SpaTeoGL: Spatiotemporal Graph Learning for Interpretable Seizure Onset Zone Analysis from Intracranial EEG

Accurate localization of the seizure onset zone (SOZ) from intracranial EEG (iEEG) is essential for epilepsy surgery but is challenged by complex spatiotemporal seizure dynamics. We propose SpaTeoGL, a spatiotemporal graph learning framework for interpretable seizure network analysis. SpaTeoGL jointly learns window-level spatial graphs capturing interactions among iEEG electrodes and a temporal graph linking time windows based on similarity of their spatial structure. The method is formulated within a smooth graph signal processing framework and solved via an alternating block coordinate descent algorithm with convergence guarantees. Experiments on a multicenter iEEG dataset with successful surgical outcomes show that SpaTeoGL is competitive with a baseline based on horizontal visibility graphs and logistic regression, while improving non-SOZ identification and providing interpretable insights into seizure onset and propagation dynamics.

💡 Research Summary

The paper introduces SpaTeoGL, a novel spatiotemporal graph learning framework designed to improve the localization of the seizure onset zone (SOZ) from intracranial EEG (iEEG) recordings, a critical step for epilepsy surgery planning. Traditional approaches either extract hand‑crafted connectivity features or convert each electrode’s time series into a graph (e.g., visibility graphs) before classification, but they treat spatial and temporal information separately, limiting their ability to capture the evolving network dynamics of seizures.

SpaTeoGL addresses this gap by jointly learning (i) a spatial graph for each time window that encodes functional interactions among the N electrodes, and (ii) a temporal graph that links the M windows based on similarity of their spatial graphs. The iEEG matrix X∈ℝ^{N×K} is first segmented into overlapping windows {X_m}_{m=1}^M. For each window, a Laplacian L^{(s)}_m is learned such that the graph‑signal smoothness term tr(X_m^T L^{(s)}_m X_m) is minimized, encouraging neighboring electrodes (as defined by the graph) to have similar activity.

To capture temporal evolution, the vectorized spatial Laplacians are stacked into a matrix \tilde X∈ℝ^{M×N^2}. Treating each row vec(L^{(s)}_m)^T as a graph signal on a temporal graph with Laplacian L^{(t)}, SpaTeoGL adds a temporal smoothness term tr(\tilde X^T L^{(t)} \tilde X). This term forces windows with similar spatial connectivity patterns to be strongly connected in the temporal graph, thereby modeling the transition from pre‑onset to propagation phases.

Both smoothness terms are regularized with Frobenius‑norm penalties β‖L^{(s)}_m‖_F^2 and β‖L^{(t)}‖_F^2, where β>0 guarantees strict convexity of each block sub‑problem. The overall objective is:

∑_{m=1}^M tr(X_m^T L^{(s)}m X_m) + β∑{m=1}^M ‖L^{(s)}_m‖_F^2 + tr(\tilde X^T L^{(t)} \tilde X) + β‖L^{(t)}‖_F^2

subject to each Laplacian belonging to the set of valid Laplacians (positive semidefinite, zero row sum, trace = N, non‑positive off‑diagonal entries).

Optimization is performed via an alternating block coordinate descent (BCD) scheme. In each iteration, (a) each spatial Laplacian L^{(s)}_m is updated by solving a convex problem that includes the spatial smoothness, regularization, and a coupling term that penalizes deviation from other spatial Laplacians weighted by the temporal adjacency; (b) the temporal Laplacian L^{(t)} is updated by solving a standard smooth‑graph‑learning problem on \tilde X. The authors prove that, under β>0 and exact solution of each sub‑problem, the objective is non‑increasing and any limit point is a stationary point, thus providing convergence guarantees.

Experimental evaluation uses the multicenter Epilepsy‑iEEG‑Multicenter dataset, focusing on nine patients with successful postoperative outcomes (31 seizures). Preprocessing includes bad‑electrode removal, 60 Hz notch, 0.5–100 Hz band‑pass, down‑sampling to 250 Hz, and segmentation into 512 ms windows with 50 % overlap, yielding M=13 windows per seizure (3 pre‑onset, 1 onset, 9 post‑onset). For each seizure SpaTeoGL learns 13 spatial graphs (one per window) and a single temporal graph linking the windows.

The baseline is a Horizontal Visibility Graph (HVG) approach: each electrode’s 1‑D series is transformed into an HVG, adjacency matrices are averaged across three selected windows, vectorized, reduced by PCA, and classified with logistic regression (LR).

Results show that SpaTeoGL achieves comparable SOZ (class 1) detection performance to HVG+LR (e.g., mean accuracy ≈0.77 vs 0.65 across patients) but significantly outperforms the baseline on non‑SOZ (class 0) discrimination (p=0.0048), leading to an overall accuracy of 70 % versus 62 % for HVG+LR. Normalized interpolated histograms confirm the statistical advantage for non‑SOZ electrodes while indicating no significant difference for SOZ electrodes (p=0.6404).

Beyond classification, SpaTeoGL provides interpretable visualizations. The learned temporal graph exhibits strong edges connecting onset and early propagation windows, clearly separating pre‑onset from post‑onset regimes. Bar plots of node degrees per window reveal that electrodes identified clinically as SOZ have higher connectivity during the onset window, and this connectivity spreads to neighboring electrodes in subsequent windows, mirroring known propagation pathways. Such graph‑based explanations are absent in the HVG+LR pipeline.

The authors acknowledge limitations: sensitivity to the regularization parameter β and windowing choices, reliance on a relatively small labeled cohort, and the need for validation on larger, more diverse datasets. Future directions include unsupervised initialization of graphs, integration of multimodal imaging (e.g., fMRI, PET), and development of lightweight, possibly online, versions of SpaTeoGL for real‑time surgical monitoring.

In summary, SpaTeoGL advances seizure network analysis by embedding spatial and temporal dynamics within a unified smooth graph‑signal framework, offering both competitive predictive performance and clinically meaningful interpretability, thereby representing a promising tool for presurgical epilepsy evaluation.