Abstracts

Rationale: Several studies have illustrated the promising utility of simultaneous EEG-fMRI for focal epilepsy presurgical evaluation (Pittau et al., 2012; Khoo et al., 2018). However, the concordance between the primary fMRI response cluster and the ictal onset zone varies across patients (Koupparis et al., 2021) mainly due to the low occurrence frequency of interictal epileptic discharges (IEDs) and the variability of hemodynamic responses across IEDs. These two inherent limitations may not be well considered in the conventional fMRI analysis that is intended for a designed task rather than observed IEDs. Even when dealing with well-controlled experiments, it has been challenged for multiple testing issues and reproducibility crisis. Hierarchical Bayesian modeling may avoid/improve these issues (Spencer et al., 2021), therefore, we apply it to better estimate IED-related fMRI responses.

Methods: IEDs were annotated by experienced neurophysiologists on processed EEGs in which gradient and ballistocardiogram artifacts had been removed. Robust anatomical and functional MRI preprocessing were performed using the fMRIPrep pipeline. Schaefer2018 atlas and Freesurfer subcortical parcellation were used as the prior knowledge to model a hierarchical structure involving local and global hemodynamic responses to IEDs (Figure 1). Hence, fMRI response in each voxel is not only estimated with its own data, but also adjusted according to the local homogeneity of responses within a region of interest (ROI). ROI level responses were further modified to take into account the global homogeneity using the resting state networks to which they belong. A null distribution was defined from all voxels’ responses, and the region with the highest confidence of being different from null was selected as the estimated IED onset zone. The IED-related fMRI response was represented by the relative percentage change of blood oxygen level dependent (BOLD) signal.

Results: The new model was tested in 5 patients. It found fMRI clusters concordant with surgical resections for patients 1-2 (Engel I), whereas the conventional model failed for patient 1. No method found a cluster concordant with the resection for patient 3 (Engel II). For patients 4-5 (no surgery), only the new model found clusters in concordance with EEG topography. The conventional model estimated IED-related BOLD percentage changes much larger than our model. Abnormally high BOLD percentage changes were found in patients 1 (17.1%) and 3 (12.3%), considering higher than 10% change may be physiologically implausible (Jia et al., 2020). Our model found more likely values, with changes lower than 5%.

Conclusions: The capability of network-ROI-voxel hierarchical Bayesian model is illustrated with preliminary results, suggesting improved estimations on both IED-related location and BOLD response amplitude. Further evaluations will be carried out using a large clinical database.

Funding: Canadian Institutes of Health Research (FDN 143208)