Authors :
Presenting Author: Elena Grazia Gado, MS – École Polytechnique Fédérale de Lausanne
Peter Hadar, MD, MS – Massachusetts General Hospital; Harvard Medical School
Sydney Cash, MD, PhD – Massachusetts General Hospital
Pariya Salami, Ph.D. – Massachusetts General Hospital
Rationale: Responsive Neurostimulation (RNS) offers a valuable alternative for patients with drug-refractory epilepsy (DRE) who are not candidates for resection. However, predicting treatment response remains a major clinical challenge, often leading to years of ineffective trial-and-error. While presurgical data show promise for outcome prediction, conventional methods fail to capture the complex network dynamics involved. To address this, we present a deep learning framework that leverages ictal presurgical intracranial EEG (iEEG) and clinical metadata to support biomarker discovery, with a focus on interpretability, generalizability, and robustness.
Methods:
We analyzed 387 seizures from 49 patients with DRE who underwent iEEG monitoring prior to RNS implantation, with electrodes targeting cortical, mesial temporal, and thalamic sites based on individual seizure localization. From 30-second windows centered on seizure onset, we computed Functional Connectivity Networks (FCNs) using magnitude-squared coherence across broadband, alpha, beta, and gamma bands. These FCNs were encoded into graph-based seizure representations, which served as input to four distinct Graph Neural Network (GNN) architectures. The response variable was the Engel surgical outcome classification, assessed at least one year post-implantation and categorized into three classes: Class I and II (14 patients), Class III (18 patients), and Class IV (17 patients), corresponding to significant seizure reduction, intermediate response, and little to no benefit, respectively. Each seizure was labeled according to the patient’s Engel class. Models were trained and evaluated using both patient-stratified and non-stratified cross-validation schemes. To enhance robustness and mitigate patient-specific bias, we integrated a Domain Adversarial Learning (DAL) framework that promotes the extraction of patient-invariant features. Model performance was benchmarked against classical statistical analyses of graph metrics.
Results: Traditional network metrics failed to differentiate RNS responders from non-responders. In contrast, the proposed GNN models achieved promising performance, reaching up to 94% accuracy in a 3-class classification task under non-stratified 5-fold cross-validation, and up to 76.7% in a non-stratified yet fully patient-invariant evaluation. Although generalization to completely held-out patients remains an unmet goal, the framework provided valuable methodological insights for future work.
Conclusions: High performance in non-stratified settings underscores the expressiveness of graph representations and the suitability of GNNs for predicting neuromodulation outcomes from intracranial recordings. The integration of domain adversarial learning enhanced both interpretability and robustness by facilitating the extraction of patient-invariant features predictive of treatment response. These findings provide a robust framework for future expansion of this work and suggest novel methodological strategies for advancing network-guided neuromodulation in epilepsy care.
Funding: W81XWH-22-1-0315