Abstracts

Rationale: High-Frequency-Oscillations (HFO) are potential biomarkers of the epileptogenic-zone and their automatic detection facilitates and increases the significance of their research. An important challenge faced by automatic detectors is the correct handling of fast-transients such as interictal-epileptic-spikes (IES), which can generate artificial-HFO when filtered but can also carry true-HFO. IES represent a valuable biomarker used to define the irritative-zone; however, HFO have been reported to better predict the Seizure-Onset-Zone (SOZ). The correct differentiation of true- from artefactual-HFO is therefore a key issue. The retrospective localization of the SOZ using true-HFO would allow obtaining their true biomarker-value without distortions stemming from IES. Methods: Automatic detectors of intracranial-IES, HFO-Ripples and HFO-FastRipples were developed based on kernelized support-vector-machines. For HFO, the selection of the support vectors was optimized by punishing artifactual detections. The detection of all event-types was tested on both real and simulated databases (FBGDAT, SIMDAT). Using automatic detections, the accumulative, per-minute and per-channel occurrence-rate was calculated for 7 different types of automatically detected biomarkers: Ripples, FastRipples, HFO, IES, coinciding Ripples and IES (IES-Ripples), coinciding FastRipples and IES (IES-FR), coinciding HFO and IES (IES-HFO). For each biomarker, their per-channel occurrence-rate was used to localize the SOZ-channels, their performance was measured by the area under the receiver-operating-characteristic curve (AUC-ROC). A total of 11 Patients were analyzed, each with 60 min. of invasive electroencephalography being at least 1 hour away from the start or end of an epileptic seizure. Results: The percentage of artefactual-detections produced by the Ripple, Fast-Ripple and HFO detectors was respectively 16%, 3% and 17% from a total of 19695 IES. The average kappa scores of the IES, Ripples, Fast-Ripples and HFO detectors on FBGDAT were: 52%, 49%, 44% and 48% respectively; the average kappa-scores on SIMDAT were: 90%, 77%, 82% and 84% respectively. The best biomarkers in descending order were: IES-Ripples and IES-HFO (both 79%), IES (74%), IES-FastRipples and HFO (both 69%), Ripples (66%), FastRipples (60%). Conclusions: The developed detectors compared favorably to other published methods(Fig.1). The parameter occurrence-rate showed an average performance-increase with time (Fig.2). None of the HFO-biomarkers were better than IES but the combined biomarkers IES-Ripples and IES-HFO performed better than IES by 5 points; moreover, only these two biomarkers localized the SOZ better than random (p < 0.001) and for all patients and minutes. The low biomarker value of FastRipples is surprising and could be caused by a low signal-to-noise-ratio (SNR) in the analyzed signals. Funding: This work was supported by the BrainLinks-BrainTools Cluster of Excellence funded by the German Research Foundation (DFG, grant number EXC CNE1086) and partly by DFG grant JA1725-2.