Abstracts

Rationale: The resection of cortical brain network hubs, derived using resting state MEG functional connectivity, was previously shown to correlate strongly with surgery outcome. However, due to difficulty recording resting state MEG signals from deep brain structures such as the hippocampus, it is likely that key information is omitted.

Structural connectivity data, supplemented with computational modelling, allows inference of MEG functional data (Messaritaki et al 2021). To date, this approach has not yet been applied in the context of epilepsy, or to deep potentially epileptogenic structures such as the hippocampus.

Here, we use a computational model to infer resting-state MEG functional connectivity across the entire brain using patient-specific MRI data. Furthermore, using individual subject coupling predictions, we identify key differences between surgical outcome groups (ILAE1 vs ILAE2+).

Methods: Diffusion-weighted, T1-weighted MRI and resting-state MEG were acquired from 25 patients with drug refractory focal epilepsy (13 temporal, 12 extratemporal) who later underwent surgery. Structural brain networks from MRI were inferred using streamline tractography, and source-localised broadband functional networks were constructed using time series amplitude correlation between regions. 128 cortical and subcortical regions of interest constituted the networks, with missing data for subcortical MEG.

The coupling weights for each individual subject were estimated using multiple linear regression models with five predictors derived from cortical and subcortical diffusion weighted structural connectivity matrices. Robust predictions of MEG node strength were obtained using ensemble and hold-out approaches. Resection of empirical cortical and estimated subcortical hubs was quantified for subjects using distinguishability measures (DRS) and subsequently used for classification of patient surgical outcome.

Results: The modelling approach performed well in predicting the MEG cortical functional connectivity across all patients (Mean Spearman rho=0.49, SD=0.14, p< 0.05). This subsequently allowed reasonable classification of patient outcomes (AUC=0.67), outperforming classification from MRI alone (AUC=0.33), and not significantly worse than when using the empirically observed MEG network (p > 0.05). Inclusion of estimated subcortical node strength paired with empirical cortical node strength achieved the best separability of outcome groups (AUC=0.81, P<0.05).