Abstracts

Rationale: High Frequency Oscillations (HFOs) are emerging as a biomarker of epileptogenic tissue. Visual marking of HFOs can be performed, but it is highly time consuming and subjectivity is inevitable. Thus, the development of automatic HFO detectors is crucial for the systematic study of HFOs and for their eventual utilization in clinical settings. At present, only a handful automatic detectors exist. In this study, a systematic comparison of the detectors on the same data set is presented. Methods: Intracerebral EEGs from 20 randomly selected patients were processed (filtered at 500Hz and sampled at 2000Hz). HFO events were identified independently by two experienced reviewers in all functioning channels during one minute of slow wave sleep. In addition, baseline segments (where it was clear that no oscillation was present) were visually marked. Channels with nearly continuous high frequency activity were excluded, resulting in 278 channels. These channels included 5238 visually identified HFO events (positive events) and 51076 visually identified baselines (negative events) that were used as the gold standard events. Four automatic HFO detectors were compared. Three are based on the comparison of the energy of the signal with the EEG epoch that includes the events. The main difference among them is the type of energy function computed on the filtered signal, either the root mean square amplitude (Staba et al., J Neurophysiol 2002; 88: 1743-52), the short-time line length (Gardner et al., Clin Neurophysiol 2007; 118: 1134-43), or the Hilbert envelope (Cr pon et al., Brain 2010; 133: 33-45). The fourth detector, the MNI detector (Zelmann et al., IEEE EMBS 2010, submitted), first detects baseline segments, where no oscillatory activity is present, and then compares the energy (root mean square amplitude) of the EEG signal with that of the detected baselines. In this way, the local characteristics of the background are considered. Receiver Operator Curves (ROC) were computed for each channel and averaged. The threshold values were computed according to the specifications of each detector. Results: The MNI detector had significantly higher sensitivity than the others, but at a cost of significantly higher false positive rate (FPR). Figure 1 presents the average ROC and performance table. The overall low FPR is not surprising, since we are only considering as reference negative events those segments of EEG visually marked as baselines. Thus, not all the EEG signal was considered in this study. The main difference in performance is observed in very active channels. Figure 2 shows an example of such channels, where one HFO is detected by all detectors whereas another is detected only by the MNI detector. Conclusions: Each automatic detector was developed for different EEG recordings and with different aims. Given the lack of a formal definition of HFOs, comparing them in a single data set is important to analyse their performance. The MNI detector performed better than the others in this dataset, but has larger FPR and was developed on channels similar to those used for testing. Supported by NSERC PGSD, CIHR MOP-10189.