Authors :
Edoardo Paolini, MS – University of Verona
Navaneethakrishna Makaram, PhD, MS – Boston Children's Hospital, Harvard Medical School, Boston, MA 02115, USA
Jeffrey Bolton, MD – Boston Childrens Hospital
Scellig Stone, MD, PhD – Boston Childrens Hospital & Harvard Medical School
Christos Papadelis, PhD – Cook Children's Health Care System
Phillip Pearl, MD – Boston Children's Hospital & Harvard Medical School
Ellen Grant, MD – Boston Childrens Hospital
Silvia F. Storti, PhD – Unversity of Verona
Presenting Author: Eleonora Tamilia, PhD – Boston Children’s Hospital, Harvard Medical School, Boston, MA, USA
Rationale:
In children with drug-resistant epilepsy (DRE), brain surgery can offer seizure control. However, 30-40% of patients continue to have seizures postoperatively. There is a critical need for noninvasive EEG biomarkers that capture intrinsic properties of the epileptogenic tissue and guide timely, personalized treatment. Interictal scalp EEG is widely available and scalable, making it a strong candidate for this purpose.
Among interictal EEG tools, the most widely used is spike localization. Yet, spikes may arise from non-epileptogenic regions, limiting their reliability for surgery. More robust interictal estimators of the epileptogenic tissue are needed.
We aim to develop a non-invasive, connectome-based EEG approach to model the epileptogenic networks of children with DRE and predict their postsurgical outcomes following the planned resection. Specifically, we developed connectome-based machine-learning (ML) classifiers using non-epileptiform versus epileptiform EEG data (i.e., with or without spikes) to test whether network analysis reveals spike-independent markers of epileptogenicity.
Methods:
We studied 66 children with DRE who had surgery with known outcome (1-year follow-up; 65% good outcome: Engel I). We performed EEG network analysis, separately, on epileptiform and non-epileptiform EEG epochs as Fig 1 outlines. We reconstructed cortical activity via electrical source imaging (Fig 1A-B), computed each child’s whole-brain functional network (amplitude envelop correlation) and then we virtually isolated the resected (targeted) network and the residual network (nodes spared by surgery; Fig 1B). We then extracted network features and compared them between outcomes (Wilcoxon rank-sum test; ROC curve; Fig 1C) and cross-validated ML classifiers (5-fold) for individual outcome prediction with parameter optimization and recursive feature elimination (Fig 1D).
To assess the added value of connectome-based analysis, we performed spike localization (three methods: dipole, beamformer, cMEM) and tested its performance using the same ML method.
Results:
Looking at individual predictors (Fig 2A), we found that resections that target weak and skewed subnetworks in the brain forecast poor seizure outcomes.
In terms of patient-level outcome prediction (Fig 2B), k-Nearest Neighbors (k-NN) classifier reached 83% accuracy (84% sensitivity; 83% specificity; p< 0.00001) using spike-free EEG, based on the properties of each patient’s targeted network.