Abstracts

Rationale: Automatic seizure detection is often used during long-term monitoring. Existing automatic detection techniques show high sensitivity and moderate specificity, and detect seizures relatively long time after onset. Analysis of high frequency activities (HFs), ranging from 80-500 Hz, indicates that these activities are prominent in many seizures and occur at seizure onset (Jirsch et al 2006, Ochi et al 2007). This study explores the use of HFs for automatic seizure detection in SEEG, as they have not been used before and may enhance the performance of existing methods. Methods: The SEEG was recorded after 500 Hz filtering with 2000 Hz sampling rate. The method was designed using 2 h of SEEG from 8 patients and a total of 10 seizures. SEEG signals were transformed into wavelet domain using the complex Morlet wavelet. The frequency ranges of interest are between 80-500 Hz. This method employed wavelet analysis of sequential 5s epochs in a sliding window to avoid edge effects. The algorithm was designed to extract features from each epoch and compare them with the features from a constantly updated background lasting 100 s and located 20 s before the current epoch. The method is aimed at detecting two features for each of 15 frequency bands with 30 Hz resolution between 80-500 Hz. The features are the number of HF discharges and the entropy. Features with values greater than 3 and 5 standard deviations above the mean were thresholded. Combinations of thresholded features were then used to mark significant HFs for detection. Results: HFs are very prominent at seizure onset. The method was evaluated on data from 8 patients and its performance was measured based on sensitivity, false detection rate and delay between seizure onset and detection. Results for the two threshold values, 3 and 5, show sensitivity of 100% and 70%, false detection rate of 0.68/h and 0.06/h and median delay of 6.1 s and 15 s, respectively. Examples are shown on figures 1 and 2. Missed seizures are characterized mainly by minimal or absent HF. False detections are mainly caused by short burst of spikes or harmonics of alternating current (60-Hz) artifact. For threshold value 3, excluding false detections caused by 60Hz, which could easily be identified, the false detection rate would become 0.43/h. Conclusions: We have demonstrated that in SEEG it is possible to detect many seizures automatically through HFs only. As some seizures show minimal amounts of this activity, HFs are not sufficient to detect all seizures. However HF detection could be combined to existing seizure detection methods due to the fact that HFs are prominent early in the discharge and are relatively specific to seizures; this would thus improve seizure detection performance. Grant support: Work supported by CIHR grant # MOP-10189 References: Jirsch et al.,(2006) Brain 129: 1593-1608 Ochi et al.,(2007) Epilepsia 48 , pp. 286-296