Authors :
Presenting Author: Vojtech Travnicek, MSc – Institute of Scientific Instruments of the CAS, v. v. i.
Petr klimes, PhD – Institute of Scientific Instruments, The Czech Academy of Sciences, Brno, Czech Republic; Petr Nejedly, MSc – Institute of Scientific Instruments of the CAS, v. v. i.; Jan Cimbalnik, PhD – International Clinical Research Center, St. Anne's University Hospital, Brno, Czech Republic; Pavel Jurak, PhD – Institute of Scientific Instruments, The Czech Academy of Sciences, Brno, Czech Republic; Vit Vsiansky, MD – 1 St Department of Neurology, Faculty of Medicine, Masaryk University, Brno, Czech Republic; Martin Pail, MD, PhD. – 1 St Department of Neurology, Faculty of Medicine, Masaryk University, Brno, Czech Republic; Milan Brazdil, MD. PhD. prof. – 1 St Department of Neurology, Faculty of Medicine, Masaryk University, Brno, Czech Republic
Rationale:
Delineation of the Seizure Onset Zone (SOZ) from interictal data has been a problem many research groups have been eager to solve in the last decades. Spikes, High-Frequency oscillations, and other signal features are used either as individual biomarkers of EZ or combined in machine learning algorithms in order to delineate SOZ from interictal data. Epilepsy, however, is a disease that can have various etymologies, and there are only some pathologies with a high likelihood of becoming seizure-free. The lack of distinction between specific pathologies and the usage of a limited amount of biomarkers can be the reason, which prevents these algorithms from the performance increase. In this study, we compare the differences in SOZ localization between different histopathologies using 5 well-established biomarkers.
Methods:
Our cohort consisted of 58 patients from the St. Anne Faculty Hospital in Brno. We grouped the histopathologies into six categories: focal cortical dysplasia (FCD), hippocampal sclerosis (HS), other malformation of cortical development, vascular malformation, other structural and negative. For our study, we kept only groups including more than 10 patients, resulting in 17 patients with FCD, 10 with HS, and 11 negative cases. These patients had 30 minutes resting state interictal recordings sampled at 5kHz, where we evaluated the following features: Spike rate with Barkmeier detector, spike rate with Janca detector, Ripples (80-250 Hz) and Fast ripples (250-500 Hz) with CS detector and relative entropy (80-250Hz). For every patient and every feature, we evaluated the performance in SOZ localization using the area under the ROC and PRC curves. We compared their performance between histopathological groups (Wilcoxon rank sum), and then we evaluated the difference in feature performance within pathological groups (Wilcoxon sign-rank).
Results:
We performed 30 comparisons evaluating the difference between histopathological groups (three combinations of pathologies,five features, two metrics) and the only difference in SOZ localization between FCD and HS using spike count detected with the Barkmeier detector (p< 0.05 for both AUROC and AUPRC, Wilcoxon rank-sum). When we compared the two spike detectors, we found that the Janca detector performed significantly better than the Barkmeier detector using AUROC in all pathologies (ps< 0.05, Wilcoxon sign-rank, Cliff's delta > medium). Then we performed 60 intraclass comparisons (ten feature combinations, three pathologies, two metrics) and found 19 significant differences (e.g., Relative entropy works better than Fast Ripples when localizing SOZ in FCD patients.)
Conclusions:
Our study containing 38 patients showed no difference in localization performance for different histopathologies. The only difference in feature performance when localizing different pathology was found with the Barkmeier detector, which can be explained by its poor precision. However, choosing the right feature extractor is crucial when localizing SOZ from interictal data.
Funding:
Czech Science Foundation, project 21-25953S