Abstracts

Rationale: We propose a novel method for the automatic detection of epileptic seizures and apply it to more than 1.000 hours of EEG-recordings from 17 patients. Automatic detection of epileptic seizures with low latency would be of great benefit in clinical practice. A reliable detection system would relieve personal from continuously monitoring the EEG during recording. At the same time it will alert medical staff of a beginning seizure in order to assist patients and to perform further testing. Furthermore, online detection of epileptic seizures is a key technology for intervention systems that can interrupt seizures by means of, e.g., electrical stimulation. Methods: Our approach focuses on the detection of rhythmic waveforms that typically characterize many epileptic seizures. In particular, it is based on a projection of the EEG into a subspace of periodic wave-forms that have a duration that is inversely proportional to their fundamental frequency. It is designed to detect periodic patterns in the delta-, theta-, alpha-, and beta-bands, whereby the durations of these patterns are assumed to decrease from the delta- to the beta-band. Normalization to the total energy in the EEG within the time window covered by each pattern is used to make the method scale independent and to detect only dominant components. Furthermore, the major portion of the signal energy is forced to be localized around the fundamental frequency. This makes the frequency estimation more robust and avoids common sub-harmonic errors. Results: We applied our method to 1.270h of EEG recordings from 17 patients containing 58 seizures. We used patient independent detection thresholds for each of the four frequency bands and manually chose one band based on a prior visual inspection of EEG data. No automatic or manual artefact reduction strategy or data pre-selection method was applied. The true positive rates (TPR) and the false alarm rates (FAR) for all 17 patients are given in Table 1. For 11 patients we obtained a TPR of 100% with FARs below 0.62 per hour. FARs for 3 of the remaining 6 patients were below 0.27 per hour although the TPR were only 71, 50 and 33%, respectively. FARs obtained for the remaining 3 patients were 0.92, 3.03 and 3.44 per hour, respectively. On average we detected 83% of the 58 seizures, with a mean FAR of 0.54 per hour. Conclusions: The proposed method for detection of epileptic seizures in scalp EEG recordings proved to be promising for application in clinical environments. Without any automatic or manual artefact reduction strategies we already achieved an adequate performance for 11 out of 17 patients. For 3 of the remaining patients, we obtained a FAR below 0.27 per hour but not all seizures were detected. These seizures did not show dominant rhythmic waveforms in the EEG and thus cannot be captured by this method alone. The remaining 3 patients showed an almost permanent occurrence of high-amplitude inter-ictal theta or alpha waves leading to a high FAR. This inter-ictal activity needs to be further addressed by introducing adaptive detection thresholds.