Abstracts

Rationale: High frequency oscillations (HFO) (gamma: 40-100 Hz, ripples: 100-200 Hz and fast ripples: 250-500 Hz) have been widely studied in health and disease. These phenomena may serve as biomarkers for epileptic brain, however a means of differentiating between pathological and normal physiological HFO is essential. We sought parameters that could separate the two phenomena, then attempted to automate the detection and classification of these events.Methods: Five patients undergoing clinical intracranial EEG for evaluation of intractable focal epilepsy were included in this study. Each patient performed either a visual recognition memory task or a finger movement task. We evaluated physiologic HFO induced by the visual or motor task (nHFO) by measuring frequency, duration and spectral amplitude of manually marked events in single trial time frequency spectra. Identification of pathologic HFO (pHFO) in the same patients was made by expert visual analysis of events marked by a line length detector from electrodes within the seizure onset zone. pHFO were manually measured for frequency, duration, and spectral amplitude in exactly the same manner as nHFO. Differences in spectral amplitude, frequency and duration between pHFO and nHFO were tested by unpaired student s t-tests. A support vector machine (SVM) classification was performed on the pHFO and nHFO using a rotating cross-validation approach. For the five patients for whom pHFO and nHFO were available a SVM classifier implemented in MATLAB (Mathworks, Natick, MA) using a linear kernel was cross-validated by training on the data for four patients and then classifying the data for the remaining patient. Using the findings from these analyses an automated detector and classifier was constructed and tested during periods of task and non-task and in areas of normal and epileptic cortex.Results: Compared to nHFO, pHFO were of higher mean spectral amplitude (M=9.10 SD 3.14 vs M=4.1, SD= 0.66, t(1042) = 33.94, p = < 1.0 x 10-6), longer duration (M=24.2 milliseconds, SD=19.5 vs. M=13.5 milliseconds, SD=8.5, t(1042) = 10.89, p < 1 x 10-6) and lower mean frequency (M = 171.0, SD = 76.0 vs. M = 200.2, SD = 98.5, t(1042) = -5.37, p < 1 x 10-06). In individual patients support vector machine analysis correctly classified pHFO with sensitivities ranging from 68-99% and specificities greater than 90% in all but one patient. Infrequent high spectral amplitude HFO arose in seizure involved cortex just before movement onset raising the possibility that nHFO may assume high power states in epileptic brain. Automated detection and classification of HFO was successfully performed and was able to detect task induced nHFO and detect and classify pHFO within the seizure onset zone.Conclusions: Pathologic and physiologic high frequency oscillations statistically separate along the feature axes of spectral amplitude, duration, and frequency. Using these features a machine learning algorithm can correctly classify the events with high accuracy. This process can be automated and allows rapid detection and classification of pHFO within epileptic brain.