Abstracts

A Machine Learning Approach to Seizure Detection in a Rat Model of Post-Traumatic Seizures

Abstract number : 2.426
Submission category : 3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year : 2022
Submission ID : 2233001
Source : www.aesnet.org
Presentation date : 12/4/2022 12:00:00 PM
Published date : Nov 22, 2022, 05:29 AM

Authors :
Robert Kotloski, MD PhD – William S Middleton Memorial Veterans Hospital; University of Wisconsin SMPH; Thomas Sutula, MD, PhD – University of Wisconsin-Madison School of Medicine and Public Health

This is a Late-Breaking abstract.

Rationale: Epilepsy is a common neurologic condition characterized by spontaneous recurrent seizures. These seizures are often diagnosed and quantified by electroencephalography (EEG). Given technological advances large datasets of EEG can be generated which are amenable to machine learning approaches to identification of patterns of interest, such as seizures.

Methods: We utilized a predesigned deep convolutional neural network (DCNN), GoogLeNet, constructed to classify images. Training images for ICTAL or BASELINE classification were generated through multiplexing of spectrograms, kurtosis, and entropy for two-second segments of EEG.

Results: Over 2200 hours of EEG data was scored for the presence of seizures. In comparison to visual scoring, accuracy for individual 2-second segments >95% and detection of >95% of seizures as compared to visual scoring. Multiplexed images combining spectrogram, kurtosis, and entropy are shown superior to scalograms alone. Use of a DCNN trained specifically for the individual animal was shown superior to using DCNNs across animals. Similar approaches have been utilized and our results produce equivalent results. Our method is novel in the use of a multiplexing approach in a predesigned DCNN to classify images generated from rodent EEG and identify seizures.

Conclusions: We present a novel use of a predesigned DCNN constructed to classify images, combined with images which multiplex spectral content, kurtosis, and entropy to rapidly and objectively identify seizures in a large dataset of rat EEG.

Funding: This work was funded by a Department of Veterans Affairs CDA-2 (IK2BX002986) (RK).
Neurophysiology