Abstracts

Rationale: High Frequency Oscillations (HFOs), divided in Ripples (R, 80-250Hz) and Fast Ripples (FR, 250-500Hz), are recorded from intracranial macroelectrodes in patients with intractable epilepsy. Studies in rats suggested that R correspond to physiologic events while FR occur only in pathological areas. In humans, R and FR have been studied as separate events but recent studies suggest that this division might be arbitrary, and that frequency alone is not sufficient to separate normal from pathologic HFOs. These raise the question of whether the separation between R and FR is well grounded. A clustering approach is proposed to classify without prior knowledge visually marked events and to analyze whether these events cluster into R and FR, or differently. Methods: Intracerebral EEGs from 10 randomly selected patients were processed (filtered at 500Hz and sampled at 2000Hz). HFOs were visually identified during one minute of EEG. Only events jointly marked by two reviewers were considered. Baseline segments free of HFOs were also marked. Channels with nearly continuous high frequency activity were excluded. The EEG segments that corresponded to R or FR occurring alone, and R and FR co-occurring were selected. We randomly selected 300 events of each type for analysis. To characterize these EEG segments, we calculated the peak frequency, power in the 70-450Hz band, and mean autocorrelation. To avoid any assumption regarding clustering of events, an unsupervised learning approach is proposed. The Fuzzy K-means algorithm was selected for clustering the segments (normalized to zero mean and one standard deviation). Results: There seems to be a single cluster when pooling R alone, FR alone and co-occurring R-FR (Figure 1). The features overlap, regardless of the original division (as shown by the histograms of each feature and the 3D distribution of events). There does not seem to be another clear subdivision of the events. To validate that the selected features characterize the events, we compared 500 randomly selected events with 500 baselines and found that clustering occurred in this case (Figure 2), where 74% of the baselines and HFOs were assigned to different clusters, 10% were incorrectly clustered, and 16% were not assigned. Conclusions: When analysing visually marked HFOs recorded with macroelectrodes in humans, we did not find a clear clustering of events according to frequency, power or autocorrelation. Thus, there seems to be a continuum of these features across the selected events rather than a clear division in two categories. Most events, including the FRs alone, seem to have a peak frequency in the R band, probably as a result of some energy in that band, despite the prominent oscillatory component in the FR band. In conjunction with other results (Jacobs et al., Epilepsia 2008;49:1893-1907) suggesting that Rs and FRs recorded in the EEG behave similarly in terms of SOZ localization, we suggest that the relevance of subdividing HFOs into R and FR in human EEGs is questionable. Supported by NSERC PGSD, CIHR MOP-10189.