DeepEZ: A Graph Convolutional Network for Automated Epileptogenic Zone Localization From Resting-State fMRI Connectivity

Objective: Epileptogenic zone (EZ) localization is a crucial step during diagnostic work up and therapeutic planning in medication refractory epilepsy. In this paper, we present the first deep learning approach to localize the EZ based on resting-state fMRI (rs-fMRI) data. Methods: Our network, called DeepEZ, uses a cascade of graph convolutions that emphasize signal propagation along expected anatomical pathways. We also integrate domain-specific information, such as an asymmetry term on the predicted EZ and a learned subject-specific bias to mitigate environmental confounds. Results: We validate DeepEZ on rs-fMRI collected from 14 patients with focal epilepsy at the University of Wisconsin Madison. Using cross validation, we demonstrate that DeepEZ achieves consistently high EZ localization performance (Accuracy: 0.88 ± 0.03; AUC: 0.73 ± 0.03) that far outstripped any of the baseline methods. This performance is notable given the variability in EZ locations and scanner type across the cohort. Conclusion: Our results highlight the promise of using DeepEZ as an accurate and noninvasive therapeutic planning tool for medication refractory epilepsy. Significance: While prior work in EZ localization focused on identifying localized aberrant signatures, there is growing evidence that epileptic seizures affect inter-regional connectivity in the brain. DeepEZ allows clinicians to harness this information from noninvasive imaging that can easily be integrated into the existing clinical workflow.


Introduction
This supplement provides further experimental results to augment our main paper. First, we display additional views (axial, sagittal and coronal) of the DeepEZ localization performance in four representative patients. We also show boxplots of the sensitivity and AUC achieved on the entire dataset across all methods considered. Finally, we stratify our patient cohort based on demographic and clinical factors and assess the performance of DeepEZ in each case.

Further Visualizations
In this section, we provide further visualizations of our main results.
SFig. 1 shows boxplots of the sensitivity (left) and AUC (right) among all methods. We include this figure to show the spread of performance based on fold membership. As described in the main text, we observe statistically higher sensitivity and AUC for the proposed DeepEZ method.
SFig. 2 shows three views of ground truth (red) and DeepEZ predictions (blue) of four representative patients. Based on clinical evaluation, the EZ SFigure 1: Boxplots for sensitivity (left) and AUC (right) across each method. The proposed DeepEZ has the best performance across methods considered.
SFigure 2: Ground truth (red) and DeepEZ predictions (blue) for four representative patients. The patients have the EZ located in left temporal, left extra-temporal, right temporal and right extra-temporal from top to bottom respectively.
for these patients are categorized as left temporal, left extra-temporal, right temporal and right extra-temporal, as organized from top to bottom in the figure. We note that DeepEZ can accurately localize the EZ in the second and fourth patients with minimal false positives. For the firth and third patients, DeepEZ exhibits high sensitivity with spurious contralateral predictions.

Patient Stratification
It is important to understand how DeepEZ performs based on different clinical and demographic factors, such as patient outcome, gender, and EZ location. We have conducted statistical evaluations based on these factors. Regarding patient outcome, ten patients are seizure free (denoted by either ILAE 1 or ILAE 2) STable 1: Mean plus or minus standard deviation performance metrics for ILAE 1-2 patients vs ILAE 3-5 using DeepEZ. The t-score compares the AUC between the two, we also note the corresponding FDR corrected p-value.

Outcome Sensitivity
Specificity Precision F1 Accuracy AUC t-score p-value ILAE 1-2 0.42 ± 0.046 0.85 ± 0.045 0.53 ± 0.048 0.46 ± 0.038 0.87 ± 0.046 0.74 ± 0.039 ILAE 3-5 0.39 ± 0.045 0.86 ± 0.037 0.47 ± 0.049 0.43 ± 0.041 0.88 ± 0.034 0.73 ± 0.036 −0.34 0.54 and four patients experience seizures but at a reduced rate (Table 1 in main text). STable 1 below shows the performance metrics using DeepEZ for ILAE 1-2 patients and ILAE 3-5 patients. While there is a slightly higher sensitivity and AUC in the seizure free cohort, a t-test on the AUC metric does not indicate significant difference in performance. We further stratified the cohort according to gender, hemisphere, and lobe of the EZ and performed the same t-test. These results are in STables 2-4. Overall, we observe a marginally significant difference in performance between left and right hemisphere SOZ (p < 0.1), but this may be due to the small sample size. We do not observe statistical difference in performance across the other stratifications of the cohort. With more data, it would be clinically valuable to see if the model performs statistically better or worse regarding patient outcome and EZ location.
STable 2: Mean plus or minus standard deviation performance metrics for male vs.
female patients using DeepEZ. The t-score compares the AUC between the two, we also note the corresponding FDR corrected p-value. STable 4: Mean plus or minus standard deviation performance metrics for temporal vs non-temporal EZ patients using DeepEZ. The t-score compares the AUC between the two, we also note the corresponding FDR corrected p-value.