Abstracts

Rationale:
For children with drug-resistant epilepsy (DRE), brain surgery can lead to seizure control; yet, 30-40% of surgical patients continue to have postoperative seizures. This uncertainty regarding the surgery outcome often delays the surgical workup initiation in good surgical candidates, or leads to redundant testing in those who will not benefit from surgery. This highlights a critical need for novel methods that enable an accurate prediction of surgical outcome early in the surgical workup.




As the primary determinant of surgical success is the presence of a confined seizure-generating network, we hypothesize that diffusely hyperconnected irritative networks predict postoperative seizure recurrence in children with DRE.




To test this hypothesis, we present a non-invasive approach designed to process interictal scalp EEG and deconstruct the epileptogenic network (via electrical source imaging, ESI, and functional connectivity, FC, analysis). Our goal is to predict epilepsy surgery success in children with DRE based on the properties of their epileptogenic network, estimated via scalp EEG.




Methods:
We analyzed interictal scalp EEG from 71 children with DRE who had non-palliative focal surgery (73% good outcome). Fig 1 outlines our methods:

A) we reconstructed each patient’s epileptiform activity around spikes via ESI (Fig 1A);

B) outlined their whole-brain irritative network via FC analysis (Fig 1B);

C) characterized the network via a set of features (Fig 1C) quantifying the network distribution and architecture, which we compared between good and poor outcome patients (Wilcoxon rank-sum);

D) cross-validated (3-fold) a machine-learning (ML) model to predict surgical success at the individual patient level (linear discriminat analysis, Fig 1D).

To assess the added value of the proposed EEG network analysis, we also performed conventional spike localization with three different methods (dipole, beamformer, cMEM) and tested their predictive performance.






Results:
The irritative networks of poor outcome patients showed a stronger (p=0.025) and less homogeneous FC (high variability, p=0.032; low kurtosis, p=0.026) than good outcomes. None of the spike localization features differed between good and poor outcomes.



The network-based ML model predicted individual patient’s outcome with an area under the ROC curve (AUC) of 79%, and positive and negative predictive values of 81% and 67% (Fig 2A). The spike localization features performed worse than network features in predicting outcome, reaching a maximum AUC of 70% using the beamformer method (Fig 2B).


Conclusions:
We present a novel method to boost the presurgical value of interictal EEG data in children with DRE by enabling the non-invasive deconstruction of their epileptogenic network. Our findings, validated on a cohort of 71 children, demonstrate that the properties of the child’s irritative network can distinguish children who will and will not gain seizure-freedom from surgery, outperforming conventional spike localization.




This tool may minimize ineffective surgeries by enabling early identification of children with epileptogenic networks that are unlikely to benefit from focal resection.




Funding: R03NS127044 (NIH)
RSZ TNC Pilot Award