Characterization of Seizure Onset Zones via Deep Learning on Brief Interictal Intracranial Recordings
Abstract number :
3.248
Submission category :
3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year :
2025
Submission ID :
693
Source :
www.aesnet.org
Presentation date :
12/8/2025 12:00:00 AM
Published date :
Authors :
Presenting Author: Sameer Sundrani, – Vanderbilt University Medical Center
Graham Johnson, MD, PhD – Mayo Clinic
Derek Doss, PhD – Vanderbilt University
Ghassan Makhoul, BS – Vanderbilt University Medical Center
Bruno Hidalgo, BS – Vanderbilt University Medical Center
Anas Reda, MS – Vanderbilt University Medical Center
Addison Cavender, BS – Vanderbilt University Medical Center
Emily Liao, BE – Vanderbilt University Medical Center
Baxter Rogers, PhD – Vanderbilt University Medical Center
Shawniqua Williams Roberson, MEng, MD – Vanderbilt University Medical Center
Sarah Bick, MD – Vanderbilt University Medical Center
Victoria Morgan, PhD – Vanderbilt University Medical Center
Dario Englot, MD PhD – Vanderbilt University Medical Center
Rationale: Epilepsy is a debilitating disorder that affects more than 50 million people worldwide, and one-third of patients are drug-resistant. If these patients’ seizures localize to a specific brain region, termed the seizure onset zone, surgical resection may be curative. Such localization is often confirmed with stereotactic electroencephalography; however, this may require patients to stay in the hospital for weeks to capture spontaneous seizures. More efficient characterization of seizure onset zones could therefore decrease morbidity and improve pre-surgical evaluation.
Methods: Over 1,000,000 interictal stereotactic electroencephalography segments were collected from 78 patients. With this data, we performed five-fold cross-validation and testing on a multichannel, multiscale 1-dimensional convolutional neural network to classify seizure onset zones (Figure 1).
Results: Across held-out test sets, our models achieved a seizure onset zone classification sensitivity of 0.702 (95% CI: [0.549, 0.805]), specificity of 0.741 (95% CI: [0.652, 0.835]), and accuracy of 0.738 (95% CI: [0.687, 0.795]), which was significantly improved compared to models trained on random labels alone (Figure 2A-B). Across regions spanning the entire brain, our models performed well, with top five location performance demonstrating accuracies between 70.0 to 88.4% (Figure 2C). The models performed significantly better on patients with favorable Engel outcomes after resection or who were RNS responders (Figure 2D). Feature analysis utilizing SHapley Additive exPlanation (SHAP) values on median normalized input data demonstrated consistently high feature importance to interictal spikes and large deflections, and analogous analyses on histogram equalized data revealed differences in feature importance assignments to low-amplitude segments (Figure 2E-F).
Conclusions: Our work demonstrates that deep learning on brief interictal intracranial data can characterize seizure onset zones across the brain. Furthermore, our findings corroborate current understandings of interictal epileptiform discharges and may help uncover novel interictal morphologies. Finally, clinical application of our models may reduce dependence on recorded seizures for localization and shorten pre-surgical evaluation time for drug-resistant epilepsy patients, improving patient experience and hospital costs.
Funding: This work was supported by NIH grants: NINDS R01NS112252, R01NS134625, F31NS131056, NIGMS T32GM007347, NIBIB T32EB021937 as well as Neil and Barbara Smit.
Neurophysiology