Abstracts

Rationale: Due to the difficulties with scalp EEG analysis including the effects of different types of artifacts and noise, most previously proposed methods for epileptic seizure prediction have been based on the intracranial recordings, and are therefore less clinically applicable. In this study, we propose two novel real-time algorithms, a wavelet-based (WB) and a zero-crossing-based (ZCB), to predict epileptic seizures using surface EEG and compare their results. Methods: With ethic approval, we applied the seizure prediction algorithms on ~34 hours of multi-channel scalp EEGs from 6 patients with focal epilepsy, recorded in Vancouver General Hospital, BC, Canada, with a sampling frequency of 256 Hz. The dataset included 27 seizures and was visually inspected by an electroencephalographer to determine the electrographic seizure onsets. EEGs from different channels were segmented into 32-second epochs with 50% overlap. The ZCB method is based on the analysis of intervals between the EEG positive zero crossings, i.e. crossing the zero level when moving from negative to positive values. The histogram of these intervals in each EEG epoch is computed, and the distribution of specific bins, selected based on interictal and preictal references, is estimated using the related histogram values from the current epoch and epochs of the last 5 min. The resultant distribution for each selected bin is then compared to two reference distributions (interictal and preictal), and a seizure prediction index is developed for the current epoch. Comparing this index with a patient-specific threshold for all EEG channels, a seizure prediction alarm is finally generated. The WB approach, on the other hand, is based on the probability distribution of the EEG relative energy in selected frequency bands, showing statistically significant difference between interictal and preictal references. The current epoch is decomposed by the wavelet packet transform (WPT), appropriate to analyze non-stationary signals such as EEG, using Daubechies-6 wavelet. The energy of each frequency band is computed using the corresponding coefficients in the last decomposition level and divided by the total energy. Similar to the ZCB method, the distribution of the relative energy of each band is then estimated and compared to the reference distributions to calculate the prediction index for each channel and finally generate an alarm. Results: Applying the proposed algorithms on the epilepsy data, the WB and ZCB revealed, respectively, average sensitivities of 77% and 88% along with average false prediction rates of 0.26/h and 0.22/h (see Table 1). A prediction alarm was considered true if a seizure happened within 40 min after the alarm; otherwise, it was a false alarm. Figure 1 shows the distribution of the prediction time for the predicted seizures. Conclusions: The proposed seizure prediction methods are able to forewarn the upcoming seizures well in advance. Compared with the ZCB, the WB resulted in less accuracy, showing that energy may not clearly reflect the EEG underlying dynamics.