Localizing the Epileptogenic Zone Using Cortico-cortical Evoked Potentials and Deep Learning
Abstract number :
2.156
Submission category :
3. Neurophysiology / 3E. Brain Stimulation
Year :
2024
Submission ID :
1089
Source :
www.aesnet.org
Presentation date :
12/8/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Zekai Qiang, MBChB – University of Sheffield, Sheffield, UK
Kelly Pu, – Duke University School of Medicine
Cindy Mei, MD – Yale University, New Haven, CT, USA
Hari McGrath, MBBS – East and North Hertfordshire NHS Trust
Robert Duckrow, MD – Yale University, New Haven, CT, USA
Eyiyemisi Damisah, MD – Yale University, New Haven, CT, USA
Dennis Spencer, MD – Yale University, New Haven, CT, USA
Adithya Sivaraju, MD, MHA – Yale School of Medicine
Hitten Zaveri, PhD – Yale University
Rationale:
Intracranial EEG monitoring is often essential to identify suitable targets for surgery or neuromodulation in patients with medically intractable epilepsy. Emerging data suggests that cortico-cortical evoked potentials (CCEPs) may hold value for localizing the seizure onset area.
Methods: We developed a recurrent neural network (RNN) architecture based on long short-term memory (LSTM) to identify brain regions proximal and distal to the seizure onset area. Contacts in the gray matter were identified using post implant CT with a CT-preimplant MRI reconstruction. All or most gray matter contacts were stimulated with 1 Hz, always starting with the ‘silent’ or non-epileptiform contacts and marching towards the ‘active’ or seizure-onset zone. CCEPs were acquired from 10 subjects who underwent bipolar, biphasic stimulations with a pulse width of 0.3 ms. Each pair of contacts was stimulated at a current of 1 mA, 5 mA, and 10 mA for a duration of 30 s to 1 min at each current strength. The RNN model consisted of a layer of 32 cells, followed by a Luong attention mechanism to enhance focus on key sequential features. The input data contained 512 timesteps, and the output layer was a single dense layer with a sigmoid activation function.
Results: The model was trained with the Adam optimizer at a learning rate of 0.001 and binary cross-entropy loss, over 16 epochs with a batch size of 128. Intracranial recordings obtained in response to each stimulation were input into the model and evaluated through two paradigms: individual-specific models, which were trained and evaluated using the same individual’s data through 5-fold cross-validation, and a pan-cohort model, which shuffled data from 8 subjects as the training set and tested on the remaining 2 subjects. The proposed architecture achieved a mean accuracy of 0.713±0.09 in classifying contacts proximal and distal to the seizure onset zone in individual subjects and achieved a mean accuracy of 0.604±0.09 when applied across the entire cohort.
Conclusions: This preliminary study demonstrates the potential of a deep neural network to extract salient temporal dynamic features of CCEPs to help localize the seizure onset area.
Funding: The Swebilius Foundation
Neurophysiology