Automatic detection of intracranially recorded High frequency oscillations using a radial basis function neural network gives reliable information about the distribution of HFOs over anatomical regions
Abstract number :
1.102
Submission category :
3. Clinical Neurophysiology
Year :
2010
Submission ID :
12302
Source :
www.aesnet.org
Presentation date :
12/3/2010 12:00:00 AM
Published date :
Dec 2, 2010, 06:00 AM
Authors :
Matthias Duempelmann, K. Kerber, J. Jacobs and A. Schulze-Bonhage
Rationale: High frequency oscillations (HFO) in the range between 80 and 500 Hz have been described to be recorded in seizure generating areas. The removal of HFO generating areas has been linked to a good postsurgical outcome. The visual identification of HFOs is a time consuming task and far from being unequivocal. Most automatic HFO detection algorithm described up to now use thresholds derived from global HFO feature statistics. Objective of the presented study was to replace these thresholds by the incorporation of human expertise in the detection algorithm using a radial basis function (RBF) neural network. Methods: Basis of this study were intracranial grid and strip recordings with commercially available electrodes of 11 patients with pathologically confirmed focal cortical dyspalsia (FCD) suffering from epilepsy. This study was restricted to HFOs in the range between (80-200Hz), termed ripples. HFOs were visually marked in a three-minute segment of slow wave sleep using high-pass filters at 80Hz The input of the RBF neural network consisted of the normalized signal energy, normalized signal line length and normalized instantaneous frequency of the high-passed filtered signal. The instantaneous frequency was computed on the basis of the Hilbert Transformation. Visually marked HFOs of three patients were used to determine the parameters of the RBF neural network. The marked HFO events of the other 8 patients were used to evaluate the detection algorithm. Results: In the segments of the 8 recordings used for evaluation 41722 HFOs were visually marked, whereas the RBF neural network detected 50606 HFOs. The percentage of overlapping detections was 30.7 %. Comparing the detections over channels the correlation of HFO count distributions over channels was significant between visual and automatic analysis for each of the data segments (min. correlation 0.366 (p<0.05), max. correlation 0.934 (p<0.001)). Rates of automatic detected HFOs were significantly higher inside the seizure onset zone SOZ (46.5/min) than outside (31.0/min, p<0.001).
Neurophysiology