Abstracts

Rationale: The visual identification of interictal high frequency oscillations (HFO) in the intracranial EEG (iEEG) of patients with pharmacoresistant epilepsy is a laborious and time-consuming process. Moreover, it greatly relies on the use of band-pass filters for EEG analysis which introduces artifacts from filtering of epileptic transients or non-sinusoidal oscillations leading to false detection of HFO (Bénar et al 2010. Clin Neurophysiol. 121(3):301-10). Band-pass filtering is also confounded by detection of harmonics of lower frequency oscillations. The characteristics of HFO described in much of the literature are largely based on the output of such filters. Automated detection of HFO in large-scale recordings is essential for investigating their potential clinical applications as biomarkers of epileptogenicity. However, most algorithms developed to date to achieve either a supervised or an unsupervised detection of HFO work by first filtering the EEG signal in a narrow high frequency band range before performing a series of complex steps based on energy measures. We aimed to develop and validate an event classifier that automatically detects HFO by using broader band-pass filters and employing different metrics to avoid the confounders. Methods: The Event Classifier developed as an open source software aims to detect prescribed patterns in EEG recordings. A processor is used to calculate baseline power which corresponds to the minimum signal power for each channel across 1 second intervals. Three metrics are produced for each interval: power, spikeyness and intermittency. Each interval appears as a point in an N-dimensional space. The Event Classifier compares the metrics of each interval with those of a library of reference events selected from 70 channels of iEEG from subdural grids, strips and depth macroelectrodes. High frequency oscillations were visually identified by two independent reviewers in 10-minute samples of slow-wave sleep in the iEEG of four patients (total of 224 channels) sampled at 512Hz. The iEEG signal was displayed as raw data on a bipolar montage and with a 60Hz high pass filter on both a bipolar and a referential montage. The timescale was stretched to display 1 second per screen to optimize visualization of HFO. The same iEEG samples were processed through the event classifier which used the reference library to automatically detect HFO. Its performance was evaluated based on the number of HFO detections. Results: We report a high concordance (R=0.85) between the numbers of HFO identified by the reviewers across channels and the automated detections from the Event Classifier; this is similar to the concordance between the two reviewers. HFO were detected in three out of four patients and were concordant with the seizure onset zone in two. Conclusions: We have developed a simple to use event classifier which reliably detects HFO from human intracranial EEG data and which avoids some of the confounders present using alternative methods.