Online High Frequency Oscillation Detection
High frequency oscillations (HFOs), or brief bursts in the high gamma band (80-500 Hz), have been studied as potential biomarkers of epileptic activity. Since the early 1990's, it has been recognized that increased high gamma power is present within the epileptogenic region at seizure onset in adults (Allen, Fish et al. 1992; Alarcon, Binnie et al. 1995) and children (Fisher, Webber et al. 1992; Traub, Whittington et al. 2001). Interictal fast ripples (Figure 1) have been detected almost exclusively in epileptogenic regions (Staba, Wilson et al. 2002; Jacobs, Levan et al. 2008; Worrell, Gardner et al. 2008), and appear to be highly specific localizing biomarkers. Additional information from studies of animal models further supports this notion. Interictal fast ripples are seen much more commonly in a rat kainate model compared to controls (Bragin, Wilson et al. 2003), and they have been seen to emerge during epileptogenesis in advance of epileptiform discharges in the same model (Bragin, Azizyan et al. 2005). Further, HFOs have been detected using electrode types that are practical for routine clinical use in neocortical epilepsies (Urrestarazu, Chander et al. 2007).
Most clinical EEG systems are incapable of recording HFOs, due to insufficient signal sampling rates or obscuration by amplifier or environmental noise. Specialized acquisition systems designed for research applications are available, but require significant technical expertise to operate, and are poorly suited to use in a clinical setting. HFOs are very brief (< 100 msec, with most less than 50 msec) and are low in amplitude relative to the low frequency components of the EEG signal; as such their detection by visual scanning alone is difficult and impractical. Automated detection methods have been developed and validated against detailed, careful, manual detection (Gardner, Worrell et al. 2007). We currently use a similar algorithm with a modified thresholding method (Schevon, Cappell et al. 2007). However, these automated methods are themselves both labor and computationally intensive, requiring computer time that is many times greater than the duration of the recording segment being tested. As such, they are only practical for short data segments, and the results of HFO analysis may be too delayed to play a role in clinical decision making. In order for HFO detection to become part of surgical epilepsy therapy, rapid automated detection over extended periods of time will be needed. It is important to recognize that the data published to date (detailed above) report on the specificity of high frequency oscillations for the epileptogenic zone, but their sensitivity is necessarily limited by the undersampling mandated by current detection techniques.
We are investigating how the process of automatic HFO detection can be accelerated using special purpose hardware designed for tasks that are easily “parallelized”, i.e. tasks that are readily subdivided into smaller ones, each of which can be performed concurrently (“in parallel”). Many graphics acceleration tasks, for example, are notably parallelizable; graphic images may be represented as arrays of pixels, and for many graphics tasks, pixels (or clusters of adjacent pixels) can be operated on concurrently, achieving massive increases in processing speed. The problem of HFO detection in multichannel iEEG is similarly parallelizable; channels can be processed concurrently, and within channels epochs (e.g. 5 seconds) can be subdivided into smaller time intervals, each of which can also be processed independently. This type of parallelization is ideally suited to a single-instruction multiple data (SIMD) parallization model, which is the simplest to implement and on low-cost hardware. Accordingly, we are investigating two possible hardware designs for high-speed HFO recognition, both based on inexpensive, off-the-shelf components: general purpose graphics processing units (GPGPUs) and field-programmable gate arrays (FPGAs).