Abstracts

There is evidence in the literature that the occurrence of epileptic seizures may not be independent or uncorrelated in a statistical sense. Statistical approaches have been used to model seizure occurrence with some success. However, these relied on patient diaries, which are inherently inaccurate, and dealt almost exclusively with daily seizure frequency. Most current approaches to prediction analyze continuously sampled EEG data using nonlinear dynamic measures to identify putative preictal states and thereby anticipate seizures. Statistical evidence that the ictal states themselves may be serially correlated has been ignored. However, inspection of sequences of durations of the interseizure interval (ISI) derived from clinical ECoG records of subjects with intractable epilepsy suggested that the time of seizure occurrence might depend on past behavior. This hypothesis was tested formally using concepts from nonlinear time series analysis: specifically, the method of analogues was used to make one-step forecasts of ISI and the prediction error compared with suitable surrogate data to test for significance. ECoGs from 60 subjects with intractable epilepsy were screened and those with at least 40 visually verified automated electrographic seizure detections ([italic]n[/italic]=24; median number of seizures=84.5; median recording length=5.6 days; median seizure index=18.6/day) were analyzed. Detections occurring within 60 s of one another were clustered. The ISI sequence (time between successive seizure onsets) was computed, differenced, delay-embedded (unit lag) and the method of analogues used to determine the embedding dimension that minimized the root-mean-squared prediction error (in-sample error, excluding the current sample) relative to the standard deviation of the data. The error was compared with that of 100 surrogate sequences generated by random permutation of the ISI sequence to test for serial correlation. A second set of linear surrogates (identical power spectrum but randomized phase compared to original sequence) was used to test for the null hypothesis of a linear stochastic process. The first test determines whether there is any serial correlation, and the second if the correlation cannot be described adequately by a linear autoregressive model. The null hypothesis (seizures are uncorrelated) was rejected ([italic]p[/italic][lt]0.01) for 17/24 subjects. The linear hypothesis was rejected for 14 of these ([italic]p[/italic][lt]0.01): in addition, eight of the 14 had a relative error [lt]0.9 and prediction correlation coefficient [gt]0.5. In about half of all cases, prediction error increased with lead-time until saturation. These findings provide a reasonable case for predictability of seizures based on discrete observations alone. Assessment of the practical utility of these results is under way. (Supported by NINDS/NIH Grant NS046060-01.)