Loading...

Investigating Brain Networks in Epileptic Patients Using Simultaneous EEG Signals and fMRI Images in Resting State using GNN Method

Noori, Salar | 2026

10 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58839 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Fatemi Zadeh, Emadeddin
  7. Abstract:
  8. Drug-resistant epilepsy is often associated with network disorders extending beyond a single focal area, and clinical decisions (including pre-surgical evaluation) depend on the accurate identification of the involved region٫network. This research presents a multimodal framework based on simultaneous EEG and resting-state fMRI data to distinguish "patient٫control" groups, aiming to analyze resting-state brain networks and extract potential biomarkers. The fMRI data was mapped to Regions of Interest (ROIs), and a sparse spatial graph was constructed based on k-nearest neighbors (kNN) adjacency between ROIs. The EEG data was segmented into short windows and aligned with the corresponding fMRI time slices. Subsequently, a spatiotemporal dual-branch model, comprising temporal EEG encoder (1D convolution) and a GAT-based ROI graph encoder with a fusion module, was trained. To prevent information leakage, the train٫validation٫test split was performed at the subject level. After refinement, the dataset included 36 subjects (22 training, 6 validation, 8 test), generating 1,160 windows for each subject (Ntrain = 25,520, Nval = 6, 960, Ntest = 9,280). The best validation loss achieved was Lval = 0.1876, and the test perfor- mance at the window level reached Acc = 1.000, F1 = 1.000, and AUC = 1.000 (after 30 training epochs). Beyond performance, interpretable analyses based on network outputs indicated that differences in rs-fMRI connectivity were primarily prominent in long- range٫inter-network patterns, manifesting qualitatively in clusters of parieto-occipital regions (with complementary roles in frontal, temporal, and midline٫cingulate areas). Similarly, a consistent spatial group pattern was observed in EEG; parieto-occipital channels (such as PO, O, and P) were more prominent in patients, while fronto-central channels (such as FC and Fz) were more prominent in the control group. This multi- modal convergence proposes a set of regions٫connections as biomarker candidates for distinguishing patients from controls
  9. Keywords:
  10. Epilepsy ; Graph Neural Network ; Drug Resistance ; Interpretability ; Resting-State Functional Magnetic Resonance Image (FMRI) ; Spatio-Temporal Graph Neural Network ; Electroencephalography ; Biomarker

 Digital Object List

 Bookmark

...see more