Abstracts

Rationale:
About a third of all epilepsy patients may experience seizures that are resistant to medication. Surgical resection may be an option for some of these patients but localization of the seizure onset zone (SOZ) is a difficult and expensive process. For this process, the seizures must be observed and recorded in a clinical setting over several days. High frequency oscillations (HFOs) have emerged as a possible new biomarker of the SOZ that may be used to map the region of interest from a brief intracranial EEG (iEEG) recording. Several studies have proposed methods to automate the detection of HFOs, but much work needs to be done before there is a consensus on criteria for defining a good quality HFO sample. For instance, physiological spikes and sharp artifacts can lead to a significant proportion of false positives due to the ringing effect of using conventional bandpass filters. Rejection of such filtering artifacts is critical if HFOs are to be used for diagnostic prediction. In this study, we have developed a novel HFO detection algorithm that is heuristic in spirit, requires no training or tuning, and reliably distinguishes HFOs from spikes and other artifacts.
Method:
With IRB approval, continuous iEEG recordings with 1000 Hz sampling rate were acquired from nine patients with pharmacorefractory epilepsy who underwent an invasive presurgical evaluation. HFO candidates were first identified using a slightly modified version of a well-known algorithm (Staba et al., 2002), which is highly sensitive to HFOs but tends to admit spikes and other artifacts as well. In addition, a nonlinear method was devised to estimate the transient baseline of candidate HFOs in a way that highlights HFOs but suppresses spikes in the residual signal without the need for bandpass filtering. Three simple criteria—for amplitude, rhythmicity, and ringing, respectively—were formulated to identify HFOs. A superset of 2700 detections made by the Staba algorithm with a relaxed amplitude criterion were gathered from 1-3 channels of each patient’s recording for the purpose of algorithm validation. Events that satisfied all three criteria were considered genuine.
Results:
The proposed algorithm was found to distinguish genuine HFOs from spikes or other transients with a sensitivity of 20%, specificity of 93%, and positive prediction value of 90%. A second version of the detector with one of the criteria relaxed distinguished HFOs with a sensitivity of 55%, specificity of 84%, and positive prediction value of 83%.
Conclusion:
An unsupervised algorithm was developed that detects HFOs with moderate sensitivity and high specificity and precision. It is hoped that this algorithm, which applies objective, event-specific criteria to test the validity of putative HFOs, will serve as a valuable aid in the diagnostic evaluation leading up to epilepsy surgery.
Funding:
:This study was supported by National Science Foundation Grant No. 1539068. Al-Bakri received scholarship support from the University of Babylon in Iraq.