Abstracts

Rationale:
Successful epilepsy surgery relies on precise localization of the epileptogenic zone (EZ), yet only about 60% of patients become seizure-free post-surgery, often due to inaccurate EZ identification. This underscores the urgent need for better diagnostic tools. This study introduces a novel method using Graph Neural Networks (GNNs) to analyze interictal biomarkers—specifically interictal epileptiform discharges (IEDs) and high-frequency oscillations (HFOs) in the frequency band of 80-250 Hz. By mapping features to a graph structure that reflects the patient-specific implantation topology, our approach aims to capture the complex dynamics of epileptic networks more accurately. We evaluate this method against traditional classifiers—Logistic Regression (LR) and Support Vector Machines (SVM)— and show that using a Graph Attention Network (GAT) on interictal stereo-EEG (SEEG) data enhances EZ localization performance, potentially improving surgical outcomes for those with drug-resistant epilepsy.




Methods:
We developed a GAT model to localize the EZ, defined in this study as seizure-onset zone contacts removed during successful epilepsy surgery. The model was trained and tested using leave-one-patient-out cross-validation on 23 seizure-free patients from St. Anne's University Hospital in Brno and the Montreal Neurological Institute & Hospital. Interictal features were detected across 30 minutes of NREM sleep SEEG using the IED Janca-detector and HFO-CS detector, then encoded into a graph structure for each patient as node features. In the graph structure, electrode contacts within 8 mm were connected by edges weighted by the Euclidean distance. Benchmark models, LR and SVM, both analyzed the IEDs and the HFOs as their two features without considering implantation topology.




Results:
In evaluating the performance of the GAT, LR, and SVM models, we assessed their median AUPRC across the patient cohort with interquartile ranges in brackets. The GAT model showed a median AUPRC of 0.685 (0.627). In contrast, the LR model achieved a median AUPRC of 0.500 (0.558), while the SVM had a median AUPRC of 0.394 (0.453). Although Wilcoxon tests showed no significant differences between GAT and LR (p-value > 0.05), comparison to SVM revealed a significant improvement for GAT in median AUPRC (p-value = 0.007), confirming GAT's superior effectiveness in localizing the epileptogenic zone.




Conclusions: The GAT model demonstrated improved performance over traditional methods like the SVM. By modelling SEEG data as graphs that incorporate implantation topology, GAT provides a more effective approach to localizing the EZ. This suggests that considering spatial relationships between electrode contacts, ignored by most traditional methods, enhances the precision of the localization. By leveraging their capability to model complex relationships within SEEG data, GNNs can improve the localization of the EZ, potentially increasing the success rate of surgical interventions and deepening our understanding of epilepsy.


Funding:
Supported by the Czech Science Foundation project n. 22-28784S, start-up funding of Duke University to B.F., and by project nr. LX22NPO5107 (MEYS): Financed by European Union – Next Generation EU.