EpiNet
Over the past few decades, the concept of the epileptogenic zone responsible for epileptic seizures has evolved significantly, particularly due to advances in electrophysiology and neuroimaging. Initially perceived as a single, localized focus, the epileptogenic zone is now better understood as a dynamic network (EpiNet): seizures are not generated by a single area but by a set of dysfunctional regions that interact with each other and can change over time in terms of location, electrical activity (areas of rapid brain activity, interictal electrical discharges), and/or their interactions. Moreover, this network is unique to each patient.
Identifying biomarkers in electrical recordings
A precise anatomo-functional description is particularly important for drug-resistant epilepsies, where surgical removal of the seizure-onset zone(s) is the only therapeutic solution. Presurgical evaluation relies on a combination of data, including electroencephalographic (EEG) recordings and neuroimaging (MRI and PET scans). Often, fine delineation of the areas to be surgically removed also requires stereo-electroencephalography (SEEG) using intracranial electrodes. Several biomarkers can be identified from SEEG data, such as biomarkers of criticality (complexity of dynamics) or functional connectivity. Machine learning models assist in identifying these biomarkers. Studies have shown that combining multiple biomarkers significantly improves network localization. However, this approach generates high-dimensional data, increasing the risk of overfitting and reducing interpretability.
Researchers from the Inria-CEA MIND team, in collaboration with the Neuroscience Center in Helsinki (Finland) and the University of Michigan (USA), hypothesized that it is possible to capture epileptogenic dynamics in a simplified representation of the aforementioned biomarkers—a low-dimensional latent space—without requiring seizure recordings (resting-state SEEG). Using interictal SEEG data from 64 patients, they extracted 260 features related to functional connectivity and criticality, which they reduced to 10 latent components. A classifier trained on these 10 components was then simplified into a probabilistic model requiring only two input components, reducing the feature space by over 99%.
The approach, described in the Journal of Neuroscience, simplifies data interpretation, facilitates the integration of additional biomarkers for presurgical diagnosis, and should enable large-scale analyses across patient cohorts.
Contact at the Frédéric-Joliot Institute for Life sciences:
This text was translated with the assistance of Mistral AI.