Abstract Details

Title Interictal Electrographic Features Differentiate MTL Responders And Nonresponders

Topic Epilepsy/Clinical Neurophysiology (EEG)

Presentation(s) S48 - Epilepsy/Clinical Neurophysiology (EEG) III (1:00 PM-1:11 PM)

Poster/Presentation
Number 001

Background Objective epilepsy biomarkers derived from interictal electrocorticographic (ECoG) data may be used to supplement patient reported clinical seizures. Interictal ambulatory ECoGs from NeuroPace RNS System clinical trials provides the opportunity to discover objective epilepsy biomarkers.

Objective Investigate whether interictal electrographic features can differentiate mesiotemporal lobe epilepsy (MTL) patients who respond to treatment with the NeuroPace® RNS® System.

Design/Methods 79 MTL patients had ≥ 6 months of clinical seizure data at year 7 post implant of the RNS System. Electrographic features from year 7 ECoGs were extracted with hand-crafted algorithms (method 1) and convolution neural networks (CNN; method 2) from upper (n = 20 patients; -96.5% median seizure change at year 7) and lower (n = 20; -17.4%) response quartiles. Hand-crafted features included interictal spikes and spectral band power in classic frequency bands. Wilcoxon rank sum test was used to test differences in hand-crafted features from the two response quartiles. CNN features were extracted from grayscale spectral images of the interictal ECoGs with pre-trained GoogLeNet Inception-V3 model. These features were used to train neural networks over 10 training epochs (LR 0.0001, SGD optimizer, batch size 32) for classifying patients into the two response quartiles.

Results In this retrospective analysis, interictal spike rate (ISR) was the most significantly different feature (p<0.0001; 0 median ISR in UQR and 0.05 median ISR in LRQ), while normalized theta and gamma power were also significantly different (p<0.05) between the two response quartiles. With interictal ECoG features extracted by a pre-trained Inception-V3 network, neural networks classified new MTL patients (i.e., patients whose data was not used for training) into the two response quartiles with 76.2% classification accuracy.

Conclusions

Two different but complementary methods, one based on hand-crafted features and another based on CNN features, have shown that interictal ECoG data may be used for differentiating MTL responders from nonresponders.

Authors/Disclosures
PRESENTER	No disclosure on file
Thomas K. Tcheng, PhD	No disclosure on file
Martha Morrell, MD, FAAN (NeuroPace/Stanford University)	Dr. Morrell has received personal compensation for serving as an employee of NeuroPace. Dr. Morrell has stock in NeuroPace. The institution of Dr. Morrell has received research support from National Institutes of Health.

��ɫ��

��ɫ��