Abstracts

Rationale: High frequency oscillations (HFOs) is a promising candidate biomarker for defining the epileptogenic zone. Visual marking of HFOs is highly time consuming and subjective. Therefore, developing accurate automatic HFOs detectors is necessary. Most existing automatic detectors use solid-state baseline computing methods, so they perform well in channels with rare HFOs but lose accuracy in very active channels. To address this problem, we propose a new algorithm for using the method of maximum distributed energy to calculate dynamic baseline to improve the accuracy of HFOs detection in both inactive and active channels.Methods: Human intracranial EEG (iEEG) data was collected from 6 patients with refractory epilepsy. We randomly chose 30 channels, and from each channel we selected a 5-minute segment during slow wave sleep for analysis. Two experienced EEG analysts visually identified and marked “ripple” HFOs separately. Activity identified by both EEG analysts as “ripple” HFOs was considered the gold standard for true HFOs. After visual marking, the automatic detector was applied. The iEEG data were band-passed between 80-200 Hz. We then calculated the absolute value of amplitude of the filtered signal and classified all points in the segment by quality threshold clustering. Points with similar energy levels were grouped into the same category. The biggest category contained the most points and had the largest distributed energy. We concluded that the points in the maximum distribution area were baseline points, and we then calculated the baseline amplitude by taking the mean amplitude of these points. Because baselines can change slightly over time, we set up a 5 second moving window. When the window moved, the baseline changed according to the signal inside the window, thus forming a dynamic baseline. “Ripple” HFOs were defined as consecutive signal activity containing at least one peak with amplitude measuring ≥ 5 standard deviations (SD) above the mean baseline and 8 consecutive peaks with amplitude measuring ≥ 3 SD above the mean baseline. Consecutive events separated by less than 25 ms were combined as one event. Final results from the automatic detector were compared with those of the iEEG analysts.Results: The sensitivity and specificity of our detector was 79.0% and 91.1%, respectively. The rank correlation between visualized and automated detection of “ripple” HFOs was significant for all recordings (r=0.9525, p<0.0001). All 6 patients eventually underwent surgical resection of the epileptogenic zone. Post-operative outcomes were good in 4 cases and poor in 2 cases. We found that a majority of the HFOs regions were resected in the 4 patients with good outcomes, but majority of the HFO regions were not removed in the 2 patients with poor prognosis.Conclusions: Our new algorithm calculates a dynamic EEG baseline using the method of maximum distributed energy. It can effectively and accurately detect ”ripple” HFOs in both inactive and active channels. It provides a promising tool for studying HFOs for future clinical application.