Abstracts

Rationale: High frequency oscillations (HFOs), encompassing ripples (frequencies between about 50 Hz to 200 Hz) and fast ripples (frequencies greater than 200 Hz), have been shown to be have some correlation with seizures both temporally and spatially, though the specificity and sensitivity are not yet high enough for clinical applications. Most work in the field has focused on very limited features of HFOs, usually either the peak frequency or rate of the HFOs. We seek to improve the specificity and sensitivity of HFOs for determining the seizure onset zone through detailed analysis of HFO features and more advanced classification algorithms.Methods: Human iEEG data was recorded at 32 kHz in eight refractory seizure patients. Over 100,000 HFOs were identified by automated analysis, and the raw signals were used to extract several signal features in both the time and frequency domain such as peak and median frequency, interquartile distance, and the standard deviation of the peak voltages per oscillation. These features were used in several different classification algorithms such as support vector machines and boosted decision trees to determine a feature set unique to the seizure onset zone. Classification was performed first on individual, single channel HFOs, then each channel is classified, and finally clustering algorithms are used to determine the final estimate of the seizure onset zone. The Matthew's Correlation Coefficient (MCC) was used to measure accuracy, and the results were also benchmarked against the common methods of HFO rate and peak frequency. The algorithms were tested by both cross validation and hold out methods on each patient, as well as training on various subsets of patients and testing on others subsets of patients.Results: The MCC score was a reliable indicator of clinical accuracy. Large disparity existed between results from different patients, both in accuracy and in the specific features that predicted the seizure onset zone. Some feature sets were found which resulted in an improved MCC score compared with the traditional rate and frequency algorithms, while some subsets preformed more poorly. There was no algorithm that performed well across multiple patients. A novel class of HFOs with broad spectral power was identified that did not appear to have any relationship with seizure onset zone.Conclusions: HFOs are complex dynamic phenomena that are only partially described by rate and peak frequency. By analyzing more sophisticated features, the specificity and sensitivity of identifying the seizure onset zone is improved. Further work is needed to identify features and algorithms of HFOs that can be used in a prospective fashion to localize epileptic networks.