Abstracts

Rationale: Numerous studies have demonstrated that seizures associated with temporal lobe epilepsy (TLE) are preceded by measurable changes in signal characteristics (linear and nonlinear) in intracranial EEGs (iEEG). These changes are detectable minutes to hours before the occurrence of a seizure. However, very few studies have applied prediction methods to scalp EEG recordings. Potential difficulties in scalp EEG analysis include the filtering effect of the skull distance from the source, recording artifacts, and rhythmic signal patterns due to normal physiological functions (e.g., sleep, chewing, etc.). A successful scalp EEG seizure prediction method must be able to utilize lower frequencies and be sufficiently robust to distinguish signal artifacts and normal EEG patterns from epileptic-related activities. In this study, we estimate the sensitivity and specificity of a seizure prediction algorithm that detects similarity of spatiotemporal dynamics among scalp EEG signals, and evaluate its performance based on statistical comparison against a random predictor. Methods: The test automated seizure prediction algorithm is based on the detection of convergence of the pattern match regularity statistic (PMRS). PMRS is a statistic that quantifies the level of signal organization (regularity) by estimating the likelihood of a pattern recurring. The algorithm was tested on continuous long-term (mean 48hrs) EEG recordings from 51 TLE patients with a total of 159 seizures recorded. PMRS was calculated sequentially (in 5.12 sec segments) for each EEG channel and the convergence was detected by the automated prediction algorithm. The performance of the test algorithm was evaluated against a random predictor that issues predictions with exponentially distributed random time intervals. In both prediction methods, a prediction was considered true only if a seizure occurred within 2.5 hours after the prediction; otherwise, it was a false prediction. Setting a minimum sensitivity requirement at 80% for each test patient, the false prediction rates (FPR) of both prediction methods were compared using the 2-sided Wilcoxon signed-rank test. Results: Under the 80% minimum sensitivity requirement for each patient, the test prediction method yielded a 95% mean sensitivity, whereas the random predictor gave an 83% mean sensitivity. The mean FPR was 1 per 8.6 hours (0.116/h) for the test algorithm, and was 1 per 2.5 hours (0.400/h) from the random predictor. The statistical analysis revealed that the mean FPR of our test algorithm was significantly smaller than that of the random predictor (p<0.001). On average, the test algorithm issued a seizure warning 58 minutes prior to the occurrence of a seizure. Conclusions: These findings suggest that convergence of signal regularity in scalp EEG can serve as a reliable pattern for seizure prediction. The ability to warn of impending seizures from analysis of scalp EEG recordings could have broad clinical applications in routine inpatient monitoring units and ambulatory EEG recordings.