By Sheng H. Wang (CEA/NeuroSpin & Inria/MIND, France; University of Helsinki & Aalto University.). November 7th, 2024.
Identifying the epileptogenic zone (EZ) for epilepsy surgery is challenging due to the unique complexity of each patient’s pathology. We have shown that individual EZs may encompass synchronized components with high-dimensional (high-D) criticality features. However, high feature dimensionality can complicate the training of machine learning (ML) models for automated EZ localization. We hypothesized that, despite its apparent high dimensionality, epileptogenicity as a construct should have a low-dimensional (low-D) representation that characterizes a gradient from low to high seizure risk. To test this hypothesis, we proposed a two-step approach: first, applying dimensionality reduction for feature selection in a low-D latent space, and second, training unsupervised ML models to identify clinically defined EZs. Hundreds of criticality features were extracted from brain regions sampled by stereo-EEG (SEEG) during interictal resting-state in 64 patients. These features were then reduced to ten eigen-features capable of distinguishing the EZ, which were used to train two ML algorithms. Across a broad parameter space, the algorithms converged on a consensus seizure-risk model. This resting-SEEG-derived model demonstrated cross-domain validity by characterizing time-varying epileptogenicity over 7–9 hours of sleep-SEEG in three patients from a different cohort, providing preliminary support for our proposed low-D representation of epileptogenicity.
Contact: Sheng H. Wang, wang.sheng.h@gmail.com
Additional authors: Morgane Marzulli, Université Paris Cité, morgan.marzulli@etu.u-paris.fr; Gabriele Arnulfo, University of Genoa, gabriele.arnulfo@dibris.unige.it; Lino Nobili, University of Genoa, lino.nobili@unige.it; Vladislav Myrov, Aalto University, vladislav.myrov@aalto.fi; David Degras, University of Massachusetts Boston, david.degras@umb.edu; Satu Palva, University of Helsinki, satu.palva@helsinki.fi; Paul Ferrari, Michigan State University, paul.ferrari@spectrumhealth.org; *Philippe Ciuciu, CEA/NeuroSpin & Inria/MIND, philippe.ciuciu@cea.fr; *J Matias Palva, Aalto University, matias.palva@aalto.fi