 70-120 Hz) allowed statistical testing within a single seizure, revealing variability among patterns of seizure propagation recorded within an individual patient. Conclusions: Bivariate smoothing with rectangular kernel is uniquely suited to test the significance of the dynamics of propagating epileptic activity, and can potentially be used in clinical applications, e.g., planning for epilepsy surgery. Funding: Supported by: NINDS R01-NS088606 and NINDS R01-NS091139