Seizure Sources Can Be Imaged from Scalp EEG by Means of Biophysically Constrained Deep Neural Networks

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate seizure localization is critical for treating drug-resistant focal epilepsy.
• Current noninvasive methods for pinpointing seizure origins can be limited.

Purpose of the Study:

• To investigate the efficacy of a novel deep learning-based source imaging framework (DeepSIF) for noninvasive seizure localization using electroencephalogram (EEG) data.
• To evaluate DeepSIF's performance in drug-resistant focal epilepsy patients.

Main Methods:

• Developed DeepSIF, a deep learning framework utilizing a neural mass model for synthetic ictal oscillation data generation.
• Trained and validated DeepSIF using computer simulations and high-density EEG recordings from 33 drug-resistant focal epilepsy patients.
• Compared DeepSIF's seizure source imaging estimates against surgical resection outcomes and seizure onset zones (SOZ).

Main Results:

• DeepSIF demonstrated superior performance in estimating spatial and temporal seizure origin information compared to other methods.
• Achieved high spatial specificity (96%) and low spatial dispersion (3.80 ± 5.74 mm) relative to resection regions.
• Showed good coverage of the seizure onset zone, with an average distance of 10.89 ± 10.14 mm.

Conclusions:

• DeepSIF shows significant potential for advancing noninvasive imaging of ictal activity origins in focal epilepsy.
• The framework aids in the management of intractable epilepsy by providing accurate seizure localization.
• DeepSIF offers a promising tool for improving clinical decision-making in epilepsy management.