Abstracts

Delineation of the Seizure Onset Zone Based on Stimulation-Evoked Responses Using Convolutional Neural Networks

Abstract number : 2.069
Submission category : 3. Neurophysiology / 3E. Brain Stimulation
Year : 2023
Submission ID : 834
Source : www.aesnet.org
Presentation date : 12/3/2023 12:00:00 AM
Published date :

Authors :
Presenting Author: Rina Zelmann, PhD – Massachusetts General Hospital

Miaolin Fan, PhD – Neurology – Massachusetts General Hospital; Jaquelin Dezha Peralta, BEng – Neurology – Massachusetts General Hospital; Tal BenOliel Berman, MD – Hadassah Medical Center; Angelique Paulk, PhD – Neurology – Massachusetts General Hospital; Sydney Cash, MD,PhD – Associate Professor, Neurology, Massachusetts General Hospital

Rationale:
The delineation of the seizure onset zone (SOZ) remains a substantial challenge in surgical and neuromodulation based approaches to intractable focal epilepsy. Several groups have suggested that single pulse electrical stimulation (SPES) could elicit physiological and pathological epileptic responses which could be used to help identify the SOZ. In this study, we differentiate SPES-evoked responses in the clinically determined seizure onset zones (SOZ) channels from those occurring outside the SOZ using a convolutional neural network (CNN).



Methods:
We delivered multi-site single pulse electrical stimulation while recording intracranial EEG in nine patients undergoing depth electrode implantation to localize the origin of their seizures. Stimulation and recording bipolar channels were labeled as in the seizure onset zone (SOZ) or outside based on the clinical report. In total 83 stimulation (range: 3-14; 31 in the SOZ) and 1349 recording (range: 114-204) channels were included. One-second post-stimulation epochs, acquired at 2000 Hz and down-sampled to 512 Hz, for a total of 205714 epochs, were used as the input for a convolutional neural network (CNN). Specifically, we adapted the 1D-ResNet (https://github.com/geekfeiw/Multi-Scale-1D-ResNet) to single-channel iEEG inputs. This model was trained with a weighted cross-entropy loss function (learning rate = 0.001, momentum = 0.9, weight decay = 0.1 with SGD optimizer, dropout rate = 0.2) and validated using 5-fold cross-validation. For post-hot analysis, N-way ANOVA was performed to examine the main effect of the stimulation and recording channels on model accuracy as well as the interactions.



Results:
The CNN model correctly classified 82% of 205714 epochs, with a mean sensitivity of 43% and specificity of 88%. Per participant, accuracy ranged from 75% to 88% (Figure 1), with sensitivity from 21% to 66% and specificity from 83% to 93%. N-way ANOVA suggested that whether the recording channels were in the SOZ or outside significantly impacted accuracy for all subjects (p < < 0.001). Whether stimulation was on the SOZ had a significant effect for 5/9 patients. The interaction of these two factors was significant in 4/9 participants. Stimulating and recording in SOZ seemed to result in the highest accuracy (example in Figure 2). These results suggest that non-SOZ channels respond to stimulation with similar morphology, while responses from SOZ channels have a more diverse morphology. This morphological variability in the responses from epileptic regions across and within participants could explain why the CNN classified non-SOZ channels more accurately than SOZ channels.
Neurophysiology