Authors :
First Author: Yorguin-Jose Mantilla-Ramos, BS – Universidad de Antioquia
Presenting Author: Wenhui Cui, – University of Southern California
Jian Li, PhD – Massachusetts General Hospital and Harvard Medical School; Dileep Nair, Dr – Cleveland Clinic Neurological Institute; Patrick Chauvel, Dr – University of Pittsburgh Medical Center; Karim Jerbi, PhD – University of Montreal; Richard Leahy, PhD – University of Southern California
Rationale:
Approximately one third of epilepsy cases are pharmacoresistant. These cases may be candidates for surgical resection or ablation but this requires identification of the Epileptogenic Zone (EZ). Currently, identification of the EZ typically includes visual inspection of intracranial electroencephalography (iEEG), a process which is challenging, time-consuming and prone to subjectivity. Therefore, finding an objective set of characteristics to localize the EZ is of great importance. For this purpose, high-frequency oscillations are often used as a potential biomarker for the EZ. A recently proposed alternative ‘fingerprint’ approach uses a combination of high-frequency oscillations, alpha-suppression and pre-ictal spiking as the biomarker. Nonetheless, accurate localization of the EZ remains a challenging problem. Beyond oscillations and the ictal fingerprint, iEEG signals contain a wealth of complexity properties that have remained underexplored in the context of epilepsy surgery. For example, the existence of power law properties in the aperiodic component of the electrophysiological spectrum.
Methods:
In the present work we evaluate classical machine learning models trained on complexity features, with the task of predicting the EZ. To this end, we used retrospective stereoelectroencephalography data from 28 pharmacoresistant focal epilepsy patients that were seizure-free after a resective surgery. The data was inspected to identify individual seizure events, each segmented into interictal and ictal periods. The following set of features were extracted from each segment and electrode location: A) The aperiodic component of the frequency spectrum, B) Long-range temporal correlation, C) Lempel–Ziv complexity, D) Different Entropies, E) Hjorth parameters, and F) Fractal dimensions. Different classical machine learning models were then trained using subject-wise cross-validation. Finally, these models were compared through their confusion matrices, and characterized with respect to their most relevant features.
Results:
Our results indicate the possibility of estimating the EZ using complexity-related features from either or both the interictal and ictal periods. Interestingly, the feature importance obtained when using both periods shows a balance between both segments (Figure 1). Moreover, some features such as Spectral Entropy and the offset of the aperiodic component are leveraged by all models. When the models are trained only on interictal features, Spectral Entropy also shows the same behavior (Figure 2). In both Figures, Fractal dimensions (e.g., Higuchi and Petrosian) are consistently relevant.
Conclusions:
These models suggest that a set of complexity features could pave the way for more efficient approaches to localize the EZ and thereby improve the outcome of surgical resection in drug-resistant epilepsy. Furthermore, the interpretation of these features may improve our understanding of the pathophysiology of pharmacoresistant focal epilepsies.Funding:
This research was supported in part by National Institutes of Health under award R01-EB026299 and R01-NS089212.