Abstracts

Rationale: Despite the regular use of multimodal data in preparation for surgical treatment of pharmacologically resistant clinical epilepsy, integrating information from multiple modalities, such as electrophysiology and magnetic resonance imaging (MRI), remains a manual process, which calls for a systematic integrative approach. Building on the approach described in Jirsa, Viktor K., Neuroimage 145 (2017): 377-388., we propose a generative brain network model, which links structural (T1-weighted MRI) and connectivity (diffusion-weighted MRI) information from magnetic resonance imaging with intracranial electrophysiological recordings of seizures. Specifically, the model combines MRI-derived anatomy with clinical hypotheses to predict the spatial and temporal profile of seizure propagation, quantifying the spatial profile or heatmap of the pathology. The link of patient-specific structure with a statistical dynamic network model allows for full Bayesian inference and testing of different hypotheses to identify those, which are best supported by the available data, in the interest of corroborating or supplementing the clinician’s existing analysis. An additional advantage of the Bayesian approach lies in the quantification of uncertainty: when the model cannot be identified properly given the available data, the results clearly state this, allowing a clinician to weight the results appropriately. Methods: A 40 patient cohort from the Marseille epilepsy clinic with intracranial electrophysiology (sEEG), T1-weighted and diffusion-weighted imaging (DWI) data were analyzed according to the following workflow: cortical and subcortical anatomy were reconstructed from the T1 image with FreeSurfer, the DWI were processed with the MRtrix3 toolkit for probabilistic tractography to obtain structural connectomes registered to the Destrieux atlas, the spatiotemporal seizure profile is extracted from the sEEG time series, and a epilepsy-specific statistical brain network model, based on the Virtual Brain brain network modeling platform, is built in the Stan software for Bayesian modeling, which infers a spatial profile of pathology by tuning predictions of seizure propagation based on structural and connectivity data. We validate the model using cross-validation across all seizure datasets in the cohort. Results: The statistical brain network model can be cross-validated satisfactorily for the majority of patients, both within datasets according to Bayesian information criteria and consensus with clinical assessment.  In general, these cases correspond to when the assumptions of model, (a) permittivity coupling and (b) propagation between and not within regions of interest in the parcellation, are satisfied.  In the minority of cases where the model is not corroborated by clinical assessment, the statistical analysis shows higher uncertainty. Conclusions: The applicability of the approach in a majority of cases is demonstrated in cohort of 40 retrospective patients, with the perspective of incorporation in the upcoming clinical trial EPINOV with 400 prospective patients. Funding: The research reported herein was supported by the Brain Network Recovery Group (220020255) through the James S. McDonnell Foundation and funding from the European Union Seventh Framework Programme Human Brain Project (grant no. 60402), Agence National de la Recherche “Vibrations” (ANR 13 PRTS 0011 01), and within the FHU EPINEXT by the A*MIDEX project (ANR-11-IDEX-0001-02) funded by the "Investissements d'Avenir" French Government program managed by the French National Research Agency (ANR).