Authors :
Neha John, MSc – Cleveland Clinic Foundation
Juan C. Bulacio, MD – Cleveland Clinic, Cleveland, United States
William Bingaman, MD – Cleveland Clinic
Imad Najm, MD – Cleveland Clinic
Balu Krishnan, PhD – Cleveland Clinic
Presenting Author: Demitre Serletis, MD, PhD – Cleveland Clinic Epilepsy Center, USA
Rationale:
Epilepsy surgery is the best treatment for medically intractable epilepsy. Finding novel mathematical biomarkers underpinning the dynamic organization of brain networks involved in seizure onset and propagation could improve localization of the epileptogenic zone (EZ), leading to better outcomes following epilepsy surgery. Multifractal formalism introduces an invaluable framework for the investigation of nonlinear, scale-invariant features across multiple time scales in non-stationary time-series data. We sought to explore multifractal features defining spatiotemporal correlations in stereoelectroencephalography (sEEG) seizure recordings.
Methods:
We retrospectively analyzed 155 sEEG-recorded seizures from 36 patients with refractory temporal lobe epilepsy treated at Cleveland Clinic. Overall, 28 patients subsequently underwent a standard anteromesial temporal lobectomy including amygdalohippocampectomy (ATL/AH), while 8 patients underwent a more limited (tailored) sEEG-guided temporal resection sparing hippocampus (LTR). All patients achieved at least 1-year of sustained seizure freedom. Multifractal Detrended Fluctuation Analysis (MFDFA) was used to test for fractal scaling in the sEEG signals. sEEG data were annotated and anatomically organized by an epileptologist and epilepsy neurosurgeon. Eleven MFDFA-derived features were subjected to a cross-validated K-means clustering analysis, the results of which were used to train a binary linear support vector machine (SVM) classifier to differentiate between the two surgical cohorts.
Results:
We identified strong multifractal signatures in human sEEG data, observing spatiotemporal differences across epileptiform states and between different anatomical networks. In particular, we measured MFDFA-derived features (e.g. Hurst exponent, multifractal width) in sEEG recordings from EZ contacts (within resected tissue), with comparisons made to non-EZ contacts outside the region of resection. We report that there is a greater degree of MFDFA feature variability across the pre-ictal, ictal and post-ictal epileptiform states for contacts within the EZ (p< 0.05). Based on cross-validation of these findings, we developed a pipeline comprised of a clustering analysis and a linear SVM classifier, to identify regional differences emerging in MFDFA features between two surgical cohorts, i.e. those patients undergoing ATL/AH versus LTR. Importantly, the Hurst Exponent emerged as a key MFDFA feature enabling successful separation of both surgical cohorts based on fractal dynamics over time, achieving an accuracy of 96.9%, precision of 93.8%, recall of 100% and F1-Score of 95.8%.