Abstracts

Rational: The basic idea underlying this study is that brain activity preceding and during an epileptic seizure changes the hidden structure of an EEG record in a stereotyped manner. Methods: We have developed a new method for automated, real-time seizure detection, characterization and anticipation from 10/20 scalp EEG recordings. The method evaluates hidden mathematical patterns in EEG recordings corresponding to "self-similar" or "fractal" structure. The terms "self-similar" and "fractal" refer to EEG signals that appear to have a similar shape when plotted on different time scales. Two computed parameters characterize self-similar and fractal structure: alpha and cf. Alpha provides information about how far back in time past EEG signal amplitudes are likely to influence future amplitudes. Cf provides information about the strength of past EEG signal amplitude influences. We retrospectively analyzed digital EEG recordings from 21 epileptic patients both awake and asleep. Results: Our primary result is that alpha and cf change in a stereotyped manner before and during a wide variety of epileptic seizures. Seizures were reliably detected in all cases by a rapid decline of alpha to values between 0 and 1 and a corresponding rise in cf. Seizure onset times computed from alpha and cf were compared to clinical onset times as determined by expert visual analysis. In most cases, the computed onset times preceded the visually determined onset times by at least several seconds. Our most important result demonstrated that computed alpha values can anticipate seizures. First, alpha values rised significantly during the preictal phase yielding an average prediction time of over 4 minutes. Second, correlations in alpha values across electrode pairs yielded new insight into the dynamics of seizure onset. Progressive synchronization of alpha values across different brain locations was observed 30 to 45 seconds before seizure onset. Discussion: Automated analysis of EEG fractal structure suggests that the preictal and seizure states can be readly identified by computer. Our preliminary results suggest that correlations of computed alpha values may aid in identifing the seizure focus.