Authors :
Artur Agaronyan, BS – Childrens National Hospital
Syeda Abeera Amir, BS – Childrens National Hospital
Marius Linguraru, DPhil – Childrens National Hospital
William Gaillard, MD – Children's National Hospital
Chima Oluigbo, MD – Children's National Hospital, George Washington University
Presenting Author: Syed Anwar, MS, PhD – Childrens National Hospital
Rationale:
The prediction of seizure outcomes after surgery has remained challenging in clinical practice, especially for patients with suspected multifocal drug-resistant epilepsy. Research suggests an active role of the thalamocortical network in seizure propagation, and while sampling of thalamic targets is a developing area in surgical epilepsy, stereo electroencephalography (sEEG) provides an exciting avenue for thalamic recordings. Given this critical need for accurate seizure outcome prediction in patients with complex seizure networks, our study introduces a deep learning graph neural network (GNN) model leveraging stereo electroencephalography (sEEG) data of patients with refractory epilepsy.
Methods:
We utilized sEEG data from 10 pediatric patients with drug-resistant, multifocal epilepsy who had electrode contacts in the thalamus. A total of 105 ictal events were used to train a graph-based deep learning model to predict seizure outcomes. A threshold of 50% reduction in seizures was used to determine the binary labels for seizure-reduction vs no significant seizure frequency reduction. We also performed network analysis to capture connectivity patterns between thalamic and cortical/subcortical regions, and report network metrics such outcome network density and eigenvector values from our results.
Results:
The model demonstrated strong classification performance across multiple evaluation metrics. It achieved an accuracy of 89.3% for patient-wise analysis, with precision of 88.4%, recall of 84.2%, and F1-score of 86.2%. In binary classification, accuracy was 92.1%, precision was 92.6%, recall was 92.8%, and F1-score was 92.7%. For multi-class analysis (high- medium-, and low-level of seizure frequency reduction), the model achieved an accuracy of 84.4%, with precision of 84.3%, recall of 84.7%, and F1-score of 83.2%. Feature importance analysis identified key brain regions contributing to seizure outcome prediction, including the anterior cingulate, thalamus, and anterior frontal pole. When utilizing only the top 10 most important channels, the model maintained an accuracy of 77.8%. Network analysis revealed that patients with higher seizure reduction had 43.71% lower average density and 72.99% less average thalamic node connection strength than patients with lower seizure reduction. However, thalamic node centrality remained similar between groups, suggesting that connectivity pattern dynamics, rather than absolute thalamic involvement, may better predict outcomes.
Conclusions:
Funding: This work is supported by Chief Academic Officer Award from Childrens National Hospital.