Imaging Seizure Sources from Scalp EEG Using Biophysically Constrained Deep Neural Networks
Abstract number :
3.294
Submission category :
3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year :
2024
Submission ID :
418
Source :
www.aesnet.org
Presentation date :
12/9/2024 12:00:00 AM
Published date :
Authors :
Rui Sun, PhD – Department of Biomedical Engineering, Carnegie Mellon University, Pittsburgh, PA
Abbas Sohrapour, PhD – Department of Biomedical Engineering, Carnegie Mellon University, Pittsburgh, PA
Presenting Author: Jesse Rong, – Department of Biomedical Engineering, Carnegie Mellon University, Pittsburgh, PA
Boney Joseph, M.B.B.S. – Department of Neurology, Mayo Clinic, Rochester, MN
Gregory Worrell, MD, PhD – Mayo Clinic
Bin He, PhD – Carnegie Mellon University
Rationale: Non-invasive EEG source imaging (ESI) is an important technique for localizing and imaging epileptic activities from EEG recordings. Efforts have been made to directly image seizure sources from scalp EEG recordings using regularization-based ESI algorithms (Sohrabpour det al, 2020; Yang et al, 2011). Recently a deep learning-based source imaging framework using biophysically constrained deep neural network, DeepSIF (Sun et al., 2022), has been proposed to improve ESI and shown promise in imaging epileptiform activity from interictal spikes. In this study, we extend the DeepSIF approach to directly image seizure sources from scalp recorded high density ictal EEG recordings.
Methods: This study expands DeepSIF to include imaging of oscillatory activities, specifically ictal oscillations. Synthetic training data were generated using ictal spatial and temporal source models for deep neural network (DNN) training. The trained network was then applied to scalp-recorded high-density ictal EEG data from 33 focal drug resistant epilepsy patients to localize seizure-generating regions. Performance evaluation included comparison with intracranial EEG-defined seizure onset zone (SOZ) and surgical resection outcomes. DeepSIF-based ictal source imaging was also evaluated in comparison to three conventional source imaging methods (sLORETA, FDI, LCMV), to provide insights into its performance.
Results: DeepSIF ictal imaging exhibited robust performance in detecting temporal dynamics variations, achieving a high linear correlation (0.98 ± 0.04) with simulated signals. Our results indicate that DeepSIF surpasses conventional methods (sLORETA, FDI, LCMV) in accurately estimating both the spatial and temporal characteristics of ictal sources. When applied to real ictal signals from focal drug resistant epilepsy patients, it attains a high spatial specificity of 96% and exhibits minimal spatial dispersion of 3.80 ± 5.74 mm compared to the resection region. Moreover, the DeepSIF-based ictal source imaging exhibits an average distance of 10.89 ± 10.14 mm from the SOZ to the reconstructed area. Fig. 1 illustrates two patient examples of seizure source imaging using DeepSIF.
Conclusions: Our study shows DeepSIF's capability in robustly estimating the extent, location, and temporal dynamics of ictal oscillations from scalp EEG, positioning it as a promising tool for advancing noninvasive imaging of seizure activities in epilepsy patients.
Funding: NIH R01NS127849, R01NS096761, and T32 EB029365.
Neurophysiology