Abstracts

RATIONALE: Signal analysis techniques from nonlinear dynamics, such as the correlation dimension (CD) and the related correlation integral (CI), continue to be applied to the problem of epileptic seizure prediction. Making the assumption that variations in these measures prior to and during ictal onset reflect characteristic changes in the dynamical properties of ECoG, studies have reported that these measures are able to predict seizure onset up to several minutes in advance. Though previous studies have assumed that CD and CI are measures of the dimensionality of the ECoG signal trajectory, this study investigates whether CD and CI are sensitive to two easily quantifiable characteristics of ECoG: amplitude and power spectral density (PSD).
METHODS: Eight seizures from 4 subjects who underwent invasive evaluation for epilepsy surgery are analyzed. The ECoG is filtered (0.5 - 70 Hz) then amplified and digitized to 240 Hz and 10 bits of precision. For a single channel of differentially recorded data, CD, CI, autocorrelation, and amplitude are computed in a 20 s window with 2 s between windows.
RESULTS: Decreases in CI are associated with increases in amplitude as demonstrated by the fact that the above-mentioned changes in CI disappear when the signal amplitude is normalized. Furthermore, CD changes also depend on the frequency variations in an amplitude normalized signal, with increases in CD corresponding to decreases of the best-fit slope in the interval [0, 25] ms of the autocorrelation function, which is the Fourier transform of the PSD by the Wiener-Khintchine theorem.
CONCLUSIONS: CD and CI are sensitive to variations in amplitude and PSD, signal changes that are indicative of seizure onset, yet also may occur during inter-ictal periods. This raises questions pertaining to the specificity of CD and CI for the task of epileptic seizure prediction, and suggests these measures may not have more discriminating power than more common signal processing measures.