Abstracts

Rationale: Scalp EEG-video monitoring has been a standard procedure in the pre-surgical evaluation of patients suffering from medically intractable epileptic seizures. Its efficiency depends on the ability to detect seizures. Because EEG is a direct correlate of brain function, an extremely complex system, it is a very complex signal, both in time and in space. Further, it is well known that paroxysmal transients that occur in normal sleep, as well as signal attenuation, poor spatial resolution, and noise or artifacts, greatly reduce the utility of the current automated techniques for EEG analysis. Therefore, to make quantitative analysis for scalp EEG clinically useful, it is essential to first identify signal characteristics that are not only sensitive to seizure activities but also robust to the noise/artifacts in scalp EEG signals. The objective of this study is to investigate the spatiotemporal dynamics of scalp EEG signals recorded from patients with temporal lobe epilepsy for seizure detection, localization and prediction. Methods: Multi-channel scalp EEG recordings obtained from 37 patients (total length ~ 2882 hours = 120 days, with a total of 88 seizures) with history of intractable epileptic seizures were analyzed in this study. Datasets were not pre-selected before the analysis. Signal complexity, frequency, and amplitude variation were estimated for each 5.12-sec non-overlapping epoch in each of the 16 recording channels. The pattern match regularity statistic (PMRS) was used to estimate the signal complexity for the detection of the rhythmic patterns in the ictal EEG. Signal frequency and amplitude variation were used for automatic rejection of artifacts caused by recording noise, movement/muscle, chewing, electrode failure, and other sources. The algorithm parameters were fixed for all patients. We evaluated the performance of the seizure onset detection by estimating the detection sensitivity and the false detection rate per hour. We also checked the identification of seizure onset zones as well as a dynamic pattern before the seizures. Results: A significant drop of PMRS values at seizure onset was observed. The values stayed low during the entire ictal period and went back to the baseline value postictally. Coupled with noise/artifact rejection rules, more than 94% (83/88) of the seizures were detected with an overall false detection rate of 0.02 per hour (i.e., 1 false detection per 50 hours). Furthermore, by comparison of PMRS values between the recording brain sites, the seizure onset zones were correctly identified and the ictal durations were properly estimated. Conclusions: The results from this study suggest that there exist changes of dynamic characteristics in scalp EEG signals that are detectable by quantitative analysis. With sophisticated pattern recognition and classification techniques, it is possible to develop clinically useful applications that can enhance the efficiency of EEG monitoring procedures.