Abstracts

RATIONALE: Based on the analysis of EEG recordings from patients with temporal lobe epilepsy, we have shown that seizures occur in a deterministic fashion (In: Nonlinear Dynamical Analysis of the EEG, World Scientific, 1993, pp.30; Epilepsy Research, 1994, 17, p. 81). More importantly, results from continuous epileptic EEG recordings indicate that, in a retrospective analysis, temporal lobe seizures are preceded by a preictal transition (J. Combinatorial optimization, 2001, 5, pp.9). Obviously, a retrospective analysis (i.e. looking backward in time after a seizure[ssquote]s occurrence to try to detect a preictal transition) does not constitute seizure prediction. In the first prospective study reported (Epilepsia, 2001, 42 S7, p.41), seizures could be predicted with sensitivity 86.5% and false positive rate 0.14 per hour about 75.4 minutes prior to their occurrence. This prediction scheme was based on the dynamics of preceding seizures. We herein report results from an improvement of that algorithm, that is, from an adaptive scheme for seizure prediction that does not depend on the occurrence and continuous detection of preceding seizures.
METHODS: The method was tested in continuous 0.76 to 5.84 days 28-channel intracranial EEG recordings from a group of 5 patients with refractory temporal lobe epilepsy. The adaptive seizure prediction system involved the following steps: (1) translate the multi-channel EEG recordings into multivariate Lyapunov (STLmax) time series, (2) select the groups of critical electrodes sites from a 10-minute window before the first recorded seizure using integer quadratic optimization (training), (3) calculate the average T-index of the selected critical sites forward in time and issue a warning of an impending seizure when STLmax converge (dynamical entrainment transition), (4) reselect the critical sites using a 10-minute window after the observed entrainment transition, (5) repeat steps 3 and 4. The warning was considered to be true if a seizure occurred within 3 hours after a transition was detected and false if it did not.
RESULTS: A fixed parameter setting (number of optimal groups and number of critical electrode sites per group selected) applied to all patients predicted 82% of seizures with a false prediction rate of 0.16 per hour. Seizure warnings occurred an average of 71.7 minutes before ictal onset. Optimizing the parameters for individual patients improved sensitivity (overall 84%) and reduced false prediction rate (0.12 per hour).
CONCLUSIONS: These findings suggest that this real time seizure prediction system, free from seizure detection ambiguities, is capable of predicting an impending seizure with performance characteristics that could have practical clinical utility. Such a device could be incorporated in an on-line EEG monitoring system. In the future, implantable systems incorporating this type of algorithm could be used to timely activate physiological or pharmacological interventions aimed at aborting an impending seizure.
[Supported by: NIH/NINDS NS039687
ASU Whitaker Seed Grant
University of Florida Division of Sponsored Research
Children[ssquote]s Miracle Network
U.S. Veterans Affairs]